sys/netinet/tcp_stacks/rack.c

   1 /*-
   2  * Copyright (c) 2016-2020 Netflix, Inc.
   3  *
   4  * Redistribution and use in source and binary forms, with or without
   5  * modification, are permitted provided that the following conditions
   6  * are met:
   7  * 1. Redistributions of source code must retain the above copyright
   8  *    notice, this list of conditions and the following disclaimer.
   9  * 2. Redistributions in binary form must reproduce the above copyright
  10  *    notice, this list of conditions and the following disclaimer in the
  11  *    documentation and/or other materials provided with the distribution.
  12  *
  13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  23  * SUCH DAMAGE.
  24  *
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include "opt_inet.h"
  31 #include "opt_inet6.h"
  32 #include "opt_ipsec.h"
  33 #include "opt_ratelimit.h"
  34 #include "opt_kern_tls.h"
  35 #if defined(INET) || defined(INET6)
  36 #include <sys/param.h>
  37 #include <sys/arb.h>
  38 #include <sys/module.h>
  39 #include <sys/kernel.h>
  40 #ifdef TCP_HHOOK
  41 #include <sys/hhook.h>
  42 #endif
  43 #include <sys/lock.h>
  44 #include <sys/malloc.h>
  45 #include <sys/lock.h>
  46 #include <sys/mutex.h>
  47 #include <sys/mbuf.h>
  48 #include <sys/proc.h>           /* for proc0 declaration */
  49 #include <sys/socket.h>
  50 #include <sys/socketvar.h>
  51 #include <sys/sysctl.h>
  52 #include <sys/systm.h>
  53 #ifdef STATS
  54 #include <sys/qmath.h>
  55 #include <sys/tree.h>
  56 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
  57 #else
  58 #include <sys/tree.h>
  59 #endif
  60 #include <sys/refcount.h>
  61 #include <sys/queue.h>
  62 #include <sys/tim_filter.h>
  63 #include <sys/smp.h>
  64 #include <sys/kthread.h>
  65 #include <sys/kern_prefetch.h>
  66 #include <sys/protosw.h>
  67 #ifdef TCP_ACCOUNTING
  68 #include <sys/sched.h>
  69 #include <machine/cpu.h>
  70 #endif
  71 #include <vm/uma.h>
  72
  73 #include <net/route.h>
  74 #include <net/route/nhop.h>
  75 #include <net/vnet.h>
  76
  77 #define TCPSTATES               /* for logging */
  78
  79 #include <netinet/in.h>
  80 #include <netinet/in_kdtrace.h>
  81 #include <netinet/in_pcb.h>
  82 #include <netinet/ip.h>
  83 #include <netinet/ip_icmp.h>    /* required for icmp_var.h */
  84 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
  85 #include <netinet/ip_var.h>
  86 #include <netinet/ip6.h>
  87 #include <netinet6/in6_pcb.h>
  88 #include <netinet6/ip6_var.h>
  89 #include <netinet/tcp.h>
  90 #define TCPOUTFLAGS
  91 #include <netinet/tcp_fsm.h>
  92 #include <netinet/tcp_seq.h>
  93 #include <netinet/tcp_timer.h>
  94 #include <netinet/tcp_var.h>
  95 #include <netinet/tcp_log_buf.h>
  96 #include <netinet/tcp_syncache.h>
  97 #include <netinet/tcp_hpts.h>
  98 #include <netinet/tcp_ratelimit.h>
  99 #include <netinet/tcp_accounting.h>
 100 #include <netinet/tcpip.h>
 101 #include <netinet/cc/cc.h>
 102 #include <netinet/cc/cc_newreno.h>
 103 #include <netinet/tcp_fastopen.h>
 104 #include <netinet/tcp_lro.h>
 105 #ifdef NETFLIX_SHARED_CWND
 106 #include <netinet/tcp_shared_cwnd.h>
 107 #endif
 108 #ifdef TCP_OFFLOAD
 109 #include <netinet/tcp_offload.h>
 110 #endif
 111 #ifdef INET6
 112 #include <netinet6/tcp6_var.h>
 113 #endif
 114 #include <netinet/tcp_ecn.h>
 115
 116 #include <netipsec/ipsec_support.h>
 117
 118 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 119 #include <netipsec/ipsec.h>
 120 #include <netipsec/ipsec6.h>
 121 #endif                          /* IPSEC */
 122
 123 #include <netinet/udp.h>
 124 #include <netinet/udp_var.h>
 125 #include <machine/in_cksum.h>
 126
 127 #ifdef MAC
 128 #include <security/mac/mac_framework.h>
 129 #endif
 130 #include "sack_filter.h"
 131 #include "tcp_rack.h"
 132 #include "tailq_hash.h"
 133 #include "rack_bbr_common.h"
 134
 135 uma_zone_t rack_zone;
 136 uma_zone_t rack_pcb_zone;
 137
 138 #ifndef TICKS2SBT
 139 #define TICKS2SBT(__t)  (tick_sbt * ((sbintime_t)(__t)))
 140 #endif
 141
 142 VNET_DECLARE(uint32_t, newreno_beta);
 143 VNET_DECLARE(uint32_t, newreno_beta_ecn);
 144 #define V_newreno_beta VNET(newreno_beta)
 145 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
 146
 147
 148 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
 149 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
 150
 151 struct sysctl_ctx_list rack_sysctl_ctx;
 152 struct sysctl_oid *rack_sysctl_root;
 153
 154 #define CUM_ACKED 1
 155 #define SACKED 2
 156
 157 /*
 158  * The RACK module incorporates a number of
 159  * TCP ideas that have been put out into the IETF
 160  * over the last few years:
 161  * - Matt Mathis's Rate Halving which slowly drops
 162  *    the congestion window so that the ack clock can
 163  *    be maintained during a recovery.
 164  * - Yuchung Cheng's RACK TCP (for which its named) that
 165  *    will stop us using the number of dup acks and instead
 166  *    use time as the gage of when we retransmit.
 167  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
 168  *    of Dukkipati et.al.
 169  * RACK depends on SACK, so if an endpoint arrives that
 170  * cannot do SACK the state machine below will shuttle the
 171  * connection back to using the "default" TCP stack that is
 172  * in FreeBSD.
 173  *
 174  * To implement RACK the original TCP stack was first decomposed
 175  * into a functional state machine with individual states
 176  * for each of the possible TCP connection states. The do_segment
 177  * functions role in life is to mandate the connection supports SACK
 178  * initially and then assure that the RACK state matches the conenction
 179  * state before calling the states do_segment function. Each
 180  * state is simplified due to the fact that the original do_segment
 181  * has been decomposed and we *know* what state we are in (no
 182  * switches on the state) and all tests for SACK are gone. This
 183  * greatly simplifies what each state does.
 184  *
 185  * TCP output is also over-written with a new version since it
 186  * must maintain the new rack scoreboard.
 187  *
 188  */
 189 static int32_t rack_tlp_thresh = 1;
 190 static int32_t rack_tlp_limit = 2;      /* No more than 2 TLPs w-out new data */
 191 static int32_t rack_tlp_use_greater = 1;
 192 static int32_t rack_reorder_thresh = 2;
 193 static int32_t rack_reorder_fade = 60000000;    /* 0 - never fade, def 60,000,000
 194                                                  * - 60 seconds */
 195 static uint32_t rack_clamp_ss_upper = 110;
 196 static uint32_t rack_clamp_ca_upper = 105;
 197 static uint32_t rack_rxt_min_rnds = 10; /* Min rounds if drastic rxt clamp is in place */
 198 static uint32_t rack_unclamp_round_thresh = 100;        /* number of perfect rounds before we unclamp */
 199 static uint32_t rack_unclamp_rxt_thresh = 5;    /* .5%  and under */
 200 static uint64_t rack_rxt_clamp_thresh = 0;      /* Do we do the rxt clamp thing */
 201 static int32_t rack_dnd_default = 0;            /* For rr_conf = 3, what is the default for dnd */
 202 static int32_t rack_rxt_controls = 0;
 203 static int32_t rack_fill_cw_state = 0;
 204 static uint8_t rack_req_measurements = 1;
 205 /* Attack threshold detections */
 206 static uint32_t rack_highest_sack_thresh_seen = 0;
 207 static uint32_t rack_highest_move_thresh_seen = 0;
 208 static uint32_t rack_merge_out_sacks_on_attack = 0;
 209 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
 210 static int32_t rack_hw_pace_extra_slots = 0;    /* 2 extra MSS time betweens */
 211 static int32_t rack_hw_rate_caps = 0; /* 1; */
 212 static int32_t rack_hw_rate_cap_per = 0;        /* 0 -- off  */
 213 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
 214 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
 215 static int32_t rack_hw_up_only = 0;
 216 static int32_t rack_stats_gets_ms_rtt = 1;
 217 static int32_t rack_prr_addbackmax = 2;
 218 static int32_t rack_do_hystart = 0;
 219 static int32_t rack_apply_rtt_with_reduced_conf = 0;
 220 static int32_t rack_hibeta_setting = 0;
 221 static int32_t rack_default_pacing_divisor = 250;
 222 static int32_t rack_uses_full_dgp_in_rec = 1;
 223 static uint16_t rack_pacing_min_seg = 0;
 224
 225
 226 static uint32_t sad_seg_size_per = 800; /* 80.0 % */
 227 static int32_t rack_pkt_delay = 1000;
 228 static int32_t rack_send_a_lot_in_prr = 1;
 229 static int32_t rack_min_to = 1000;      /* Number of microsecond  min timeout */
 230 static int32_t rack_verbose_logging = 0;
 231 static int32_t rack_ignore_data_after_close = 1;
 232 static int32_t rack_enable_shared_cwnd = 1;
 233 static int32_t rack_use_cmp_acks = 1;
 234 static int32_t rack_use_fsb = 1;
 235 static int32_t rack_use_rfo = 1;
 236 static int32_t rack_use_rsm_rfo = 1;
 237 static int32_t rack_max_abc_post_recovery = 2;
 238 static int32_t rack_client_low_buf = 0;
 239 static int32_t rack_dsack_std_based = 0x3;      /* bit field bit 1 sets rc_rack_tmr_std_based and bit 2 sets rc_rack_use_dsack */
 240 static int32_t rack_bw_multipler = 2;           /* Limit on fill cw's jump up to be this x gp_est */
 241 #ifdef TCP_ACCOUNTING
 242 static int32_t rack_tcp_accounting = 0;
 243 #endif
 244 static int32_t rack_limits_scwnd = 1;
 245 static int32_t rack_enable_mqueue_for_nonpaced = 0;
 246 static int32_t rack_hybrid_allow_set_maxseg = 0;
 247 static int32_t rack_disable_prr = 0;
 248 static int32_t use_rack_rr = 1;
 249 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
 250 static int32_t rack_persist_min = 250000;       /* 250usec */
 251 static int32_t rack_persist_max = 2000000;      /* 2 Second in usec's */
 252 static int32_t rack_sack_not_required = 1;      /* set to one to allow non-sack to use rack */
 253 static int32_t rack_default_init_window = 0;    /* Use system default */
 254 static int32_t rack_limit_time_with_srtt = 0;
 255 static int32_t rack_autosndbuf_inc = 20;        /* In percentage form */
 256 static int32_t rack_enobuf_hw_boost_mult = 0;   /* How many times the hw rate we boost slot using time_between */
 257 static int32_t rack_enobuf_hw_max = 12000;      /* 12 ms in usecs */
 258 static int32_t rack_enobuf_hw_min = 10000;      /* 10 ms in usecs */
 259 static int32_t rack_hw_rwnd_factor = 2;         /* How many max_segs the rwnd must be before we hold off sending */
 260 static int32_t rack_hw_check_queue = 0;         /* Do we always pre-check queue depth of a hw queue */
 261 static int32_t rack_full_buffer_discount = 10;
 262 /*
 263  * Currently regular tcp has a rto_min of 30ms
 264  * the backoff goes 12 times so that ends up
 265  * being a total of 122.850 seconds before a
 266  * connection is killed.
 267  */
 268 static uint32_t rack_def_data_window = 20;
 269 static uint32_t rack_goal_bdp = 2;
 270 static uint32_t rack_min_srtts = 1;
 271 static uint32_t rack_min_measure_usec = 0;
 272 static int32_t rack_tlp_min = 10000;    /* 10ms */
 273 static int32_t rack_rto_min = 30000;    /* 30,000 usec same as main freebsd */
 274 static int32_t rack_rto_max = 4000000;  /* 4 seconds in usec's */
 275 static const int32_t rack_free_cache = 2;
 276 static int32_t rack_hptsi_segments = 40;
 277 static int32_t rack_rate_sample_method = USE_RTT_LOW;
 278 static int32_t rack_pace_every_seg = 0;
 279 static int32_t rack_delayed_ack_time = 40000;   /* 40ms in usecs */
 280 static int32_t rack_slot_reduction = 4;
 281 static int32_t rack_wma_divisor = 8;            /* For WMA calculation */
 282 static int32_t rack_cwnd_block_ends_measure = 0;
 283 static int32_t rack_rwnd_block_ends_measure = 0;
 284 static int32_t rack_def_profile = 0;
 285
 286 static int32_t rack_lower_cwnd_at_tlp = 0;
 287 static int32_t rack_limited_retran = 0;
 288 static int32_t rack_always_send_oldest = 0;
 289 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
 290
 291 static uint16_t rack_per_of_gp_ss = 250;        /* 250 % slow-start */
 292 static uint16_t rack_per_of_gp_ca = 200;        /* 200 % congestion-avoidance */
 293 static uint16_t rack_per_of_gp_rec = 200;       /* 200 % of bw */
 294
 295 /* Probertt */
 296 static uint16_t rack_per_of_gp_probertt = 60;   /* 60% of bw */
 297 static uint16_t rack_per_of_gp_lowthresh = 40;  /* 40% is bottom */
 298 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
 299 static uint16_t rack_atexit_prtt_hbp = 130;     /* Clamp to 130% on exit prtt if highly buffered path */
 300 static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */
 301
 302 static uint32_t rack_max_drain_wait = 2;        /* How man gp srtt's before we give up draining */
 303 static uint32_t rack_must_drain = 1;            /* How many GP srtt's we *must* wait */
 304 static uint32_t rack_probertt_use_min_rtt_entry = 1;    /* Use the min to calculate the goal else gp_srtt */
 305 static uint32_t rack_probertt_use_min_rtt_exit = 0;
 306 static uint32_t rack_probe_rtt_sets_cwnd = 0;
 307 static uint32_t rack_probe_rtt_safety_val = 2000000;    /* No more than 2 sec in probe-rtt */
 308 static uint32_t rack_time_between_probertt = 9600000;   /* 9.6 sec in usecs */
 309 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;       /* How many srtt periods does probe-rtt last top fraction */
 310 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;       /* How many srtt periods does probe-rtt last bottom fraction */
 311 static uint32_t rack_min_probertt_hold = 40000;         /* Equal to delayed ack time */
 312 static uint32_t rack_probertt_filter_life = 10000000;
 313 static uint32_t rack_probertt_lower_within = 10;
 314 static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds)  to count as a lowering */
 315 static int32_t rack_pace_one_seg = 0;           /* Shall we pace for less than 1.4Meg 1MSS at a time */
 316 static int32_t rack_probertt_clear_is = 1;
 317 static int32_t rack_max_drain_hbp = 1;          /* Extra drain times gpsrtt for highly buffered paths */
 318 static int32_t rack_hbp_thresh = 3;             /* what is the divisor max_rtt/min_rtt to decided a hbp */
 319
 320 /* Part of pacing */
 321 static int32_t rack_max_per_above = 30;         /* When we go to increment stop if above 100+this% */
 322
 323 /* Timely information */
 324 /* Combine these two gives the range of 'no change' to bw */
 325 /* ie the up/down provide the upper and lower bound */
 326 static int32_t rack_gp_per_bw_mul_up = 2;       /* 2% */
 327 static int32_t rack_gp_per_bw_mul_down = 4;     /* 4% */
 328 static int32_t rack_gp_rtt_maxmul = 3;          /* 3 x maxmin */
 329 static int32_t rack_gp_rtt_minmul = 1;          /* minrtt + (minrtt/mindiv) is lower rtt */
 330 static int32_t rack_gp_rtt_mindiv = 4;          /* minrtt + (minrtt * minmul/mindiv) is lower rtt */
 331 static int32_t rack_gp_decrease_per = 20;       /* 20% decrease in multiplier */
 332 static int32_t rack_gp_increase_per = 2;        /* 2% increase in multiplier */
 333 static int32_t rack_per_lower_bound = 50;       /* Don't allow to drop below this multiplier */
 334 static int32_t rack_per_upper_bound_ss = 0;     /* Don't allow SS to grow above this */
 335 static int32_t rack_per_upper_bound_ca = 0;     /* Don't allow CA to grow above this */
 336 static int32_t rack_do_dyn_mul = 0;             /* Are the rack gp multipliers dynamic */
 337 static int32_t rack_gp_no_rec_chg = 1;          /* Prohibit recovery from reducing it's multiplier */
 338 static int32_t rack_timely_dec_clear = 6;       /* Do we clear decrement count at a value (6)? */
 339 static int32_t rack_timely_max_push_rise = 3;   /* One round of pushing */
 340 static int32_t rack_timely_max_push_drop = 3;   /* Three round of pushing */
 341 static int32_t rack_timely_min_segs = 4;        /* 4 segment minimum */
 342 static int32_t rack_use_max_for_nobackoff = 0;
 343 static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */
 344 static int32_t rack_timely_no_stopping = 0;
 345 static int32_t rack_down_raise_thresh = 100;
 346 static int32_t rack_req_segs = 1;
 347 static uint64_t rack_bw_rate_cap = 0;
 348
 349
 350 /* Rack specific counters */
 351 counter_u64_t rack_saw_enobuf;
 352 counter_u64_t rack_saw_enobuf_hw;
 353 counter_u64_t rack_saw_enetunreach;
 354 counter_u64_t rack_persists_sends;
 355 counter_u64_t rack_persists_acks;
 356 counter_u64_t rack_persists_loss;
 357 counter_u64_t rack_persists_lost_ends;
 358 counter_u64_t rack_total_bytes;
 359 #ifdef INVARIANTS
 360 counter_u64_t rack_adjust_map_bw;
 361 #endif
 362 /* Tail loss probe counters */
 363 counter_u64_t rack_tlp_tot;
 364 counter_u64_t rack_tlp_newdata;
 365 counter_u64_t rack_tlp_retran;
 366 counter_u64_t rack_tlp_retran_bytes;
 367 counter_u64_t rack_to_tot;
 368 counter_u64_t rack_hot_alloc;
 369 counter_u64_t rack_to_alloc;
 370 counter_u64_t rack_to_alloc_hard;
 371 counter_u64_t rack_to_alloc_emerg;
 372 counter_u64_t rack_to_alloc_limited;
 373 counter_u64_t rack_alloc_limited_conns;
 374 counter_u64_t rack_split_limited;
 375 counter_u64_t rack_rxt_clamps_cwnd;
 376 counter_u64_t rack_rxt_clamps_cwnd_uniq;
 377
 378 counter_u64_t rack_multi_single_eq;
 379 counter_u64_t rack_proc_non_comp_ack;
 380
 381 counter_u64_t rack_fto_send;
 382 counter_u64_t rack_fto_rsm_send;
 383 counter_u64_t rack_nfto_resend;
 384 counter_u64_t rack_non_fto_send;
 385 counter_u64_t rack_extended_rfo;
 386
 387 counter_u64_t rack_sack_proc_all;
 388 counter_u64_t rack_sack_proc_short;
 389 counter_u64_t rack_sack_proc_restart;
 390 counter_u64_t rack_sack_attacks_detected;
 391 counter_u64_t rack_sack_attacks_reversed;
 392 counter_u64_t rack_sack_attacks_suspect;
 393 counter_u64_t rack_sack_used_next_merge;
 394 counter_u64_t rack_sack_splits;
 395 counter_u64_t rack_sack_used_prev_merge;
 396 counter_u64_t rack_sack_skipped_acked;
 397 counter_u64_t rack_ack_total;
 398 counter_u64_t rack_express_sack;
 399 counter_u64_t rack_sack_total;
 400 counter_u64_t rack_move_none;
 401 counter_u64_t rack_move_some;
 402
 403 counter_u64_t rack_input_idle_reduces;
 404 counter_u64_t rack_collapsed_win;
 405 counter_u64_t rack_collapsed_win_seen;
 406 counter_u64_t rack_collapsed_win_rxt;
 407 counter_u64_t rack_collapsed_win_rxt_bytes;
 408 counter_u64_t rack_try_scwnd;
 409 counter_u64_t rack_hw_pace_init_fail;
 410 counter_u64_t rack_hw_pace_lost;
 411
 412 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
 413 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
 414
 415
 416 #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
 417
 418 #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {  \
 419         (tv) = (value) + slop;   \
 420         if ((u_long)(tv) < (u_long)(tvmin)) \
 421                 (tv) = (tvmin); \
 422         if ((u_long)(tv) > (u_long)(tvmax)) \
 423                 (tv) = (tvmax); \
 424 } while (0)
 425
 426 static void
 427 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
 428
 429 static int
 430 rack_process_ack(struct mbuf *m, struct tcphdr *th,
 431     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
 432     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
 433 static int
 434 rack_process_data(struct mbuf *m, struct tcphdr *th,
 435     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 436     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
 437 static void
 438 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
 439    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
 440 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
 441 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
 442     uint8_t limit_type);
 443 static struct rack_sendmap *
 444 rack_check_recovery_mode(struct tcpcb *tp,
 445     uint32_t tsused);
 446 static void
 447 rack_cong_signal(struct tcpcb *tp,
 448                  uint32_t type, uint32_t ack, int );
 449 static void rack_counter_destroy(void);
 450 static int
 451 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
 452 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
 453 static void
 454 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
 455 static void
 456 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
 457     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos);
 458 static void rack_dtor(void *mem, int32_t size, void *arg);
 459 static void
 460 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
 461     uint32_t flex1, uint32_t flex2,
 462     uint32_t flex3, uint32_t flex4,
 463     uint32_t flex5, uint32_t flex6,
 464     uint16_t flex7, uint8_t mod);
 465
 466 static void
 467 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
 468    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
 469    struct rack_sendmap *rsm, uint8_t quality);
 470 static struct rack_sendmap *
 471 rack_find_high_nonack(struct tcp_rack *rack,
 472     struct rack_sendmap *rsm);
 473 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
 474 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
 475 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
 476 static int rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt);
 477 static void
 478 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
 479                             tcp_seq th_ack, int line, uint8_t quality);
 480 static void
 481 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm);
 482
 483 static uint32_t
 484 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
 485 static int32_t rack_handoff_ok(struct tcpcb *tp);
 486 static int32_t rack_init(struct tcpcb *tp, void **ptr);
 487 static void rack_init_sysctls(void);
 488
 489 static void
 490 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
 491     struct tcphdr *th, int entered_rec, int dup_ack_struck,
 492     int *dsack_seen, int *sacks_seen);
 493 static void
 494 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
 495     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
 496     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls, int segsiz);
 497
 498 static uint64_t rack_get_gp_est(struct tcp_rack *rack);
 499
 500 static void
 501 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
 502     struct rack_sendmap *rsm);
 503 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
 504 static int32_t rack_output(struct tcpcb *tp);
 505
 506 static uint32_t
 507 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
 508     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
 509     uint32_t cts, int *no_extra, int *moved_two, uint32_t segsiz);
 510 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
 511 static void rack_remxt_tmr(struct tcpcb *tp);
 512 static int rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt);
 513 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
 514 static int32_t rack_stopall(struct tcpcb *tp);
 515 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
 516 static uint32_t
 517 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
 518     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag, int segsiz);
 519 static void
 520 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
 521     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag, int segsiz);
 522 static int
 523 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
 524     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
 525 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
 526 static int
 527 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
 528     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 529     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 530 static int
 531 rack_do_closing(struct mbuf *m, struct tcphdr *th,
 532     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 533     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 534 static int
 535 rack_do_established(struct mbuf *m, struct tcphdr *th,
 536     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 537     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 538 static int
 539 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
 540     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 541     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
 542 static int
 543 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
 544     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 545     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 546 static int
 547 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
 548     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 549     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 550 static int
 551 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
 552     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 553     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 554 static int
 555 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
 556     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 557     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 558 static int
 559 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
 560     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 561     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 562 static void rack_chk_http_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts);
 563 struct rack_sendmap *
 564 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
 565     uint32_t tsused);
 566 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
 567     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
 568 static void
 569      tcp_rack_partialack(struct tcpcb *tp);
 570 static int
 571 rack_set_profile(struct tcp_rack *rack, int prof);
 572 static void
 573 rack_apply_deferred_options(struct tcp_rack *rack);
 574
 575 int32_t rack_clear_counter=0;
 576
 577 static uint64_t
 578 rack_get_lt_bw(struct tcp_rack *rack)
 579 {
 580         struct timeval tv;
 581         uint64_t tim, bytes;
 582
 583         tim = rack->r_ctl.lt_bw_time;
 584         bytes = rack->r_ctl.lt_bw_bytes;
 585         if (rack->lt_bw_up) {
 586                 /* Include all the current bytes too */
 587                 microuptime(&tv);
 588                 bytes += (rack->rc_tp->snd_una - rack->r_ctl.lt_seq);
 589                 tim += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
 590         }
 591         if ((bytes != 0) && (tim != 0))
 592                 return ((bytes * (uint64_t)1000000) / tim);
 593         else
 594                 return (0);
 595 }
 596
 597 static void
 598 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
 599 {
 600         struct sockopt sopt;
 601         struct cc_newreno_opts opt;
 602         struct newreno old;
 603         struct tcpcb *tp;
 604         int error, failed = 0;
 605
 606         tp = rack->rc_tp;
 607         if (tp->t_cc == NULL) {
 608                 /* Tcb is leaving */
 609                 return;
 610         }
 611         rack->rc_pacing_cc_set = 1;
 612         if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
 613                 /* Not new-reno we can't play games with beta! */
 614                 failed = 1;
 615                 goto out;
 616
 617         }
 618         if (CC_ALGO(tp)->ctl_output == NULL)  {
 619                 /* Huh, not using new-reno so no swaps.? */
 620                 failed = 2;
 621                 goto out;
 622         }
 623         /* Get the current values out */
 624         sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
 625         sopt.sopt_dir = SOPT_GET;
 626         opt.name = CC_NEWRENO_BETA;
 627         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 628         if (error)  {
 629                 failed = 3;
 630                 goto out;
 631         }
 632         old.beta = opt.val;
 633         opt.name = CC_NEWRENO_BETA_ECN;
 634         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 635         if (error)  {
 636                 failed = 4;
 637                 goto out;
 638         }
 639         old.beta_ecn = opt.val;
 640
 641         /* Now lets set in the values we have stored */
 642         sopt.sopt_dir = SOPT_SET;
 643         opt.name = CC_NEWRENO_BETA;
 644         opt.val = rack->r_ctl.rc_saved_beta.beta;
 645         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 646         if (error)  {
 647                 failed = 5;
 648                 goto out;
 649         }
 650         opt.name = CC_NEWRENO_BETA_ECN;
 651         opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
 652         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 653         if (error) {
 654                 failed = 6;
 655                 goto out;
 656         }
 657         /* Save off the values for restoral */
 658         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 659 out:
 660         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
 661                 union tcp_log_stackspecific log;
 662                 struct timeval tv;
 663                 struct newreno *ptr;
 664
 665                 ptr = ((struct newreno *)tp->t_ccv.cc_data);
 666                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 667                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 668                 log.u_bbr.flex1 = ptr->beta;
 669                 log.u_bbr.flex2 = ptr->beta_ecn;
 670                 log.u_bbr.flex3 = ptr->newreno_flags;
 671                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 672                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 673                 log.u_bbr.flex6 = failed;
 674                 log.u_bbr.flex7 = rack->gp_ready;
 675                 log.u_bbr.flex7 <<= 1;
 676                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 677                 log.u_bbr.flex7 <<= 1;
 678                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 679                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 680                 log.u_bbr.flex8 = flex8;
 681                 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
 682                                0, &log, false, NULL, NULL, 0, &tv);
 683         }
 684 }
 685
 686 static void
 687 rack_set_cc_pacing(struct tcp_rack *rack)
 688 {
 689         if (rack->rc_pacing_cc_set)
 690                 return;
 691         /*
 692          * Use the swap utility placing in 3 for flex8 to id a
 693          * set of a new set of values.
 694          */
 695         rack->rc_pacing_cc_set = 1;
 696         rack_swap_beta_values(rack, 3);
 697 }
 698
 699 static void
 700 rack_undo_cc_pacing(struct tcp_rack *rack)
 701 {
 702         if (rack->rc_pacing_cc_set == 0)
 703                 return;
 704         /*
 705          * Use the swap utility placing in 4 for flex8 to id a
 706          * restoral of the old values.
 707          */
 708         rack->rc_pacing_cc_set = 0;
 709         rack_swap_beta_values(rack, 4);
 710 }
 711
 712 static void
 713 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
 714                uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
 715 {
 716         if (tcp_bblogging_on(rack->rc_tp)) {
 717                 union tcp_log_stackspecific log;
 718                 struct timeval tv;
 719
 720                 memset(&log, 0, sizeof(log));
 721                 log.u_bbr.flex1 = seq_end;
 722                 log.u_bbr.flex2 = rack->rc_tp->gput_seq;
 723                 log.u_bbr.flex3 = ack_end_t;
 724                 log.u_bbr.flex4 = rack->rc_tp->gput_ts;
 725                 log.u_bbr.flex5 = send_end_t;
 726                 log.u_bbr.flex6 = rack->rc_tp->gput_ack;
 727                 log.u_bbr.flex7 = mode;
 728                 log.u_bbr.flex8 = 69;
 729                 log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
 730                 log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
 731                 log.u_bbr.pkts_out = line;
 732                 log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
 733                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
 734                 if (rsm != NULL) {
 735                         log.u_bbr.applimited = rsm->r_start;
 736                         log.u_bbr.delivered = rsm->r_end;
 737                         log.u_bbr.epoch = rsm->r_flags;
 738                 }
 739                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 740                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
 741                     &rack->rc_inp->inp_socket->so_rcv,
 742                     &rack->rc_inp->inp_socket->so_snd,
 743                     BBR_LOG_HPTSI_CALC, 0,
 744                     0, &log, false, &tv);
 745         }
 746 }
 747
 748 static int
 749 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
 750 {
 751         uint32_t stat;
 752         int32_t error;
 753
 754         error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
 755         if (error || req->newptr == NULL)
 756                 return error;
 757
 758         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
 759         if (error)
 760                 return (error);
 761         if (stat == 1) {
 762 #ifdef INVARIANTS
 763                 printf("Clearing RACK counters\n");
 764 #endif
 765                 counter_u64_zero(rack_tlp_tot);
 766                 counter_u64_zero(rack_tlp_newdata);
 767                 counter_u64_zero(rack_tlp_retran);
 768                 counter_u64_zero(rack_tlp_retran_bytes);
 769                 counter_u64_zero(rack_to_tot);
 770                 counter_u64_zero(rack_saw_enobuf);
 771                 counter_u64_zero(rack_saw_enobuf_hw);
 772                 counter_u64_zero(rack_saw_enetunreach);
 773                 counter_u64_zero(rack_persists_sends);
 774                 counter_u64_zero(rack_total_bytes);
 775                 counter_u64_zero(rack_persists_acks);
 776                 counter_u64_zero(rack_persists_loss);
 777                 counter_u64_zero(rack_persists_lost_ends);
 778 #ifdef INVARIANTS
 779                 counter_u64_zero(rack_adjust_map_bw);
 780 #endif
 781                 counter_u64_zero(rack_to_alloc_hard);
 782                 counter_u64_zero(rack_to_alloc_emerg);
 783                 counter_u64_zero(rack_sack_proc_all);
 784                 counter_u64_zero(rack_fto_send);
 785                 counter_u64_zero(rack_fto_rsm_send);
 786                 counter_u64_zero(rack_extended_rfo);
 787                 counter_u64_zero(rack_hw_pace_init_fail);
 788                 counter_u64_zero(rack_hw_pace_lost);
 789                 counter_u64_zero(rack_non_fto_send);
 790                 counter_u64_zero(rack_nfto_resend);
 791                 counter_u64_zero(rack_sack_proc_short);
 792                 counter_u64_zero(rack_sack_proc_restart);
 793                 counter_u64_zero(rack_to_alloc);
 794                 counter_u64_zero(rack_to_alloc_limited);
 795                 counter_u64_zero(rack_alloc_limited_conns);
 796                 counter_u64_zero(rack_split_limited);
 797                 counter_u64_zero(rack_rxt_clamps_cwnd);
 798                 counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
 799                 counter_u64_zero(rack_multi_single_eq);
 800                 counter_u64_zero(rack_proc_non_comp_ack);
 801                 counter_u64_zero(rack_sack_attacks_detected);
 802                 counter_u64_zero(rack_sack_attacks_reversed);
 803                 counter_u64_zero(rack_sack_attacks_suspect);
 804                 counter_u64_zero(rack_sack_used_next_merge);
 805                 counter_u64_zero(rack_sack_used_prev_merge);
 806                 counter_u64_zero(rack_sack_splits);
 807                 counter_u64_zero(rack_sack_skipped_acked);
 808                 counter_u64_zero(rack_ack_total);
 809                 counter_u64_zero(rack_express_sack);
 810                 counter_u64_zero(rack_sack_total);
 811                 counter_u64_zero(rack_move_none);
 812                 counter_u64_zero(rack_move_some);
 813                 counter_u64_zero(rack_try_scwnd);
 814                 counter_u64_zero(rack_collapsed_win);
 815                 counter_u64_zero(rack_collapsed_win_rxt);
 816                 counter_u64_zero(rack_collapsed_win_seen);
 817                 counter_u64_zero(rack_collapsed_win_rxt_bytes);
 818         } else if (stat == 2) {
 819 #ifdef INVARIANTS
 820                 printf("Clearing RACK option array\n");
 821 #endif
 822                 COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
 823         } else if (stat == 3) {
 824                 printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
 825         } else if (stat == 4) {
 826 #ifdef INVARIANTS
 827                 printf("Clearing RACK out size array\n");
 828 #endif
 829                 COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
 830         }
 831         rack_clear_counter = 0;
 832         return (0);
 833 }
 834
 835 static void
 836 rack_init_sysctls(void)
 837 {
 838         struct sysctl_oid *rack_counters;
 839         struct sysctl_oid *rack_attack;
 840         struct sysctl_oid *rack_pacing;
 841         struct sysctl_oid *rack_timely;
 842         struct sysctl_oid *rack_timers;
 843         struct sysctl_oid *rack_tlp;
 844         struct sysctl_oid *rack_misc;
 845         struct sysctl_oid *rack_features;
 846         struct sysctl_oid *rack_measure;
 847         struct sysctl_oid *rack_probertt;
 848         struct sysctl_oid *rack_hw_pacing;
 849
 850         rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 851             SYSCTL_CHILDREN(rack_sysctl_root),
 852             OID_AUTO,
 853             "sack_attack",
 854             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 855             "Rack Sack Attack Counters and Controls");
 856         rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 857             SYSCTL_CHILDREN(rack_sysctl_root),
 858             OID_AUTO,
 859             "stats",
 860             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 861             "Rack Counters");
 862         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 863             SYSCTL_CHILDREN(rack_sysctl_root),
 864             OID_AUTO, "rate_sample_method", CTLFLAG_RW,
 865             &rack_rate_sample_method , USE_RTT_LOW,
 866             "What method should we use for rate sampling 0=high, 1=low ");
 867         /* Probe rtt related controls */
 868         rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 869             SYSCTL_CHILDREN(rack_sysctl_root),
 870             OID_AUTO,
 871             "probertt",
 872             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 873             "ProbeRTT related Controls");
 874         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 875             SYSCTL_CHILDREN(rack_probertt),
 876             OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
 877             &rack_atexit_prtt_hbp, 130,
 878             "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
 879         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 880             SYSCTL_CHILDREN(rack_probertt),
 881             OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
 882             &rack_atexit_prtt, 130,
 883             "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
 884         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 885             SYSCTL_CHILDREN(rack_probertt),
 886             OID_AUTO, "gp_per_mul", CTLFLAG_RW,
 887             &rack_per_of_gp_probertt, 60,
 888             "What percentage of goodput do we pace at in probertt");
 889         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 890             SYSCTL_CHILDREN(rack_probertt),
 891             OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
 892             &rack_per_of_gp_probertt_reduce, 10,
 893             "What percentage of goodput do we reduce every gp_srtt");
 894         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 895             SYSCTL_CHILDREN(rack_probertt),
 896             OID_AUTO, "gp_per_low", CTLFLAG_RW,
 897             &rack_per_of_gp_lowthresh, 40,
 898             "What percentage of goodput do we allow the multiplier to fall to");
 899         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 900             SYSCTL_CHILDREN(rack_probertt),
 901             OID_AUTO, "time_between", CTLFLAG_RW,
 902             & rack_time_between_probertt, 96000000,
 903             "How many useconds between the lowest rtt falling must past before we enter probertt");
 904         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 905             SYSCTL_CHILDREN(rack_probertt),
 906             OID_AUTO, "safety", CTLFLAG_RW,
 907             &rack_probe_rtt_safety_val, 2000000,
 908             "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
 909         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 910             SYSCTL_CHILDREN(rack_probertt),
 911             OID_AUTO, "sets_cwnd", CTLFLAG_RW,
 912             &rack_probe_rtt_sets_cwnd, 0,
 913             "Do we set the cwnd too (if always_lower is on)");
 914         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 915             SYSCTL_CHILDREN(rack_probertt),
 916             OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
 917             &rack_max_drain_wait, 2,
 918             "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
 919         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 920             SYSCTL_CHILDREN(rack_probertt),
 921             OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
 922             &rack_must_drain, 1,
 923             "We must drain this many gp_srtt's waiting for flight to reach goal");
 924         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 925             SYSCTL_CHILDREN(rack_probertt),
 926             OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
 927             &rack_probertt_use_min_rtt_entry, 1,
 928             "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
 929         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 930             SYSCTL_CHILDREN(rack_probertt),
 931             OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
 932             &rack_probertt_use_min_rtt_exit, 0,
 933             "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
 934         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 935             SYSCTL_CHILDREN(rack_probertt),
 936             OID_AUTO, "length_div", CTLFLAG_RW,
 937             &rack_probertt_gpsrtt_cnt_div, 0,
 938             "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
 939         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 940             SYSCTL_CHILDREN(rack_probertt),
 941             OID_AUTO, "length_mul", CTLFLAG_RW,
 942             &rack_probertt_gpsrtt_cnt_mul, 0,
 943             "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
 944         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 945             SYSCTL_CHILDREN(rack_probertt),
 946             OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
 947             &rack_min_probertt_hold, 200000,
 948             "What is the minimum time we hold probertt at target");
 949         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 950             SYSCTL_CHILDREN(rack_probertt),
 951             OID_AUTO, "filter_life", CTLFLAG_RW,
 952             &rack_probertt_filter_life, 10000000,
 953             "What is the time for the filters life in useconds");
 954         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 955             SYSCTL_CHILDREN(rack_probertt),
 956             OID_AUTO, "lower_within", CTLFLAG_RW,
 957             &rack_probertt_lower_within, 10,
 958             "If the rtt goes lower within this percentage of the time, go into probe-rtt");
 959         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 960             SYSCTL_CHILDREN(rack_probertt),
 961             OID_AUTO, "must_move", CTLFLAG_RW,
 962             &rack_min_rtt_movement, 250,
 963             "How much is the minimum movement in rtt to count as a drop for probertt purposes");
 964         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 965             SYSCTL_CHILDREN(rack_probertt),
 966             OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
 967             &rack_probertt_clear_is, 1,
 968             "Do we clear I/S counts on exiting probe-rtt");
 969         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 970             SYSCTL_CHILDREN(rack_probertt),
 971             OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
 972             &rack_max_drain_hbp, 1,
 973             "How many extra drain gpsrtt's do we get in highly buffered paths");
 974         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 975             SYSCTL_CHILDREN(rack_probertt),
 976             OID_AUTO, "hbp_threshold", CTLFLAG_RW,
 977             &rack_hbp_thresh, 3,
 978             "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
 979         /* Pacing related sysctls */
 980         rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 981             SYSCTL_CHILDREN(rack_sysctl_root),
 982             OID_AUTO,
 983             "pacing",
 984             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 985             "Pacing related Controls");
 986         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 987             SYSCTL_CHILDREN(rack_pacing),
 988             OID_AUTO, "fulldgpinrec", CTLFLAG_RW,
 989             &rack_uses_full_dgp_in_rec, 1,
 990             "Do we use all DGP features in recovery (fillcw, timely et.al.)?");
 991         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 992             SYSCTL_CHILDREN(rack_pacing),
 993             OID_AUTO, "fullbufdisc", CTLFLAG_RW,
 994             &rack_full_buffer_discount, 10,
 995             "What percentage b/w reduction over the GP estimate for a full buffer (default=0 off)?");
 996         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 997             SYSCTL_CHILDREN(rack_pacing),
 998             OID_AUTO, "fillcw", CTLFLAG_RW,
 999             &rack_fill_cw_state, 0,
1000             "Enable fillcw on new connections (default=0 off)?");
1001         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1002             SYSCTL_CHILDREN(rack_pacing),
1003             OID_AUTO, "min_burst", CTLFLAG_RW,
1004             &rack_pacing_min_seg, 0,
1005             "What is the min burst size for pacing (0 disables)?");
1006         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1007             SYSCTL_CHILDREN(rack_pacing),
1008             OID_AUTO, "divisor", CTLFLAG_RW,
1009             &rack_default_pacing_divisor, 4,
1010             "What is the default divisor given to the rl code?");
1011         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1012             SYSCTL_CHILDREN(rack_pacing),
1013             OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1014             &rack_bw_multipler, 2,
1015             "What is the multiplier of the current gp_est that fillcw can increase the b/w too?");
1016         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1017             SYSCTL_CHILDREN(rack_pacing),
1018             OID_AUTO, "max_pace_over", CTLFLAG_RW,
1019             &rack_max_per_above, 30,
1020             "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1021         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1022             SYSCTL_CHILDREN(rack_pacing),
1023             OID_AUTO, "allow1mss", CTLFLAG_RW,
1024             &rack_pace_one_seg, 0,
1025             "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1026         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1027             SYSCTL_CHILDREN(rack_pacing),
1028             OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1029             &rack_limit_time_with_srtt, 0,
1030             "Do we limit pacing time based on srtt");
1031         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1032             SYSCTL_CHILDREN(rack_pacing),
1033             OID_AUTO, "init_win", CTLFLAG_RW,
1034             &rack_default_init_window, 0,
1035             "Do we have a rack initial window 0 = system default");
1036         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1037             SYSCTL_CHILDREN(rack_pacing),
1038             OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1039             &rack_per_of_gp_ss, 250,
1040             "If non zero, what percentage of goodput to pace at in slow start");
1041         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1042             SYSCTL_CHILDREN(rack_pacing),
1043             OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1044             &rack_per_of_gp_ca, 150,
1045             "If non zero, what percentage of goodput to pace at in congestion avoidance");
1046         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1047             SYSCTL_CHILDREN(rack_pacing),
1048             OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1049             &rack_per_of_gp_rec, 200,
1050             "If non zero, what percentage of goodput to pace at in recovery");
1051         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1052             SYSCTL_CHILDREN(rack_pacing),
1053             OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1054             &rack_hptsi_segments, 40,
1055             "What size is the max for TSO segments in pacing and burst mitigation");
1056         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1057             SYSCTL_CHILDREN(rack_pacing),
1058             OID_AUTO, "burst_reduces", CTLFLAG_RW,
1059             &rack_slot_reduction, 4,
1060             "When doing only burst mitigation what is the reduce divisor");
1061         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1062             SYSCTL_CHILDREN(rack_sysctl_root),
1063             OID_AUTO, "use_pacing", CTLFLAG_RW,
1064             &rack_pace_every_seg, 0,
1065             "If set we use pacing, if clear we use only the original burst mitigation");
1066         SYSCTL_ADD_U64(&rack_sysctl_ctx,
1067             SYSCTL_CHILDREN(rack_pacing),
1068             OID_AUTO, "rate_cap", CTLFLAG_RW,
1069             &rack_bw_rate_cap, 0,
1070             "If set we apply this value to the absolute rate cap used by pacing");
1071         SYSCTL_ADD_U8(&rack_sysctl_ctx,
1072             SYSCTL_CHILDREN(rack_sysctl_root),
1073             OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1074             &rack_req_measurements, 1,
1075             "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1076         /* Hardware pacing */
1077         rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1078             SYSCTL_CHILDREN(rack_sysctl_root),
1079             OID_AUTO,
1080             "hdwr_pacing",
1081             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1082             "Pacing related Controls");
1083         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1084             SYSCTL_CHILDREN(rack_hw_pacing),
1085             OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1086             &rack_hw_rwnd_factor, 2,
1087             "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1088         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1089             SYSCTL_CHILDREN(rack_hw_pacing),
1090             OID_AUTO, "precheck", CTLFLAG_RW,
1091             &rack_hw_check_queue, 0,
1092             "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1093         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1094             SYSCTL_CHILDREN(rack_hw_pacing),
1095             OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1096             &rack_enobuf_hw_boost_mult, 0,
1097             "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1098         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1099             SYSCTL_CHILDREN(rack_hw_pacing),
1100             OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1101             &rack_enobuf_hw_max, 2,
1102             "What is the max boost the pacing time if we see a ENOBUFS?");
1103         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1104             SYSCTL_CHILDREN(rack_hw_pacing),
1105             OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1106             &rack_enobuf_hw_min, 2,
1107             "What is the min boost the pacing time if we see a ENOBUFS?");
1108         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1109             SYSCTL_CHILDREN(rack_hw_pacing),
1110             OID_AUTO, "enable", CTLFLAG_RW,
1111             &rack_enable_hw_pacing, 0,
1112             "Should RACK attempt to use hw pacing?");
1113         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1114             SYSCTL_CHILDREN(rack_hw_pacing),
1115             OID_AUTO, "rate_cap", CTLFLAG_RW,
1116             &rack_hw_rate_caps, 0,
1117             "Does the highest hardware pacing rate cap the rate we will send at??");
1118         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1119             SYSCTL_CHILDREN(rack_hw_pacing),
1120             OID_AUTO, "uncap_per", CTLFLAG_RW,
1121             &rack_hw_rate_cap_per, 0,
1122             "If you go over b/w by this amount you will be uncapped (0 = never)");
1123         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1124             SYSCTL_CHILDREN(rack_hw_pacing),
1125             OID_AUTO, "rate_min", CTLFLAG_RW,
1126             &rack_hw_rate_min, 0,
1127             "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1128         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1129             SYSCTL_CHILDREN(rack_hw_pacing),
1130             OID_AUTO, "rate_to_low", CTLFLAG_RW,
1131             &rack_hw_rate_to_low, 0,
1132             "If we fall below this rate, dis-engage hw pacing?");
1133         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1134             SYSCTL_CHILDREN(rack_hw_pacing),
1135             OID_AUTO, "up_only", CTLFLAG_RW,
1136             &rack_hw_up_only, 0,
1137             "Do we allow hw pacing to lower the rate selected?");
1138         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1139             SYSCTL_CHILDREN(rack_hw_pacing),
1140             OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1141             &rack_hw_pace_extra_slots, 0,
1142             "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1143         rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1144             SYSCTL_CHILDREN(rack_sysctl_root),
1145             OID_AUTO,
1146             "timely",
1147             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1148             "Rack Timely RTT Controls");
1149         /* Timely based GP dynmics */
1150         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1151             SYSCTL_CHILDREN(rack_timely),
1152             OID_AUTO, "upper", CTLFLAG_RW,
1153             &rack_gp_per_bw_mul_up, 2,
1154             "Rack timely upper range for equal b/w (in percentage)");
1155         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1156             SYSCTL_CHILDREN(rack_timely),
1157             OID_AUTO, "lower", CTLFLAG_RW,
1158             &rack_gp_per_bw_mul_down, 4,
1159             "Rack timely lower range for equal b/w (in percentage)");
1160         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1161             SYSCTL_CHILDREN(rack_timely),
1162             OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1163             &rack_gp_rtt_maxmul, 3,
1164             "Rack timely multiplier of lowest rtt for rtt_max");
1165         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1166             SYSCTL_CHILDREN(rack_timely),
1167             OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1168             &rack_gp_rtt_mindiv, 4,
1169             "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1170         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1171             SYSCTL_CHILDREN(rack_timely),
1172             OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1173             &rack_gp_rtt_minmul, 1,
1174             "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1175         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1176             SYSCTL_CHILDREN(rack_timely),
1177             OID_AUTO, "decrease", CTLFLAG_RW,
1178             &rack_gp_decrease_per, 20,
1179             "Rack timely decrease percentage of our GP multiplication factor");
1180         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1181             SYSCTL_CHILDREN(rack_timely),
1182             OID_AUTO, "increase", CTLFLAG_RW,
1183             &rack_gp_increase_per, 2,
1184             "Rack timely increase perentage of our GP multiplication factor");
1185         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1186             SYSCTL_CHILDREN(rack_timely),
1187             OID_AUTO, "lowerbound", CTLFLAG_RW,
1188             &rack_per_lower_bound, 50,
1189             "Rack timely lowest percentage we allow GP multiplier to fall to");
1190         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1191             SYSCTL_CHILDREN(rack_timely),
1192             OID_AUTO, "upperboundss", CTLFLAG_RW,
1193             &rack_per_upper_bound_ss, 0,
1194             "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1195         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1196             SYSCTL_CHILDREN(rack_timely),
1197             OID_AUTO, "upperboundca", CTLFLAG_RW,
1198             &rack_per_upper_bound_ca, 0,
1199             "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1200         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1201             SYSCTL_CHILDREN(rack_timely),
1202             OID_AUTO, "dynamicgp", CTLFLAG_RW,
1203             &rack_do_dyn_mul, 0,
1204             "Rack timely do we enable dynmaic timely goodput by default");
1205         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1206             SYSCTL_CHILDREN(rack_timely),
1207             OID_AUTO, "no_rec_red", CTLFLAG_RW,
1208             &rack_gp_no_rec_chg, 1,
1209             "Rack timely do we prohibit the recovery multiplier from being lowered");
1210         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1211             SYSCTL_CHILDREN(rack_timely),
1212             OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1213             &rack_timely_dec_clear, 6,
1214             "Rack timely what threshold do we count to before another boost during b/w decent");
1215         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1216             SYSCTL_CHILDREN(rack_timely),
1217             OID_AUTO, "max_push_rise", CTLFLAG_RW,
1218             &rack_timely_max_push_rise, 3,
1219             "Rack timely how many times do we push up with b/w increase");
1220         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1221             SYSCTL_CHILDREN(rack_timely),
1222             OID_AUTO, "max_push_drop", CTLFLAG_RW,
1223             &rack_timely_max_push_drop, 3,
1224             "Rack timely how many times do we push back on b/w decent");
1225         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1226             SYSCTL_CHILDREN(rack_timely),
1227             OID_AUTO, "min_segs", CTLFLAG_RW,
1228             &rack_timely_min_segs, 4,
1229             "Rack timely when setting the cwnd what is the min num segments");
1230         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1231             SYSCTL_CHILDREN(rack_timely),
1232             OID_AUTO, "noback_max", CTLFLAG_RW,
1233             &rack_use_max_for_nobackoff, 0,
1234             "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1235         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1236             SYSCTL_CHILDREN(rack_timely),
1237             OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1238             &rack_timely_int_timely_only, 0,
1239             "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1240         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1241             SYSCTL_CHILDREN(rack_timely),
1242             OID_AUTO, "nonstop", CTLFLAG_RW,
1243             &rack_timely_no_stopping, 0,
1244             "Rack timely don't stop increase");
1245         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1246             SYSCTL_CHILDREN(rack_timely),
1247             OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1248             &rack_down_raise_thresh, 100,
1249             "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1250         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1251             SYSCTL_CHILDREN(rack_timely),
1252             OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1253             &rack_req_segs, 1,
1254             "Bottom dragging if not these many segments outstanding and room");
1255
1256         /* TLP and Rack related parameters */
1257         rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1258             SYSCTL_CHILDREN(rack_sysctl_root),
1259             OID_AUTO,
1260             "tlp",
1261             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1262             "TLP and Rack related Controls");
1263         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1264             SYSCTL_CHILDREN(rack_tlp),
1265             OID_AUTO, "use_rrr", CTLFLAG_RW,
1266             &use_rack_rr, 1,
1267             "Do we use Rack Rapid Recovery");
1268         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1269             SYSCTL_CHILDREN(rack_tlp),
1270             OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1271             &rack_max_abc_post_recovery, 2,
1272             "Since we do early recovery, do we override the l_abc to a value, if so what?");
1273         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1274             SYSCTL_CHILDREN(rack_tlp),
1275             OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1276             &rack_non_rxt_use_cr, 0,
1277             "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1278         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1279             SYSCTL_CHILDREN(rack_tlp),
1280             OID_AUTO, "tlpmethod", CTLFLAG_RW,
1281             &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1282             "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1283         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1284             SYSCTL_CHILDREN(rack_tlp),
1285             OID_AUTO, "limit", CTLFLAG_RW,
1286             &rack_tlp_limit, 2,
1287             "How many TLP's can be sent without sending new data");
1288         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1289             SYSCTL_CHILDREN(rack_tlp),
1290             OID_AUTO, "use_greater", CTLFLAG_RW,
1291             &rack_tlp_use_greater, 1,
1292             "Should we use the rack_rtt time if its greater than srtt");
1293         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1294             SYSCTL_CHILDREN(rack_tlp),
1295             OID_AUTO, "tlpminto", CTLFLAG_RW,
1296             &rack_tlp_min, 10000,
1297             "TLP minimum timeout per the specification (in microseconds)");
1298         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1299             SYSCTL_CHILDREN(rack_tlp),
1300             OID_AUTO, "send_oldest", CTLFLAG_RW,
1301             &rack_always_send_oldest, 0,
1302             "Should we always send the oldest TLP and RACK-TLP");
1303         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1304             SYSCTL_CHILDREN(rack_tlp),
1305             OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1306             &rack_limited_retran, 0,
1307             "How many times can a rack timeout drive out sends");
1308         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1309             SYSCTL_CHILDREN(rack_tlp),
1310             OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1311             &rack_lower_cwnd_at_tlp, 0,
1312             "When a TLP completes a retran should we enter recovery");
1313         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1314             SYSCTL_CHILDREN(rack_tlp),
1315             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1316             &rack_reorder_thresh, 2,
1317             "What factor for rack will be added when seeing reordering (shift right)");
1318         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1319             SYSCTL_CHILDREN(rack_tlp),
1320             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1321             &rack_tlp_thresh, 1,
1322             "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1323         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1324             SYSCTL_CHILDREN(rack_tlp),
1325             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1326             &rack_reorder_fade, 60000000,
1327             "Does reorder detection fade, if so how many microseconds (0 means never)");
1328         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1329             SYSCTL_CHILDREN(rack_tlp),
1330             OID_AUTO, "pktdelay", CTLFLAG_RW,
1331             &rack_pkt_delay, 1000,
1332             "Extra RACK time (in microseconds) besides reordering thresh");
1333
1334         /* Timer related controls */
1335         rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1336             SYSCTL_CHILDREN(rack_sysctl_root),
1337             OID_AUTO,
1338             "timers",
1339             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1340             "Timer related controls");
1341         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1342             SYSCTL_CHILDREN(rack_timers),
1343             OID_AUTO, "persmin", CTLFLAG_RW,
1344             &rack_persist_min, 250000,
1345             "What is the minimum time in microseconds between persists");
1346         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1347             SYSCTL_CHILDREN(rack_timers),
1348             OID_AUTO, "persmax", CTLFLAG_RW,
1349             &rack_persist_max, 2000000,
1350             "What is the largest delay in microseconds between persists");
1351         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1352             SYSCTL_CHILDREN(rack_timers),
1353             OID_AUTO, "delayed_ack", CTLFLAG_RW,
1354             &rack_delayed_ack_time, 40000,
1355             "Delayed ack time (40ms in microseconds)");
1356         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1357             SYSCTL_CHILDREN(rack_timers),
1358             OID_AUTO, "minrto", CTLFLAG_RW,
1359             &rack_rto_min, 30000,
1360             "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1361         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1362             SYSCTL_CHILDREN(rack_timers),
1363             OID_AUTO, "maxrto", CTLFLAG_RW,
1364             &rack_rto_max, 4000000,
1365             "Maximum RTO in microseconds -- should be at least as large as min_rto");
1366         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1367             SYSCTL_CHILDREN(rack_timers),
1368             OID_AUTO, "minto", CTLFLAG_RW,
1369             &rack_min_to, 1000,
1370             "Minimum rack timeout in microseconds");
1371         /* Measure controls */
1372         rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1373             SYSCTL_CHILDREN(rack_sysctl_root),
1374             OID_AUTO,
1375             "measure",
1376             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1377             "Measure related controls");
1378         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1379             SYSCTL_CHILDREN(rack_measure),
1380             OID_AUTO, "wma_divisor", CTLFLAG_RW,
1381             &rack_wma_divisor, 8,
1382             "When doing b/w calculation what is the  divisor for the WMA");
1383         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1384             SYSCTL_CHILDREN(rack_measure),
1385             OID_AUTO, "end_cwnd", CTLFLAG_RW,
1386             &rack_cwnd_block_ends_measure, 0,
1387             "Does a cwnd just-return end the measurement window (app limited)");
1388         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1389             SYSCTL_CHILDREN(rack_measure),
1390             OID_AUTO, "end_rwnd", CTLFLAG_RW,
1391             &rack_rwnd_block_ends_measure, 0,
1392             "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1393         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1394             SYSCTL_CHILDREN(rack_measure),
1395             OID_AUTO, "min_target", CTLFLAG_RW,
1396             &rack_def_data_window, 20,
1397             "What is the minimum target window (in mss) for a GP measurements");
1398         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1399             SYSCTL_CHILDREN(rack_measure),
1400             OID_AUTO, "goal_bdp", CTLFLAG_RW,
1401             &rack_goal_bdp, 2,
1402             "What is the goal BDP to measure");
1403         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1404             SYSCTL_CHILDREN(rack_measure),
1405             OID_AUTO, "min_srtts", CTLFLAG_RW,
1406             &rack_min_srtts, 1,
1407             "What is the goal BDP to measure");
1408         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1409             SYSCTL_CHILDREN(rack_measure),
1410             OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1411             &rack_min_measure_usec, 0,
1412             "What is the Minimum time time for a measurement if 0, this is off");
1413         /* Features */
1414         rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1415             SYSCTL_CHILDREN(rack_sysctl_root),
1416             OID_AUTO,
1417             "features",
1418             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1419             "Feature controls");
1420         SYSCTL_ADD_U64(&rack_sysctl_ctx,
1421             SYSCTL_CHILDREN(rack_features),
1422             OID_AUTO, "rxt_clamp_thresh", CTLFLAG_RW,
1423             &rack_rxt_clamp_thresh, 0,
1424             "Bit encoded clamping setup bits CCCC CCCCC UUUU UULF PPPP PPPP PPPP PPPP");
1425         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1426             SYSCTL_CHILDREN(rack_features),
1427             OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1428             &rack_hybrid_allow_set_maxseg, 0,
1429             "Should hybrid pacing allow the setmss command");
1430         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1431             SYSCTL_CHILDREN(rack_features),
1432             OID_AUTO, "cmpack", CTLFLAG_RW,
1433             &rack_use_cmp_acks, 1,
1434             "Should RACK have LRO send compressed acks");
1435         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1436             SYSCTL_CHILDREN(rack_features),
1437             OID_AUTO, "fsb", CTLFLAG_RW,
1438             &rack_use_fsb, 1,
1439             "Should RACK use the fast send block?");
1440         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1441             SYSCTL_CHILDREN(rack_features),
1442             OID_AUTO, "rfo", CTLFLAG_RW,
1443             &rack_use_rfo, 1,
1444             "Should RACK use rack_fast_output()?");
1445         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1446             SYSCTL_CHILDREN(rack_features),
1447             OID_AUTO, "rsmrfo", CTLFLAG_RW,
1448             &rack_use_rsm_rfo, 1,
1449             "Should RACK use rack_fast_rsm_output()?");
1450         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1451             SYSCTL_CHILDREN(rack_features),
1452             OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1453             &rack_enable_mqueue_for_nonpaced, 0,
1454             "Should RACK use mbuf queuing for non-paced connections");
1455         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1456             SYSCTL_CHILDREN(rack_features),
1457             OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1458             &rack_do_hystart, 0,
1459             "Should RACK enable HyStart++ on connections?");
1460         /* Misc rack controls */
1461         rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1462             SYSCTL_CHILDREN(rack_sysctl_root),
1463             OID_AUTO,
1464             "misc",
1465             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1466             "Misc related controls");
1467 #ifdef TCP_ACCOUNTING
1468         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1469             SYSCTL_CHILDREN(rack_misc),
1470             OID_AUTO, "tcp_acct", CTLFLAG_RW,
1471             &rack_tcp_accounting, 0,
1472             "Should we turn on TCP accounting for all rack sessions?");
1473 #endif
1474         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1475             SYSCTL_CHILDREN(rack_misc),
1476             OID_AUTO, "dnd", CTLFLAG_RW,
1477             &rack_dnd_default, 0,
1478             "Do not disturb default for rack_rrr = 3");
1479         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1480             SYSCTL_CHILDREN(rack_misc),
1481             OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1482             &sad_seg_size_per, 800,
1483             "Percentage of segment size needed in a sack 800 = 80.0?");
1484         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1485             SYSCTL_CHILDREN(rack_misc),
1486             OID_AUTO, "rxt_controls", CTLFLAG_RW,
1487             &rack_rxt_controls, 0,
1488             "Retransmit sending size controls (valid  values 0, 1, 2 default=1)?");
1489         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1490             SYSCTL_CHILDREN(rack_misc),
1491             OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1492             &rack_hibeta_setting, 0,
1493             "Do we ue a high beta (80 instead of 50)?");
1494         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1495             SYSCTL_CHILDREN(rack_misc),
1496             OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1497             &rack_apply_rtt_with_reduced_conf, 0,
1498             "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1499         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1500             SYSCTL_CHILDREN(rack_misc),
1501             OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1502             &rack_dsack_std_based, 3,
1503             "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1504         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1505             SYSCTL_CHILDREN(rack_misc),
1506             OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1507             &rack_prr_addbackmax, 2,
1508             "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1509         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1510             SYSCTL_CHILDREN(rack_misc),
1511             OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1512             &rack_stats_gets_ms_rtt, 1,
1513             "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1514         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1515             SYSCTL_CHILDREN(rack_misc),
1516             OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1517             &rack_client_low_buf, 0,
1518             "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1519         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1520             SYSCTL_CHILDREN(rack_misc),
1521             OID_AUTO, "defprofile", CTLFLAG_RW,
1522             &rack_def_profile, 0,
1523             "Should RACK use a default profile (0=no, num == profile num)?");
1524         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1525             SYSCTL_CHILDREN(rack_misc),
1526             OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1527             &rack_enable_shared_cwnd, 1,
1528             "Should RACK try to use the shared cwnd on connections where allowed");
1529         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1530             SYSCTL_CHILDREN(rack_misc),
1531             OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1532             &rack_limits_scwnd, 1,
1533             "Should RACK place low end time limits on the shared cwnd feature");
1534         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1535             SYSCTL_CHILDREN(rack_misc),
1536             OID_AUTO, "no_prr", CTLFLAG_RW,
1537             &rack_disable_prr, 0,
1538             "Should RACK not use prr and only pace (must have pacing on)");
1539         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1540             SYSCTL_CHILDREN(rack_misc),
1541             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1542             &rack_verbose_logging, 0,
1543             "Should RACK black box logging be verbose");
1544         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1545             SYSCTL_CHILDREN(rack_misc),
1546             OID_AUTO, "data_after_close", CTLFLAG_RW,
1547             &rack_ignore_data_after_close, 1,
1548             "Do we hold off sending a RST until all pending data is ack'd");
1549         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1550             SYSCTL_CHILDREN(rack_misc),
1551             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1552             &rack_sack_not_required, 1,
1553             "Do we allow rack to run on connections not supporting SACK");
1554         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1555             SYSCTL_CHILDREN(rack_misc),
1556             OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1557             &rack_send_a_lot_in_prr, 1,
1558             "Send a lot in prr");
1559         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1560             SYSCTL_CHILDREN(rack_misc),
1561             OID_AUTO, "autoscale", CTLFLAG_RW,
1562             &rack_autosndbuf_inc, 20,
1563             "What percentage should rack scale up its snd buffer by?");
1564         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1565             SYSCTL_CHILDREN(rack_misc),
1566             OID_AUTO, "rnds_for_rxt_clamp", CTLFLAG_RW,
1567             &rack_rxt_min_rnds, 10,
1568             "Number of rounds needed between RTT clamps due to high loss rates");
1569         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1570             SYSCTL_CHILDREN(rack_misc),
1571             OID_AUTO, "rnds_for_unclamp", CTLFLAG_RW,
1572             &rack_unclamp_round_thresh, 100,
1573             "Number of rounds needed with no loss to unclamp");
1574         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1575             SYSCTL_CHILDREN(rack_misc),
1576             OID_AUTO, "rxt_threshs_for_unclamp", CTLFLAG_RW,
1577             &rack_unclamp_rxt_thresh, 5,
1578            "Percentage of retransmits we need to be under to unclamp (5 = .5 percent)\n");
1579         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1580             SYSCTL_CHILDREN(rack_misc),
1581             OID_AUTO, "clamp_ss_upper", CTLFLAG_RW,
1582             &rack_clamp_ss_upper, 110,
1583             "Clamp percentage ceiling in SS?");
1584         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1585             SYSCTL_CHILDREN(rack_misc),
1586             OID_AUTO, "clamp_ca_upper", CTLFLAG_RW,
1587             &rack_clamp_ca_upper, 110,
1588             "Clamp percentage ceiling in CA?");
1589         /* Sack Attacker detection stuff */
1590         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1591             SYSCTL_CHILDREN(rack_attack),
1592             OID_AUTO, "merge_out", CTLFLAG_RW,
1593             &rack_merge_out_sacks_on_attack, 0,
1594             "Do we merge the sendmap when we decide we are being attacked?");
1595
1596         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1597             SYSCTL_CHILDREN(rack_attack),
1598             OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1599             &rack_highest_sack_thresh_seen, 0,
1600             "Highest sack to ack ratio seen");
1601         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1602             SYSCTL_CHILDREN(rack_attack),
1603             OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1604             &rack_highest_move_thresh_seen, 0,
1605             "Highest move to non-move ratio seen");
1606         rack_ack_total = counter_u64_alloc(M_WAITOK);
1607         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1608             SYSCTL_CHILDREN(rack_attack),
1609             OID_AUTO, "acktotal", CTLFLAG_RD,
1610             &rack_ack_total,
1611             "Total number of Ack's");
1612         rack_express_sack = counter_u64_alloc(M_WAITOK);
1613         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1614             SYSCTL_CHILDREN(rack_attack),
1615             OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1616             &rack_express_sack,
1617             "Total expresss number of Sack's");
1618         rack_sack_total = counter_u64_alloc(M_WAITOK);
1619         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1620             SYSCTL_CHILDREN(rack_attack),
1621             OID_AUTO, "sacktotal", CTLFLAG_RD,
1622             &rack_sack_total,
1623             "Total number of SACKs");
1624         rack_move_none = counter_u64_alloc(M_WAITOK);
1625         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1626             SYSCTL_CHILDREN(rack_attack),
1627             OID_AUTO, "move_none", CTLFLAG_RD,
1628             &rack_move_none,
1629             "Total number of SACK index reuse of positions under threshold");
1630         rack_move_some = counter_u64_alloc(M_WAITOK);
1631         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1632             SYSCTL_CHILDREN(rack_attack),
1633             OID_AUTO, "move_some", CTLFLAG_RD,
1634             &rack_move_some,
1635             "Total number of SACK index reuse of positions over threshold");
1636         rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1637         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1638             SYSCTL_CHILDREN(rack_attack),
1639             OID_AUTO, "attacks", CTLFLAG_RD,
1640             &rack_sack_attacks_detected,
1641             "Total number of SACK attackers that had sack disabled");
1642         rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1643         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1644             SYSCTL_CHILDREN(rack_attack),
1645             OID_AUTO, "reversed", CTLFLAG_RD,
1646             &rack_sack_attacks_reversed,
1647             "Total number of SACK attackers that were later determined false positive");
1648         rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1649         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1650             SYSCTL_CHILDREN(rack_attack),
1651             OID_AUTO, "suspect", CTLFLAG_RD,
1652             &rack_sack_attacks_suspect,
1653             "Total number of SACKs that triggered early detection");
1654
1655         rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1656         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1657             SYSCTL_CHILDREN(rack_attack),
1658             OID_AUTO, "nextmerge", CTLFLAG_RD,
1659             &rack_sack_used_next_merge,
1660             "Total number of times we used the next merge");
1661         rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1662         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1663             SYSCTL_CHILDREN(rack_attack),
1664             OID_AUTO, "prevmerge", CTLFLAG_RD,
1665             &rack_sack_used_prev_merge,
1666             "Total number of times we used the prev merge");
1667         /* Counters */
1668         rack_total_bytes = counter_u64_alloc(M_WAITOK);
1669         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1670             SYSCTL_CHILDREN(rack_counters),
1671             OID_AUTO, "totalbytes", CTLFLAG_RD,
1672             &rack_total_bytes,
1673             "Total number of bytes sent");
1674         rack_fto_send = counter_u64_alloc(M_WAITOK);
1675         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1676             SYSCTL_CHILDREN(rack_counters),
1677             OID_AUTO, "fto_send", CTLFLAG_RD,
1678             &rack_fto_send, "Total number of rack_fast_output sends");
1679         rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1680         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1681             SYSCTL_CHILDREN(rack_counters),
1682             OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1683             &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1684         rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1685         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1686             SYSCTL_CHILDREN(rack_counters),
1687             OID_AUTO, "nfto_resend", CTLFLAG_RD,
1688             &rack_nfto_resend, "Total number of rack_output retransmissions");
1689         rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1690         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1691             SYSCTL_CHILDREN(rack_counters),
1692             OID_AUTO, "nfto_send", CTLFLAG_RD,
1693             &rack_non_fto_send, "Total number of rack_output first sends");
1694         rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1695         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1696             SYSCTL_CHILDREN(rack_counters),
1697             OID_AUTO, "rfo_extended", CTLFLAG_RD,
1698             &rack_extended_rfo, "Total number of times we extended rfo");
1699
1700         rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1701         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1702             SYSCTL_CHILDREN(rack_counters),
1703             OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1704             &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1705         rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1706
1707         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1708             SYSCTL_CHILDREN(rack_counters),
1709             OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1710             &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1711         rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1712         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1713             SYSCTL_CHILDREN(rack_counters),
1714             OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1715             &rack_tlp_tot,
1716             "Total number of tail loss probe expirations");
1717         rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1718         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1719             SYSCTL_CHILDREN(rack_counters),
1720             OID_AUTO, "tlp_new", CTLFLAG_RD,
1721             &rack_tlp_newdata,
1722             "Total number of tail loss probe sending new data");
1723         rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1724         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1725             SYSCTL_CHILDREN(rack_counters),
1726             OID_AUTO, "tlp_retran", CTLFLAG_RD,
1727             &rack_tlp_retran,
1728             "Total number of tail loss probe sending retransmitted data");
1729         rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1730         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1731             SYSCTL_CHILDREN(rack_counters),
1732             OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1733             &rack_tlp_retran_bytes,
1734             "Total bytes of tail loss probe sending retransmitted data");
1735         rack_to_tot = counter_u64_alloc(M_WAITOK);
1736         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1737             SYSCTL_CHILDREN(rack_counters),
1738             OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1739             &rack_to_tot,
1740             "Total number of times the rack to expired");
1741         rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1742         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1743             SYSCTL_CHILDREN(rack_counters),
1744             OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1745             &rack_saw_enobuf,
1746             "Total number of times a sends returned enobuf for non-hdwr paced connections");
1747         rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1748         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1749             SYSCTL_CHILDREN(rack_counters),
1750             OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1751             &rack_saw_enobuf_hw,
1752             "Total number of times a send returned enobuf for hdwr paced connections");
1753         rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1754         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1755             SYSCTL_CHILDREN(rack_counters),
1756             OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1757             &rack_saw_enetunreach,
1758             "Total number of times a send received a enetunreachable");
1759         rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1760         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1761             SYSCTL_CHILDREN(rack_counters),
1762             OID_AUTO, "alloc_hot", CTLFLAG_RD,
1763             &rack_hot_alloc,
1764             "Total allocations from the top of our list");
1765         rack_to_alloc = counter_u64_alloc(M_WAITOK);
1766         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1767             SYSCTL_CHILDREN(rack_counters),
1768             OID_AUTO, "allocs", CTLFLAG_RD,
1769             &rack_to_alloc,
1770             "Total allocations of tracking structures");
1771         rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1772         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1773             SYSCTL_CHILDREN(rack_counters),
1774             OID_AUTO, "allochard", CTLFLAG_RD,
1775             &rack_to_alloc_hard,
1776             "Total allocations done with sleeping the hard way");
1777         rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1778         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1779             SYSCTL_CHILDREN(rack_counters),
1780             OID_AUTO, "allocemerg", CTLFLAG_RD,
1781             &rack_to_alloc_emerg,
1782             "Total allocations done from emergency cache");
1783         rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1784         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1785             SYSCTL_CHILDREN(rack_counters),
1786             OID_AUTO, "alloc_limited", CTLFLAG_RD,
1787             &rack_to_alloc_limited,
1788             "Total allocations dropped due to limit");
1789         rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1790         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1791             SYSCTL_CHILDREN(rack_counters),
1792             OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1793             &rack_alloc_limited_conns,
1794             "Connections with allocations dropped due to limit");
1795         rack_split_limited = counter_u64_alloc(M_WAITOK);
1796         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1797             SYSCTL_CHILDREN(rack_counters),
1798             OID_AUTO, "split_limited", CTLFLAG_RD,
1799             &rack_split_limited,
1800             "Split allocations dropped due to limit");
1801         rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1802         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1803             SYSCTL_CHILDREN(rack_counters),
1804             OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1805             &rack_rxt_clamps_cwnd,
1806             "Number of times that excessive rxt clamped the cwnd down");
1807         rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1808         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1809             SYSCTL_CHILDREN(rack_counters),
1810             OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1811             &rack_rxt_clamps_cwnd_uniq,
1812             "Number of connections that have had excessive rxt clamped the cwnd down");
1813         rack_persists_sends = counter_u64_alloc(M_WAITOK);
1814         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1815             SYSCTL_CHILDREN(rack_counters),
1816             OID_AUTO, "persist_sends", CTLFLAG_RD,
1817             &rack_persists_sends,
1818             "Number of times we sent a persist probe");
1819         rack_persists_acks = counter_u64_alloc(M_WAITOK);
1820         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1821             SYSCTL_CHILDREN(rack_counters),
1822             OID_AUTO, "persist_acks", CTLFLAG_RD,
1823             &rack_persists_acks,
1824             "Number of times a persist probe was acked");
1825         rack_persists_loss = counter_u64_alloc(M_WAITOK);
1826         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1827             SYSCTL_CHILDREN(rack_counters),
1828             OID_AUTO, "persist_loss", CTLFLAG_RD,
1829             &rack_persists_loss,
1830             "Number of times we detected a lost persist probe (no ack)");
1831         rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1832         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1833             SYSCTL_CHILDREN(rack_counters),
1834             OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1835             &rack_persists_lost_ends,
1836             "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1837 #ifdef INVARIANTS
1838         rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1839         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1840             SYSCTL_CHILDREN(rack_counters),
1841             OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1842             &rack_adjust_map_bw,
1843             "Number of times we hit the case where the sb went up and down on a sendmap entry");
1844 #endif
1845         rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1846         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1847             SYSCTL_CHILDREN(rack_counters),
1848             OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1849             &rack_multi_single_eq,
1850             "Number of compressed acks total represented");
1851         rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1852         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1853             SYSCTL_CHILDREN(rack_counters),
1854             OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1855             &rack_proc_non_comp_ack,
1856             "Number of non compresseds acks that we processed");
1857
1858
1859         rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1860         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1861             SYSCTL_CHILDREN(rack_counters),
1862             OID_AUTO, "sack_long", CTLFLAG_RD,
1863             &rack_sack_proc_all,
1864             "Total times we had to walk whole list for sack processing");
1865         rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1866         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1867             SYSCTL_CHILDREN(rack_counters),
1868             OID_AUTO, "sack_restart", CTLFLAG_RD,
1869             &rack_sack_proc_restart,
1870             "Total times we had to walk whole list due to a restart");
1871         rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1872         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1873             SYSCTL_CHILDREN(rack_counters),
1874             OID_AUTO, "sack_short", CTLFLAG_RD,
1875             &rack_sack_proc_short,
1876             "Total times we took shortcut for sack processing");
1877         rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1878         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1879             SYSCTL_CHILDREN(rack_attack),
1880             OID_AUTO, "skipacked", CTLFLAG_RD,
1881             &rack_sack_skipped_acked,
1882             "Total number of times we skipped previously sacked");
1883         rack_sack_splits = counter_u64_alloc(M_WAITOK);
1884         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1885             SYSCTL_CHILDREN(rack_attack),
1886             OID_AUTO, "ofsplit", CTLFLAG_RD,
1887             &rack_sack_splits,
1888             "Total number of times we did the old fashion tree split");
1889         rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1890         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1891             SYSCTL_CHILDREN(rack_counters),
1892             OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1893             &rack_input_idle_reduces,
1894             "Total number of idle reductions on input");
1895         rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1896         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1897             SYSCTL_CHILDREN(rack_counters),
1898             OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1899             &rack_collapsed_win_seen,
1900             "Total number of collapsed window events seen (where our window shrinks)");
1901
1902         rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1903         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1904             SYSCTL_CHILDREN(rack_counters),
1905             OID_AUTO, "collapsed_win", CTLFLAG_RD,
1906             &rack_collapsed_win,
1907             "Total number of collapsed window events where we mark packets");
1908         rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1909         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1910             SYSCTL_CHILDREN(rack_counters),
1911             OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1912             &rack_collapsed_win_rxt,
1913             "Total number of packets that were retransmitted");
1914         rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1915         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1916             SYSCTL_CHILDREN(rack_counters),
1917             OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1918             &rack_collapsed_win_rxt_bytes,
1919             "Total number of bytes that were retransmitted");
1920         rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1921         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1922             SYSCTL_CHILDREN(rack_counters),
1923             OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1924             &rack_try_scwnd,
1925             "Total number of scwnd attempts");
1926         COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1927         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1928             OID_AUTO, "outsize", CTLFLAG_RD,
1929             rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1930         COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1931         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1932             OID_AUTO, "opts", CTLFLAG_RD,
1933             rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1934         SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1935             SYSCTL_CHILDREN(rack_sysctl_root),
1936             OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1937             &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1938 }
1939
1940 static uint32_t
1941 rc_init_window(struct tcp_rack *rack)
1942 {
1943         uint32_t win;
1944
1945         if (rack->rc_init_win == 0) {
1946                 /*
1947                  * Nothing set by the user, use the system stack
1948                  * default.
1949                  */
1950                 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1951         }
1952         win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1953         return (win);
1954 }
1955
1956 static uint64_t
1957 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1958 {
1959         if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1960                 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1961         else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1962                 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1963         else
1964                 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1965 }
1966
1967 static void
1968 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
1969         uint64_t data, uint8_t mod, uint16_t aux,
1970         struct http_sendfile_track *cur)
1971 {
1972 #ifdef TCP_REQUEST_TRK
1973         int do_log = 0;
1974
1975         /*
1976          * The rate cap one is noisy and only should come out when normal BB logging
1977          * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
1978          * once per chunk and make up the BBpoint that can be turned on by the client.
1979          */
1980         if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
1981                 /*
1982                  * The very noisy two need to only come out when
1983                  * we have verbose logging on.
1984                  */
1985                 if (rack_verbose_logging != 0)
1986                         do_log = tcp_bblogging_on(rack->rc_tp);
1987                 else
1988                         do_log = 0;
1989         } else if (mod != HYBRID_LOG_BW_MEASURE) {
1990                 /*
1991                  * All other less noisy logs here except the measure which
1992                  * also needs to come out on the point and the log.
1993                  */
1994                 do_log = tcp_bblogging_on(rack->rc_tp);
1995         } else {
1996                 do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
1997         }
1998
1999         if (do_log) {
2000                 union tcp_log_stackspecific log;
2001                 struct timeval tv;
2002                 uint64_t lt_bw;
2003
2004                 /* Convert our ms to a microsecond */
2005                 memset(&log, 0, sizeof(log));
2006
2007                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2008                 log.u_bbr.rttProp = tim;
2009                 log.u_bbr.bw_inuse = cbw;
2010                 log.u_bbr.delRate = rack_get_gp_est(rack);
2011                 lt_bw = rack_get_lt_bw(rack);
2012                 log.u_bbr.flex1 = seq;
2013                 log.u_bbr.pacing_gain = aux;
2014                 /* lt_bw = < flex3 | flex2 > */
2015                 log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2016                 log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2017                 /* Record the last obtained us rtt in inflight */
2018                 if (cur == NULL) {
2019                         /* Make sure we are looking at the right log if an overide comes in */
2020                         cur = rack->r_ctl.rc_last_sft;
2021                 }
2022                 if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2023                         log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2024                 else {
2025                         /* Use the last known rtt i.e. the rack-rtt */
2026                         log.u_bbr.inflight = rack->rc_rack_rtt;
2027                 }
2028                 if (cur != NULL) {
2029                         uint64_t off;
2030
2031                         log.u_bbr.cur_del_rate = cur->deadline;
2032                         if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2033                                 /* start = < lost | pkt_epoch > */
2034                                 log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2035                                 log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2036                                 log.u_bbr.flex6 = cur->start_seq;
2037                                 log.u_bbr.pkts_out = cur->end_seq;
2038                         } else {
2039                                 /* start = < lost | pkt_epoch > */
2040                                 log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2041                                 log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2042                                 /* end = < pkts_out | flex6 > */
2043                                 log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2044                                 log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2045                         }
2046                         /* first_send = <lt_epoch | epoch> */
2047                         log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2048                         log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2049                         /* localtime = <delivered | applimited>*/
2050                         log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2051                         log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2052                         off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_http_info[0]);
2053                         log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct http_sendfile_track));
2054                         log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2055                         log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2056                         log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2057                 } else {
2058                         log.u_bbr.flex7 = 0xffff;
2059                         log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2060                 }
2061                 /*
2062                  * Compose bbr_state to be a bit wise 0000ADHF
2063                  * where A is the always_pace flag
2064                  * where D is the dgp_on flag
2065                  * where H is the hybrid_mode on flag
2066                  * where F is the use_fixed_rate flag.
2067                  */
2068                 log.u_bbr.bbr_state = rack->rc_always_pace;
2069                 log.u_bbr.bbr_state <<= 1;
2070                 log.u_bbr.bbr_state |= rack->dgp_on;
2071                 log.u_bbr.bbr_state <<= 1;
2072                 log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2073                 log.u_bbr.bbr_state <<= 1;
2074                 log.u_bbr.bbr_state |= rack->use_fixed_rate;
2075                 log.u_bbr.flex8 = mod;
2076                 tcp_log_event(rack->rc_tp, NULL,
2077                     &rack->rc_inp->inp_socket->so_rcv,
2078                     &rack->rc_inp->inp_socket->so_snd,
2079                     TCP_HYBRID_PACING_LOG, 0,
2080                     0, &log, false, NULL, __func__, __LINE__, &tv);
2081
2082         }
2083 #endif
2084 }
2085
2086 static inline uint64_t
2087 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2088 {
2089         uint64_t ret_bw, ether;
2090         uint64_t u_segsiz;
2091
2092         ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2093         if (rack->r_is_v6){
2094 #ifdef INET6
2095                 ether += sizeof(struct ip6_hdr);
2096 #endif
2097                 ether += 14;    /* eheader size 6+6+2 */
2098         } else {
2099 #ifdef INET
2100                 ether += sizeof(struct ip);
2101 #endif
2102                 ether += 14;    /* eheader size 6+6+2 */
2103         }
2104         u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2105         ret_bw = bw;
2106         ret_bw *= ether;
2107         ret_bw /= u_segsiz;
2108         return (ret_bw);
2109 }
2110
2111 static void
2112 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2113 {
2114 #ifdef TCP_REQUEST_TRK
2115         struct timeval tv;
2116         uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2117 #endif
2118
2119         if (rack->r_ctl.bw_rate_cap == 0)
2120                 return;
2121 #ifdef TCP_REQUEST_TRK
2122         if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2123             (rack->r_ctl.rc_last_sft != NULL)) {
2124                 /*
2125                  * We have a dynamic cap. The original target
2126                  * is in bw_rate_cap, but we need to look at
2127                  * how long it is until we hit the deadline.
2128                  */
2129                 struct http_sendfile_track *ent;
2130
2131                 ent = rack->r_ctl.rc_last_sft;
2132                 microuptime(&tv);
2133                 timenow = tcp_tv_to_lusectick(&tv);
2134                 if (timenow >= ent->deadline) {
2135                         /* No time left we do DGP only */
2136                         rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2137                                            0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent);
2138                         rack->r_ctl.bw_rate_cap = 0;
2139                         return;
2140                 }
2141                 /* We have the time */
2142                 timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2143                 if (timeleft < HPTS_MSEC_IN_SEC) {
2144                         /* If there is less than a ms left just use DGPs rate */
2145                         rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2146                                            0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent);
2147                         rack->r_ctl.bw_rate_cap = 0;
2148                         return;
2149                 }
2150                 /*
2151                  * Now lets find the amount of data left to send.
2152                  *
2153                  * Now ideally we want to use the end_seq to figure out how much more
2154                  * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2155                  */
2156                 if (ent->flags & TCP_HTTP_TRACK_FLG_COMP) {
2157                         if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2158                                 lenleft = ent->end_seq - rack->rc_tp->snd_una;
2159                         else {
2160                                 /* TSNH, we should catch it at the send */
2161                                 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2162                                                    0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent);
2163                                 rack->r_ctl.bw_rate_cap = 0;
2164                                 return;
2165                         }
2166                 } else {
2167                         /*
2168                          * The hard way, figure out how much is gone and then
2169                          * take that away from the total the client asked for
2170                          * (thats off by tls overhead if this is tls).
2171                          */
2172                         if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2173                                 lengone = rack->rc_tp->snd_una - ent->start_seq;
2174                         else
2175                                 lengone = 0;
2176                         if (lengone < (ent->end - ent->start))
2177                                 lenleft = (ent->end - ent->start) - lengone;
2178                         else {
2179                                 /* TSNH, we should catch it at the send */
2180                                 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2181                                                    0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent);
2182                                 rack->r_ctl.bw_rate_cap = 0;
2183                                 return;
2184                         }
2185                 }
2186                 if (lenleft == 0) {
2187                         /* We have it all sent */
2188                         rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2189                                            0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent);
2190                         if (rack->r_ctl.bw_rate_cap)
2191                                 goto normal_ratecap;
2192                         else
2193                                 return;
2194                 }
2195                 calcbw = lenleft * HPTS_USEC_IN_SEC;
2196                 calcbw /= timeleft;
2197                 /* Now we must compensate for IP/TCP overhead */
2198                 calcbw = rack_compensate_for_linerate(rack, calcbw);
2199                 /* Update the bit rate cap */
2200                 rack->r_ctl.bw_rate_cap = calcbw;
2201                 if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2202                     (rack_hybrid_allow_set_maxseg == 1) &&
2203                     ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2204                         /* Lets set in a smaller mss possibly here to match our rate-cap */
2205                         uint32_t orig_max;
2206
2207                         orig_max = rack->r_ctl.rc_pace_max_segs;
2208                         rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2209                         rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2210                         rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2211                 }
2212                 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2213                                    calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent);
2214                 if ((calcbw > 0) && (*bw > calcbw)) {
2215                         rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2216                                            *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent);
2217                         *capped = 1;
2218                         *bw = calcbw;
2219                 }
2220                 return;
2221         }
2222 normal_ratecap:
2223 #endif
2224         if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2225 #ifdef TCP_REQUEST_TRK
2226                 if (rack->rc_hybrid_mode &&
2227                     rack->rc_catch_up &&
2228                     (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2229                     (rack_hybrid_allow_set_maxseg == 1) &&
2230                     ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2231                         /* Lets set in a smaller mss possibly here to match our rate-cap */
2232                         uint32_t orig_max;
2233
2234                         orig_max = rack->r_ctl.rc_pace_max_segs;
2235                         rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2236                         rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2237                         rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2238                 }
2239 #endif
2240                 *capped = 1;
2241                 *bw = rack->r_ctl.bw_rate_cap;
2242                 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2243                                    *bw, 0, 0,
2244                                    HYBRID_LOG_RATE_CAP, 1, NULL);
2245         }
2246 }
2247
2248 static uint64_t
2249 rack_get_gp_est(struct tcp_rack *rack)
2250 {
2251         uint64_t bw, lt_bw, ret_bw;
2252
2253         if (rack->rc_gp_filled == 0) {
2254                 /*
2255                  * We have yet no b/w measurement,
2256                  * if we have a user set initial bw
2257                  * return it. If we don't have that and
2258                  * we have an srtt, use the tcp IW (10) to
2259                  * calculate a fictional b/w over the SRTT
2260                  * which is more or less a guess. Note
2261                  * we don't use our IW from rack on purpose
2262                  * so if we have like IW=30, we are not
2263                  * calculating a "huge" b/w.
2264                  */
2265                 uint64_t srtt;
2266
2267                 lt_bw = rack_get_lt_bw(rack);
2268                 if (lt_bw) {
2269                         /*
2270                          * No goodput bw but a long-term b/w does exist
2271                          * lets use that.
2272                          */
2273                         ret_bw = lt_bw;
2274                         goto compensate;
2275                 }
2276                 if (rack->r_ctl.init_rate)
2277                         return (rack->r_ctl.init_rate);
2278
2279                 /* Ok lets come up with the IW guess, if we have a srtt */
2280                 if (rack->rc_tp->t_srtt == 0) {
2281                         /*
2282                          * Go with old pacing method
2283                          * i.e. burst mitigation only.
2284                          */
2285                         return (0);
2286                 }
2287                 /* Ok lets get the initial TCP win (not racks) */
2288                 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2289                 srtt = (uint64_t)rack->rc_tp->t_srtt;
2290                 bw *= (uint64_t)USECS_IN_SECOND;
2291                 bw /= srtt;
2292                 ret_bw = bw;
2293                 goto compensate;
2294
2295         }
2296         if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2297                 /* Averaging is done, we can return the value */
2298                 bw = rack->r_ctl.gp_bw;
2299         } else {
2300                 /* Still doing initial average must calculate */
2301                 bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2302         }
2303         lt_bw = rack_get_lt_bw(rack);
2304         if (lt_bw == 0) {
2305                 /* If we don't have one then equate it to the gp_bw */
2306                 lt_bw = rack->r_ctl.gp_bw;
2307         }
2308         if ((rack->r_cwnd_was_clamped == 1) && (rack->r_clamped_gets_lower > 0)){
2309                 /*  if clamped take the lowest */
2310                 if (lt_bw < bw)
2311                         ret_bw = lt_bw;
2312                 else
2313                         ret_bw = bw;
2314         } else {
2315                 /* If not set for clamped to get lowest, take the highest */
2316                 if (lt_bw > bw)
2317                         ret_bw = lt_bw;
2318                 else
2319                         ret_bw = bw;
2320         }
2321         /*
2322          * Now lets compensate based on the TCP/IP overhead. Our
2323          * Goodput estimate does not include this so we must pace out
2324          * a bit faster since our pacing calculations do. The pacing
2325          * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2326          * we are using to do this, so we do that here in the opposite
2327          * direction as well. This means that if we are tunneled and the
2328          * segsiz is say 1200 bytes we will get quite a boost, but its
2329          * compensated for in the pacing time the opposite way.
2330          */
2331 compensate:
2332         ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2333         return(ret_bw);
2334 }
2335
2336
2337 static uint64_t
2338 rack_get_bw(struct tcp_rack *rack)
2339 {
2340         uint64_t bw;
2341
2342         if (rack->use_fixed_rate) {
2343                 /* Return the fixed pacing rate */
2344                 return (rack_get_fixed_pacing_bw(rack));
2345         }
2346         bw = rack_get_gp_est(rack);
2347         return (bw);
2348 }
2349
2350 static uint16_t
2351 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2352 {
2353         if (rack->use_fixed_rate) {
2354                 return (100);
2355         } else if (rack->in_probe_rtt && (rsm == NULL))
2356                 return (rack->r_ctl.rack_per_of_gp_probertt);
2357         else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2358                   rack->r_ctl.rack_per_of_gp_rec)) {
2359                 if (rsm) {
2360                         /* a retransmission always use the recovery rate */
2361                         return (rack->r_ctl.rack_per_of_gp_rec);
2362                 } else if (rack->rack_rec_nonrxt_use_cr) {
2363                         /* Directed to use the configured rate */
2364                         goto configured_rate;
2365                 } else if (rack->rack_no_prr &&
2366                            (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2367                         /* No PRR, lets just use the b/w estimate only */
2368                         return (100);
2369                 } else {
2370                         /*
2371                          * Here we may have a non-retransmit but we
2372                          * have no overrides, so just use the recovery
2373                          * rate (prr is in effect).
2374                          */
2375                         return (rack->r_ctl.rack_per_of_gp_rec);
2376                 }
2377         }
2378 configured_rate:
2379         /* For the configured rate we look at our cwnd vs the ssthresh */
2380         if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2381                 return (rack->r_ctl.rack_per_of_gp_ss);
2382         else
2383                 return (rack->r_ctl.rack_per_of_gp_ca);
2384 }
2385
2386 static void
2387 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2388 {
2389         /*
2390          * Types of logs (mod value)
2391          * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2392          * 2 = a dsack round begins, persist is reset to 16.
2393          * 3 = a dsack round ends
2394          * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2395          * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2396          * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2397          */
2398         if (tcp_bblogging_on(rack->rc_tp)) {
2399                 union tcp_log_stackspecific log;
2400                 struct timeval tv;
2401
2402                 memset(&log, 0, sizeof(log));
2403                 log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2404                 log.u_bbr.flex1 <<= 1;
2405                 log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2406                 log.u_bbr.flex1 <<= 1;
2407                 log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2408                 log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2409                 log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2410                 log.u_bbr.flex4 = flex4;
2411                 log.u_bbr.flex5 = flex5;
2412                 log.u_bbr.flex6 = flex6;
2413                 log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2414                 log.u_bbr.flex8 = mod;
2415                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2416                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2417                     &rack->rc_inp->inp_socket->so_rcv,
2418                     &rack->rc_inp->inp_socket->so_snd,
2419                     RACK_DSACK_HANDLING, 0,
2420                     0, &log, false, &tv);
2421         }
2422 }
2423
2424 static void
2425 rack_log_hdwr_pacing(struct tcp_rack *rack,
2426                      uint64_t rate, uint64_t hw_rate, int line,
2427                      int error, uint16_t mod)
2428 {
2429         if (tcp_bblogging_on(rack->rc_tp)) {
2430                 union tcp_log_stackspecific log;
2431                 struct timeval tv;
2432                 const struct ifnet *ifp;
2433
2434                 memset(&log, 0, sizeof(log));
2435                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2436                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2437                 if (rack->r_ctl.crte) {
2438                         ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2439                 } else if (rack->rc_inp->inp_route.ro_nh &&
2440                            rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2441                         ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2442                 } else
2443                         ifp = NULL;
2444                 if (ifp) {
2445                         log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2446                         log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2447                 }
2448                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2449                 log.u_bbr.bw_inuse = rate;
2450                 log.u_bbr.flex5 = line;
2451                 log.u_bbr.flex6 = error;
2452                 log.u_bbr.flex7 = mod;
2453                 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2454                 log.u_bbr.flex8 = rack->use_fixed_rate;
2455                 log.u_bbr.flex8 <<= 1;
2456                 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2457                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2458                 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2459                 if (rack->r_ctl.crte)
2460                         log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2461                 else
2462                         log.u_bbr.cur_del_rate = 0;
2463                 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2464                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2465                     &rack->rc_inp->inp_socket->so_rcv,
2466                     &rack->rc_inp->inp_socket->so_snd,
2467                     BBR_LOG_HDWR_PACE, 0,
2468                     0, &log, false, &tv);
2469         }
2470 }
2471
2472 static uint64_t
2473 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2474 {
2475         /*
2476          * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2477          */
2478         uint64_t bw_est, high_rate;
2479         uint64_t gain;
2480
2481         if ((rack->r_pacing_discount == 0) ||
2482             (rack_full_buffer_discount == 0)) {
2483                 /*
2484                  * No buffer level based discount from client buffer
2485                  * level is enabled or the feature is disabled.
2486                  */
2487                 gain = (uint64_t)rack_get_output_gain(rack, rsm);
2488                 bw_est = bw * gain;
2489                 bw_est /= (uint64_t)100;
2490         } else {
2491                 /*
2492                  * We have a discount in place apply it with
2493                  * just a 100% gain (we get no boost if the buffer
2494                  * is full).
2495                  */
2496                 uint64_t discount;
2497
2498                 discount = bw * (uint64_t)(rack_full_buffer_discount * rack->r_ctl.pacing_discount_amm);
2499                 discount /= 100;
2500                 /* What %% of the b/w do we discount */
2501                 bw_est = bw - discount;
2502         }
2503         /* Never fall below the minimum (def 64kbps) */
2504         if (bw_est < RACK_MIN_BW)
2505                 bw_est = RACK_MIN_BW;
2506         if (rack->r_rack_hw_rate_caps) {
2507                 /* Rate caps are in place */
2508                 if (rack->r_ctl.crte != NULL) {
2509                         /* We have a hdwr rate already */
2510                         high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2511                         if (bw_est >= high_rate) {
2512                                 /* We are capping bw at the highest rate table entry */
2513                                 if (rack_hw_rate_cap_per &&
2514                                     (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2515                                         rack->r_rack_hw_rate_caps = 0;
2516                                         goto done;
2517                                 }
2518                                 rack_log_hdwr_pacing(rack,
2519                                                      bw_est, high_rate, __LINE__,
2520                                                      0, 3);
2521                                 bw_est = high_rate;
2522                                 if (capped)
2523                                         *capped = 1;
2524                         }
2525                 } else if ((rack->rack_hdrw_pacing == 0) &&
2526                            (rack->rack_hdw_pace_ena) &&
2527                            (rack->rack_attempt_hdwr_pace == 0) &&
2528                            (rack->rc_inp->inp_route.ro_nh != NULL) &&
2529                            (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2530                         /*
2531                          * Special case, we have not yet attempted hardware
2532                          * pacing, and yet we may, when we do, find out if we are
2533                          * above the highest rate. We need to know the maxbw for the interface
2534                          * in question (if it supports ratelimiting). We get back
2535                          * a 0, if the interface is not found in the RL lists.
2536                          */
2537                         high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2538                         if (high_rate) {
2539                                 /* Yep, we have a rate is it above this rate? */
2540                                 if (bw_est > high_rate) {
2541                                         bw_est = high_rate;
2542                                         if (capped)
2543                                                 *capped = 1;
2544                                 }
2545                         }
2546                 }
2547         }
2548 done:
2549         return (bw_est);
2550 }
2551
2552 static void
2553 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2554 {
2555         if (tcp_bblogging_on(rack->rc_tp)) {
2556                 union tcp_log_stackspecific log;
2557                 struct timeval tv;
2558
2559                 if (rack->sack_attack_disable > 0)
2560                         goto log_anyway;
2561                 if ((mod != 1) && (rack_verbose_logging == 0))  {
2562                         /*
2563                          * We get 3 values currently for mod
2564                          * 1 - We are retransmitting and this tells the reason.
2565                          * 2 - We are clearing a dup-ack count.
2566                          * 3 - We are incrementing a dup-ack count.
2567                          *
2568                          * The clear/increment are only logged
2569                          * if you have BBverbose on.
2570                          */
2571                         return;
2572                 }
2573 log_anyway:
2574                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2575                 log.u_bbr.flex1 = tsused;
2576                 log.u_bbr.flex2 = thresh;
2577                 log.u_bbr.flex3 = rsm->r_flags;
2578                 log.u_bbr.flex4 = rsm->r_dupack;
2579                 log.u_bbr.flex5 = rsm->r_start;
2580                 log.u_bbr.flex6 = rsm->r_end;
2581                 log.u_bbr.flex8 = mod;
2582                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2583                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2584                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2585                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2586                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2587                 log.u_bbr.pacing_gain = rack->r_must_retran;
2588                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2589                     &rack->rc_inp->inp_socket->so_rcv,
2590                     &rack->rc_inp->inp_socket->so_snd,
2591                     BBR_LOG_SETTINGS_CHG, 0,
2592                     0, &log, false, &tv);
2593         }
2594 }
2595
2596 static void
2597 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2598 {
2599         if (tcp_bblogging_on(rack->rc_tp)) {
2600                 union tcp_log_stackspecific log;
2601                 struct timeval tv;
2602
2603                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2604                 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2605                 log.u_bbr.flex2 = to;
2606                 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2607                 log.u_bbr.flex4 = slot;
2608                 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2609                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2610                 log.u_bbr.flex7 = rack->rc_in_persist;
2611                 log.u_bbr.flex8 = which;
2612                 if (rack->rack_no_prr)
2613                         log.u_bbr.pkts_out = 0;
2614                 else
2615                         log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2616                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2617                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2618                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2619                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2620                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2621                 log.u_bbr.pacing_gain = rack->r_must_retran;
2622                 log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2623                 log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2624                 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2625                 log.u_bbr.lost = rack_rto_min;
2626                 log.u_bbr.epoch = rack->r_ctl.roundends;
2627                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2628                     &rack->rc_inp->inp_socket->so_rcv,
2629                     &rack->rc_inp->inp_socket->so_snd,
2630                     BBR_LOG_TIMERSTAR, 0,
2631                     0, &log, false, &tv);
2632         }
2633 }
2634
2635 static void
2636 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2637 {
2638         if (tcp_bblogging_on(rack->rc_tp)) {
2639                 union tcp_log_stackspecific log;
2640                 struct timeval tv;
2641
2642                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2643                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2644                 log.u_bbr.flex8 = to_num;
2645                 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2646                 log.u_bbr.flex2 = rack->rc_rack_rtt;
2647                 if (rsm == NULL)
2648                         log.u_bbr.flex3 = 0;
2649                 else
2650                         log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2651                 if (rack->rack_no_prr)
2652                         log.u_bbr.flex5 = 0;
2653                 else
2654                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2655                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2656                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2657                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2658                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2659                 log.u_bbr.pacing_gain = rack->r_must_retran;
2660                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2661                     &rack->rc_inp->inp_socket->so_rcv,
2662                     &rack->rc_inp->inp_socket->so_snd,
2663                     BBR_LOG_RTO, 0,
2664                     0, &log, false, &tv);
2665         }
2666 }
2667
2668 static void
2669 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2670                  struct rack_sendmap *prev,
2671                  struct rack_sendmap *rsm,
2672                  struct rack_sendmap *next,
2673                  int flag, uint32_t th_ack, int line)
2674 {
2675         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2676                 union tcp_log_stackspecific log;
2677                 struct timeval tv;
2678
2679                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2680                 log.u_bbr.flex8 = flag;
2681                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2682                 log.u_bbr.cur_del_rate = (uint64_t)prev;
2683                 log.u_bbr.delRate = (uint64_t)rsm;
2684                 log.u_bbr.rttProp = (uint64_t)next;
2685                 log.u_bbr.flex7 = 0;
2686                 if (prev) {
2687                         log.u_bbr.flex1 = prev->r_start;
2688                         log.u_bbr.flex2 = prev->r_end;
2689                         log.u_bbr.flex7 |= 0x4;
2690                 }
2691                 if (rsm) {
2692                         log.u_bbr.flex3 = rsm->r_start;
2693                         log.u_bbr.flex4 = rsm->r_end;
2694                         log.u_bbr.flex7 |= 0x2;
2695                 }
2696                 if (next) {
2697                         log.u_bbr.flex5 = next->r_start;
2698                         log.u_bbr.flex6 = next->r_end;
2699                         log.u_bbr.flex7 |= 0x1;
2700                 }
2701                 log.u_bbr.applimited = line;
2702                 log.u_bbr.pkts_out = th_ack;
2703                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2704                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2705                 if (rack->rack_no_prr)
2706                         log.u_bbr.lost = 0;
2707                 else
2708                         log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2709                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2710                     &rack->rc_inp->inp_socket->so_rcv,
2711                     &rack->rc_inp->inp_socket->so_snd,
2712                     TCP_LOG_MAPCHG, 0,
2713                     0, &log, false, &tv);
2714         }
2715 }
2716
2717 static void
2718 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2719                  struct rack_sendmap *rsm, int conf)
2720 {
2721         if (tcp_bblogging_on(tp)) {
2722                 union tcp_log_stackspecific log;
2723                 struct timeval tv;
2724                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2725                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2726                 log.u_bbr.flex1 = t;
2727                 log.u_bbr.flex2 = len;
2728                 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2729                 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2730                 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2731                 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2732                 log.u_bbr.flex7 = conf;
2733                 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2734                 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2735                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2736                 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2737                 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2738                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2739                 if (rsm) {
2740                         log.u_bbr.pkt_epoch = rsm->r_start;
2741                         log.u_bbr.lost = rsm->r_end;
2742                         log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2743                         /* We loose any upper of the 24 bits */
2744                         log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2745                 } else {
2746                         /* Its a SYN */
2747                         log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2748                         log.u_bbr.lost = 0;
2749                         log.u_bbr.cwnd_gain = 0;
2750                         log.u_bbr.pacing_gain = 0;
2751                 }
2752                 /* Write out general bits of interest rrs here */
2753                 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2754                 log.u_bbr.use_lt_bw <<= 1;
2755                 log.u_bbr.use_lt_bw |= rack->forced_ack;
2756                 log.u_bbr.use_lt_bw <<= 1;
2757                 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2758                 log.u_bbr.use_lt_bw <<= 1;
2759                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2760                 log.u_bbr.use_lt_bw <<= 1;
2761                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2762                 log.u_bbr.use_lt_bw <<= 1;
2763                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2764                 log.u_bbr.use_lt_bw <<= 1;
2765                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2766                 log.u_bbr.use_lt_bw <<= 1;
2767                 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2768                 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2769                 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2770                 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2771                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2772                 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2773                 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2774                 log.u_bbr.bw_inuse <<= 32;
2775                 if (rsm)
2776                         log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2777                 TCP_LOG_EVENTP(tp, NULL,
2778                     &rack->rc_inp->inp_socket->so_rcv,
2779                     &rack->rc_inp->inp_socket->so_snd,
2780                     BBR_LOG_BBRRTT, 0,
2781                     0, &log, false, &tv);
2782
2783
2784         }
2785 }
2786
2787 static void
2788 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2789 {
2790         /*
2791          * Log the rtt sample we are
2792          * applying to the srtt algorithm in
2793          * useconds.
2794          */
2795         if (tcp_bblogging_on(rack->rc_tp)) {
2796                 union tcp_log_stackspecific log;
2797                 struct timeval tv;
2798
2799                 /* Convert our ms to a microsecond */
2800                 memset(&log, 0, sizeof(log));
2801                 log.u_bbr.flex1 = rtt;
2802                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2803                 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2804                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2805                 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2806                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2807                 log.u_bbr.flex7 = 1;
2808                 log.u_bbr.flex8 = rack->sack_attack_disable;
2809                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2810                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2811                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2812                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2813                 log.u_bbr.pacing_gain = rack->r_must_retran;
2814                 /*
2815                  * We capture in delRate the upper 32 bits as
2816                  * the confidence level we had declared, and the
2817                  * lower 32 bits as the actual RTT using the arrival
2818                  * timestamp.
2819                  */
2820                 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2821                 log.u_bbr.delRate <<= 32;
2822                 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2823                 /* Lets capture all the things that make up t_rtxcur */
2824                 log.u_bbr.applimited = rack_rto_min;
2825                 log.u_bbr.epoch = rack_rto_max;
2826                 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2827                 log.u_bbr.lost = rack_rto_min;
2828                 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2829                 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2830                 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2831                 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2832                 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2833                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2834                     &rack->rc_inp->inp_socket->so_rcv,
2835                     &rack->rc_inp->inp_socket->so_snd,
2836                     TCP_LOG_RTT, 0,
2837                     0, &log, false, &tv);
2838         }
2839 }
2840
2841 static void
2842 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2843 {
2844         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2845                 union tcp_log_stackspecific log;
2846                 struct timeval tv;
2847
2848                 /* Convert our ms to a microsecond */
2849                 memset(&log, 0, sizeof(log));
2850                 log.u_bbr.flex1 = rtt;
2851                 log.u_bbr.flex2 = send_time;
2852                 log.u_bbr.flex3 = ack_time;
2853                 log.u_bbr.flex4 = where;
2854                 log.u_bbr.flex7 = 2;
2855                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2856                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2857                     &rack->rc_inp->inp_socket->so_rcv,
2858                     &rack->rc_inp->inp_socket->so_snd,
2859                     TCP_LOG_RTT, 0,
2860                     0, &log, false, &tv);
2861         }
2862 }
2863
2864
2865 static void
2866 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
2867 {
2868         if (tcp_bblogging_on(rack->rc_tp)) {
2869                 union tcp_log_stackspecific log;
2870                 struct timeval tv;
2871
2872                 /* Convert our ms to a microsecond */
2873                 memset(&log, 0, sizeof(log));
2874                 log.u_bbr.flex1 = idx;
2875                 log.u_bbr.flex2 = rack_ts_to_msec(tsv);
2876                 log.u_bbr.flex3 = tsecho;
2877                 log.u_bbr.flex7 = 3;
2878                 log.u_bbr.rttProp = tsv;
2879                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2880                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2881                     &rack->rc_inp->inp_socket->so_rcv,
2882                     &rack->rc_inp->inp_socket->so_snd,
2883                     TCP_LOG_RTT, 0,
2884                     0, &log, false, &tv);
2885         }
2886 }
2887
2888
2889 static inline void
2890 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2891 {
2892         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2893                 union tcp_log_stackspecific log;
2894                 struct timeval tv;
2895
2896                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2897                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2898                 log.u_bbr.flex1 = line;
2899                 log.u_bbr.flex2 = tick;
2900                 log.u_bbr.flex3 = tp->t_maxunacktime;
2901                 log.u_bbr.flex4 = tp->t_acktime;
2902                 log.u_bbr.flex8 = event;
2903                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2904                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2905                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2906                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2907                 log.u_bbr.pacing_gain = rack->r_must_retran;
2908                 TCP_LOG_EVENTP(tp, NULL,
2909                     &rack->rc_inp->inp_socket->so_rcv,
2910                     &rack->rc_inp->inp_socket->so_snd,
2911                     BBR_LOG_PROGRESS, 0,
2912                     0, &log, false, &tv);
2913         }
2914 }
2915
2916 static void
2917 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
2918 {
2919         if (tcp_bblogging_on(rack->rc_tp)) {
2920                 union tcp_log_stackspecific log;
2921
2922                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2923                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2924                 log.u_bbr.flex1 = slot;
2925                 if (rack->rack_no_prr)
2926                         log.u_bbr.flex2 = 0;
2927                 else
2928                         log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2929                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2930                 log.u_bbr.flex5 = rack->r_ctl.ack_during_sd;
2931                 log.u_bbr.flex6 = line;
2932                 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2933                 log.u_bbr.flex8 = rack->rc_in_persist;
2934                 log.u_bbr.timeStamp = cts;
2935                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2936                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2937                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2938                 log.u_bbr.pacing_gain = rack->r_must_retran;
2939                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2940                     &rack->rc_inp->inp_socket->so_rcv,
2941                     &rack->rc_inp->inp_socket->so_snd,
2942                     BBR_LOG_BBRSND, 0,
2943                     0, &log, false, tv);
2944         }
2945 }
2946
2947 static void
2948 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2949 {
2950         if (tcp_bblogging_on(rack->rc_tp)) {
2951                 union tcp_log_stackspecific log;
2952                 struct timeval tv;
2953
2954                 memset(&log, 0, sizeof(log));
2955                 log.u_bbr.flex1 = did_out;
2956                 log.u_bbr.flex2 = nxt_pkt;
2957                 log.u_bbr.flex3 = way_out;
2958                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2959                 if (rack->rack_no_prr)
2960                         log.u_bbr.flex5 = 0;
2961                 else
2962                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2963                 log.u_bbr.flex6 = nsegs;
2964                 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2965                 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;        /* Do we have ack-can-send set */
2966                 log.u_bbr.flex7 <<= 1;
2967                 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2968                 log.u_bbr.flex7 <<= 1;
2969                 log.u_bbr.flex7 |= rack->r_wanted_output;       /* Do we want output */
2970                 log.u_bbr.flex8 = rack->rc_in_persist;
2971                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2972                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2973                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2974                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2975                 log.u_bbr.use_lt_bw <<= 1;
2976                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2977                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2978                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2979                 log.u_bbr.pacing_gain = rack->r_must_retran;
2980                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2981                     &rack->rc_inp->inp_socket->so_rcv,
2982                     &rack->rc_inp->inp_socket->so_snd,
2983                     BBR_LOG_DOSEG_DONE, 0,
2984                     0, &log, false, &tv);
2985         }
2986 }
2987
2988 static void
2989 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2990 {
2991         if (tcp_bblogging_on(rack->rc_tp)) {
2992                 union tcp_log_stackspecific log;
2993                 struct timeval tv;
2994
2995                 memset(&log, 0, sizeof(log));
2996                 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2997                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2998                 log.u_bbr.flex4 = arg1;
2999                 log.u_bbr.flex5 = arg2;
3000                 log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3001                 log.u_bbr.flex6 = arg3;
3002                 log.u_bbr.flex8 = frm;
3003                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3004                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3005                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3006                 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3007                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3008                 log.u_bbr.pacing_gain = rack->r_must_retran;
3009                 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3010                     &tptosocket(tp)->so_snd,
3011                     TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3012         }
3013 }
3014
3015 static void
3016 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3017                           uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3018 {
3019         if (tcp_bblogging_on(rack->rc_tp)) {
3020                 union tcp_log_stackspecific log;
3021                 struct timeval tv;
3022
3023                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3024                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
3025                 log.u_bbr.flex1 = slot;
3026                 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3027                 log.u_bbr.flex4 = reason;
3028                 if (rack->rack_no_prr)
3029                         log.u_bbr.flex5 = 0;
3030                 else
3031                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3032                 log.u_bbr.flex7 = hpts_calling;
3033                 log.u_bbr.flex8 = rack->rc_in_persist;
3034                 log.u_bbr.lt_epoch = cwnd_to_use;
3035                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3036                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3037                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3038                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3039                 log.u_bbr.pacing_gain = rack->r_must_retran;
3040                 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
3041                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3042                     &rack->rc_inp->inp_socket->so_rcv,
3043                     &rack->rc_inp->inp_socket->so_snd,
3044                     BBR_LOG_JUSTRET, 0,
3045                     tlen, &log, false, &tv);
3046         }
3047 }
3048
3049 static void
3050 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3051                    struct timeval *tv, uint32_t flags_on_entry)
3052 {
3053         if (tcp_bblogging_on(rack->rc_tp)) {
3054                 union tcp_log_stackspecific log;
3055
3056                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3057                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
3058                 log.u_bbr.flex1 = line;
3059                 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3060                 log.u_bbr.flex3 = flags_on_entry;
3061                 log.u_bbr.flex4 = us_cts;
3062                 if (rack->rack_no_prr)
3063                         log.u_bbr.flex5 = 0;
3064                 else
3065                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3066                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3067                 log.u_bbr.flex7 = hpts_removed;
3068                 log.u_bbr.flex8 = 1;
3069                 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3070                 log.u_bbr.timeStamp = us_cts;
3071                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3072                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3073                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3074                 log.u_bbr.pacing_gain = rack->r_must_retran;
3075                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3076                     &rack->rc_inp->inp_socket->so_rcv,
3077                     &rack->rc_inp->inp_socket->so_snd,
3078                     BBR_LOG_TIMERCANC, 0,
3079                     0, &log, false, tv);
3080         }
3081 }
3082
3083 static void
3084 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3085                           uint32_t flex1, uint32_t flex2,
3086                           uint32_t flex3, uint32_t flex4,
3087                           uint32_t flex5, uint32_t flex6,
3088                           uint16_t flex7, uint8_t mod)
3089 {
3090         if (tcp_bblogging_on(rack->rc_tp)) {
3091                 union tcp_log_stackspecific log;
3092                 struct timeval tv;
3093
3094                 if (mod == 1) {
3095                         /* No you can't use 1, its for the real to cancel */
3096                         return;
3097                 }
3098                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3099                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3100                 log.u_bbr.flex1 = flex1;
3101                 log.u_bbr.flex2 = flex2;
3102                 log.u_bbr.flex3 = flex3;
3103                 log.u_bbr.flex4 = flex4;
3104                 log.u_bbr.flex5 = flex5;
3105                 log.u_bbr.flex6 = flex6;
3106                 log.u_bbr.flex7 = flex7;
3107                 log.u_bbr.flex8 = mod;
3108                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3109                     &rack->rc_inp->inp_socket->so_rcv,
3110                     &rack->rc_inp->inp_socket->so_snd,
3111                     BBR_LOG_TIMERCANC, 0,
3112                     0, &log, false, &tv);
3113         }
3114 }
3115
3116 static void
3117 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3118 {
3119         if (tcp_bblogging_on(rack->rc_tp)) {
3120                 union tcp_log_stackspecific log;
3121                 struct timeval tv;
3122
3123                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3124                 log.u_bbr.flex1 = timers;
3125                 log.u_bbr.flex2 = ret;
3126                 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3127                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3128                 log.u_bbr.flex5 = cts;
3129                 if (rack->rack_no_prr)
3130                         log.u_bbr.flex6 = 0;
3131                 else
3132                         log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3133                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3134                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3135                 log.u_bbr.pacing_gain = rack->r_must_retran;
3136                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3137                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3138                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3139                     &rack->rc_inp->inp_socket->so_rcv,
3140                     &rack->rc_inp->inp_socket->so_snd,
3141                     BBR_LOG_TO_PROCESS, 0,
3142                     0, &log, false, &tv);
3143         }
3144 }
3145
3146 static void
3147 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3148 {
3149         if (tcp_bblogging_on(rack->rc_tp)) {
3150                 union tcp_log_stackspecific log;
3151                 struct timeval tv;
3152
3153                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3154                 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3155                 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3156                 if (rack->rack_no_prr)
3157                         log.u_bbr.flex3 = 0;
3158                 else
3159                         log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3160                 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3161                 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3162                 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3163                 log.u_bbr.flex7 = line;
3164                 log.u_bbr.flex8 = frm;
3165                 log.u_bbr.pkts_out = orig_cwnd;
3166                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3167                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3168                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3169                 log.u_bbr.use_lt_bw <<= 1;
3170                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
3171                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3172                     &rack->rc_inp->inp_socket->so_rcv,
3173                     &rack->rc_inp->inp_socket->so_snd,
3174                     BBR_LOG_BBRUPD, 0,
3175                     0, &log, false, &tv);
3176         }
3177 }
3178
3179 #ifdef TCP_SAD_DETECTION
3180 static void
3181 rack_log_sad(struct tcp_rack *rack, int event)
3182 {
3183         if (tcp_bblogging_on(rack->rc_tp)) {
3184                 union tcp_log_stackspecific log;
3185                 struct timeval tv;
3186
3187                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3188                 log.u_bbr.flex1 = rack->r_ctl.sack_count;
3189                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
3190                 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
3191                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
3192                 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
3193                 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
3194                 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
3195                 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
3196                 log.u_bbr.lt_epoch |= rack->do_detection;
3197                 log.u_bbr.applimited = tcp_map_minimum;
3198                 log.u_bbr.flex7 = rack->sack_attack_disable;
3199                 log.u_bbr.flex8 = event;
3200                 log.u_bbr.bbr_state = rack->rc_suspicious;
3201                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3202                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3203                 log.u_bbr.delivered = tcp_sad_decay_val;
3204                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3205                     &rack->rc_inp->inp_socket->so_rcv,
3206                     &rack->rc_inp->inp_socket->so_snd,
3207                     TCP_SAD_DETECT, 0,
3208                     0, &log, false, &tv);
3209         }
3210 }
3211 #endif
3212
3213 static void
3214 rack_counter_destroy(void)
3215 {
3216         counter_u64_free(rack_total_bytes);
3217         counter_u64_free(rack_fto_send);
3218         counter_u64_free(rack_fto_rsm_send);
3219         counter_u64_free(rack_nfto_resend);
3220         counter_u64_free(rack_hw_pace_init_fail);
3221         counter_u64_free(rack_hw_pace_lost);
3222         counter_u64_free(rack_non_fto_send);
3223         counter_u64_free(rack_extended_rfo);
3224         counter_u64_free(rack_ack_total);
3225         counter_u64_free(rack_express_sack);
3226         counter_u64_free(rack_sack_total);
3227         counter_u64_free(rack_move_none);
3228         counter_u64_free(rack_move_some);
3229         counter_u64_free(rack_sack_attacks_detected);
3230         counter_u64_free(rack_sack_attacks_reversed);
3231         counter_u64_free(rack_sack_attacks_suspect);
3232         counter_u64_free(rack_sack_used_next_merge);
3233         counter_u64_free(rack_sack_used_prev_merge);
3234         counter_u64_free(rack_tlp_tot);
3235         counter_u64_free(rack_tlp_newdata);
3236         counter_u64_free(rack_tlp_retran);
3237         counter_u64_free(rack_tlp_retran_bytes);
3238         counter_u64_free(rack_to_tot);
3239         counter_u64_free(rack_saw_enobuf);
3240         counter_u64_free(rack_saw_enobuf_hw);
3241         counter_u64_free(rack_saw_enetunreach);
3242         counter_u64_free(rack_hot_alloc);
3243         counter_u64_free(rack_to_alloc);
3244         counter_u64_free(rack_to_alloc_hard);
3245         counter_u64_free(rack_to_alloc_emerg);
3246         counter_u64_free(rack_to_alloc_limited);
3247         counter_u64_free(rack_alloc_limited_conns);
3248         counter_u64_free(rack_split_limited);
3249         counter_u64_free(rack_multi_single_eq);
3250         counter_u64_free(rack_rxt_clamps_cwnd);
3251         counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3252         counter_u64_free(rack_proc_non_comp_ack);
3253         counter_u64_free(rack_sack_proc_all);
3254         counter_u64_free(rack_sack_proc_restart);
3255         counter_u64_free(rack_sack_proc_short);
3256         counter_u64_free(rack_sack_skipped_acked);
3257         counter_u64_free(rack_sack_splits);
3258         counter_u64_free(rack_input_idle_reduces);
3259         counter_u64_free(rack_collapsed_win);
3260         counter_u64_free(rack_collapsed_win_rxt);
3261         counter_u64_free(rack_collapsed_win_rxt_bytes);
3262         counter_u64_free(rack_collapsed_win_seen);
3263         counter_u64_free(rack_try_scwnd);
3264         counter_u64_free(rack_persists_sends);
3265         counter_u64_free(rack_persists_acks);
3266         counter_u64_free(rack_persists_loss);
3267         counter_u64_free(rack_persists_lost_ends);
3268 #ifdef INVARIANTS
3269         counter_u64_free(rack_adjust_map_bw);
3270 #endif
3271         COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3272         COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3273 }
3274
3275 static struct rack_sendmap *
3276 rack_alloc(struct tcp_rack *rack)
3277 {
3278         struct rack_sendmap *rsm;
3279
3280         /*
3281          * First get the top of the list it in
3282          * theory is the "hottest" rsm we have,
3283          * possibly just freed by ack processing.
3284          */
3285         if (rack->rc_free_cnt > rack_free_cache) {
3286                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3287                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3288                 counter_u64_add(rack_hot_alloc, 1);
3289                 rack->rc_free_cnt--;
3290                 return (rsm);
3291         }
3292         /*
3293          * Once we get under our free cache we probably
3294          * no longer have a "hot" one available. Lets
3295          * get one from UMA.
3296          */
3297         rsm = uma_zalloc(rack_zone, M_NOWAIT);
3298         if (rsm) {
3299                 rack->r_ctl.rc_num_maps_alloced++;
3300                 counter_u64_add(rack_to_alloc, 1);
3301                 return (rsm);
3302         }
3303         /*
3304          * Dig in to our aux rsm's (the last two) since
3305          * UMA failed to get us one.
3306          */
3307         if (rack->rc_free_cnt) {
3308                 counter_u64_add(rack_to_alloc_emerg, 1);
3309                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3310                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3311                 rack->rc_free_cnt--;
3312                 return (rsm);
3313         }
3314         return (NULL);
3315 }
3316
3317 static struct rack_sendmap *
3318 rack_alloc_full_limit(struct tcp_rack *rack)
3319 {
3320         if ((V_tcp_map_entries_limit > 0) &&
3321             (rack->do_detection == 0) &&
3322             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3323                 counter_u64_add(rack_to_alloc_limited, 1);
3324                 if (!rack->alloc_limit_reported) {
3325                         rack->alloc_limit_reported = 1;
3326                         counter_u64_add(rack_alloc_limited_conns, 1);
3327                 }
3328                 return (NULL);
3329         }
3330         return (rack_alloc(rack));
3331 }
3332
3333 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3334 static struct rack_sendmap *
3335 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3336 {
3337         struct rack_sendmap *rsm;
3338
3339         if (limit_type) {
3340                 /* currently there is only one limit type */
3341                 if (rack->r_ctl.rc_split_limit > 0 &&
3342                     (rack->do_detection == 0) &&
3343                     rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3344                         counter_u64_add(rack_split_limited, 1);
3345                         if (!rack->alloc_limit_reported) {
3346                                 rack->alloc_limit_reported = 1;
3347                                 counter_u64_add(rack_alloc_limited_conns, 1);
3348                         }
3349                         return (NULL);
3350 #ifdef TCP_SAD_DETECTION
3351                 } else if ((tcp_sad_limit != 0) &&
3352                            (rack->do_detection == 1) &&
3353                            (rack->r_ctl.rc_num_split_allocs >= tcp_sad_limit)) {
3354                         counter_u64_add(rack_split_limited, 1);
3355                         if (!rack->alloc_limit_reported) {
3356                                 rack->alloc_limit_reported = 1;
3357                                 counter_u64_add(rack_alloc_limited_conns, 1);
3358                         }
3359                         return (NULL);
3360 #endif
3361                 }
3362         }
3363
3364         /* allocate and mark in the limit type, if set */
3365         rsm = rack_alloc(rack);
3366         if (rsm != NULL && limit_type) {
3367                 rsm->r_limit_type = limit_type;
3368                 rack->r_ctl.rc_num_split_allocs++;
3369         }
3370         return (rsm);
3371 }
3372
3373 static void
3374 rack_free_trim(struct tcp_rack *rack)
3375 {
3376         struct rack_sendmap *rsm;
3377
3378         /*
3379          * Free up all the tail entries until
3380          * we get our list down to the limit.
3381          */
3382         while (rack->rc_free_cnt > rack_free_cache) {
3383                 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3384                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3385                 rack->rc_free_cnt--;
3386                 rack->r_ctl.rc_num_maps_alloced--;
3387                 uma_zfree(rack_zone, rsm);
3388         }
3389 }
3390
3391 static void
3392 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3393 {
3394         if (rsm->r_flags & RACK_APP_LIMITED) {
3395                 if (rack->r_ctl.rc_app_limited_cnt > 0) {
3396                         rack->r_ctl.rc_app_limited_cnt--;
3397                 }
3398         }
3399         if (rsm->r_limit_type) {
3400                 /* currently there is only one limit type */
3401                 rack->r_ctl.rc_num_split_allocs--;
3402         }
3403         if (rsm == rack->r_ctl.rc_first_appl) {
3404                 if (rack->r_ctl.rc_app_limited_cnt == 0)
3405                         rack->r_ctl.rc_first_appl = NULL;
3406                 else
3407                         rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3408         }
3409         if (rsm == rack->r_ctl.rc_resend)
3410                 rack->r_ctl.rc_resend = NULL;
3411         if (rsm == rack->r_ctl.rc_end_appl)
3412                 rack->r_ctl.rc_end_appl = NULL;
3413         if (rack->r_ctl.rc_tlpsend == rsm)
3414                 rack->r_ctl.rc_tlpsend = NULL;
3415         if (rack->r_ctl.rc_sacklast == rsm)
3416                 rack->r_ctl.rc_sacklast = NULL;
3417         memset(rsm, 0, sizeof(struct rack_sendmap));
3418         /* Make sure we are not going to overrun our count limit of 0xff */
3419         if ((rack->rc_free_cnt + 1) > 0xff) {
3420                 rack_free_trim(rack);
3421         }
3422         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3423         rack->rc_free_cnt++;
3424 }
3425
3426 static uint32_t
3427 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3428 {
3429         uint64_t srtt, bw, len, tim;
3430         uint32_t segsiz, def_len, minl;
3431
3432         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3433         def_len = rack_def_data_window * segsiz;
3434         if (rack->rc_gp_filled == 0) {
3435                 /*
3436                  * We have no measurement (IW is in flight?) so
3437                  * we can only guess using our data_window sysctl
3438                  * value (usually 20MSS).
3439                  */
3440                 return (def_len);
3441         }
3442         /*
3443          * Now we have a number of factors to consider.
3444          *
3445          * 1) We have a desired BDP which is usually
3446          *    at least 2.
3447          * 2) We have a minimum number of rtt's usually 1 SRTT
3448          *    but we allow it too to be more.
3449          * 3) We want to make sure a measurement last N useconds (if
3450          *    we have set rack_min_measure_usec.
3451          *
3452          * We handle the first concern here by trying to create a data
3453          * window of max(rack_def_data_window, DesiredBDP). The
3454          * second concern we handle in not letting the measurement
3455          * window end normally until at least the required SRTT's
3456          * have gone by which is done further below in
3457          * rack_enough_for_measurement(). Finally the third concern
3458          * we also handle here by calculating how long that time
3459          * would take at the current BW and then return the
3460          * max of our first calculation and that length. Note
3461          * that if rack_min_measure_usec is 0, we don't deal
3462          * with concern 3. Also for both Concern 1 and 3 an
3463          * application limited period could end the measurement
3464          * earlier.
3465          *
3466          * So lets calculate the BDP with the "known" b/w using
3467          * the SRTT has our rtt and then multiply it by the
3468          * goal.
3469          */
3470         bw = rack_get_bw(rack);
3471         srtt = (uint64_t)tp->t_srtt;
3472         len = bw * srtt;
3473         len /= (uint64_t)HPTS_USEC_IN_SEC;
3474         len *= max(1, rack_goal_bdp);
3475         /* Now we need to round up to the nearest MSS */
3476         len = roundup(len, segsiz);
3477         if (rack_min_measure_usec) {
3478                 /* Now calculate our min length for this b/w */
3479                 tim = rack_min_measure_usec;
3480                 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3481                 if (minl == 0)
3482                         minl = 1;
3483                 minl = roundup(minl, segsiz);
3484                 if (len < minl)
3485                         len = minl;
3486         }
3487         /*
3488          * Now if we have a very small window we want
3489          * to attempt to get the window that is
3490          * as small as possible. This happens on
3491          * low b/w connections and we don't want to
3492          * span huge numbers of rtt's between measurements.
3493          *
3494          * We basically include 2 over our "MIN window" so
3495          * that the measurement can be shortened (possibly) by
3496          * an ack'ed packet.
3497          */
3498         if (len < def_len)
3499                 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3500         else
3501                 return (max((uint32_t)len, def_len));
3502
3503 }
3504
3505 static int
3506 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3507 {
3508         uint32_t tim, srtts, segsiz;
3509
3510         /*
3511          * Has enough time passed for the GP measurement to be valid?
3512          */
3513         if (SEQ_LT(th_ack, tp->gput_seq)) {
3514                 /* Not enough bytes yet */
3515                 return (0);
3516         }
3517         if ((tp->snd_max == tp->snd_una) ||
3518             (th_ack == tp->snd_max)){
3519                 /*
3520                  * All is acked quality of all acked is
3521                  * usually low or medium, but we in theory could split
3522                  * all acked into two cases, where you got
3523                  * a signifigant amount of your window and
3524                  * where you did not. For now we leave it
3525                  * but it is something to contemplate in the
3526                  * future. The danger here is that delayed ack
3527                  * is effecting the last byte (which is a 50:50 chance).
3528                  */
3529                 *quality = RACK_QUALITY_ALLACKED;
3530                 return (1);
3531         }
3532         if (SEQ_GEQ(th_ack,  tp->gput_ack)) {
3533                 /*
3534                  * We obtained our entire window of data we wanted
3535                  * no matter if we are in recovery or not then
3536                  * its ok since expanding the window does not
3537                  * make things fuzzy (or at least not as much).
3538                  */
3539                 *quality = RACK_QUALITY_HIGH;
3540                 return (1);
3541         }
3542         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3543         if (SEQ_LT(th_ack, tp->gput_ack) &&
3544             ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3545                 /* Not enough bytes yet */
3546                 return (0);
3547         }
3548         if (rack->r_ctl.rc_first_appl &&
3549             (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3550                 /*
3551                  * We are up to the app limited send point
3552                  * we have to measure irrespective of the time..
3553                  */
3554                 *quality = RACK_QUALITY_APPLIMITED;
3555                 return (1);
3556         }
3557         /* Now what about time? */
3558         srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3559         tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3560         if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3561                 /*
3562                  * We do not allow a measurement if we are in recovery
3563                  * that would shrink the goodput window we wanted.
3564                  * This is to prevent cloudyness of when the last send
3565                  * was actually made.
3566                  */
3567                 *quality = RACK_QUALITY_HIGH;
3568                 return (1);
3569         }
3570         /* Nope not even a full SRTT has passed */
3571         return (0);
3572 }
3573
3574 static void
3575 rack_log_timely(struct tcp_rack *rack,
3576                 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3577                 uint64_t up_bnd, int line, uint8_t method)
3578 {
3579         if (tcp_bblogging_on(rack->rc_tp)) {
3580                 union tcp_log_stackspecific log;
3581                 struct timeval tv;
3582
3583                 memset(&log, 0, sizeof(log));
3584                 log.u_bbr.flex1 = logged;
3585                 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3586                 log.u_bbr.flex2 <<= 4;
3587                 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3588                 log.u_bbr.flex2 <<= 4;
3589                 log.u_bbr.flex2 |= rack->rc_gp_incr;
3590                 log.u_bbr.flex2 <<= 4;
3591                 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3592                 log.u_bbr.flex3 = rack->rc_gp_incr;
3593                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3594                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3595                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3596                 log.u_bbr.flex7 = rack->rc_gp_bwred;
3597                 log.u_bbr.flex8 = method;
3598                 log.u_bbr.cur_del_rate = cur_bw;
3599                 log.u_bbr.delRate = low_bnd;
3600                 log.u_bbr.bw_inuse = up_bnd;
3601                 log.u_bbr.rttProp = rack_get_bw(rack);
3602                 log.u_bbr.pkt_epoch = line;
3603                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3604                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3605                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3606                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3607                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3608                 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3609                 log.u_bbr.cwnd_gain <<= 1;
3610                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3611                 log.u_bbr.cwnd_gain <<= 1;
3612                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3613                 log.u_bbr.cwnd_gain <<= 1;
3614                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3615                 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3616                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3617                     &rack->rc_inp->inp_socket->so_rcv,
3618                     &rack->rc_inp->inp_socket->so_snd,
3619                     TCP_TIMELY_WORK, 0,
3620                     0, &log, false, &tv);
3621         }
3622 }
3623
3624 static int
3625 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3626 {
3627         /*
3628          * Before we increase we need to know if
3629          * the estimate just made was less than
3630          * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3631          *
3632          * If we already are pacing at a fast enough
3633          * rate to push us faster there is no sense of
3634          * increasing.
3635          *
3636          * We first caculate our actual pacing rate (ss or ca multiplier
3637          * times our cur_bw).
3638          *
3639          * Then we take the last measured rate and multipy by our
3640          * maximum pacing overage to give us a max allowable rate.
3641          *
3642          * If our act_rate is smaller than our max_allowable rate
3643          * then we should increase. Else we should hold steady.
3644          *
3645          */
3646         uint64_t act_rate, max_allow_rate;
3647
3648         if (rack_timely_no_stopping)
3649                 return (1);
3650
3651         if ((cur_bw == 0) || (last_bw_est == 0)) {
3652                 /*
3653                  * Initial startup case or
3654                  * everything is acked case.
3655                  */
3656                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3657                                 __LINE__, 9);
3658                 return (1);
3659         }
3660         if (mult <= 100) {
3661                 /*
3662                  * We can always pace at or slightly above our rate.
3663                  */
3664                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3665                                 __LINE__, 9);
3666                 return (1);
3667         }
3668         act_rate = cur_bw * (uint64_t)mult;
3669         act_rate /= 100;
3670         max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3671         max_allow_rate /= 100;
3672         if (act_rate < max_allow_rate) {
3673                 /*
3674                  * Here the rate we are actually pacing at
3675                  * is smaller than 10% above our last measurement.
3676                  * This means we are pacing below what we would
3677                  * like to try to achieve (plus some wiggle room).
3678                  */
3679                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3680                                 __LINE__, 9);
3681                 return (1);
3682         } else {
3683                 /*
3684                  * Here we are already pacing at least rack_max_per_above(10%)
3685                  * what we are getting back. This indicates most likely
3686                  * that we are being limited (cwnd/rwnd/app) and can't
3687                  * get any more b/w. There is no sense of trying to
3688                  * raise up the pacing rate its not speeding us up
3689                  * and we already are pacing faster than we are getting.
3690                  */
3691                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3692                                 __LINE__, 8);
3693                 return (0);
3694         }
3695 }
3696
3697 static void
3698 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3699 {
3700         /*
3701          * When we drag bottom, we want to assure
3702          * that no multiplier is below 1.0, if so
3703          * we want to restore it to at least that.
3704          */
3705         if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3706                 /* This is unlikely we usually do not touch recovery */
3707                 rack->r_ctl.rack_per_of_gp_rec = 100;
3708         }
3709         if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3710                 rack->r_ctl.rack_per_of_gp_ca = 100;
3711         }
3712         if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3713                 rack->r_ctl.rack_per_of_gp_ss = 100;
3714         }
3715 }
3716
3717 static void
3718 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3719 {
3720         if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3721                 rack->r_ctl.rack_per_of_gp_ca = 100;
3722         }
3723         if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3724                 rack->r_ctl.rack_per_of_gp_ss = 100;
3725         }
3726 }
3727
3728 static void
3729 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3730 {
3731         int32_t  calc, logged, plus;
3732
3733         logged = 0;
3734
3735         if (override) {
3736                 /*
3737                  * override is passed when we are
3738                  * loosing b/w and making one last
3739                  * gasp at trying to not loose out
3740                  * to a new-reno flow.
3741                  */
3742                 goto extra_boost;
3743         }
3744         /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3745         if (rack->rc_gp_incr &&
3746             ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3747                 /*
3748                  * Reset and get 5 strokes more before the boost. Note
3749                  * that the count is 0 based so we have to add one.
3750                  */
3751 extra_boost:
3752                 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3753                 rack->rc_gp_timely_inc_cnt = 0;
3754         } else
3755                 plus = (uint32_t)rack_gp_increase_per;
3756         /* Must be at least 1% increase for true timely increases */
3757         if ((plus < 1) &&
3758             ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3759                 plus = 1;
3760         if (rack->rc_gp_saw_rec &&
3761             (rack->rc_gp_no_rec_chg == 0) &&
3762             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3763                                   rack->r_ctl.rack_per_of_gp_rec)) {
3764                 /* We have been in recovery ding it too */
3765                 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3766                 if (calc > 0xffff)
3767                         calc = 0xffff;
3768                 logged |= 1;
3769                 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3770                 if (rack->r_ctl.rack_per_upper_bound_ca &&
3771                     (rack->rc_dragged_bottom == 0) &&
3772                     (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
3773                         rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
3774         }
3775         if (rack->rc_gp_saw_ca &&
3776             (rack->rc_gp_saw_ss == 0) &&
3777             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3778                                   rack->r_ctl.rack_per_of_gp_ca)) {
3779                 /* In CA */
3780                 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3781                 if (calc > 0xffff)
3782                         calc = 0xffff;
3783                 logged |= 2;
3784                 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3785                 if (rack->r_ctl.rack_per_upper_bound_ca &&
3786                     (rack->rc_dragged_bottom == 0) &&
3787                     (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
3788                         rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
3789         }
3790         if (rack->rc_gp_saw_ss &&
3791             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3792                                   rack->r_ctl.rack_per_of_gp_ss)) {
3793                 /* In SS */
3794                 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3795                 if (calc > 0xffff)
3796                         calc = 0xffff;
3797                 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3798                 if (rack->r_ctl.rack_per_upper_bound_ss &&
3799                     (rack->rc_dragged_bottom == 0) &&
3800                     (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
3801                         rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
3802                 logged |= 4;
3803         }
3804         if (logged &&
3805             (rack->rc_gp_incr == 0)){
3806                 /* Go into increment mode */
3807                 rack->rc_gp_incr = 1;
3808                 rack->rc_gp_timely_inc_cnt = 0;
3809         }
3810         if (rack->rc_gp_incr &&
3811             logged &&
3812             (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3813                 rack->rc_gp_timely_inc_cnt++;
3814         }
3815         rack_log_timely(rack,  logged, plus, 0, 0,
3816                         __LINE__, 1);
3817 }
3818
3819 static uint32_t
3820 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3821 {
3822         /*
3823          * norm_grad = rtt_diff / minrtt;
3824          * new_per = curper * (1 - B * norm_grad)
3825          *
3826          * B = rack_gp_decrease_per (default 10%)
3827          * rtt_dif = input var current rtt-diff
3828          * curper = input var current percentage
3829          * minrtt = from rack filter
3830          *
3831          */
3832         uint64_t perf;
3833
3834         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3835                     ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3836                      (((uint64_t)rtt_diff * (uint64_t)1000000)/
3837                       (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3838                      (uint64_t)1000000)) /
3839                 (uint64_t)1000000);
3840         if (perf > curper) {
3841                 /* TSNH */
3842                 perf = curper - 1;
3843         }
3844         return ((uint32_t)perf);
3845 }
3846
3847 static uint32_t
3848 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3849 {
3850         /*
3851          *                                   highrttthresh
3852          * result = curper * (1 - (B * ( 1 -  ------          ))
3853          *                                     gp_srtt
3854          *
3855          * B = rack_gp_decrease_per (default 10%)
3856          * highrttthresh = filter_min * rack_gp_rtt_maxmul
3857          */
3858         uint64_t perf;
3859         uint32_t highrttthresh;
3860
3861         highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3862
3863         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3864                                      ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3865                                         ((uint64_t)highrttthresh * (uint64_t)1000000) /
3866                                                     (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3867         return (perf);
3868 }
3869
3870 static void
3871 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3872 {
3873         uint64_t logvar, logvar2, logvar3;
3874         uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3875
3876         if (rack->rc_gp_incr) {
3877                 /* Turn off increment counting */
3878                 rack->rc_gp_incr = 0;
3879                 rack->rc_gp_timely_inc_cnt = 0;
3880         }
3881         ss_red = ca_red = rec_red = 0;
3882         logged = 0;
3883         /* Calculate the reduction value */
3884         if (rtt_diff < 0) {
3885                 rtt_diff *= -1;
3886         }
3887         /* Must be at least 1% reduction */
3888         if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3889                 /* We have been in recovery ding it too */
3890                 if (timely_says == 2) {
3891                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3892                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3893                         if (alt < new_per)
3894                                 val = alt;
3895                         else
3896                                 val = new_per;
3897                 } else
3898                          val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3899                 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3900                         rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3901                         rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3902                 } else {
3903                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3904                         rec_red = 0;
3905                 }
3906                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3907                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3908                 logged |= 1;
3909         }
3910         if (rack->rc_gp_saw_ss) {
3911                 /* Sent in SS */
3912                 if (timely_says == 2) {
3913                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3914                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3915                         if (alt < new_per)
3916                                 val = alt;
3917                         else
3918                                 val = new_per;
3919                 } else
3920                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3921                 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3922                         ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3923                         rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3924                 } else {
3925                         ss_red = new_per;
3926                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3927                         logvar = new_per;
3928                         logvar <<= 32;
3929                         logvar |= alt;
3930                         logvar2 = (uint32_t)rtt;
3931                         logvar2 <<= 32;
3932                         logvar2 |= (uint32_t)rtt_diff;
3933                         logvar3 = rack_gp_rtt_maxmul;
3934                         logvar3 <<= 32;
3935                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3936                         rack_log_timely(rack, timely_says,
3937                                         logvar2, logvar3,
3938                                         logvar, __LINE__, 10);
3939                 }
3940                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3941                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3942                 logged |= 4;
3943         } else if (rack->rc_gp_saw_ca) {
3944                 /* Sent in CA */
3945                 if (timely_says == 2) {
3946                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3947                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3948                         if (alt < new_per)
3949                                 val = alt;
3950                         else
3951                                 val = new_per;
3952                 } else
3953                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3954                 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3955                         ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3956                         rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3957                 } else {
3958                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3959                         ca_red = 0;
3960                         logvar = new_per;
3961                         logvar <<= 32;
3962                         logvar |= alt;
3963                         logvar2 = (uint32_t)rtt;
3964                         logvar2 <<= 32;
3965                         logvar2 |= (uint32_t)rtt_diff;
3966                         logvar3 = rack_gp_rtt_maxmul;
3967                         logvar3 <<= 32;
3968                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3969                         rack_log_timely(rack, timely_says,
3970                                         logvar2, logvar3,
3971                                         logvar, __LINE__, 10);
3972                 }
3973                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3974                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3975                 logged |= 2;
3976         }
3977         if (rack->rc_gp_timely_dec_cnt < 0x7) {
3978                 rack->rc_gp_timely_dec_cnt++;
3979                 if (rack_timely_dec_clear &&
3980                     (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3981                         rack->rc_gp_timely_dec_cnt = 0;
3982         }
3983         logvar = ss_red;
3984         logvar <<= 32;
3985         logvar |= ca_red;
3986         rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3987                         __LINE__, 2);
3988 }
3989
3990 static void
3991 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3992                      uint32_t rtt, uint32_t line, uint8_t reas)
3993 {
3994         if (tcp_bblogging_on(rack->rc_tp)) {
3995                 union tcp_log_stackspecific log;
3996                 struct timeval tv;
3997
3998                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3999                 log.u_bbr.flex1 = line;
4000                 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4001                 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4002                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4003                 log.u_bbr.flex5 = rtt;
4004                 log.u_bbr.flex6 = rack->rc_highly_buffered;
4005                 log.u_bbr.flex6 <<= 1;
4006                 log.u_bbr.flex6 |= rack->forced_ack;
4007                 log.u_bbr.flex6 <<= 1;
4008                 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4009                 log.u_bbr.flex6 <<= 1;
4010                 log.u_bbr.flex6 |= rack->in_probe_rtt;
4011                 log.u_bbr.flex6 <<= 1;
4012                 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4013                 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4014                 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4015                 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4016                 log.u_bbr.flex8 = reas;
4017                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4018                 log.u_bbr.delRate = rack_get_bw(rack);
4019                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4020                 log.u_bbr.cur_del_rate <<= 32;
4021                 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4022                 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4023                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4024                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4025                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4026                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4027                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4028                 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4029                 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4030                 log.u_bbr.rttProp = us_cts;
4031                 log.u_bbr.rttProp <<= 32;
4032                 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4033                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
4034                     &rack->rc_inp->inp_socket->so_rcv,
4035                     &rack->rc_inp->inp_socket->so_snd,
4036                     BBR_LOG_RTT_SHRINKS, 0,
4037                     0, &log, false, &rack->r_ctl.act_rcv_time);
4038         }
4039 }
4040
4041 static void
4042 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4043 {
4044         uint64_t bwdp;
4045
4046         bwdp = rack_get_bw(rack);
4047         bwdp *= (uint64_t)rtt;
4048         bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4049         rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4050         if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4051                 /*
4052                  * A window protocol must be able to have 4 packets
4053                  * outstanding as the floor in order to function
4054                  * (especially considering delayed ack :D).
4055                  */
4056                 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4057         }
4058 }
4059
4060 static void
4061 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4062 {
4063         /**
4064          * ProbeRTT is a bit different in rack_pacing than in
4065          * BBR. It is like BBR in that it uses the lowering of
4066          * the RTT as a signal that we saw something new and
4067          * counts from there for how long between. But it is
4068          * different in that its quite simple. It does not
4069          * play with the cwnd and wait until we get down
4070          * to N segments outstanding and hold that for
4071          * 200ms. Instead it just sets the pacing reduction
4072          * rate to a set percentage (70 by default) and hold
4073          * that for a number of recent GP Srtt's.
4074          */
4075         uint32_t segsiz;
4076
4077         if (rack->rc_gp_dyn_mul == 0)
4078                 return;
4079
4080         if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4081                 /* We are idle */
4082                 return;
4083         }
4084         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4085             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4086                 /*
4087                  * Stop the goodput now, the idea here is
4088                  * that future measurements with in_probe_rtt
4089                  * won't register if they are not greater so
4090                  * we want to get what info (if any) is available
4091                  * now.
4092                  */
4093                 rack_do_goodput_measurement(rack->rc_tp, rack,
4094                                             rack->rc_tp->snd_una, __LINE__,
4095                                             RACK_QUALITY_PROBERTT);
4096         }
4097         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4098         rack->r_ctl.rc_time_probertt_entered = us_cts;
4099         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4100                      rack->r_ctl.rc_pace_min_segs);
4101         rack->in_probe_rtt = 1;
4102         rack->measure_saw_probe_rtt = 1;
4103         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4104         rack->r_ctl.rc_time_probertt_starts = 0;
4105         rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4106         if (rack_probertt_use_min_rtt_entry)
4107                 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4108         else
4109                 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4110         rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4111                              __LINE__, RACK_RTTS_ENTERPROBE);
4112 }
4113
4114 static void
4115 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4116 {
4117         struct rack_sendmap *rsm;
4118         uint32_t segsiz;
4119
4120         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4121                      rack->r_ctl.rc_pace_min_segs);
4122         rack->in_probe_rtt = 0;
4123         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4124             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4125                 /*
4126                  * Stop the goodput now, the idea here is
4127                  * that future measurements with in_probe_rtt
4128                  * won't register if they are not greater so
4129                  * we want to get what info (if any) is available
4130                  * now.
4131                  */
4132                 rack_do_goodput_measurement(rack->rc_tp, rack,
4133                                             rack->rc_tp->snd_una, __LINE__,
4134                                             RACK_QUALITY_PROBERTT);
4135         } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4136                 /*
4137                  * We don't have enough data to make a measurement.
4138                  * So lets just stop and start here after exiting
4139                  * probe-rtt. We probably are not interested in
4140                  * the results anyway.
4141                  */
4142                 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4143         }
4144         /*
4145          * Measurements through the current snd_max are going
4146          * to be limited by the slower pacing rate.
4147          *
4148          * We need to mark these as app-limited so we
4149          * don't collapse the b/w.
4150          */
4151         rsm = tqhash_max(rack->r_ctl.tqh);
4152         if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4153                 if (rack->r_ctl.rc_app_limited_cnt == 0)
4154                         rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4155                 else {
4156                         /*
4157                          * Go out to the end app limited and mark
4158                          * this new one as next and move the end_appl up
4159                          * to this guy.
4160                          */
4161                         if (rack->r_ctl.rc_end_appl)
4162                                 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4163                         rack->r_ctl.rc_end_appl = rsm;
4164                 }
4165                 rsm->r_flags |= RACK_APP_LIMITED;
4166                 rack->r_ctl.rc_app_limited_cnt++;
4167         }
4168         /*
4169          * Now, we need to examine our pacing rate multipliers.
4170          * If its under 100%, we need to kick it back up to
4171          * 100%. We also don't let it be over our "max" above
4172          * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4173          * Note setting clamp_atexit_prtt to 0 has the effect
4174          * of setting CA/SS to 100% always at exit (which is
4175          * the default behavior).
4176          */
4177         if (rack_probertt_clear_is) {
4178                 rack->rc_gp_incr = 0;
4179                 rack->rc_gp_bwred = 0;
4180                 rack->rc_gp_timely_inc_cnt = 0;
4181                 rack->rc_gp_timely_dec_cnt = 0;
4182         }
4183         /* Do we do any clamping at exit? */
4184         if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4185                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4186                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4187         }
4188         if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4189                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4190                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4191         }
4192         /*
4193          * Lets set rtt_diff to 0, so that we will get a "boost"
4194          * after exiting.
4195          */
4196         rack->r_ctl.rc_rtt_diff = 0;
4197
4198         /* Clear all flags so we start fresh */
4199         rack->rc_tp->t_bytes_acked = 0;
4200         rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4201         /*
4202          * If configured to, set the cwnd and ssthresh to
4203          * our targets.
4204          */
4205         if (rack_probe_rtt_sets_cwnd) {
4206                 uint64_t ebdp;
4207                 uint32_t setto;
4208
4209                 /* Set ssthresh so we get into CA once we hit our target */
4210                 if (rack_probertt_use_min_rtt_exit == 1) {
4211                         /* Set to min rtt */
4212                         rack_set_prtt_target(rack, segsiz,
4213                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4214                 } else if (rack_probertt_use_min_rtt_exit == 2) {
4215                         /* Set to current gp rtt */
4216                         rack_set_prtt_target(rack, segsiz,
4217                                              rack->r_ctl.rc_gp_srtt);
4218                 } else if (rack_probertt_use_min_rtt_exit == 3) {
4219                         /* Set to entry gp rtt */
4220                         rack_set_prtt_target(rack, segsiz,
4221                                              rack->r_ctl.rc_entry_gp_rtt);
4222                 } else {
4223                         uint64_t sum;
4224                         uint32_t setval;
4225
4226                         sum = rack->r_ctl.rc_entry_gp_rtt;
4227                         sum *= 10;
4228                         sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4229                         if (sum >= 20) {
4230                                 /*
4231                                  * A highly buffered path needs
4232                                  * cwnd space for timely to work.
4233                                  * Lets set things up as if
4234                                  * we are heading back here again.
4235                                  */
4236                                 setval = rack->r_ctl.rc_entry_gp_rtt;
4237                         } else if (sum >= 15) {
4238                                 /*
4239                                  * Lets take the smaller of the
4240                                  * two since we are just somewhat
4241                                  * buffered.
4242                                  */
4243                                 setval = rack->r_ctl.rc_gp_srtt;
4244                                 if (setval > rack->r_ctl.rc_entry_gp_rtt)
4245                                         setval = rack->r_ctl.rc_entry_gp_rtt;
4246                         } else {
4247                                 /*
4248                                  * Here we are not highly buffered
4249                                  * and should pick the min we can to
4250                                  * keep from causing loss.
4251                                  */
4252                                 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4253                         }
4254                         rack_set_prtt_target(rack, segsiz,
4255                                              setval);
4256                 }
4257                 if (rack_probe_rtt_sets_cwnd > 1) {
4258                         /* There is a percentage here to boost */
4259                         ebdp = rack->r_ctl.rc_target_probertt_flight;
4260                         ebdp *= rack_probe_rtt_sets_cwnd;
4261                         ebdp /= 100;
4262                         setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4263                 } else
4264                         setto = rack->r_ctl.rc_target_probertt_flight;
4265                 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4266                 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4267                         /* Enforce a min */
4268                         rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4269                 }
4270                 /* If we set in the cwnd also set the ssthresh point so we are in CA */
4271                 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4272         }
4273         rack_log_rtt_shrinks(rack,  us_cts,
4274                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4275                              __LINE__, RACK_RTTS_EXITPROBE);
4276         /* Clear times last so log has all the info */
4277         rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4278         rack->r_ctl.rc_time_probertt_entered = us_cts;
4279         rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4280         rack->r_ctl.rc_time_of_last_probertt = us_cts;
4281 }
4282
4283 static void
4284 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4285 {
4286         /* Check in on probe-rtt */
4287         if (rack->rc_gp_filled == 0) {
4288                 /* We do not do p-rtt unless we have gp measurements */
4289                 return;
4290         }
4291         if (rack->in_probe_rtt) {
4292                 uint64_t no_overflow;
4293                 uint32_t endtime, must_stay;
4294
4295                 if (rack->r_ctl.rc_went_idle_time &&
4296                     ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4297                         /*
4298                          * We went idle during prtt, just exit now.
4299                          */
4300                         rack_exit_probertt(rack, us_cts);
4301                 } else if (rack_probe_rtt_safety_val &&
4302                     TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4303                     ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4304                         /*
4305                          * Probe RTT safety value triggered!
4306                          */
4307                         rack_log_rtt_shrinks(rack,  us_cts,
4308                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4309                                              __LINE__, RACK_RTTS_SAFETY);
4310                         rack_exit_probertt(rack, us_cts);
4311                 }
4312                 /* Calculate the max we will wait */
4313                 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4314                 if (rack->rc_highly_buffered)
4315                         endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4316                 /* Calculate the min we must wait */
4317                 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4318                 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4319                     TSTMP_LT(us_cts, endtime)) {
4320                         uint32_t calc;
4321                         /* Do we lower more? */
4322 no_exit:
4323                         if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4324                                 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4325                         else
4326                                 calc = 0;
4327                         calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4328                         if (calc) {
4329                                 /* Maybe */
4330                                 calc *= rack_per_of_gp_probertt_reduce;
4331                                 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4332                                 /* Limit it too */
4333                                 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4334                                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4335                         }
4336                         /* We must reach target or the time set */
4337                         return;
4338                 }
4339                 if (rack->r_ctl.rc_time_probertt_starts == 0) {
4340                         if ((TSTMP_LT(us_cts, must_stay) &&
4341                              rack->rc_highly_buffered) ||
4342                              (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4343                               rack->r_ctl.rc_target_probertt_flight)) {
4344                                 /* We are not past the must_stay time */
4345                                 goto no_exit;
4346                         }
4347                         rack_log_rtt_shrinks(rack,  us_cts,
4348                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4349                                              __LINE__, RACK_RTTS_REACHTARGET);
4350                         rack->r_ctl.rc_time_probertt_starts = us_cts;
4351                         if (rack->r_ctl.rc_time_probertt_starts == 0)
4352                                 rack->r_ctl.rc_time_probertt_starts = 1;
4353                         /* Restore back to our rate we want to pace at in prtt */
4354                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4355                 }
4356                 /*
4357                  * Setup our end time, some number of gp_srtts plus 200ms.
4358                  */
4359                 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4360                                (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4361                 if (rack_probertt_gpsrtt_cnt_div)
4362                         endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4363                 else
4364                         endtime = 0;
4365                 endtime += rack_min_probertt_hold;
4366                 endtime += rack->r_ctl.rc_time_probertt_starts;
4367                 if (TSTMP_GEQ(us_cts,  endtime)) {
4368                         /* yes, exit probertt */
4369                         rack_exit_probertt(rack, us_cts);
4370                 }
4371
4372         } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
4373                 /* Go into probertt, its been too long since we went lower */
4374                 rack_enter_probertt(rack, us_cts);
4375         }
4376 }
4377
4378 static void
4379 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4380                        uint32_t rtt, int32_t rtt_diff)
4381 {
4382         uint64_t cur_bw, up_bnd, low_bnd, subfr;
4383         uint32_t losses;
4384
4385         if ((rack->rc_gp_dyn_mul == 0) ||
4386             (rack->use_fixed_rate) ||
4387             (rack->in_probe_rtt) ||
4388             (rack->rc_always_pace == 0)) {
4389                 /* No dynamic GP multiplier in play */
4390                 return;
4391         }
4392         losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4393         cur_bw = rack_get_bw(rack);
4394         /* Calculate our up and down range */
4395         up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4396         up_bnd /= 100;
4397         up_bnd += rack->r_ctl.last_gp_comp_bw;
4398
4399         subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4400         subfr /= 100;
4401         low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4402         if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4403                 /*
4404                  * This is the case where our RTT is above
4405                  * the max target and we have been configured
4406                  * to just do timely no bonus up stuff in that case.
4407                  *
4408                  * There are two configurations, set to 1, and we
4409                  * just do timely if we are over our max. If its
4410                  * set above 1 then we slam the multipliers down
4411                  * to 100 and then decrement per timely.
4412                  */
4413                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4414                                 __LINE__, 3);
4415                 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4416                         rack_validate_multipliers_at_or_below_100(rack);
4417                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4418         } else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4419                 /*
4420                  * We are decreasing this is a bit complicated this
4421                  * means we are loosing ground. This could be
4422                  * because another flow entered and we are competing
4423                  * for b/w with it. This will push the RTT up which
4424                  * makes timely unusable unless we want to get shoved
4425                  * into a corner and just be backed off (the age
4426                  * old problem with delay based CC).
4427                  *
4428                  * On the other hand if it was a route change we
4429                  * would like to stay somewhat contained and not
4430                  * blow out the buffers.
4431                  */
4432                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4433                                 __LINE__, 3);
4434                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4435                 if (rack->rc_gp_bwred == 0) {
4436                         /* Go into reduction counting */
4437                         rack->rc_gp_bwred = 1;
4438                         rack->rc_gp_timely_dec_cnt = 0;
4439                 }
4440                 if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4441                         /*
4442                          * Push another time with a faster pacing
4443                          * to try to gain back (we include override to
4444                          * get a full raise factor).
4445                          */
4446                         if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4447                             (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4448                             (timely_says == 0) ||
4449                             (rack_down_raise_thresh == 0)) {
4450                                 /*
4451                                  * Do an override up in b/w if we were
4452                                  * below the threshold or if the threshold
4453                                  * is zero we always do the raise.
4454                                  */
4455                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4456                         } else {
4457                                 /* Log it stays the same */
4458                                 rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4459                                                 __LINE__, 11);
4460                         }
4461                         rack->rc_gp_timely_dec_cnt++;
4462                         /* We are not incrementing really no-count */
4463                         rack->rc_gp_incr = 0;
4464                         rack->rc_gp_timely_inc_cnt = 0;
4465                 } else {
4466                         /*
4467                          * Lets just use the RTT
4468                          * information and give up
4469                          * pushing.
4470                          */
4471                         goto use_timely;
4472                 }
4473         } else if ((timely_says != 2) &&
4474                     !losses &&
4475                     (last_bw_est > up_bnd)) {
4476                 /*
4477                  * We are increasing b/w lets keep going, updating
4478                  * our b/w and ignoring any timely input, unless
4479                  * of course we are at our max raise (if there is one).
4480                  */
4481
4482                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4483                                 __LINE__, 3);
4484                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4485                 if (rack->rc_gp_saw_ss &&
4486                     rack->r_ctl.rack_per_upper_bound_ss &&
4487                      (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4488                             /*
4489                              * In cases where we can't go higher
4490                              * we should just use timely.
4491                              */
4492                             goto use_timely;
4493                 }
4494                 if (rack->rc_gp_saw_ca &&
4495                     rack->r_ctl.rack_per_upper_bound_ca &&
4496                     (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4497                             /*
4498                              * In cases where we can't go higher
4499                              * we should just use timely.
4500                              */
4501                             goto use_timely;
4502                 }
4503                 rack->rc_gp_bwred = 0;
4504                 rack->rc_gp_timely_dec_cnt = 0;
4505                 /* You get a set number of pushes if timely is trying to reduce */
4506                 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4507                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4508                 } else {
4509                         /* Log it stays the same */
4510                         rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4511                             __LINE__, 12);
4512                 }
4513                 return;
4514         } else {
4515                 /*
4516                  * We are staying between the lower and upper range bounds
4517                  * so use timely to decide.
4518                  */
4519                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4520                                 __LINE__, 3);
4521 use_timely:
4522                 if (timely_says) {
4523                         rack->rc_gp_incr = 0;
4524                         rack->rc_gp_timely_inc_cnt = 0;
4525                         if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4526                             !losses &&
4527                             (last_bw_est < low_bnd)) {
4528                                 /* We are loosing ground */
4529                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4530                                 rack->rc_gp_timely_dec_cnt++;
4531                                 /* We are not incrementing really no-count */
4532                                 rack->rc_gp_incr = 0;
4533                                 rack->rc_gp_timely_inc_cnt = 0;
4534                         } else
4535                                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4536                 } else {
4537                         rack->rc_gp_bwred = 0;
4538                         rack->rc_gp_timely_dec_cnt = 0;
4539                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4540                 }
4541         }
4542 }
4543
4544 static int32_t
4545 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4546 {
4547         int32_t timely_says;
4548         uint64_t log_mult, log_rtt_a_diff;
4549
4550         log_rtt_a_diff = rtt;
4551         log_rtt_a_diff <<= 32;
4552         log_rtt_a_diff |= (uint32_t)rtt_diff;
4553         if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4554                     rack_gp_rtt_maxmul)) {
4555                 /* Reduce the b/w multiplier */
4556                 timely_says = 2;
4557                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4558                 log_mult <<= 32;
4559                 log_mult |= prev_rtt;
4560                 rack_log_timely(rack,  timely_says, log_mult,
4561                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4562                                 log_rtt_a_diff, __LINE__, 4);
4563         } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4564                            ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4565                             max(rack_gp_rtt_mindiv , 1)))) {
4566                 /* Increase the b/w multiplier */
4567                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4568                         ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4569                          max(rack_gp_rtt_mindiv , 1));
4570                 log_mult <<= 32;
4571                 log_mult |= prev_rtt;
4572                 timely_says = 0;
4573                 rack_log_timely(rack,  timely_says, log_mult ,
4574                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4575                                 log_rtt_a_diff, __LINE__, 5);
4576         } else {
4577                 /*
4578                  * Use a gradient to find it the timely gradient
4579                  * is:
4580                  * grad = rc_rtt_diff / min_rtt;
4581                  *
4582                  * anything below or equal to 0 will be
4583                  * a increase indication. Anything above
4584                  * zero is a decrease. Note we take care
4585                  * of the actual gradient calculation
4586                  * in the reduction (its not needed for
4587                  * increase).
4588                  */
4589                 log_mult = prev_rtt;
4590                 if (rtt_diff <= 0) {
4591                         /*
4592                          * Rttdiff is less than zero, increase the
4593                          * b/w multiplier (its 0 or negative)
4594                          */
4595                         timely_says = 0;
4596                         rack_log_timely(rack,  timely_says, log_mult,
4597                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4598                 } else {
4599                         /* Reduce the b/w multiplier */
4600                         timely_says = 1;
4601                         rack_log_timely(rack,  timely_says, log_mult,
4602                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4603                 }
4604         }
4605         return (timely_says);
4606 }
4607
4608 static __inline int
4609 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4610 {
4611         if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4612             SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4613                 /**
4614                  * This covers the case that the
4615                  * resent is completely inside
4616                  * the gp range or up to it.
4617                  *      |----------------|
4618                  *      |-----| <or>
4619                  *            |----|
4620                  *            <or>   |---|
4621                  */
4622                 return (1);
4623         } else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4624                    SEQ_GT(rsm->r_end, tp->gput_seq)){
4625                 /**
4626                  * This covers the case of
4627                  *      |--------------|
4628                  *  |-------->|
4629                  */
4630                 return (1);
4631         } else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4632                    SEQ_LT(rsm->r_start, tp->gput_ack) &&
4633                    SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4634
4635                 /**
4636                  * This covers the case of
4637                  *      |--------------|
4638                  *              |-------->|
4639                  */
4640                 return (1);
4641         }
4642         return (0);
4643 }
4644
4645 static __inline void
4646 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4647 {
4648
4649         if ((tp->t_flags & TF_GPUTINPROG) == 0)
4650                 return;
4651         /*
4652          * We have a Goodput measurement in progress. Mark
4653          * the send if its within the window. If its not
4654          * in the window make sure it does not have the mark.
4655          */
4656         if (rack_in_gp_window(tp, rsm))
4657                 rsm->r_flags |= RACK_IN_GP_WIN;
4658         else
4659                 rsm->r_flags &= ~RACK_IN_GP_WIN;
4660 }
4661
4662 static __inline void
4663 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4664 {
4665         /* A GP measurement is ending, clear all marks on the send map*/
4666         struct rack_sendmap *rsm = NULL;
4667
4668         rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4669         if (rsm == NULL) {
4670                 rsm = tqhash_min(rack->r_ctl.tqh);
4671         }
4672         /* Nothing left? */
4673         while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4674                 rsm->r_flags &= ~RACK_IN_GP_WIN;
4675                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4676         }
4677 }
4678
4679
4680 static __inline void
4681 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4682 {
4683         struct rack_sendmap *rsm = NULL;
4684
4685         if (tp->snd_una == tp->snd_max) {
4686                 /* Nothing outstanding yet, nothing to do here */
4687                 return;
4688         }
4689         if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4690                 /*
4691                  * We are measuring ahead of some outstanding
4692                  * data. We need to walk through up until we get
4693                  * to gp_seq marking so that no rsm is set incorrectly
4694                  * with RACK_IN_GP_WIN.
4695                  */
4696                 rsm = tqhash_min(rack->r_ctl.tqh);
4697                 while (rsm != NULL) {
4698                         rack_mark_in_gp_win(tp, rsm);
4699                         if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4700                                 break;
4701                         rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4702                 }
4703         }
4704         if (rsm == NULL) {
4705                 /*
4706                  * Need to find the GP seq, if rsm is
4707                  * set we stopped as we hit it.
4708                  */
4709                 rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4710                 if (rsm == NULL)
4711                         return;
4712                 rack_mark_in_gp_win(tp, rsm);
4713         }
4714         /*
4715          * Now we may need to mark already sent rsm, ahead of
4716          * gput_seq in the window since they may have been sent
4717          * *before* we started our measurment. The rsm, if non-null
4718          * has been marked (note if rsm would have been NULL we would have
4719          * returned in the previous block). So we go to the next, and continue
4720          * until we run out of entries or we exceed the gp_ack value.
4721          */
4722         rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4723         while (rsm) {
4724                 rack_mark_in_gp_win(tp, rsm);
4725                 if (SEQ_GT(rsm->r_end, tp->gput_ack))
4726                         break;
4727                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4728         }
4729 }
4730
4731 static void
4732 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4733                             tcp_seq th_ack, int line, uint8_t quality)
4734 {
4735         uint64_t tim, bytes_ps, stim, utim;
4736         uint32_t segsiz, bytes, reqbytes, us_cts;
4737         int32_t gput, new_rtt_diff, timely_says;
4738         uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4739         int did_add = 0;
4740
4741         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4742         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4743         if (TSTMP_GEQ(us_cts, tp->gput_ts))
4744                 tim = us_cts - tp->gput_ts;
4745         else
4746                 tim = 0;
4747         if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4748                 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4749         else
4750                 stim = 0;
4751         /*
4752          * Use the larger of the send time or ack time. This prevents us
4753          * from being influenced by ack artifacts to come up with too
4754          * high of measurement. Note that since we are spanning over many more
4755          * bytes in most of our measurements hopefully that is less likely to
4756          * occur.
4757          */
4758         if (tim > stim)
4759                 utim = max(tim, 1);
4760         else
4761                 utim = max(stim, 1);
4762         reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4763         rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
4764         if ((tim == 0) && (stim == 0)) {
4765                 /*
4766                  * Invalid measurement time, maybe
4767                  * all on one ack/one send?
4768                  */
4769                 bytes = 0;
4770                 bytes_ps = 0;
4771                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4772                                            0, 0, 0, 10, __LINE__, NULL, quality);
4773                 goto skip_measurement;
4774         }
4775         if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4776                 /* We never made a us_rtt measurement? */
4777                 bytes = 0;
4778                 bytes_ps = 0;
4779                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4780                                            0, 0, 0, 10, __LINE__, NULL, quality);
4781                 goto skip_measurement;
4782         }
4783         /*
4784          * Calculate the maximum possible b/w this connection
4785          * could have. We base our calculation on the lowest
4786          * rtt we have seen during the measurement and the
4787          * largest rwnd the client has given us in that time. This
4788          * forms a BDP that is the maximum that we could ever
4789          * get to the client. Anything larger is not valid.
4790          *
4791          * I originally had code here that rejected measurements
4792          * where the time was less than 1/2 the latest us_rtt.
4793          * But after thinking on that I realized its wrong since
4794          * say you had a 150Mbps or even 1Gbps link, and you
4795          * were a long way away.. example I am in Europe (100ms rtt)
4796          * talking to my 1Gbps link in S.C. Now measuring say 150,000
4797          * bytes my time would be 1.2ms, and yet my rtt would say
4798          * the measurement was invalid the time was < 50ms. The
4799          * same thing is true for 150Mb (8ms of time).
4800          *
4801          * A better way I realized is to look at what the maximum
4802          * the connection could possibly do. This is gated on
4803          * the lowest RTT we have seen and the highest rwnd.
4804          * We should in theory never exceed that, if we are
4805          * then something on the path is storing up packets
4806          * and then feeding them all at once to our endpoint
4807          * messing up our measurement.
4808          */
4809         rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4810         rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4811         rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4812         if (SEQ_LT(th_ack, tp->gput_seq)) {
4813                 /* No measurement can be made */
4814                 bytes = 0;
4815                 bytes_ps = 0;
4816                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4817                                            0, 0, 0, 10, __LINE__, NULL, quality);
4818                 goto skip_measurement;
4819         } else
4820                 bytes = (th_ack - tp->gput_seq);
4821         bytes_ps = (uint64_t)bytes;
4822         /*
4823          * Don't measure a b/w for pacing unless we have gotten at least
4824          * an initial windows worth of data in this measurement interval.
4825          *
4826          * Small numbers of bytes get badly influenced by delayed ack and
4827          * other artifacts. Note we take the initial window or our
4828          * defined minimum GP (defaulting to 10 which hopefully is the
4829          * IW).
4830          */
4831         if (rack->rc_gp_filled == 0) {
4832                 /*
4833                  * The initial estimate is special. We
4834                  * have blasted out an IW worth of packets
4835                  * without a real valid ack ts results. We
4836                  * then setup the app_limited_needs_set flag,
4837                  * this should get the first ack in (probably 2
4838                  * MSS worth) to be recorded as the timestamp.
4839                  * We thus allow a smaller number of bytes i.e.
4840                  * IW - 2MSS.
4841                  */
4842                 reqbytes -= (2 * segsiz);
4843                 /* Also lets fill previous for our first measurement to be neutral */
4844                 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4845         }
4846         if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4847                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4848                                            rack->r_ctl.rc_app_limited_cnt,
4849                                            0, 0, 10, __LINE__, NULL, quality);
4850                 goto skip_measurement;
4851         }
4852         /*
4853          * We now need to calculate the Timely like status so
4854          * we can update (possibly) the b/w multipliers.
4855          */
4856         new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4857         if (rack->rc_gp_filled == 0) {
4858                 /* No previous reading */
4859                 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4860         } else {
4861                 if (rack->measure_saw_probe_rtt == 0) {
4862                         /*
4863                          * We don't want a probertt to be counted
4864                          * since it will be negative incorrectly. We
4865                          * expect to be reducing the RTT when we
4866                          * pace at a slower rate.
4867                          */
4868                         rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4869                         rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4870                 }
4871         }
4872         timely_says = rack_make_timely_judgement(rack,
4873             rack->r_ctl.rc_gp_srtt,
4874             rack->r_ctl.rc_rtt_diff,
4875             rack->r_ctl.rc_prev_gp_srtt
4876         );
4877         bytes_ps *= HPTS_USEC_IN_SEC;
4878         bytes_ps /= utim;
4879         if (bytes_ps > rack->r_ctl.last_max_bw) {
4880                 /*
4881                  * Something is on path playing
4882                  * since this b/w is not possible based
4883                  * on our BDP (highest rwnd and lowest rtt
4884                  * we saw in the measurement window).
4885                  *
4886                  * Another option here would be to
4887                  * instead skip the measurement.
4888                  */
4889                 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4890                                            bytes_ps, rack->r_ctl.last_max_bw, 0,
4891                                            11, __LINE__, NULL, quality);
4892                 bytes_ps = rack->r_ctl.last_max_bw;
4893         }
4894         /* We store gp for b/w in bytes per second */
4895         if (rack->rc_gp_filled == 0) {
4896                 /* Initial measurement */
4897                 if (bytes_ps) {
4898                         rack->r_ctl.gp_bw = bytes_ps;
4899                         rack->rc_gp_filled = 1;
4900                         rack->r_ctl.num_measurements = 1;
4901                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4902                 } else {
4903                         rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4904                                                    rack->r_ctl.rc_app_limited_cnt,
4905                                                    0, 0, 10, __LINE__, NULL, quality);
4906                 }
4907                 if (tcp_in_hpts(rack->rc_inp) &&
4908                     (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4909                         /*
4910                          * Ok we can't trust the pacer in this case
4911                          * where we transition from un-paced to paced.
4912                          * Or for that matter when the burst mitigation
4913                          * was making a wild guess and got it wrong.
4914                          * Stop the pacer and clear up all the aggregate
4915                          * delays etc.
4916                          */
4917                         tcp_hpts_remove(rack->rc_inp);
4918                         rack->r_ctl.rc_hpts_flags = 0;
4919                         rack->r_ctl.rc_last_output_to = 0;
4920                 }
4921                 did_add = 2;
4922         } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4923                 /* Still a small number run an average */
4924                 rack->r_ctl.gp_bw += bytes_ps;
4925                 addpart = rack->r_ctl.num_measurements;
4926                 rack->r_ctl.num_measurements++;
4927                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4928                         /* We have collected enough to move forward */
4929                         rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4930                 }
4931                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
4932                 did_add = 3;
4933         } else {
4934                 /*
4935                  * We want to take 1/wma of the goodput and add in to 7/8th
4936                  * of the old value weighted by the srtt. So if your measurement
4937                  * period is say 2 SRTT's long you would get 1/4 as the
4938                  * value, if it was like 1/2 SRTT then you would get 1/16th.
4939                  *
4940                  * But we must be careful not to take too much i.e. if the
4941                  * srtt is say 20ms and the measurement is taken over
4942                  * 400ms our weight would be 400/20 i.e. 20. On the
4943                  * other hand if we get a measurement over 1ms with a
4944                  * 10ms rtt we only want to take a much smaller portion.
4945                  */
4946                 if (rack->r_ctl.num_measurements < 0xff) {
4947                         rack->r_ctl.num_measurements++;
4948                 }
4949                 srtt = (uint64_t)tp->t_srtt;
4950                 if (srtt == 0) {
4951                         /*
4952                          * Strange why did t_srtt go back to zero?
4953                          */
4954                         if (rack->r_ctl.rc_rack_min_rtt)
4955                                 srtt = rack->r_ctl.rc_rack_min_rtt;
4956                         else
4957                                 srtt = HPTS_USEC_IN_MSEC;
4958                 }
4959                 /*
4960                  * XXXrrs: Note for reviewers, in playing with
4961                  * dynamic pacing I discovered this GP calculation
4962                  * as done originally leads to some undesired results.
4963                  * Basically you can get longer measurements contributing
4964                  * too much to the WMA. Thus I changed it if you are doing
4965                  * dynamic adjustments to only do the aportioned adjustment
4966                  * if we have a very small (time wise) measurement. Longer
4967                  * measurements just get there weight (defaulting to 1/8)
4968                  * add to the WMA. We may want to think about changing
4969                  * this to always do that for both sides i.e. dynamic
4970                  * and non-dynamic... but considering lots of folks
4971                  * were playing with this I did not want to change the
4972                  * calculation per.se. without your thoughts.. Lawerence?
4973                  * Peter??
4974                  */
4975                 if (rack->rc_gp_dyn_mul == 0) {
4976                         subpart = rack->r_ctl.gp_bw * utim;
4977                         subpart /= (srtt * 8);
4978                         if (subpart < (rack->r_ctl.gp_bw / 2)) {
4979                                 /*
4980                                  * The b/w update takes no more
4981                                  * away then 1/2 our running total
4982                                  * so factor it in.
4983                                  */
4984                                 addpart = bytes_ps * utim;
4985                                 addpart /= (srtt * 8);
4986                         } else {
4987                                 /*
4988                                  * Don't allow a single measurement
4989                                  * to account for more than 1/2 of the
4990                                  * WMA. This could happen on a retransmission
4991                                  * where utim becomes huge compared to
4992                                  * srtt (multiple retransmissions when using
4993                                  * the sending rate which factors in all the
4994                                  * transmissions from the first one).
4995                                  */
4996                                 subpart = rack->r_ctl.gp_bw / 2;
4997                                 addpart = bytes_ps / 2;
4998                         }
4999                         resid_bw = rack->r_ctl.gp_bw - subpart;
5000                         rack->r_ctl.gp_bw = resid_bw + addpart;
5001                         did_add = 1;
5002                 } else {
5003                         if ((utim / srtt) <= 1) {
5004                                 /*
5005                                  * The b/w update was over a small period
5006                                  * of time. The idea here is to prevent a small
5007                                  * measurement time period from counting
5008                                  * too much. So we scale it based on the
5009                                  * time so it attributes less than 1/rack_wma_divisor
5010                                  * of its measurement.
5011                                  */
5012                                 subpart = rack->r_ctl.gp_bw * utim;
5013                                 subpart /= (srtt * rack_wma_divisor);
5014                                 addpart = bytes_ps * utim;
5015                                 addpart /= (srtt * rack_wma_divisor);
5016                         } else {
5017                                 /*
5018                                  * The scaled measurement was long
5019                                  * enough so lets just add in the
5020                                  * portion of the measurement i.e. 1/rack_wma_divisor
5021                                  */
5022                                 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5023                                 addpart = bytes_ps / rack_wma_divisor;
5024                         }
5025                         if ((rack->measure_saw_probe_rtt == 0) ||
5026                             (bytes_ps > rack->r_ctl.gp_bw)) {
5027                                 /*
5028                                  * For probe-rtt we only add it in
5029                                  * if its larger, all others we just
5030                                  * add in.
5031                                  */
5032                                 did_add = 1;
5033                                 resid_bw = rack->r_ctl.gp_bw - subpart;
5034                                 rack->r_ctl.gp_bw = resid_bw + addpart;
5035                         }
5036                 }
5037                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
5038         }
5039         if ((rack->gp_ready == 0) &&
5040             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5041                 /* We have enough measurements now */
5042                 rack->gp_ready = 1;
5043                 if ((rack->rc_always_pace && (rack->use_fixed_rate == 0)) ||
5044                     rack->rack_hibeta)
5045                         rack_set_cc_pacing(rack);
5046                 if (rack->defer_options)
5047                         rack_apply_deferred_options(rack);
5048         }
5049         rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5050                                    rack_get_bw(rack), 22, did_add, NULL, quality);
5051         /* We do not update any multipliers if we are in or have seen a probe-rtt */
5052         if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
5053                 rack_update_multiplier(rack, timely_says, bytes_ps,
5054                                        rack->r_ctl.rc_gp_srtt,
5055                                        rack->r_ctl.rc_rtt_diff);
5056         rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5057                                    rack_get_bw(rack), 3, line, NULL, quality);
5058         rack_log_pacing_delay_calc(rack,
5059                                    bytes, /* flex2 */
5060                                    tim, /* flex1 */
5061                                    bytes_ps, /* bw_inuse */
5062                                    rack->r_ctl.gp_bw, /* delRate */
5063                                    rack_get_lt_bw(rack), /* rttProp */
5064                                    20, line, NULL, 0);
5065         /* reset the gp srtt and setup the new prev */
5066         rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5067         /* Record the lost count for the next measurement */
5068         rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5069 skip_measurement:
5070         /*
5071          * We restart our diffs based on the gpsrtt in the
5072          * measurement window.
5073          */
5074         rack->rc_gp_rtt_set = 0;
5075         rack->rc_gp_saw_rec = 0;
5076         rack->rc_gp_saw_ca = 0;
5077         rack->rc_gp_saw_ss = 0;
5078         rack->rc_dragged_bottom = 0;
5079
5080         if (quality == RACK_QUALITY_HIGH) {
5081                 /*
5082                  * Gput in the stats world is in kbps where bytes_ps is
5083                  * bytes per second so we do ((x * 8)/ 1000).
5084                  */
5085                 gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5086 #ifdef STATS
5087                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5088                                          gput);
5089                 /*
5090                  * XXXLAS: This is a temporary hack, and should be
5091                  * chained off VOI_TCP_GPUT when stats(9) grows an
5092                  * API to deal with chained VOIs.
5093                  */
5094                 if (tp->t_stats_gput_prev > 0)
5095                         stats_voi_update_abs_s32(tp->t_stats,
5096                                                  VOI_TCP_GPUT_ND,
5097                                                  ((gput - tp->t_stats_gput_prev) * 100) /
5098                                                  tp->t_stats_gput_prev);
5099 #endif
5100                 tp->t_stats_gput_prev = gput;
5101         }
5102         tp->t_flags &= ~TF_GPUTINPROG;
5103         /*
5104          * Now are we app limited now and there is space from where we
5105          * were to where we want to go?
5106          *
5107          * We don't do the other case i.e. non-applimited here since
5108          * the next send will trigger us picking up the missing data.
5109          */
5110         if (rack->r_ctl.rc_first_appl &&
5111             TCPS_HAVEESTABLISHED(tp->t_state) &&
5112             rack->r_ctl.rc_app_limited_cnt &&
5113             (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5114             ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5115              max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5116                 /*
5117                  * Yep there is enough outstanding to make a measurement here.
5118                  */
5119                 struct rack_sendmap *rsm;
5120
5121                 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5122                 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5123                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
5124                 rack->app_limited_needs_set = 0;
5125                 tp->gput_seq = th_ack;
5126                 if (rack->in_probe_rtt)
5127                         rack->measure_saw_probe_rtt = 1;
5128                 else if ((rack->measure_saw_probe_rtt) &&
5129                          (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5130                         rack->measure_saw_probe_rtt = 0;
5131                 if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5132                         /* There is a full window to gain info from */
5133                         tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5134                 } else {
5135                         /* We can only measure up to the applimited point */
5136                         tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5137                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5138                                 /*
5139                                  * We don't have enough to make a measurement.
5140                                  */
5141                                 tp->t_flags &= ~TF_GPUTINPROG;
5142                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5143                                                            0, 0, 0, 6, __LINE__, NULL, quality);
5144                                 return;
5145                         }
5146                 }
5147                 if (tp->t_state >= TCPS_FIN_WAIT_1) {
5148                         /*
5149                          * We will get no more data into the SB
5150                          * this means we need to have the data available
5151                          * before we start a measurement.
5152                          */
5153                         if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5154                                 /* Nope not enough data. */
5155                                 return;
5156                         }
5157                 }
5158                 tp->t_flags |= TF_GPUTINPROG;
5159                 /*
5160                  * Now we need to find the timestamp of the send at tp->gput_seq
5161                  * for the send based measurement.
5162                  */
5163                 rack->r_ctl.rc_gp_cumack_ts = 0;
5164                 rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5165                 if (rsm) {
5166                         /* Ok send-based limit is set */
5167                         if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5168                                 /*
5169                                  * Move back to include the earlier part
5170                                  * so our ack time lines up right (this may
5171                                  * make an overlapping measurement but thats
5172                                  * ok).
5173                                  */
5174                                 tp->gput_seq = rsm->r_start;
5175                         }
5176                         if (rsm->r_flags & RACK_ACKED) {
5177                                 struct rack_sendmap *nrsm;
5178
5179                                 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5180                                 tp->gput_seq = rsm->r_end;
5181                                 nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5182                                 if (nrsm)
5183                                         rsm = nrsm;
5184                                 else {
5185                                         rack->app_limited_needs_set = 1;
5186                                 }
5187                         } else
5188                                 rack->app_limited_needs_set = 1;
5189                         /* We always go from the first send */
5190                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5191                 } else {
5192                         /*
5193                          * If we don't find the rsm due to some
5194                          * send-limit set the current time, which
5195                          * basically disables the send-limit.
5196                          */
5197                         struct timeval tv;
5198
5199                         microuptime(&tv);
5200                         rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5201                 }
5202                 rack_tend_gp_marks(tp, rack);
5203                 rack_log_pacing_delay_calc(rack,
5204                                            tp->gput_seq,
5205                                            tp->gput_ack,
5206                                            (uint64_t)rsm,
5207                                            tp->gput_ts,
5208                                            (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5209                                            9,
5210                                            __LINE__, rsm, quality);
5211                 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5212         } else {
5213                 /*
5214                  * To make sure proper timestamp merging occurs, we need to clear
5215                  * all GP marks if we don't start a measurement.
5216                  */
5217                 rack_clear_gp_marks(tp, rack);
5218         }
5219 }
5220
5221 /*
5222  * CC wrapper hook functions
5223  */
5224 static void
5225 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5226     uint16_t type, int32_t recovery)
5227 {
5228         uint32_t prior_cwnd, acked;
5229         struct tcp_log_buffer *lgb = NULL;
5230         uint8_t labc_to_use, quality;
5231
5232         INP_WLOCK_ASSERT(tptoinpcb(tp));
5233         tp->t_ccv.nsegs = nsegs;
5234         acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5235         if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5236                 uint32_t max;
5237
5238                 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5239                 if (tp->t_ccv.bytes_this_ack > max) {
5240                         tp->t_ccv.bytes_this_ack = max;
5241                 }
5242         }
5243 #ifdef STATS
5244         stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5245             ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5246 #endif
5247         if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5248                 /* We will ack all, time
5249                  * to end any lt_bw_up we
5250                  * have running until something
5251                  * new is sent.
5252                  */
5253                 struct timeval tv;
5254
5255                 rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5256                 rack->r_ctl.lt_seq = tp->snd_max;
5257                 (void)tcp_get_usecs(&tv);
5258                 rack->r_ctl.lt_bw_time += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
5259                 rack->lt_bw_up = 0;
5260         }
5261         quality = RACK_QUALITY_NONE;
5262         if ((tp->t_flags & TF_GPUTINPROG) &&
5263             rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5264                 /* Measure the Goodput */
5265                 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5266         }
5267         /* Which way our we limited, if not cwnd limited no advance in CA */
5268         if (tp->snd_cwnd <= tp->snd_wnd)
5269                 tp->t_ccv.flags |= CCF_CWND_LIMITED;
5270         else
5271                 tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5272         if (tp->snd_cwnd > tp->snd_ssthresh) {
5273                 tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5274                          nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5275                 /* For the setting of a window past use the actual scwnd we are using */
5276                 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5277                         tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5278                         tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5279                 }
5280         } else {
5281                 tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5282                 tp->t_bytes_acked = 0;
5283         }
5284         prior_cwnd = tp->snd_cwnd;
5285         if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5286             (rack_client_low_buf && rack->client_bufferlvl &&
5287             (rack->client_bufferlvl < rack_client_low_buf)))
5288                 labc_to_use = rack->rc_labc;
5289         else
5290                 labc_to_use = rack_max_abc_post_recovery;
5291         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5292                 union tcp_log_stackspecific log;
5293                 struct timeval tv;
5294
5295                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5296                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5297                 log.u_bbr.flex1 = th_ack;
5298                 log.u_bbr.flex2 = tp->t_ccv.flags;
5299                 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5300                 log.u_bbr.flex4 = tp->t_ccv.nsegs;
5301                 log.u_bbr.flex5 = labc_to_use;
5302                 log.u_bbr.flex6 = prior_cwnd;
5303                 log.u_bbr.flex7 = V_tcp_do_newsack;
5304                 log.u_bbr.flex8 = 1;
5305                 lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5306                                      0, &log, false, NULL, __func__, __LINE__,&tv);
5307         }
5308         if (CC_ALGO(tp)->ack_received != NULL) {
5309                 /* XXXLAS: Find a way to live without this */
5310                 tp->t_ccv.curack = th_ack;
5311                 tp->t_ccv.labc = labc_to_use;
5312                 tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5313                 CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5314         }
5315         if (lgb) {
5316                 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5317         }
5318         if (rack->r_must_retran) {
5319                 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5320                         /*
5321                          * We now are beyond the rxt point so lets disable
5322                          * the flag.
5323                          */
5324                         rack->r_ctl.rc_out_at_rto = 0;
5325                         rack->r_must_retran = 0;
5326                 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5327                         /*
5328                          * Only decrement the rc_out_at_rto if the cwnd advances
5329                          * at least a whole segment. Otherwise next time the peer
5330                          * acks, we won't be able to send this generaly happens
5331                          * when we are in Congestion Avoidance.
5332                          */
5333                         if (acked <= rack->r_ctl.rc_out_at_rto){
5334                                 rack->r_ctl.rc_out_at_rto -= acked;
5335                         } else {
5336                                 rack->r_ctl.rc_out_at_rto = 0;
5337                         }
5338                 }
5339         }
5340 #ifdef STATS
5341         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5342 #endif
5343         if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5344                 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5345         }
5346 }
5347
5348 static void
5349 tcp_rack_partialack(struct tcpcb *tp)
5350 {
5351         struct tcp_rack *rack;
5352
5353         rack = (struct tcp_rack *)tp->t_fb_ptr;
5354         INP_WLOCK_ASSERT(tptoinpcb(tp));
5355         /*
5356          * If we are doing PRR and have enough
5357          * room to send <or> we are pacing and prr
5358          * is disabled we will want to see if we
5359          * can send data (by setting r_wanted_output to
5360          * true).
5361          */
5362         if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5363             rack->rack_no_prr)
5364                 rack->r_wanted_output = 1;
5365 }
5366
5367 static inline void
5368 rack_set_most_aggr(struct tcp_rack *rack)
5369 {
5370         rack->r_fill_less_agg = 0;
5371         /* Once the cwnd as been clamped we don't do fill_cw */
5372         if (rack->r_cwnd_was_clamped == 0)
5373                 rack->rc_pace_to_cwnd = 1;
5374         rack->r_pacing_discount = 0;
5375 }
5376
5377 static inline void
5378 rack_limit_fillcw(struct tcp_rack *rack)
5379 {
5380         rack->r_fill_less_agg = 1;
5381         /* Once the cwnd as been clamped we don't do fill_cw */
5382         if (rack->r_cwnd_was_clamped == 0)
5383                 rack->rc_pace_to_cwnd = 1;
5384         rack->r_pacing_discount = 0;
5385 }
5386
5387 static inline void
5388 rack_disable_fillcw(struct tcp_rack *rack)
5389 {
5390         rack->r_fill_less_agg = 1;
5391         rack->rc_pace_to_cwnd = 0;
5392         rack->r_pacing_discount = 0;
5393 }
5394
5395 static void
5396 rack_client_buffer_level_set(struct tcp_rack *rack)
5397 {
5398         /*
5399          * Only if DGP is on do we do anything that
5400          * changes stack behavior. If DGP is off all
5401          * we will do is issue a BB log (if BB logging is
5402          * on) and return.
5403          */
5404         if (rack->dgp_on == 0) {
5405                 rack_log_pacing_delay_calc(rack, 0, rack->client_bufferlvl,
5406                                            0, 0, 0, 30, __LINE__, NULL, 0);
5407                 return;
5408         }
5409         if (IN_RECOVERY(rack->rc_tp->t_flags) && rack->r_ctl.full_dgp_in_rec) {
5410                 goto set_most_agg;
5411         }
5412         /*
5413          * We are in DGP so what setting should we
5414          * apply based on where the client is?
5415          */
5416         switch(rack->r_ctl.rc_dgp_bl_agg) {
5417         default:
5418         case DGP_LEVEL0:
5419 set_most_agg:
5420                 rack_set_most_aggr(rack);
5421                 break;
5422         case DGP_LEVEL1:
5423                 if (rack->client_bufferlvl == 4)
5424                         rack_limit_fillcw(rack);
5425                 else if (rack->client_bufferlvl == 5)
5426                         rack_disable_fillcw(rack);
5427                 else
5428                         rack_set_most_aggr(rack);
5429                 break;
5430         case DGP_LEVEL2:
5431                 if (rack->client_bufferlvl == 3)
5432                         rack_limit_fillcw(rack);
5433                 else if (rack->client_bufferlvl == 4)
5434                         rack_disable_fillcw(rack);
5435                 else if (rack->client_bufferlvl == 5) {
5436                         rack_disable_fillcw(rack);
5437                         rack->r_pacing_discount = 1;
5438                         rack->r_ctl.pacing_discount_amm = 1;
5439                 } else
5440                         rack_set_most_aggr(rack);
5441                 break;
5442         case DGP_LEVEL3:
5443                 if (rack->client_bufferlvl == 2)
5444                         rack_limit_fillcw(rack);
5445                 else if (rack->client_bufferlvl == 3)
5446                         rack_disable_fillcw(rack);
5447                 else if (rack->client_bufferlvl == 4) {
5448                         rack_disable_fillcw(rack);
5449                         rack->r_pacing_discount = 1;
5450                         rack->r_ctl.pacing_discount_amm = 1;
5451                 } else if (rack->client_bufferlvl == 5) {
5452                         rack_disable_fillcw(rack);
5453                         rack->r_pacing_discount = 1;
5454                         rack->r_ctl.pacing_discount_amm = 2;
5455                 } else
5456                         rack_set_most_aggr(rack);
5457                 break;
5458         }
5459         rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_dgp_bl_agg, rack->client_bufferlvl, 0,
5460                                    0, 0, 30, __LINE__, NULL, 0);
5461 }
5462
5463 static void
5464 do_rack_check_for_unclamp(struct tcpcb *tp, struct tcp_rack *rack)
5465 {
5466         /*
5467          * Can we unclamp. We unclamp if more than
5468          * N rounds have transpired with no loss.
5469          */
5470         uint64_t snds, rxts, rxt_per;
5471         uint32_t rnds;
5472
5473         rnds = rack->r_ctl.current_round - rack->r_ctl.last_rnd_rxt_clamped;
5474         if ((rack_unclamp_round_thresh > 0) &&
5475             (rnds >= rack_unclamp_round_thresh)) {
5476                 snds = tp->t_sndbytes - rack->r_ctl.last_sndbytes;
5477                 KASSERT ((snds > 0), ("rack:%p tp:%p snds:%ju is 0", rack, tp,
5478                     (uintmax_t)snds));
5479                 rxts = tp->t_snd_rxt_bytes - rack->r_ctl.last_snd_rxt_bytes;
5480                 rxt_per = rxts * 1000;
5481                 rxt_per /= snds;
5482                 if ((uint32_t)rxt_per <= rack_unclamp_rxt_thresh) {
5483                         /* Unclamp */
5484                         if (tcp_bblogging_on(rack->rc_tp)) {
5485                                 union tcp_log_stackspecific log;
5486                                 struct timeval tv;
5487
5488                                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5489                                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5490                                 log.u_bbr.flex3 = rnds;
5491                                 log.u_bbr.flex4 = rack_unclamp_round_thresh;
5492                                 log.u_bbr.flex5 = (uint32_t)rxt_per;
5493                                 log.u_bbr.flex8 = 6;
5494                                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_pace_max_segs;
5495                                 log.u_bbr.bbr_state = rack->rc_pace_to_cwnd;
5496                                 log.u_bbr.delivered = rack->r_ctl.num_of_clamps_applied;
5497                                 log.u_bbr.applimited = rack->r_ctl.max_clamps;
5498                                 log.u_bbr.epoch = rack->r_ctl.clamp_options;
5499                                 log.u_bbr.cur_del_rate = rxts;
5500                                 log.u_bbr.bw_inuse = rack_get_lt_bw(rack);
5501                                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5502                                 log.u_bbr.lt_epoch = (uint32_t)((rack->r_ctl.gp_bw >> 32) & 0x00000000ffffffff);
5503                                 log.u_bbr.pkts_out = (uint32_t)(rack->r_ctl.gp_bw & 0x00000000ffffffff);
5504                                 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5505                                               0, &log, false, NULL, NULL, 0, &tv);
5506                         }
5507                         rack->r_ctl.num_of_clamps_applied = 0;
5508                         rack->r_cwnd_was_clamped = 0;
5509                         rack->excess_rxt_on = 1;
5510                         if (rack->r_ctl.clamp_options) {
5511                                 /*
5512                                  * We only allow fillcw to be toggled
5513                                  * if you are setting a max seg too.
5514                                  */
5515                                 if (rack->r_ctl.clamp_options & 0x1) {
5516                                         if ((rack->rc_pace_to_cwnd == 0) && (rack->dgp_on == 0)) {
5517                                                 /* turn on fill cw  for non-dgp*/
5518                                                 rack->rc_pace_to_cwnd = 0;
5519                                         } else if ((rack->dgp_on == 1) && (rack->rc_pace_to_cwnd == 1)) {
5520                                                 /* For DGP we want it off */
5521                                                 rack->rc_pace_to_cwnd = 1;
5522                                         }
5523                                 }
5524                         }
5525                         if (rack->dgp_on) {
5526                                 /* Reset all multipliers to 100.0 so just the measured bw */
5527                                 /* Crash any per boosts down to 100% */
5528                                 rack->r_ctl.rack_per_of_gp_rec = 100;
5529                                 rack->r_ctl.rack_per_of_gp_ss = 100;
5530                                 rack->r_ctl.rack_per_of_gp_ca = 100;
5531                                 /* Set in an upper bound for ss/ca % increase */
5532                                 rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
5533                                 rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
5534                         }
5535                 }
5536         }
5537 }
5538
5539 static void
5540 do_rack_excess_rxt(struct tcpcb *tp, struct tcp_rack *rack)
5541 {
5542         /*
5543          * Rack excess rxt accounting is turned on. If we
5544          * are above a threshold of rxt's in at least N
5545          * rounds, then back off the cwnd and ssthresh
5546          * to fit into the long-term b/w.
5547          */
5548         uint64_t snds, rxts, rxt_per, lt_bw, bdp;
5549         uint32_t rnds, new_cwnd, new_ssthresh, rtt, shared_cwnd_was_enabled = 0;
5550
5551         /* Is it shut off by 0 rounds? */
5552         if (rack_rxt_min_rnds == 0)
5553                 return;
5554         if ((rack->r_ctl.max_clamps > 0) &&
5555             (rack->r_ctl.num_of_clamps_applied >= rack->r_ctl.max_clamps)) {
5556                 /*
5557                  * The idea, if max_clamps is set, is that if clamping it
5558                  * N times did not work again, then there is no sense
5559                  * clamping it again. The link is just a lossy link and
5560                  * our clamps are doing no good. Turn it off so we don't come
5561                  * back here again.
5562                  */
5563                 rack->excess_rxt_on = 0;
5564                 rack->r_cwnd_was_clamped = 0;
5565                 rack->r_ctl.num_of_clamps_applied = 0;
5566                 return;
5567         }
5568         snds = tp->t_sndbytes - rack->r_ctl.last_sndbytes;
5569         rxts = tp->t_snd_rxt_bytes - rack->r_ctl.last_snd_rxt_bytes;
5570         rnds = rack->r_ctl.current_round - rack->r_ctl.last_rnd_rxt_clamped;
5571         /* Has enough rounds progressed for us to re-measure? */
5572         if ((rnds >= rack_rxt_min_rnds) &&
5573             (rack->r_ctl.rxt_threshold > 0)){
5574                 rxt_per = rxts * 1000;
5575                 rxt_per /= snds;
5576                 if (rxt_per >= rack->r_ctl.rxt_threshold) {
5577                         /*
5578                          * Action required:
5579                          *  We are above our excess retransmit level, lets
5580                          *  cut down the cwnd and ssthresh to match the long-term
5581                          *  b/w we are getting.
5582                          */
5583                         /* First disable scwnd if enabled */
5584 #ifdef NETFLIX_SHARED_CWND
5585                         rack->rack_enable_scwnd = 0;
5586                         if (rack->r_ctl.rc_scw) {
5587                                 uint32_t limit;
5588
5589                                 shared_cwnd_was_enabled = 1;
5590                                 if (rack->r_limit_scw)
5591                                         limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
5592                                 else
5593                                         limit = 0;
5594                                 tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
5595                                                           rack->r_ctl.rc_scw_index,
5596                                                           limit);
5597                                 rack->r_ctl.rc_scw = NULL;
5598                         }
5599
5600 #endif
5601                         /* Calculate what the cwnd and ssthresh should be */
5602                         tcp_trace_point(rack->rc_tp, TCP_TP_EXCESS_RXT);
5603                         lt_bw = rack_get_lt_bw(rack);
5604                         if (lt_bw == 0) {
5605                                 /*
5606                                  * No lt_bw, lets chop things to one MSS
5607                                  * and the ssthresh to the iwnd.
5608                                  */
5609 reset_to_iw:
5610                                 new_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5611                                 new_ssthresh = tcp_compute_initwnd(tcp_maxseg(tp));
5612                         } else {
5613                                 rtt = rack->rc_rack_rtt;
5614                                 if (rtt == 0) {
5615                                         /* If we have no rack_rtt drop to the IW situation */
5616                                         goto reset_to_iw;
5617                                 }
5618                                 bdp = lt_bw * (uint64_t)rtt;
5619                                 bdp /= HPTS_USEC_IN_SEC;
5620                                 new_cwnd = (uint32_t)bdp;
5621                                 new_ssthresh = new_cwnd - 1;
5622                                 if (new_cwnd < ctf_fixed_maxseg(tp)) {
5623                                         /* Rock bottom, goto IW settings  */
5624                                         goto reset_to_iw;
5625                                 }
5626                         }
5627                         rack->r_cwnd_was_clamped = 1;
5628                         rack->r_ctl.num_of_clamps_applied++;
5629                         /* Reset the counter fromn now */
5630                         tp->t_bytes_acked = 0;
5631                         /*
5632                          * Now what about options?
5633                          * We look at the bottom  8 bits:
5634                          * F = fill cw bit (toggle it if set)
5635                          * S = Segment bits
5636                          * M = set max segment bit
5637                          *
5638                          * SSSS SSMF
5639                          */
5640                         if (rack->r_ctl.clamp_options) {
5641                                 if (rack->r_ctl.clamp_options & 0x1) {
5642                                         if ((rack->rc_pace_to_cwnd == 0) && (rack->dgp_on == 0)) {
5643                                                 /* turn on fill cw  for non-dgp*/
5644                                                 rack->rc_pace_to_cwnd = 1;
5645                                         } else if ((rack->dgp_on == 1) && (rack->rc_pace_to_cwnd == 1)) {
5646                                                 /* For DGP we want it off */
5647                                                 rack->rc_pace_to_cwnd = 0;
5648                                         }
5649                                 }
5650                         }
5651                         if (rack->dgp_on) {
5652                                 /* Reset all multipliers to 100.0 so just the measured bw */
5653                                 /* Crash any per boosts down to 100% */
5654                                 rack->r_ctl.rack_per_of_gp_rec = 100;
5655                                 rack->r_ctl.rack_per_of_gp_ss = 100;
5656                                 rack->r_ctl.rack_per_of_gp_ca = 100;
5657                                 /* Set in an upper bound for ss/ca % increase */
5658                                 rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_clamp_ss_upper;
5659                                 rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_clamp_ca_upper;
5660                                 /* Now move to the lt_bw */
5661                                 rack->r_ctl.gp_bw = lt_bw;
5662                                 rack->rc_gp_filled = 1;
5663                                 rack->r_ctl.num_measurements = RACK_REQ_AVG;
5664                         }
5665                         if (tcp_bblogging_on(rack->rc_tp)) {
5666                                 union tcp_log_stackspecific log;
5667                                 struct timeval tv;
5668
5669                                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5670                                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5671                                 log.u_bbr.flex1 = new_cwnd;
5672                                 log.u_bbr.flex2 = new_ssthresh;
5673                                 log.u_bbr.flex3 = rnds;
5674                                 log.u_bbr.flex4 = rack_rxt_min_rnds;
5675                                 log.u_bbr.flex5 = rtt;
5676                                 log.u_bbr.flex6 = shared_cwnd_was_enabled;
5677                                 log.u_bbr.flex8 = 5;
5678                                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_pace_max_segs;
5679                                 log.u_bbr.bbr_state = rack->rc_pace_to_cwnd;
5680                                 log.u_bbr.delivered = rack->r_ctl.num_of_clamps_applied;
5681                                 log.u_bbr.applimited = rack->r_ctl.max_clamps;
5682                                 log.u_bbr.epoch = rack->r_ctl.clamp_options;
5683                                 log.u_bbr.cur_del_rate = rxts;
5684                                 log.u_bbr.delRate = snds;
5685                                 log.u_bbr.rttProp = rack->r_ctl.rxt_threshold;
5686                                 log.u_bbr.bw_inuse = lt_bw;
5687                                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5688                                 log.u_bbr.lt_epoch = (uint32_t)((rack->r_ctl.gp_bw >> 32) & 0x00000000ffffffff);
5689                                 log.u_bbr.pkts_out = (uint32_t)(rack->r_ctl.gp_bw & 0x00000000ffffffff);
5690                                 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5691                                                0, &log, false, NULL, NULL, 0, &tv);
5692                         }
5693                         /* Update our point where we did it */
5694                         if (rack->r_ctl.already_had_a_excess == 0) {
5695                                 rack->r_ctl.already_had_a_excess = 1;
5696                                 counter_u64_add(rack_rxt_clamps_cwnd_uniq, 1);
5697                         }
5698                         counter_u64_add(rack_rxt_clamps_cwnd, 1);
5699                         rack->r_ctl.last_sndbytes = tp->t_sndbytes;
5700                         rack->r_ctl.last_snd_rxt_bytes = tp->t_snd_rxt_bytes;
5701                         rack->r_ctl.last_rnd_rxt_clamped = rack->r_ctl.current_round;
5702                         if (new_cwnd < tp->snd_cwnd)
5703                                 tp->snd_cwnd = new_cwnd;
5704                         if (new_ssthresh < tp->snd_ssthresh)
5705                                 tp->snd_ssthresh = new_ssthresh;
5706                 }
5707         }
5708 }
5709
5710 static void
5711 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
5712 {
5713         struct tcp_rack *rack;
5714         uint32_t orig_cwnd;
5715
5716         orig_cwnd = tp->snd_cwnd;
5717         INP_WLOCK_ASSERT(tptoinpcb(tp));
5718         rack = (struct tcp_rack *)tp->t_fb_ptr;
5719         /* only alert CC if we alerted when we entered */
5720         if (CC_ALGO(tp)->post_recovery != NULL) {
5721                 tp->t_ccv.curack = th_ack;
5722                 CC_ALGO(tp)->post_recovery(&tp->t_ccv);
5723                 if (tp->snd_cwnd < tp->snd_ssthresh) {
5724                         /*
5725                          * Rack has burst control and pacing
5726                          * so lets not set this any lower than
5727                          * snd_ssthresh per RFC-6582 (option 2).
5728                          */
5729                         tp->snd_cwnd = tp->snd_ssthresh;
5730                 }
5731         }
5732         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5733                 union tcp_log_stackspecific log;
5734                 struct timeval tv;
5735
5736                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5737                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5738                 log.u_bbr.flex1 = th_ack;
5739                 log.u_bbr.flex2 = tp->t_ccv.flags;
5740                 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5741                 log.u_bbr.flex4 = tp->t_ccv.nsegs;
5742                 log.u_bbr.flex5 = V_tcp_abc_l_var;
5743                 log.u_bbr.flex6 = orig_cwnd;
5744                 log.u_bbr.flex7 = V_tcp_do_newsack;
5745                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
5746                 log.u_bbr.flex8 = 2;
5747                 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5748                                0, &log, false, NULL, __func__, __LINE__, &tv);
5749         }
5750         if ((rack->rack_no_prr == 0) &&
5751             (rack->no_prr_addback == 0) &&
5752             (rack->r_ctl.rc_prr_sndcnt > 0)) {
5753                 /*
5754                  * Suck the next prr cnt back into cwnd, but
5755                  * only do that if we are not application limited.
5756                  */
5757                 if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
5758                         /*
5759                          * We are allowed to add back to the cwnd the amount we did
5760                          * not get out if:
5761                          * a) no_prr_addback is off.
5762                          * b) we are not app limited
5763                          * c) we are doing prr
5764                          * <and>
5765                          * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
5766                          */
5767                         tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
5768                                             rack->r_ctl.rc_prr_sndcnt);
5769                 }
5770                 rack->r_ctl.rc_prr_sndcnt = 0;
5771                 rack_log_to_prr(rack, 1, 0, __LINE__);
5772         }
5773         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
5774         tp->snd_recover = tp->snd_una;
5775         if (rack->r_ctl.dsack_persist) {
5776                 rack->r_ctl.dsack_persist--;
5777                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
5778                         rack->r_ctl.num_dsack = 0;
5779                 }
5780                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
5781         }
5782         EXIT_RECOVERY(tp->t_flags);
5783         if (rack->r_ctl.full_dgp_in_rec)
5784                 rack_client_buffer_level_set(rack);
5785 }
5786
5787 static void
5788 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
5789 {
5790         struct tcp_rack *rack;
5791         uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
5792
5793         INP_WLOCK_ASSERT(tptoinpcb(tp));
5794 #ifdef STATS
5795         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
5796 #endif
5797         if (IN_RECOVERY(tp->t_flags) == 0) {
5798                 in_rec_at_entry = 0;
5799                 ssthresh_enter = tp->snd_ssthresh;
5800                 cwnd_enter = tp->snd_cwnd;
5801         } else
5802                 in_rec_at_entry = 1;
5803         rack = (struct tcp_rack *)tp->t_fb_ptr;
5804         switch (type) {
5805         case CC_NDUPACK:
5806                 tp->t_flags &= ~TF_WASFRECOVERY;
5807                 tp->t_flags &= ~TF_WASCRECOVERY;
5808                 if (!IN_FASTRECOVERY(tp->t_flags)) {
5809                         if (rack->dgp_on && rack->r_cwnd_was_clamped) {
5810                                 /* Reset the gains so that on exit we will be softer longer */
5811                                 rack->r_ctl.rack_per_of_gp_rec = 100;
5812                                 rack->r_ctl.rack_per_of_gp_ss = 98;
5813                                 rack->r_ctl.rack_per_of_gp_ca = 98;
5814                         }
5815                         rack->r_ctl.rc_prr_delivered = 0;
5816                         rack->r_ctl.rc_prr_out = 0;
5817                         rack->r_fast_output = 0;
5818                         if (rack->rack_no_prr == 0) {
5819                                 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5820                                 rack_log_to_prr(rack, 2, in_rec_at_entry, line);
5821                         }
5822                         rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
5823                         tp->snd_recover = tp->snd_max;
5824                         if (tp->t_flags2 & TF2_ECN_PERMIT)
5825                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
5826                 }
5827                 break;
5828         case CC_ECN:
5829                 if (!IN_CONGRECOVERY(tp->t_flags) ||
5830                     /*
5831                      * Allow ECN reaction on ACK to CWR, if
5832                      * that data segment was also CE marked.
5833                      */
5834                     SEQ_GEQ(ack, tp->snd_recover)) {
5835                         EXIT_CONGRECOVERY(tp->t_flags);
5836                         KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
5837                         rack->r_fast_output = 0;
5838                         tp->snd_recover = tp->snd_max + 1;
5839                         if (tp->t_flags2 & TF2_ECN_PERMIT)
5840                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
5841                 }
5842                 break;
5843         case CC_RTO:
5844                 tp->t_dupacks = 0;
5845                 tp->t_bytes_acked = 0;
5846                 rack->r_fast_output = 0;
5847                 EXIT_RECOVERY(tp->t_flags);
5848                 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
5849                     ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
5850                 orig_cwnd = tp->snd_cwnd;
5851                 tp->snd_cwnd = ctf_fixed_maxseg(tp);
5852                 rack_log_to_prr(rack, 16, orig_cwnd, line);
5853                 if (tp->t_flags2 & TF2_ECN_PERMIT)
5854                         tp->t_flags2 |= TF2_ECN_SND_CWR;
5855                 break;
5856         case CC_RTO_ERR:
5857                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
5858                 /* RTO was unnecessary, so reset everything. */
5859                 tp->snd_cwnd = tp->snd_cwnd_prev;
5860                 tp->snd_ssthresh = tp->snd_ssthresh_prev;
5861                 tp->snd_recover = tp->snd_recover_prev;
5862                 if (tp->t_flags & TF_WASFRECOVERY) {
5863                         ENTER_FASTRECOVERY(tp->t_flags);
5864                         tp->t_flags &= ~TF_WASFRECOVERY;
5865                 }
5866                 if (tp->t_flags & TF_WASCRECOVERY) {
5867                         ENTER_CONGRECOVERY(tp->t_flags);
5868                         tp->t_flags &= ~TF_WASCRECOVERY;
5869                 }
5870                 tp->snd_nxt = tp->snd_max;
5871                 tp->t_badrxtwin = 0;
5872                 break;
5873         }
5874         if ((CC_ALGO(tp)->cong_signal != NULL)  &&
5875             (type != CC_RTO)){
5876                 tp->t_ccv.curack = ack;
5877                 CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
5878         }
5879         if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5880                 rack_log_to_prr(rack, 15, cwnd_enter, line);
5881                 if (rack->r_ctl.full_dgp_in_rec)
5882                         rack_client_buffer_level_set(rack);
5883                 rack->r_ctl.dsack_byte_cnt = 0;
5884                 rack->r_ctl.retran_during_recovery = 0;
5885                 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5886                 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5887                 rack->r_ent_rec_ns = 1;
5888         }
5889 }
5890
5891 static inline void
5892 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5893 {
5894         uint32_t i_cwnd;
5895
5896         INP_WLOCK_ASSERT(tptoinpcb(tp));
5897
5898         if (CC_ALGO(tp)->after_idle != NULL)
5899                 CC_ALGO(tp)->after_idle(&tp->t_ccv);
5900
5901         if (tp->snd_cwnd == 1)
5902                 i_cwnd = tp->t_maxseg;          /* SYN(-ACK) lost */
5903         else
5904                 i_cwnd = rc_init_window(rack);
5905
5906         /*
5907          * Being idle is no different than the initial window. If the cc
5908          * clamps it down below the initial window raise it to the initial
5909          * window.
5910          */
5911         if (tp->snd_cwnd < i_cwnd) {
5912                 tp->snd_cwnd = i_cwnd;
5913         }
5914 }
5915
5916 /*
5917  * Indicate whether this ack should be delayed.  We can delay the ack if
5918  * following conditions are met:
5919  *      - There is no delayed ack timer in progress.
5920  *      - Our last ack wasn't a 0-sized window. We never want to delay
5921  *        the ack that opens up a 0-sized window.
5922  *      - LRO wasn't used for this segment. We make sure by checking that the
5923  *        segment size is not larger than the MSS.
5924  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
5925  *        connection.
5926  */
5927 #define DELAY_ACK(tp, tlen)                      \
5928         (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
5929         ((tp->t_flags & TF_DELACK) == 0) &&      \
5930         (tlen <= tp->t_maxseg) &&                \
5931         (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
5932
5933 static struct rack_sendmap *
5934 rack_find_lowest_rsm(struct tcp_rack *rack)
5935 {
5936         struct rack_sendmap *rsm;
5937
5938         /*
5939          * Walk the time-order transmitted list looking for an rsm that is
5940          * not acked. This will be the one that was sent the longest time
5941          * ago that is still outstanding.
5942          */
5943         TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
5944                 if (rsm->r_flags & RACK_ACKED) {
5945                         continue;
5946                 }
5947                 goto finish;
5948         }
5949 finish:
5950         return (rsm);
5951 }
5952
5953 static struct rack_sendmap *
5954 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5955 {
5956         struct rack_sendmap *prsm;
5957
5958         /*
5959          * Walk the sequence order list backward until we hit and arrive at
5960          * the highest seq not acked. In theory when this is called it
5961          * should be the last segment (which it was not).
5962          */
5963         prsm = rsm;
5964
5965         TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
5966                 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5967                         continue;
5968                 }
5969                 return (prsm);
5970         }
5971         return (NULL);
5972 }
5973
5974 static uint32_t
5975 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
5976 {
5977         int32_t lro;
5978         uint32_t thresh;
5979
5980         /*
5981          * lro is the flag we use to determine if we have seen reordering.
5982          * If it gets set we have seen reordering. The reorder logic either
5983          * works in one of two ways:
5984          *
5985          * If reorder-fade is configured, then we track the last time we saw
5986          * re-ordering occur. If we reach the point where enough time as
5987          * passed we no longer consider reordering has occuring.
5988          *
5989          * Or if reorder-face is 0, then once we see reordering we consider
5990          * the connection to alway be subject to reordering and just set lro
5991          * to 1.
5992          *
5993          * In the end if lro is non-zero we add the extra time for
5994          * reordering in.
5995          */
5996         if (srtt == 0)
5997                 srtt = 1;
5998         if (rack->r_ctl.rc_reorder_ts) {
5999                 if (rack->r_ctl.rc_reorder_fade) {
6000                         if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
6001                                 lro = cts - rack->r_ctl.rc_reorder_ts;
6002                                 if (lro == 0) {
6003                                         /*
6004                                          * No time as passed since the last
6005                                          * reorder, mark it as reordering.
6006                                          */
6007                                         lro = 1;
6008                                 }
6009                         } else {
6010                                 /* Negative time? */
6011                                 lro = 0;
6012                         }
6013                         if (lro > rack->r_ctl.rc_reorder_fade) {
6014                                 /* Turn off reordering seen too */
6015                                 rack->r_ctl.rc_reorder_ts = 0;
6016                                 lro = 0;
6017                         }
6018                 } else {
6019                         /* Reodering does not fade */
6020                         lro = 1;
6021                 }
6022         } else {
6023                 lro = 0;
6024         }
6025         if (rack->rc_rack_tmr_std_based == 0) {
6026                 thresh = srtt + rack->r_ctl.rc_pkt_delay;
6027         } else {
6028                 /* Standards based pkt-delay is 1/4 srtt */
6029                 thresh = srtt +  (srtt >> 2);
6030         }
6031         if (lro && (rack->rc_rack_tmr_std_based == 0)) {
6032                 /* It must be set, if not you get 1/4 rtt */
6033                 if (rack->r_ctl.rc_reorder_shift)
6034                         thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
6035                 else
6036                         thresh += (srtt >> 2);
6037         }
6038         if (rack->rc_rack_use_dsack &&
6039             lro &&
6040             (rack->r_ctl.num_dsack > 0)) {
6041                 /*
6042                  * We only increase the reordering window if we
6043                  * have seen reordering <and> we have a DSACK count.
6044                  */
6045                 thresh += rack->r_ctl.num_dsack * (srtt >> 2);
6046                 rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
6047         }
6048         /* SRTT * 2 is the ceiling */
6049         if (thresh > (srtt * 2)) {
6050                 thresh = srtt * 2;
6051         }
6052         /* And we don't want it above the RTO max either */
6053         if (thresh > rack_rto_max) {
6054                 thresh = rack_rto_max;
6055         }
6056         rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
6057         return (thresh);
6058 }
6059
6060 static uint32_t
6061 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
6062                      struct rack_sendmap *rsm, uint32_t srtt)
6063 {
6064         struct rack_sendmap *prsm;
6065         uint32_t thresh, len;
6066         int segsiz;
6067
6068         if (srtt == 0)
6069                 srtt = 1;
6070         if (rack->r_ctl.rc_tlp_threshold)
6071                 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
6072         else
6073                 thresh = (srtt * 2);
6074
6075         /* Get the previous sent packet, if any */
6076         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
6077         len = rsm->r_end - rsm->r_start;
6078         if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
6079                 /* Exactly like the ID */
6080                 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
6081                         uint32_t alt_thresh;
6082                         /*
6083                          * Compensate for delayed-ack with the d-ack time.
6084                          */
6085                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6086                         if (alt_thresh > thresh)
6087                                 thresh = alt_thresh;
6088                 }
6089         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6090                 /* 2.1 behavior */
6091                 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6092                 if (prsm && (len <= segsiz)) {
6093                         /*
6094                          * Two packets outstanding, thresh should be (2*srtt) +
6095                          * possible inter-packet delay (if any).
6096                          */
6097                         uint32_t inter_gap = 0;
6098                         int idx, nidx;
6099
6100                         idx = rsm->r_rtr_cnt - 1;
6101                         nidx = prsm->r_rtr_cnt - 1;
6102                         if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6103                                 /* Yes it was sent later (or at the same time) */
6104                                 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6105                         }
6106                         thresh += inter_gap;
6107                 } else if (len <= segsiz) {
6108                         /*
6109                          * Possibly compensate for delayed-ack.
6110                          */
6111                         uint32_t alt_thresh;
6112
6113                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6114                         if (alt_thresh > thresh)
6115                                 thresh = alt_thresh;
6116                 }
6117         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6118                 /* 2.2 behavior */
6119                 if (len <= segsiz) {
6120                         uint32_t alt_thresh;
6121                         /*
6122                          * Compensate for delayed-ack with the d-ack time.
6123                          */
6124                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6125                         if (alt_thresh > thresh)
6126                                 thresh = alt_thresh;
6127                 }
6128         }
6129         /* Not above an RTO */
6130         if (thresh > tp->t_rxtcur) {
6131                 thresh = tp->t_rxtcur;
6132         }
6133         /* Not above a RTO max */
6134         if (thresh > rack_rto_max) {
6135                 thresh = rack_rto_max;
6136         }
6137         /* Apply user supplied min TLP */
6138         if (thresh < rack_tlp_min) {
6139                 thresh = rack_tlp_min;
6140         }
6141         return (thresh);
6142 }
6143
6144 static uint32_t
6145 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6146 {
6147         /*
6148          * We want the rack_rtt which is the
6149          * last rtt we measured. However if that
6150          * does not exist we fallback to the srtt (which
6151          * we probably will never do) and then as a last
6152          * resort we use RACK_INITIAL_RTO if no srtt is
6153          * yet set.
6154          */
6155         if (rack->rc_rack_rtt)
6156                 return (rack->rc_rack_rtt);
6157         else if (tp->t_srtt == 0)
6158                 return (RACK_INITIAL_RTO);
6159         return (tp->t_srtt);
6160 }
6161
6162 static struct rack_sendmap *
6163 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6164 {
6165         /*
6166          * Check to see that we don't need to fall into recovery. We will
6167          * need to do so if our oldest transmit is past the time we should
6168          * have had an ack.
6169          */
6170         struct tcp_rack *rack;
6171         struct rack_sendmap *rsm;
6172         int32_t idx;
6173         uint32_t srtt, thresh;
6174
6175         rack = (struct tcp_rack *)tp->t_fb_ptr;
6176         if (tqhash_empty(rack->r_ctl.tqh)) {
6177                 return (NULL);
6178         }
6179         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6180         if (rsm == NULL)
6181                 return (NULL);
6182
6183
6184         if (rsm->r_flags & RACK_ACKED) {
6185                 rsm = rack_find_lowest_rsm(rack);
6186                 if (rsm == NULL)
6187                         return (NULL);
6188         }
6189         idx = rsm->r_rtr_cnt - 1;
6190         srtt = rack_grab_rtt(tp, rack);
6191         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
6192         if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6193                 return (NULL);
6194         }
6195         if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6196                 return (NULL);
6197         }
6198         /* Ok if we reach here we are over-due and this guy can be sent */
6199         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6200         return (rsm);
6201 }
6202
6203 static uint32_t
6204 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6205 {
6206         int32_t t;
6207         int32_t tt;
6208         uint32_t ret_val;
6209
6210         t = (tp->t_srtt + (tp->t_rttvar << 2));
6211         RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6212             rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6213         rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6214         ret_val = (uint32_t)tt;
6215         return (ret_val);
6216 }
6217
6218 static uint32_t
6219 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6220 {
6221         /*
6222          * Start the FR timer, we do this based on getting the first one in
6223          * the rc_tmap. Note that if its NULL we must stop the timer. in all
6224          * events we need to stop the running timer (if its running) before
6225          * starting the new one.
6226          */
6227         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6228         uint32_t srtt_cur;
6229         int32_t idx;
6230         int32_t is_tlp_timer = 0;
6231         struct rack_sendmap *rsm;
6232
6233         if (rack->t_timers_stopped) {
6234                 /* All timers have been stopped none are to run */
6235                 return (0);
6236         }
6237         if (rack->rc_in_persist) {
6238                 /* We can't start any timer in persists */
6239                 return (rack_get_persists_timer_val(tp, rack));
6240         }
6241         rack->rc_on_min_to = 0;
6242         if ((tp->t_state < TCPS_ESTABLISHED) ||
6243             (rack->sack_attack_disable > 0) ||
6244             ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6245                 goto activate_rxt;
6246         }
6247         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6248         if ((rsm == NULL) || sup_rack) {
6249                 /* Nothing on the send map or no rack */
6250 activate_rxt:
6251                 time_since_sent = 0;
6252                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6253                 if (rsm) {
6254                         /*
6255                          * Should we discount the RTX timer any?
6256                          *
6257                          * We want to discount it the smallest amount.
6258                          * If a timer (Rack/TLP or RXT) has gone off more
6259                          * recently thats the discount we want to use (now - timer time).
6260                          * If the retransmit of the oldest packet was more recent then
6261                          * we want to use that (now - oldest-packet-last_transmit_time).
6262                          *
6263                          */
6264                         idx = rsm->r_rtr_cnt - 1;
6265                         if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6266                                 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6267                         else
6268                                 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6269                         if (TSTMP_GT(cts, tstmp_touse))
6270                             time_since_sent = cts - tstmp_touse;
6271                 }
6272                 if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6273                     sbavail(&tptosocket(tp)->so_snd)) {
6274                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6275                         to = tp->t_rxtcur;
6276                         if (to > time_since_sent)
6277                                 to -= time_since_sent;
6278                         else
6279                                 to = rack->r_ctl.rc_min_to;
6280                         if (to == 0)
6281                                 to = 1;
6282                         /* Special case for KEEPINIT */
6283                         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6284                             (TP_KEEPINIT(tp) != 0) &&
6285                             rsm) {
6286                                 /*
6287                                  * We have to put a ceiling on the rxt timer
6288                                  * of the keep-init timeout.
6289                                  */
6290                                 uint32_t max_time, red;
6291
6292                                 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6293                                 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6294                                         red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6295                                         if (red < max_time)
6296                                                 max_time -= red;
6297                                         else
6298                                                 max_time = 1;
6299                                 }
6300                                 /* Reduce timeout to the keep value if needed */
6301                                 if (max_time < to)
6302                                         to = max_time;
6303                         }
6304                         return (to);
6305                 }
6306                 return (0);
6307         }
6308         if (rsm->r_flags & RACK_ACKED) {
6309                 rsm = rack_find_lowest_rsm(rack);
6310                 if (rsm == NULL) {
6311                         /* No lowest? */
6312                         goto activate_rxt;
6313                 }
6314         }
6315         if (rack->sack_attack_disable) {
6316                 /*
6317                  * We don't want to do
6318                  * any TLP's if you are an attacker.
6319                  * Though if you are doing what
6320                  * is expected you may still have
6321                  * SACK-PASSED marks.
6322                  */
6323                 goto activate_rxt;
6324         }
6325         /* Convert from ms to usecs */
6326         if ((rsm->r_flags & RACK_SACK_PASSED) ||
6327             (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6328             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6329                 if ((tp->t_flags & TF_SENTFIN) &&
6330                     ((tp->snd_max - tp->snd_una) == 1) &&
6331                     (rsm->r_flags & RACK_HAS_FIN)) {
6332                         /*
6333                          * We don't start a rack timer if all we have is a
6334                          * FIN outstanding.
6335                          */
6336                         goto activate_rxt;
6337                 }
6338                 if ((rack->use_rack_rr == 0) &&
6339                     (IN_FASTRECOVERY(tp->t_flags)) &&
6340                     (rack->rack_no_prr == 0) &&
6341                      (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
6342                         /*
6343                          * We are not cheating, in recovery  and
6344                          * not enough ack's to yet get our next
6345                          * retransmission out.
6346                          *
6347                          * Note that classified attackers do not
6348                          * get to use the rack-cheat.
6349                          */
6350                         goto activate_tlp;
6351                 }
6352                 srtt = rack_grab_rtt(tp, rack);
6353                 thresh = rack_calc_thresh_rack(rack, srtt, cts);
6354                 idx = rsm->r_rtr_cnt - 1;
6355                 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6356                 if (SEQ_GEQ(exp, cts)) {
6357                         to = exp - cts;
6358                         if (to < rack->r_ctl.rc_min_to) {
6359                                 to = rack->r_ctl.rc_min_to;
6360                                 if (rack->r_rr_config == 3)
6361                                         rack->rc_on_min_to = 1;
6362                         }
6363                 } else {
6364                         to = rack->r_ctl.rc_min_to;
6365                         if (rack->r_rr_config == 3)
6366                                 rack->rc_on_min_to = 1;
6367                 }
6368         } else {
6369                 /* Ok we need to do a TLP not RACK */
6370 activate_tlp:
6371                 if ((rack->rc_tlp_in_progress != 0) &&
6372                     (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6373                         /*
6374                          * The previous send was a TLP and we have sent
6375                          * N TLP's without sending new data.
6376                          */
6377                         goto activate_rxt;
6378                 }
6379                 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6380                 if (rsm == NULL) {
6381                         /* We found no rsm to TLP with. */
6382                         goto activate_rxt;
6383                 }
6384                 if (rsm->r_flags & RACK_HAS_FIN) {
6385                         /* If its a FIN we dont do TLP */
6386                         rsm = NULL;
6387                         goto activate_rxt;
6388                 }
6389                 idx = rsm->r_rtr_cnt - 1;
6390                 time_since_sent = 0;
6391                 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6392                         tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6393                 else
6394                         tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6395                 if (TSTMP_GT(cts, tstmp_touse))
6396                     time_since_sent = cts - tstmp_touse;
6397                 is_tlp_timer = 1;
6398                 if (tp->t_srtt) {
6399                         if ((rack->rc_srtt_measure_made == 0) &&
6400                             (tp->t_srtt == 1)) {
6401                                 /*
6402                                  * If another stack as run and set srtt to 1,
6403                                  * then the srtt was 0, so lets use the initial.
6404                                  */
6405                                 srtt = RACK_INITIAL_RTO;
6406                         } else {
6407                                 srtt_cur = tp->t_srtt;
6408                                 srtt = srtt_cur;
6409                         }
6410                 } else
6411                         srtt = RACK_INITIAL_RTO;
6412                 /*
6413                  * If the SRTT is not keeping up and the
6414                  * rack RTT has spiked we want to use
6415                  * the last RTT not the smoothed one.
6416                  */
6417                 if (rack_tlp_use_greater &&
6418                     tp->t_srtt &&
6419                     (srtt < rack_grab_rtt(tp, rack))) {
6420                         srtt = rack_grab_rtt(tp, rack);
6421                 }
6422                 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6423                 if (thresh > time_since_sent) {
6424                         to = thresh - time_since_sent;
6425                 } else {
6426                         to = rack->r_ctl.rc_min_to;
6427                         rack_log_alt_to_to_cancel(rack,
6428                                                   thresh,               /* flex1 */
6429                                                   time_since_sent,      /* flex2 */
6430                                                   tstmp_touse,          /* flex3 */
6431                                                   rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
6432                                                   (uint32_t)rsm->r_tim_lastsent[idx],
6433                                                   srtt,
6434                                                   idx, 99);
6435                 }
6436                 if (to < rack_tlp_min) {
6437                         to = rack_tlp_min;
6438                 }
6439                 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
6440                         /*
6441                          * If the TLP time works out to larger than the max
6442                          * RTO lets not do TLP.. just RTO.
6443                          */
6444                         goto activate_rxt;
6445                 }
6446         }
6447         if (is_tlp_timer == 0) {
6448                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
6449         } else {
6450                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
6451         }
6452         if (to == 0)
6453                 to = 1;
6454         return (to);
6455 }
6456
6457 static void
6458 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
6459 {
6460         struct timeval tv;
6461
6462         if (rack->rc_in_persist == 0) {
6463                 if (tp->t_flags & TF_GPUTINPROG) {
6464                         /*
6465                          * Stop the goodput now, the calling of the
6466                          * measurement function clears the flag.
6467                          */
6468                         rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
6469                                                     RACK_QUALITY_PERSIST);
6470                 }
6471 #ifdef NETFLIX_SHARED_CWND
6472                 if (rack->r_ctl.rc_scw) {
6473                         tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6474                         rack->rack_scwnd_is_idle = 1;
6475                 }
6476 #endif
6477                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(&tv);
6478                 if (rack->lt_bw_up) {
6479                         /* Suspend our LT BW measurement */
6480                         uint64_t tmark;
6481
6482                         rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
6483                         rack->r_ctl.lt_seq = snd_una;
6484                         tmark = tcp_tv_to_lusectick(&tv);
6485                         rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
6486                         rack->r_ctl.lt_timemark = tmark;
6487                         rack->lt_bw_up = 0;
6488                         rack->r_persist_lt_bw_off = 1;
6489                 }
6490                 if (rack->r_ctl.rc_went_idle_time == 0)
6491                         rack->r_ctl.rc_went_idle_time = 1;
6492                 rack_timer_cancel(tp, rack, cts, __LINE__);
6493                 rack->r_ctl.persist_lost_ends = 0;
6494                 rack->probe_not_answered = 0;
6495                 rack->forced_ack = 0;
6496                 tp->t_rxtshift = 0;
6497                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6498                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6499                 rack->rc_in_persist = 1;
6500         }
6501 }
6502
6503 static void
6504 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6505 {
6506         struct timeval tv;
6507         uint32_t t_time;
6508
6509         if (tcp_in_hpts(rack->rc_inp)) {
6510                 tcp_hpts_remove(rack->rc_inp);
6511                 rack->r_ctl.rc_hpts_flags = 0;
6512         }
6513 #ifdef NETFLIX_SHARED_CWND
6514         if (rack->r_ctl.rc_scw) {
6515                 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6516                 rack->rack_scwnd_is_idle = 0;
6517         }
6518 #endif
6519         t_time = tcp_get_usecs(&tv);
6520         if (rack->rc_gp_dyn_mul &&
6521             (rack->use_fixed_rate == 0) &&
6522             (rack->rc_always_pace)) {
6523                 /*
6524                  * Do we count this as if a probe-rtt just
6525                  * finished?
6526                  */
6527                 uint32_t time_idle, idle_min;
6528
6529                 time_idle = t_time - rack->r_ctl.rc_went_idle_time;
6530                 idle_min = rack_min_probertt_hold;
6531                 if (rack_probertt_gpsrtt_cnt_div) {
6532                         uint64_t extra;
6533                         extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
6534                                 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
6535                         extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
6536                         idle_min += (uint32_t)extra;
6537                 }
6538                 if (time_idle >= idle_min) {
6539                         /* Yes, we count it as a probe-rtt. */
6540                         uint32_t us_cts;
6541
6542                         us_cts = tcp_get_usecs(NULL);
6543                         if (rack->in_probe_rtt == 0) {
6544                                 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6545                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
6546                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
6547                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
6548                         } else {
6549                                 rack_exit_probertt(rack, us_cts);
6550                         }
6551                 }
6552         }
6553         if (rack->r_persist_lt_bw_off) {
6554                 /* Continue where we left off */
6555                 rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
6556                 rack->lt_bw_up = 1;
6557                 rack->r_persist_lt_bw_off = 0;
6558         }
6559         rack->rc_in_persist = 0;
6560         rack->r_ctl.rc_went_idle_time = 0;
6561         tp->t_rxtshift = 0;
6562         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6563            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6564         rack->r_ctl.rc_agg_delayed = 0;
6565         rack->r_early = 0;
6566         rack->r_late = 0;
6567         rack->r_ctl.rc_agg_early = 0;
6568 }
6569
6570 static void
6571 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
6572                    struct hpts_diag *diag, struct timeval *tv)
6573 {
6574         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6575                 union tcp_log_stackspecific log;
6576
6577                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6578                 log.u_bbr.flex1 = diag->p_nxt_slot;
6579                 log.u_bbr.flex2 = diag->p_cur_slot;
6580                 log.u_bbr.flex3 = diag->slot_req;
6581                 log.u_bbr.flex4 = diag->inp_hptsslot;
6582                 log.u_bbr.flex5 = diag->slot_remaining;
6583                 log.u_bbr.flex6 = diag->need_new_to;
6584                 log.u_bbr.flex7 = diag->p_hpts_active;
6585                 log.u_bbr.flex8 = diag->p_on_min_sleep;
6586                 /* Hijack other fields as needed */
6587                 log.u_bbr.epoch = diag->have_slept;
6588                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
6589                 log.u_bbr.pkts_out = diag->co_ret;
6590                 log.u_bbr.applimited = diag->hpts_sleep_time;
6591                 log.u_bbr.delivered = diag->p_prev_slot;
6592                 log.u_bbr.inflight = diag->p_runningslot;
6593                 log.u_bbr.bw_inuse = diag->wheel_slot;
6594                 log.u_bbr.rttProp = diag->wheel_cts;
6595                 log.u_bbr.timeStamp = cts;
6596                 log.u_bbr.delRate = diag->maxslots;
6597                 log.u_bbr.cur_del_rate = diag->p_curtick;
6598                 log.u_bbr.cur_del_rate <<= 32;
6599                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
6600                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
6601                     &rack->rc_inp->inp_socket->so_rcv,
6602                     &rack->rc_inp->inp_socket->so_snd,
6603                     BBR_LOG_HPTSDIAG, 0,
6604                     0, &log, false, tv);
6605         }
6606
6607 }
6608
6609 static void
6610 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
6611 {
6612         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6613                 union tcp_log_stackspecific log;
6614                 struct timeval tv;
6615
6616                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6617                 log.u_bbr.flex1 = sb->sb_flags;
6618                 log.u_bbr.flex2 = len;
6619                 log.u_bbr.flex3 = sb->sb_state;
6620                 log.u_bbr.flex8 = type;
6621                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6622                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
6623                     &rack->rc_inp->inp_socket->so_rcv,
6624                     &rack->rc_inp->inp_socket->so_snd,
6625                     TCP_LOG_SB_WAKE, 0,
6626                     len, &log, false, &tv);
6627         }
6628 }
6629
6630 static void
6631 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
6632       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
6633 {
6634         struct hpts_diag diag;
6635         struct inpcb *inp = tptoinpcb(tp);
6636         struct timeval tv;
6637         uint32_t delayed_ack = 0;
6638         uint32_t hpts_timeout;
6639         uint32_t entry_slot = slot;
6640         uint8_t stopped;
6641         uint32_t left = 0;
6642         uint32_t us_cts;
6643
6644         if ((tp->t_state == TCPS_CLOSED) ||
6645             (tp->t_state == TCPS_LISTEN)) {
6646                 return;
6647         }
6648         if (tcp_in_hpts(inp)) {
6649                 /* Already on the pacer */
6650                 return;
6651         }
6652         stopped = rack->rc_tmr_stopped;
6653         if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
6654                 left = rack->r_ctl.rc_timer_exp - cts;
6655         }
6656         rack->r_ctl.rc_timer_exp = 0;
6657         rack->r_ctl.rc_hpts_flags = 0;
6658         us_cts = tcp_get_usecs(&tv);
6659         /* Now early/late accounting */
6660         rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
6661         if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
6662                 /*
6663                  * We have a early carry over set,
6664                  * we can always add more time so we
6665                  * can always make this compensation.
6666                  *
6667                  * Note if ack's are allowed to wake us do not
6668                  * penalize the next timer for being awoke
6669                  * by an ack aka the rc_agg_early (non-paced mode).
6670                  */
6671                 slot += rack->r_ctl.rc_agg_early;
6672                 rack->r_early = 0;
6673                 rack->r_ctl.rc_agg_early = 0;
6674         }
6675         if (rack->r_late) {
6676                 /*
6677                  * This is harder, we can
6678                  * compensate some but it
6679                  * really depends on what
6680                  * the current pacing time is.
6681                  */
6682                 if (rack->r_ctl.rc_agg_delayed >= slot) {
6683                         /*
6684                          * We can't compensate for it all.
6685                          * And we have to have some time
6686                          * on the clock. We always have a min
6687                          * 10 slots (10 x 10 i.e. 100 usecs).
6688                          */
6689                         if (slot <= HPTS_TICKS_PER_SLOT) {
6690                                 /* We gain delay */
6691                                 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
6692                                 slot = HPTS_TICKS_PER_SLOT;
6693                         } else {
6694                                 /* We take off some */
6695                                 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
6696                                 slot = HPTS_TICKS_PER_SLOT;
6697                         }
6698                 } else {
6699                         slot -= rack->r_ctl.rc_agg_delayed;
6700                         rack->r_ctl.rc_agg_delayed = 0;
6701                         /* Make sure we have 100 useconds at minimum */
6702                         if (slot < HPTS_TICKS_PER_SLOT) {
6703                                 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
6704                                 slot = HPTS_TICKS_PER_SLOT;
6705                         }
6706                         if (rack->r_ctl.rc_agg_delayed == 0)
6707                                 rack->r_late = 0;
6708                 }
6709         }
6710         hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
6711 #ifdef TCP_SAD_DETECTION
6712         if (rack->sack_attack_disable &&
6713             (rack->r_ctl.ack_during_sd > 0) &&
6714             (slot < tcp_sad_pacing_interval)) {
6715                 /*
6716                  * We have a potential attacker on
6717                  * the line. We have possibly some
6718                  * (or now) pacing time set. We want to
6719                  * slow down the processing of sacks by some
6720                  * amount (if it is an attacker). Set the default
6721                  * slot for attackers in place (unless the original
6722                  * interval is longer). Its stored in
6723                  * micro-seconds, so lets convert to msecs.
6724                  */
6725                 slot = tcp_sad_pacing_interval;
6726                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6727                 rack->r_ctl.ack_during_sd = 0;
6728         }
6729 #endif
6730         if (tp->t_flags & TF_DELACK) {
6731                 delayed_ack = TICKS_2_USEC(tcp_delacktime);
6732                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
6733         }
6734         if (delayed_ack && ((hpts_timeout == 0) ||
6735                             (delayed_ack < hpts_timeout)))
6736                 hpts_timeout = delayed_ack;
6737         else
6738                 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6739         /*
6740          * If no timers are going to run and we will fall off the hptsi
6741          * wheel, we resort to a keep-alive timer if its configured.
6742          */
6743         if ((hpts_timeout == 0) &&
6744             (slot == 0)) {
6745                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6746                     (tp->t_state <= TCPS_CLOSING)) {
6747                         /*
6748                          * Ok we have no timer (persists, rack, tlp, rxt  or
6749                          * del-ack), we don't have segments being paced. So
6750                          * all that is left is the keepalive timer.
6751                          */
6752                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6753                                 /* Get the established keep-alive time */
6754                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
6755                         } else {
6756                                 /*
6757                                  * Get the initial setup keep-alive time,
6758                                  * note that this is probably not going to
6759                                  * happen, since rack will be running a rxt timer
6760                                  * if a SYN of some sort is outstanding. It is
6761                                  * actually handled in rack_timeout_rxt().
6762                                  */
6763                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
6764                         }
6765                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
6766                         if (rack->in_probe_rtt) {
6767                                 /*
6768                                  * We want to instead not wake up a long time from
6769                                  * now but to wake up about the time we would
6770                                  * exit probe-rtt and initiate a keep-alive ack.
6771                                  * This will get us out of probe-rtt and update
6772                                  * our min-rtt.
6773                                  */
6774                                 hpts_timeout = rack_min_probertt_hold;
6775                         }
6776                 }
6777         }
6778         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
6779             (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
6780                 /*
6781                  * RACK, TLP, persists and RXT timers all are restartable
6782                  * based on actions input .. i.e we received a packet (ack
6783                  * or sack) and that changes things (rw, or snd_una etc).
6784                  * Thus we can restart them with a new value. For
6785                  * keep-alive, delayed_ack we keep track of what was left
6786                  * and restart the timer with a smaller value.
6787                  */
6788                 if (left < hpts_timeout)
6789                         hpts_timeout = left;
6790         }
6791         if (hpts_timeout) {
6792                 /*
6793                  * Hack alert for now we can't time-out over 2,147,483
6794                  * seconds (a bit more than 596 hours), which is probably ok
6795                  * :).
6796                  */
6797                 if (hpts_timeout > 0x7ffffffe)
6798                         hpts_timeout = 0x7ffffffe;
6799                 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
6800         }
6801         rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
6802         if ((rack->gp_ready == 0) &&
6803             (rack->use_fixed_rate == 0) &&
6804             (hpts_timeout < slot) &&
6805             (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
6806                 /*
6807                  * We have no good estimate yet for the
6808                  * old clunky burst mitigation or the
6809                  * real pacing. And the tlp or rxt is smaller
6810                  * than the pacing calculation. Lets not
6811                  * pace that long since we know the calculation
6812                  * so far is not accurate.
6813                  */
6814                 slot = hpts_timeout;
6815         }
6816         /**
6817          * Turn off all the flags for queuing by default. The
6818          * flags have important meanings to what happens when
6819          * LRO interacts with the transport. Most likely (by default now)
6820          * mbuf_queueing and ack compression are on. So the transport
6821          * has a couple of flags that control what happens (if those
6822          * are not on then these flags won't have any effect since it
6823          * won't go through the queuing LRO path).
6824          *
6825          * INP_MBUF_QUEUE_READY - This flags says that I am busy
6826          *                        pacing output, so don't disturb. But
6827          *                        it also means LRO can wake me if there
6828          *                        is a SACK arrival.
6829          *
6830          * INP_DONT_SACK_QUEUE - This flag is used in conjunction
6831          *                       with the above flag (QUEUE_READY) and
6832          *                       when present it says don't even wake me
6833          *                       if a SACK arrives.
6834          *
6835          * The idea behind these flags is that if we are pacing we
6836          * set the MBUF_QUEUE_READY and only get woken up if
6837          * a SACK arrives (which could change things) or if
6838          * our pacing timer expires. If, however, we have a rack
6839          * timer running, then we don't even want a sack to wake
6840          * us since the rack timer has to expire before we can send.
6841          *
6842          * Other cases should usually have none of the flags set
6843          * so LRO can call into us.
6844          */
6845         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
6846         if (slot) {
6847                 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
6848                 rack->r_ctl.rc_last_output_to = us_cts + slot;
6849                 /*
6850                  * A pacing timer (slot) is being set, in
6851                  * such a case we cannot send (we are blocked by
6852                  * the timer). So lets tell LRO that it should not
6853                  * wake us unless there is a SACK. Note this only
6854                  * will be effective if mbuf queueing is on or
6855                  * compressed acks are being processed.
6856                  */
6857                 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
6858                 /*
6859                  * But wait if we have a Rack timer running
6860                  * even a SACK should not disturb us (with
6861                  * the exception of r_rr_config 3).
6862                  */
6863                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) {
6864                         if (rack->r_rr_config != 3)
6865                                 inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
6866                         else if (rack->rc_pace_dnd) {
6867                                 if (IN_RECOVERY(tp->t_flags)) {
6868                                         /*
6869                                          * When DND is on, we only let a sack
6870                                          * interrupt us if we are not in recovery.
6871                                          *
6872                                          * If DND is off, then we never hit here
6873                                          * and let all sacks wake us up.
6874                                          *
6875                                          */
6876                                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
6877                                 }
6878                         }
6879                 }
6880                 /* For sack attackers we want to ignore sack */
6881                 if (rack->sack_attack_disable == 1) {
6882                         inp->inp_flags2 |= (INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
6883                 } else if (rack->rc_ack_can_sendout_data) {
6884                         /*
6885                          * Ahh but wait, this is that special case
6886                          * where the pacing timer can be disturbed
6887                          * backout the changes (used for non-paced
6888                          * burst limiting).
6889                          */
6890                         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
6891                 }
6892                 if ((rack->use_rack_rr) &&
6893                     (rack->r_rr_config < 2) &&
6894                     ((hpts_timeout) && (hpts_timeout < slot))) {
6895                         /*
6896                          * Arrange for the hpts to kick back in after the
6897                          * t-o if the t-o does not cause a send.
6898                          */
6899                         (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6900                                                    __LINE__, &diag);
6901                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6902                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6903                 } else {
6904                         (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(slot),
6905                                                    __LINE__, &diag);
6906                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6907                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
6908                 }
6909         } else if (hpts_timeout) {
6910                 /*
6911                  * With respect to inp_flags2 here, lets let any new acks wake
6912                  * us up here. Since we are not pacing (no pacing timer), output
6913                  * can happen so we should let it. If its a Rack timer, then any inbound
6914                  * packet probably won't change the sending (we will be blocked)
6915                  * but it may change the prr stats so letting it in (the set defaults
6916                  * at the start of this block) are good enough.
6917                  */
6918                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6919                 (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6920                                            __LINE__, &diag);
6921                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6922                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6923         } else {
6924                 /* No timer starting */
6925 #ifdef INVARIANTS
6926                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
6927                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
6928                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
6929                 }
6930 #endif
6931         }
6932         rack->rc_tmr_stopped = 0;
6933         if (slot)
6934                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6935 }
6936
6937 /*
6938  * RACK Timer, here we simply do logging and house keeping.
6939  * the normal rack_output() function will call the
6940  * appropriate thing to check if we need to do a RACK retransmit.
6941  * We return 1, saying don't proceed with rack_output only
6942  * when all timers have been stopped (destroyed PCB?).
6943  */
6944 static int
6945 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6946 {
6947         /*
6948          * This timer simply provides an internal trigger to send out data.
6949          * The check_recovery_mode call will see if there are needed
6950          * retransmissions, if so we will enter fast-recovery. The output
6951          * call may or may not do the same thing depending on sysctl
6952          * settings.
6953          */
6954         struct rack_sendmap *rsm;
6955
6956         counter_u64_add(rack_to_tot, 1);
6957         if (rack->r_state && (rack->r_state != tp->t_state))
6958                 rack_set_state(tp, rack);
6959         rack->rc_on_min_to = 0;
6960         rsm = rack_check_recovery_mode(tp, cts);
6961         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
6962         if (rsm) {
6963                 rack->r_ctl.rc_resend = rsm;
6964                 rack->r_timer_override = 1;
6965                 if (rack->use_rack_rr) {
6966                         /*
6967                          * Don't accumulate extra pacing delay
6968                          * we are allowing the rack timer to
6969                          * over-ride pacing i.e. rrr takes precedence
6970                          * if the pacing interval is longer than the rrr
6971                          * time (in other words we get the min pacing
6972                          * time versus rrr pacing time).
6973                          */
6974                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6975                 }
6976         }
6977         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
6978         if (rsm == NULL) {
6979                 /* restart a timer and return 1 */
6980                 rack_start_hpts_timer(rack, tp, cts,
6981                                       0, 0, 0);
6982                 return (1);
6983         }
6984         return (0);
6985 }
6986
6987
6988
6989 static void
6990 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
6991 {
6992
6993         if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
6994                 /*
6995                  * The trailing space changed, mbufs can grow
6996                  * at the tail but they can't shrink from
6997                  * it, KASSERT that. Adjust the orig_m_len to
6998                  * compensate for this change.
6999                  */
7000                 KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
7001                         ("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
7002                          rsm->m,
7003                          rsm,
7004                          (intmax_t)M_TRAILINGROOM(rsm->m),
7005                          rsm->orig_t_space,
7006                          rsm->orig_m_len,
7007                          rsm->m->m_len));
7008                 rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
7009                 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7010         }
7011         if (rsm->m->m_len < rsm->orig_m_len) {
7012                 /*
7013                  * Mbuf shrank, trimmed off the top by an ack, our
7014                  * offset changes.
7015                  */
7016                 KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
7017                         ("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
7018                          rsm->m, rsm->m->m_len,
7019                          rsm, rsm->orig_m_len,
7020                          rsm->soff));
7021                 if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
7022                         rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
7023                 else
7024                         rsm->soff = 0;
7025                 rsm->orig_m_len = rsm->m->m_len;
7026 #ifdef INVARIANTS
7027         } else if (rsm->m->m_len > rsm->orig_m_len) {
7028                 panic("rsm:%p m:%p m_len grew outside of t_space compensation",
7029                       rsm, rsm->m);
7030 #endif
7031         }
7032 }
7033
7034 static void
7035 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
7036 {
7037         struct mbuf *m;
7038         uint32_t soff;
7039
7040         if (src_rsm->m &&
7041             ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
7042              (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
7043                 /* Fix up the orig_m_len and possibly the mbuf offset */
7044                 rack_adjust_orig_mlen(src_rsm);
7045         }
7046         m = src_rsm->m;
7047         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
7048         while (soff >= m->m_len) {
7049                 /* Move out past this mbuf */
7050                 soff -= m->m_len;
7051                 m = m->m_next;
7052                 KASSERT((m != NULL),
7053                         ("rsm:%p nrsm:%p hit at soff:%u null m",
7054                          src_rsm, rsm, soff));
7055                 if (m == NULL) {
7056                         /* This should *not* happen which is why there is a kassert */
7057                         src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7058                                                (src_rsm->r_start - rack->rc_tp->snd_una),
7059                                                &src_rsm->soff);
7060                         src_rsm->orig_m_len = src_rsm->m->m_len;
7061                         src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7062                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7063                                            (rsm->r_start - rack->rc_tp->snd_una),
7064                                            &rsm->soff);
7065                         rsm->orig_m_len = rsm->m->m_len;
7066                         rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7067                         return;
7068                 }
7069         }
7070         rsm->m = m;
7071         rsm->soff = soff;
7072         rsm->orig_m_len = m->m_len;
7073         rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7074 }
7075
7076 static __inline void
7077 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7078                struct rack_sendmap *rsm, uint32_t start)
7079 {
7080         int idx;
7081
7082         nrsm->r_start = start;
7083         nrsm->r_end = rsm->r_end;
7084         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7085         nrsm->r_flags = rsm->r_flags;
7086         nrsm->r_dupack = rsm->r_dupack;
7087         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7088         nrsm->r_rtr_bytes = 0;
7089         nrsm->r_fas = rsm->r_fas;
7090         nrsm->r_bas = rsm->r_bas;
7091         rsm->r_end = nrsm->r_start;
7092         nrsm->r_just_ret = rsm->r_just_ret;
7093         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7094                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7095         }
7096         /* Now if we have SYN flag we keep it on the left edge */
7097         if (nrsm->r_flags & RACK_HAS_SYN)
7098                 nrsm->r_flags &= ~RACK_HAS_SYN;
7099         /* Now if we have a FIN flag we keep it on the right edge */
7100         if (rsm->r_flags & RACK_HAS_FIN)
7101                 rsm->r_flags &= ~RACK_HAS_FIN;
7102         /* Push bit must go to the right edge as well */
7103         if (rsm->r_flags & RACK_HAD_PUSH)
7104                 rsm->r_flags &= ~RACK_HAD_PUSH;
7105         /* Clone over the state of the hw_tls flag */
7106         nrsm->r_hw_tls = rsm->r_hw_tls;
7107         /*
7108          * Now we need to find nrsm's new location in the mbuf chain
7109          * we basically calculate a new offset, which is soff +
7110          * how much is left in original rsm. Then we walk out the mbuf
7111          * chain to find the righ position, it may be the same mbuf
7112          * or maybe not.
7113          */
7114         KASSERT(((rsm->m != NULL) ||
7115                  (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7116                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7117         if (rsm->m)
7118                 rack_setup_offset_for_rsm(rack, rsm, nrsm);
7119 }
7120
7121 static struct rack_sendmap *
7122 rack_merge_rsm(struct tcp_rack *rack,
7123                struct rack_sendmap *l_rsm,
7124                struct rack_sendmap *r_rsm)
7125 {
7126         /*
7127          * We are merging two ack'd RSM's,
7128          * the l_rsm is on the left (lower seq
7129          * values) and the r_rsm is on the right
7130          * (higher seq value). The simplest way
7131          * to merge these is to move the right
7132          * one into the left. I don't think there
7133          * is any reason we need to try to find
7134          * the oldest (or last oldest retransmitted).
7135          */
7136         rack_log_map_chg(rack->rc_tp, rack, NULL,
7137                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7138         l_rsm->r_end = r_rsm->r_end;
7139         if (l_rsm->r_dupack < r_rsm->r_dupack)
7140                 l_rsm->r_dupack = r_rsm->r_dupack;
7141         if (r_rsm->r_rtr_bytes)
7142                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7143         if (r_rsm->r_in_tmap) {
7144                 /* This really should not happen */
7145                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7146                 r_rsm->r_in_tmap = 0;
7147         }
7148
7149         /* Now the flags */
7150         if (r_rsm->r_flags & RACK_HAS_FIN)
7151                 l_rsm->r_flags |= RACK_HAS_FIN;
7152         if (r_rsm->r_flags & RACK_TLP)
7153                 l_rsm->r_flags |= RACK_TLP;
7154         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7155                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7156         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
7157             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7158                 /*
7159                  * If both are app-limited then let the
7160                  * free lower the count. If right is app
7161                  * limited and left is not, transfer.
7162                  */
7163                 l_rsm->r_flags |= RACK_APP_LIMITED;
7164                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
7165                 if (r_rsm == rack->r_ctl.rc_first_appl)
7166                         rack->r_ctl.rc_first_appl = l_rsm;
7167         }
7168         tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7169         /*
7170          * We keep the largest value, which is the newest
7171          * send. We do this in case a segment that is
7172          * joined together and not part of a GP estimate
7173          * later gets expanded into the GP estimate.
7174          *
7175          * We prohibit the merging of unlike kinds i.e.
7176          * all pieces that are in the GP estimate can be
7177          * merged and all pieces that are not in a GP estimate
7178          * can be merged, but not disimilar pieces. Combine
7179          * this with taking the highest here and we should
7180          * be ok unless of course the client reneges. Then
7181          * all bets are off.
7182          */
7183         if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7184            r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7185                 l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7186         }
7187         /*
7188          * When merging two RSM's we also need to consider the ack time and keep
7189          * newest. If the ack gets merged into a measurement then that is the
7190          * one we will want to be using.
7191          */
7192         if(l_rsm->r_ack_arrival  < r_rsm->r_ack_arrival)
7193                 l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7194
7195         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7196                 /* Transfer the split limit to the map we free */
7197                 r_rsm->r_limit_type = l_rsm->r_limit_type;
7198                 l_rsm->r_limit_type = 0;
7199         }
7200         rack_free(rack, r_rsm);
7201         l_rsm->r_flags |= RACK_MERGED;
7202         return (l_rsm);
7203 }
7204
7205 /*
7206  * TLP Timer, here we simply setup what segment we want to
7207  * have the TLP expire on, the normal rack_output() will then
7208  * send it out.
7209  *
7210  * We return 1, saying don't proceed with rack_output only
7211  * when all timers have been stopped (destroyed PCB?).
7212  */
7213 static int
7214 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7215 {
7216         /*
7217          * Tail Loss Probe.
7218          */
7219         struct rack_sendmap *rsm = NULL;
7220         int insret __diagused;
7221         struct socket *so = tptosocket(tp);
7222         uint32_t amm;
7223         uint32_t out, avail;
7224         int collapsed_win = 0;
7225
7226         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7227                 /* Its not time yet */
7228                 return (0);
7229         }
7230         if (ctf_progress_timeout_check(tp, true)) {
7231                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7232                 return (-ETIMEDOUT);    /* tcp_drop() */
7233         }
7234         /*
7235          * A TLP timer has expired. We have been idle for 2 rtts. So we now
7236          * need to figure out how to force a full MSS segment out.
7237          */
7238         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7239         rack->r_ctl.retran_during_recovery = 0;
7240         rack->r_ctl.dsack_byte_cnt = 0;
7241         counter_u64_add(rack_tlp_tot, 1);
7242         if (rack->r_state && (rack->r_state != tp->t_state))
7243                 rack_set_state(tp, rack);
7244         avail = sbavail(&so->so_snd);
7245         out = tp->snd_max - tp->snd_una;
7246         if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7247                 /* special case, we need a retransmission */
7248                 collapsed_win = 1;
7249                 goto need_retran;
7250         }
7251         if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7252                 rack->r_ctl.dsack_persist--;
7253                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7254                         rack->r_ctl.num_dsack = 0;
7255                 }
7256                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7257         }
7258         if ((tp->t_flags & TF_GPUTINPROG) &&
7259             (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7260                 /*
7261                  * If this is the second in a row
7262                  * TLP and we are doing a measurement
7263                  * its time to abandon the measurement.
7264                  * Something is likely broken on
7265                  * the clients network and measuring a
7266                  * broken network does us no good.
7267                  */
7268                 tp->t_flags &= ~TF_GPUTINPROG;
7269                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7270                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
7271                                            tp->gput_seq,
7272                                            0, 0, 18, __LINE__, NULL, 0);
7273         }
7274         /*
7275          * Check our send oldest always settings, and if
7276          * there is an oldest to send jump to the need_retran.
7277          */
7278         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7279                 goto need_retran;
7280
7281         if (avail > out) {
7282                 /* New data is available */
7283                 amm = avail - out;
7284                 if (amm > ctf_fixed_maxseg(tp)) {
7285                         amm = ctf_fixed_maxseg(tp);
7286                         if ((amm + out) > tp->snd_wnd) {
7287                                 /* We are rwnd limited */
7288                                 goto need_retran;
7289                         }
7290                 } else if (amm < ctf_fixed_maxseg(tp)) {
7291                         /* not enough to fill a MTU */
7292                         goto need_retran;
7293                 }
7294                 if (IN_FASTRECOVERY(tp->t_flags)) {
7295                         /* Unlikely */
7296                         if (rack->rack_no_prr == 0) {
7297                                 if (out + amm <= tp->snd_wnd) {
7298                                         rack->r_ctl.rc_prr_sndcnt = amm;
7299                                         rack->r_ctl.rc_tlp_new_data = amm;
7300                                         rack_log_to_prr(rack, 4, 0, __LINE__);
7301                                 }
7302                         } else
7303                                 goto need_retran;
7304                 } else {
7305                         /* Set the send-new override */
7306                         if (out + amm <= tp->snd_wnd)
7307                                 rack->r_ctl.rc_tlp_new_data = amm;
7308                         else
7309                                 goto need_retran;
7310                 }
7311                 rack->r_ctl.rc_tlpsend = NULL;
7312                 counter_u64_add(rack_tlp_newdata, 1);
7313                 goto send;
7314         }
7315 need_retran:
7316         /*
7317          * Ok we need to arrange the last un-acked segment to be re-sent, or
7318          * optionally the first un-acked segment.
7319          */
7320         if (collapsed_win == 0) {
7321                 if (rack_always_send_oldest)
7322                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7323                 else {
7324                         rsm = tqhash_max(rack->r_ctl.tqh);
7325                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7326                                 rsm = rack_find_high_nonack(rack, rsm);
7327                         }
7328                 }
7329                 if (rsm == NULL) {
7330 #ifdef TCP_BLACKBOX
7331                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7332 #endif
7333                         goto out;
7334                 }
7335         } else {
7336                 /*
7337                  * We had a collapsed window, lets find
7338                  * the point before the collapse.
7339                  */
7340                 if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7341                         rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7342                 else {
7343                         rsm = tqhash_min(rack->r_ctl.tqh);
7344                 }
7345                 if (rsm == NULL) {
7346                         /* Huh */
7347                         goto out;
7348                 }
7349         }
7350         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7351                 /*
7352                  * We need to split this the last segment in two.
7353                  */
7354                 struct rack_sendmap *nrsm;
7355
7356                 nrsm = rack_alloc_full_limit(rack);
7357                 if (nrsm == NULL) {
7358                         /*
7359                          * No memory to split, we will just exit and punt
7360                          * off to the RXT timer.
7361                          */
7362                         goto out;
7363                 }
7364                 rack_clone_rsm(rack, nrsm, rsm,
7365                                (rsm->r_end - ctf_fixed_maxseg(tp)));
7366                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7367 #ifndef INVARIANTS
7368                 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7369 #else
7370                 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7371                         panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
7372                               nrsm, insret, rack, rsm);
7373                 }
7374 #endif
7375                 if (rsm->r_in_tmap) {
7376                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7377                         nrsm->r_in_tmap = 1;
7378                 }
7379                 rsm = nrsm;
7380         }
7381         rack->r_ctl.rc_tlpsend = rsm;
7382 send:
7383         /* Make sure output path knows we are doing a TLP */
7384         *doing_tlp = 1;
7385         rack->r_timer_override = 1;
7386         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7387         return (0);
7388 out:
7389         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7390         return (0);
7391 }
7392
7393 /*
7394  * Delayed ack Timer, here we simply need to setup the
7395  * ACK_NOW flag and remove the DELACK flag. From there
7396  * the output routine will send the ack out.
7397  *
7398  * We only return 1, saying don't proceed, if all timers
7399  * are stopped (destroyed PCB?).
7400  */
7401 static int
7402 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7403 {
7404
7405         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7406         tp->t_flags &= ~TF_DELACK;
7407         tp->t_flags |= TF_ACKNOW;
7408         KMOD_TCPSTAT_INC(tcps_delack);
7409         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7410         return (0);
7411 }
7412
7413 /*
7414  * Persists timer, here we simply send the
7415  * same thing as a keepalive will.
7416  * the one byte send.
7417  *
7418  * We only return 1, saying don't proceed, if all timers
7419  * are stopped (destroyed PCB?).
7420  */
7421 static int
7422 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7423 {
7424         struct tcptemp *t_template;
7425         int32_t retval = 1;
7426
7427         if (rack->rc_in_persist == 0)
7428                 return (0);
7429         if (ctf_progress_timeout_check(tp, false)) {
7430                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7431                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7432                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7433                 return (-ETIMEDOUT);    /* tcp_drop() */
7434         }
7435         /*
7436          * Persistence timer into zero window. Force a byte to be output, if
7437          * possible.
7438          */
7439         KMOD_TCPSTAT_INC(tcps_persisttimeo);
7440         /*
7441          * Hack: if the peer is dead/unreachable, we do not time out if the
7442          * window is closed.  After a full backoff, drop the connection if
7443          * the idle time (no responses to probes) reaches the maximum
7444          * backoff that we would use if retransmitting.
7445          */
7446         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
7447             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
7448              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
7449                 KMOD_TCPSTAT_INC(tcps_persistdrop);
7450                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7451                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7452                 retval = -ETIMEDOUT;    /* tcp_drop() */
7453                 goto out;
7454         }
7455         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
7456             tp->snd_una == tp->snd_max)
7457                 rack_exit_persist(tp, rack, cts);
7458         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
7459         /*
7460          * If the user has closed the socket then drop a persisting
7461          * connection after a much reduced timeout.
7462          */
7463         if (tp->t_state > TCPS_CLOSE_WAIT &&
7464             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
7465                 KMOD_TCPSTAT_INC(tcps_persistdrop);
7466                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7467                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7468                 retval = -ETIMEDOUT;    /* tcp_drop() */
7469                 goto out;
7470         }
7471         t_template = tcpip_maketemplate(rack->rc_inp);
7472         if (t_template) {
7473                 /* only set it if we were answered */
7474                 if (rack->forced_ack == 0) {
7475                         rack->forced_ack = 1;
7476                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7477                 } else {
7478                         rack->probe_not_answered = 1;
7479                         counter_u64_add(rack_persists_loss, 1);
7480                         rack->r_ctl.persist_lost_ends++;
7481                 }
7482                 counter_u64_add(rack_persists_sends, 1);
7483                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
7484                 tcp_respond(tp, t_template->tt_ipgen,
7485                             &t_template->tt_t, (struct mbuf *)NULL,
7486                             tp->rcv_nxt, tp->snd_una - 1, 0);
7487                 /* This sends an ack */
7488                 if (tp->t_flags & TF_DELACK)
7489                         tp->t_flags &= ~TF_DELACK;
7490                 free(t_template, M_TEMP);
7491         }
7492         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
7493                 tp->t_rxtshift++;
7494 out:
7495         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
7496         rack_start_hpts_timer(rack, tp, cts,
7497                               0, 0, 0);
7498         return (retval);
7499 }
7500
7501 /*
7502  * If a keepalive goes off, we had no other timers
7503  * happening. We always return 1 here since this
7504  * routine either drops the connection or sends
7505  * out a segment with respond.
7506  */
7507 static int
7508 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7509 {
7510         struct tcptemp *t_template;
7511         struct inpcb *inp = tptoinpcb(tp);
7512
7513         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
7514         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
7515         /*
7516          * Keep-alive timer went off; send something or drop connection if
7517          * idle for too long.
7518          */
7519         KMOD_TCPSTAT_INC(tcps_keeptimeo);
7520         if (tp->t_state < TCPS_ESTABLISHED)
7521                 goto dropit;
7522         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7523             tp->t_state <= TCPS_CLOSING) {
7524                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
7525                         goto dropit;
7526                 /*
7527                  * Send a packet designed to force a response if the peer is
7528                  * up and reachable: either an ACK if the connection is
7529                  * still alive, or an RST if the peer has closed the
7530                  * connection due to timeout or reboot. Using sequence
7531                  * number tp->snd_una-1 causes the transmitted zero-length
7532                  * segment to lie outside the receive window; by the
7533                  * protocol spec, this requires the correspondent TCP to
7534                  * respond.
7535                  */
7536                 KMOD_TCPSTAT_INC(tcps_keepprobe);
7537                 t_template = tcpip_maketemplate(inp);
7538                 if (t_template) {
7539                         if (rack->forced_ack == 0) {
7540                                 rack->forced_ack = 1;
7541                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7542                         } else {
7543                                 rack->probe_not_answered = 1;
7544                         }
7545                         tcp_respond(tp, t_template->tt_ipgen,
7546                             &t_template->tt_t, (struct mbuf *)NULL,
7547                             tp->rcv_nxt, tp->snd_una - 1, 0);
7548                         free(t_template, M_TEMP);
7549                 }
7550         }
7551         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
7552         return (1);
7553 dropit:
7554         KMOD_TCPSTAT_INC(tcps_keepdrops);
7555         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7556         return (-ETIMEDOUT);    /* tcp_drop() */
7557 }
7558
7559 /*
7560  * Retransmit helper function, clear up all the ack
7561  * flags and take care of important book keeping.
7562  */
7563 static void
7564 rack_remxt_tmr(struct tcpcb *tp)
7565 {
7566         /*
7567          * The retransmit timer went off, all sack'd blocks must be
7568          * un-acked.
7569          */
7570         struct rack_sendmap *rsm, *trsm = NULL;
7571         struct tcp_rack *rack;
7572
7573         rack = (struct tcp_rack *)tp->t_fb_ptr;
7574         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
7575         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
7576         if (rack->r_state && (rack->r_state != tp->t_state))
7577                 rack_set_state(tp, rack);
7578         /*
7579          * Ideally we would like to be able to
7580          * mark SACK-PASS on anything not acked here.
7581          *
7582          * However, if we do that we would burst out
7583          * all that data 1ms apart. This would be unwise,
7584          * so for now we will just let the normal rxt timer
7585          * and tlp timer take care of it.
7586          *
7587          * Also we really need to stick them back in sequence
7588          * order. This way we send in the proper order and any
7589          * sacks that come floating in will "re-ack" the data.
7590          * To do this we zap the tmap with an INIT and then
7591          * walk through and place every rsm in the RB tree
7592          * back in its seq ordered place.
7593          */
7594         TAILQ_INIT(&rack->r_ctl.rc_tmap);
7595
7596         TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
7597                 rsm->r_dupack = 0;
7598                 if (rack_verbose_logging)
7599                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7600                 /* We must re-add it back to the tlist */
7601                 if (trsm == NULL) {
7602                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7603                 } else {
7604                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
7605                 }
7606                 rsm->r_in_tmap = 1;
7607                 trsm = rsm;
7608                 if (rsm->r_flags & RACK_ACKED)
7609                         rsm->r_flags |= RACK_WAS_ACKED;
7610                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
7611                 rsm->r_flags |= RACK_MUST_RXT;
7612         }
7613         /* Clear the count (we just un-acked them) */
7614         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
7615         rack->r_ctl.rc_sacked = 0;
7616         rack->r_ctl.rc_sacklast = NULL;
7617         rack->r_ctl.rc_agg_delayed = 0;
7618         rack->r_early = 0;
7619         rack->r_ctl.rc_agg_early = 0;
7620         rack->r_late = 0;
7621         /* Clear the tlp rtx mark */
7622         rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7623         if (rack->r_ctl.rc_resend != NULL)
7624                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7625         rack->r_ctl.rc_prr_sndcnt = 0;
7626         rack_log_to_prr(rack, 6, 0, __LINE__);
7627         rack->r_timer_override = 1;
7628         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
7629 #ifdef TCP_SAD_DETECTION
7630             || (rack->sack_attack_disable != 0)
7631 #endif
7632                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
7633                 /*
7634                  * For non-sack customers new data
7635                  * needs to go out as retransmits until
7636                  * we retransmit up to snd_max.
7637                  */
7638                 rack->r_must_retran = 1;
7639                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
7640                                                 rack->r_ctl.rc_sacked);
7641         }
7642         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
7643 }
7644
7645 static void
7646 rack_convert_rtts(struct tcpcb *tp)
7647 {
7648         tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
7649         tp->t_rxtcur = RACK_REXMTVAL(tp);
7650         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7651                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
7652         }
7653         if (tp->t_rxtcur > rack_rto_max) {
7654                 tp->t_rxtcur = rack_rto_max;
7655         }
7656 }
7657
7658 static void
7659 rack_cc_conn_init(struct tcpcb *tp)
7660 {
7661         struct tcp_rack *rack;
7662         uint32_t srtt;
7663
7664         rack = (struct tcp_rack *)tp->t_fb_ptr;
7665         srtt = tp->t_srtt;
7666         cc_conn_init(tp);
7667         /*
7668          * Now convert to rack's internal format,
7669          * if required.
7670          */
7671         if ((srtt == 0) && (tp->t_srtt != 0))
7672                 rack_convert_rtts(tp);
7673         /*
7674          * We want a chance to stay in slowstart as
7675          * we create a connection. TCP spec says that
7676          * initially ssthresh is infinite. For our
7677          * purposes that is the snd_wnd.
7678          */
7679         if (tp->snd_ssthresh < tp->snd_wnd) {
7680                 tp->snd_ssthresh = tp->snd_wnd;
7681         }
7682         /*
7683          * We also want to assure a IW worth of
7684          * data can get inflight.
7685          */
7686         if (rc_init_window(rack) < tp->snd_cwnd)
7687                 tp->snd_cwnd = rc_init_window(rack);
7688 }
7689
7690 /*
7691  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
7692  * we will setup to retransmit the lowest seq number outstanding.
7693  */
7694 static int
7695 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7696 {
7697         struct inpcb *inp = tptoinpcb(tp);
7698         int32_t rexmt;
7699         int32_t retval = 0;
7700         bool isipv6;
7701
7702         if ((tp->t_flags & TF_GPUTINPROG) &&
7703             (tp->t_rxtshift)) {
7704                 /*
7705                  * We have had a second timeout
7706                  * measurements on successive rxt's are not profitable.
7707                  * It is unlikely to be of any use (the network is
7708                  * broken or the client went away).
7709                  */
7710                 tp->t_flags &= ~TF_GPUTINPROG;
7711                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7712                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
7713                                            tp->gput_seq,
7714                                            0, 0, 18, __LINE__, NULL, 0);
7715         }
7716         if (ctf_progress_timeout_check(tp, false)) {
7717                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7718                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7719                 return (-ETIMEDOUT);    /* tcp_drop() */
7720         }
7721         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
7722         rack->r_ctl.retran_during_recovery = 0;
7723         rack->rc_ack_required = 1;
7724         rack->r_ctl.dsack_byte_cnt = 0;
7725         if (IN_FASTRECOVERY(tp->t_flags))
7726                 tp->t_flags |= TF_WASFRECOVERY;
7727         else
7728                 tp->t_flags &= ~TF_WASFRECOVERY;
7729         if (IN_CONGRECOVERY(tp->t_flags))
7730                 tp->t_flags |= TF_WASCRECOVERY;
7731         else
7732                 tp->t_flags &= ~TF_WASCRECOVERY;
7733         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7734             (tp->snd_una == tp->snd_max)) {
7735                 /* Nothing outstanding .. nothing to do */
7736                 return (0);
7737         }
7738         if (rack->r_ctl.dsack_persist) {
7739                 rack->r_ctl.dsack_persist--;
7740                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7741                         rack->r_ctl.num_dsack = 0;
7742                 }
7743                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7744         }
7745         /*
7746          * Rack can only run one timer  at a time, so we cannot
7747          * run a KEEPINIT (gating SYN sending) and a retransmit
7748          * timer for the SYN. So if we are in a front state and
7749          * have a KEEPINIT timer we need to check the first transmit
7750          * against now to see if we have exceeded the KEEPINIT time
7751          * (if one is set).
7752          */
7753         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
7754             (TP_KEEPINIT(tp) != 0)) {
7755                 struct rack_sendmap *rsm;
7756
7757                 rsm = tqhash_min(rack->r_ctl.tqh);
7758                 if (rsm) {
7759                         /* Ok we have something outstanding to test keepinit with */
7760                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
7761                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
7762                                 /* We have exceeded the KEEPINIT time */
7763                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7764                                 goto drop_it;
7765                         }
7766                 }
7767         }
7768         /*
7769          * Retransmission timer went off.  Message has not been acked within
7770          * retransmit interval.  Back off to a longer retransmit interval
7771          * and retransmit one segment.
7772          */
7773         rack_remxt_tmr(tp);
7774         if ((rack->r_ctl.rc_resend == NULL) ||
7775             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
7776                 /*
7777                  * If the rwnd collapsed on
7778                  * the one we are retransmitting
7779                  * it does not count against the
7780                  * rxt count.
7781                  */
7782                 tp->t_rxtshift++;
7783         }
7784         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
7785                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7786 drop_it:
7787                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
7788                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
7789                 /* XXXGL: previously t_softerror was casted to uint16_t */
7790                 MPASS(tp->t_softerror >= 0);
7791                 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
7792                 goto out;       /* tcp_drop() */
7793         }
7794         if (tp->t_state == TCPS_SYN_SENT) {
7795                 /*
7796                  * If the SYN was retransmitted, indicate CWND to be limited
7797                  * to 1 segment in cc_conn_init().
7798                  */
7799                 tp->snd_cwnd = 1;
7800         } else if (tp->t_rxtshift == 1) {
7801                 /*
7802                  * first retransmit; record ssthresh and cwnd so they can be
7803                  * recovered if this turns out to be a "bad" retransmit. A
7804                  * retransmit is considered "bad" if an ACK for this segment
7805                  * is received within RTT/2 interval; the assumption here is
7806                  * that the ACK was already in flight.  See "On Estimating
7807                  * End-to-End Network Path Properties" by Allman and Paxson
7808                  * for more details.
7809                  */
7810                 tp->snd_cwnd_prev = tp->snd_cwnd;
7811                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
7812                 tp->snd_recover_prev = tp->snd_recover;
7813                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
7814                 tp->t_flags |= TF_PREVVALID;
7815         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
7816                 tp->t_flags &= ~TF_PREVVALID;
7817         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
7818         if ((tp->t_state == TCPS_SYN_SENT) ||
7819             (tp->t_state == TCPS_SYN_RECEIVED))
7820                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
7821         else
7822                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
7823
7824         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
7825            max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
7826         /*
7827          * We enter the path for PLMTUD if connection is established or, if
7828          * connection is FIN_WAIT_1 status, reason for the last is that if
7829          * amount of data we send is very small, we could send it in couple
7830          * of packets and process straight to FIN. In that case we won't
7831          * catch ESTABLISHED state.
7832          */
7833 #ifdef INET6
7834         isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
7835 #else
7836         isipv6 = false;
7837 #endif
7838         if (((V_tcp_pmtud_blackhole_detect == 1) ||
7839             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
7840             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
7841             ((tp->t_state == TCPS_ESTABLISHED) ||
7842             (tp->t_state == TCPS_FIN_WAIT_1))) {
7843                 /*
7844                  * Idea here is that at each stage of mtu probe (usually,
7845                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
7846                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
7847                  * should take care of that.
7848                  */
7849                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
7850                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
7851                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
7852                     tp->t_rxtshift % 2 == 0)) {
7853                         /*
7854                          * Enter Path MTU Black-hole Detection mechanism: -
7855                          * Disable Path MTU Discovery (IP "DF" bit). -
7856                          * Reduce MTU to lower value than what we negotiated
7857                          * with peer.
7858                          */
7859                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7860                                 /* Record that we may have found a black hole. */
7861                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7862                                 /* Keep track of previous MSS. */
7863                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7864                         }
7865
7866                         /*
7867                          * Reduce the MSS to blackhole value or to the
7868                          * default in an attempt to retransmit.
7869                          */
7870 #ifdef INET6
7871                         if (isipv6 &&
7872                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7873                                 /* Use the sysctl tuneable blackhole MSS. */
7874                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7875                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7876                         } else if (isipv6) {
7877                                 /* Use the default MSS. */
7878                                 tp->t_maxseg = V_tcp_v6mssdflt;
7879                                 /*
7880                                  * Disable Path MTU Discovery when we switch
7881                                  * to minmss.
7882                                  */
7883                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7884                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7885                         }
7886 #endif
7887 #if defined(INET6) && defined(INET)
7888                         else
7889 #endif
7890 #ifdef INET
7891                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7892                                 /* Use the sysctl tuneable blackhole MSS. */
7893                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
7894                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7895                         } else {
7896                                 /* Use the default MSS. */
7897                                 tp->t_maxseg = V_tcp_mssdflt;
7898                                 /*
7899                                  * Disable Path MTU Discovery when we switch
7900                                  * to minmss.
7901                                  */
7902                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7903                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7904                         }
7905 #endif
7906                 } else {
7907                         /*
7908                          * If further retransmissions are still unsuccessful
7909                          * with a lowered MTU, maybe this isn't a blackhole
7910                          * and we restore the previous MSS and blackhole
7911                          * detection flags. The limit '6' is determined by
7912                          * giving each probe stage (1448, 1188, 524) 2
7913                          * chances to recover.
7914                          */
7915                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
7916                             (tp->t_rxtshift >= 6)) {
7917                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
7918                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
7919                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
7920                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
7921                         }
7922                 }
7923         }
7924         /*
7925          * Disable RFC1323 and SACK if we haven't got any response to
7926          * our third SYN to work-around some broken terminal servers
7927          * (most of which have hopefully been retired) that have bad VJ
7928          * header compression code which trashes TCP segments containing
7929          * unknown-to-them TCP options.
7930          */
7931         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
7932             (tp->t_rxtshift == 3))
7933                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
7934         /*
7935          * If we backed off this far, our srtt estimate is probably bogus.
7936          * Clobber it so we'll take the next rtt measurement as our srtt;
7937          * move the current srtt into rttvar to keep the current retransmit
7938          * times until then.
7939          */
7940         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
7941 #ifdef INET6
7942                 if ((inp->inp_vflag & INP_IPV6) != 0)
7943                         in6_losing(inp);
7944                 else
7945 #endif
7946                         in_losing(inp);
7947                 tp->t_rttvar += tp->t_srtt;
7948                 tp->t_srtt = 0;
7949         }
7950         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
7951         tp->snd_recover = tp->snd_max;
7952         tp->t_flags |= TF_ACKNOW;
7953         tp->t_rtttime = 0;
7954         rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
7955 out:
7956         return (retval);
7957 }
7958
7959 static int
7960 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
7961 {
7962         int32_t ret = 0;
7963         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
7964
7965         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7966             (tp->t_flags & TF_GPUTINPROG)) {
7967                 /*
7968                  * We have a goodput in progress
7969                  * and we have entered a late state.
7970                  * Do we have enough data in the sb
7971                  * to handle the GPUT request?
7972                  */
7973                 uint32_t bytes;
7974
7975                 bytes = tp->gput_ack - tp->gput_seq;
7976                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
7977                         bytes += tp->gput_seq - tp->snd_una;
7978                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
7979                         /*
7980                          * There are not enough bytes in the socket
7981                          * buffer that have been sent to cover this
7982                          * measurement. Cancel it.
7983                          */
7984                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7985                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
7986                                                    tp->gput_seq,
7987                                                    0, 0, 18, __LINE__, NULL, 0);
7988                         tp->t_flags &= ~TF_GPUTINPROG;
7989                 }
7990         }
7991         if (timers == 0) {
7992                 return (0);
7993         }
7994         if (tp->t_state == TCPS_LISTEN) {
7995                 /* no timers on listen sockets */
7996                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
7997                         return (0);
7998                 return (1);
7999         }
8000         if ((timers & PACE_TMR_RACK) &&
8001             rack->rc_on_min_to) {
8002                 /*
8003                  * For the rack timer when we
8004                  * are on a min-timeout (which means rrr_conf = 3)
8005                  * we don't want to check the timer. It may
8006                  * be going off for a pace and thats ok we
8007                  * want to send the retransmit (if its ready).
8008                  *
8009                  * If its on a normal rack timer (non-min) then
8010                  * we will check if its expired.
8011                  */
8012                 goto skip_time_check;
8013         }
8014         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
8015                 uint32_t left;
8016
8017                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
8018                         ret = -1;
8019                         rack_log_to_processing(rack, cts, ret, 0);
8020                         return (0);
8021                 }
8022                 if (hpts_calling == 0) {
8023                         /*
8024                          * A user send or queued mbuf (sack) has called us? We
8025                          * return 0 and let the pacing guards
8026                          * deal with it if they should or
8027                          * should not cause a send.
8028                          */
8029                         ret = -2;
8030                         rack_log_to_processing(rack, cts, ret, 0);
8031                         return (0);
8032                 }
8033                 /*
8034                  * Ok our timer went off early and we are not paced false
8035                  * alarm, go back to sleep. We make sure we don't have
8036                  * no-sack wakeup on since we no longer have a PKT_OUTPUT
8037                  * flag in place.
8038                  */
8039                 rack->rc_inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
8040                 ret = -3;
8041                 left = rack->r_ctl.rc_timer_exp - cts;
8042                 tcp_hpts_insert(tptoinpcb(tp), HPTS_MS_TO_SLOTS(left));
8043                 rack_log_to_processing(rack, cts, ret, left);
8044                 return (1);
8045         }
8046 skip_time_check:
8047         rack->rc_tmr_stopped = 0;
8048         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8049         if (timers & PACE_TMR_DELACK) {
8050                 ret = rack_timeout_delack(tp, rack, cts);
8051         } else if (timers & PACE_TMR_RACK) {
8052                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8053                 rack->r_fast_output = 0;
8054                 ret = rack_timeout_rack(tp, rack, cts);
8055         } else if (timers & PACE_TMR_TLP) {
8056                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8057                 ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8058         } else if (timers & PACE_TMR_RXT) {
8059                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8060                 rack->r_fast_output = 0;
8061                 ret = rack_timeout_rxt(tp, rack, cts);
8062         } else if (timers & PACE_TMR_PERSIT) {
8063                 ret = rack_timeout_persist(tp, rack, cts);
8064         } else if (timers & PACE_TMR_KEEP) {
8065                 ret = rack_timeout_keepalive(tp, rack, cts);
8066         }
8067         rack_log_to_processing(rack, cts, ret, timers);
8068         return (ret);
8069 }
8070
8071 static void
8072 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8073 {
8074         struct timeval tv;
8075         uint32_t us_cts, flags_on_entry;
8076         uint8_t hpts_removed = 0;
8077
8078         flags_on_entry = rack->r_ctl.rc_hpts_flags;
8079         us_cts = tcp_get_usecs(&tv);
8080         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8081             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8082              ((tp->snd_max - tp->snd_una) == 0))) {
8083                 tcp_hpts_remove(rack->rc_inp);
8084                 hpts_removed = 1;
8085                 /* If we were not delayed cancel out the flag. */
8086                 if ((tp->snd_max - tp->snd_una) == 0)
8087                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8088                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8089         }
8090         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8091                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8092                 if (tcp_in_hpts(rack->rc_inp) &&
8093                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8094                         /*
8095                          * Canceling timer's when we have no output being
8096                          * paced. We also must remove ourselves from the
8097                          * hpts.
8098                          */
8099                         tcp_hpts_remove(rack->rc_inp);
8100                         hpts_removed = 1;
8101                 }
8102                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8103         }
8104         if (hpts_removed == 0)
8105                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8106 }
8107
8108 static int
8109 rack_stopall(struct tcpcb *tp)
8110 {
8111         struct tcp_rack *rack;
8112         rack = (struct tcp_rack *)tp->t_fb_ptr;
8113         rack->t_timers_stopped = 1;
8114         return (0);
8115 }
8116
8117 static void
8118 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8119 {
8120         /*
8121          * Assure no timers are running.
8122          */
8123         if (tcp_timer_active(tp, TT_PERSIST)) {
8124                 /* We enter in persists, set the flag appropriately */
8125                 rack->rc_in_persist = 1;
8126         }
8127         if (tcp_in_hpts(rack->rc_inp)) {
8128                 tcp_hpts_remove(rack->rc_inp);
8129         }
8130 }
8131
8132 static void
8133 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8134     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag, int segsiz)
8135 {
8136         int32_t idx;
8137
8138         rsm->r_rtr_cnt++;
8139         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8140         rsm->r_dupack = 0;
8141         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8142                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8143                 rsm->r_flags |= RACK_OVERMAX;
8144         }
8145         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8146                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8147                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8148         }
8149         idx = rsm->r_rtr_cnt - 1;
8150         rsm->r_tim_lastsent[idx] = ts;
8151         /*
8152          * Here we don't add in the len of send, since its already
8153          * in snduna <->snd_max.
8154          */
8155         rsm->r_fas = ctf_flight_size(rack->rc_tp,
8156                                      rack->r_ctl.rc_sacked);
8157         if (rsm->r_flags & RACK_ACKED) {
8158                 /* Problably MTU discovery messing with us */
8159                 rsm->r_flags &= ~RACK_ACKED;
8160                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8161         }
8162         if (rsm->r_in_tmap) {
8163                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8164                 rsm->r_in_tmap = 0;
8165         }
8166         /* Lets make sure it really is in or not the GP window */
8167         rack_mark_in_gp_win(tp, rsm);
8168         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8169         rsm->r_in_tmap = 1;
8170         rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8171         /* Take off the must retransmit flag, if its on */
8172         if (rsm->r_flags & RACK_MUST_RXT) {
8173                 if (rack->r_must_retran)
8174                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8175                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8176                         /*
8177                          * We have retransmitted all we need. Clear
8178                          * any must retransmit flags.
8179                          */
8180                         rack->r_must_retran = 0;
8181                         rack->r_ctl.rc_out_at_rto = 0;
8182                 }
8183                 rsm->r_flags &= ~RACK_MUST_RXT;
8184         }
8185         /* Remove any collapsed flag */
8186         rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8187         if (rsm->r_flags & RACK_SACK_PASSED) {
8188                 /* We have retransmitted due to the SACK pass */
8189                 rsm->r_flags &= ~RACK_SACK_PASSED;
8190                 rsm->r_flags |= RACK_WAS_SACKPASS;
8191         }
8192 }
8193
8194 static uint32_t
8195 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8196     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag, int segsiz)
8197 {
8198         /*
8199          * We (re-)transmitted starting at rsm->r_start for some length
8200          * (possibly less than r_end.
8201          */
8202         struct rack_sendmap *nrsm;
8203         int insret __diagused;
8204         uint32_t c_end;
8205         int32_t len;
8206
8207         len = *lenp;
8208         c_end = rsm->r_start + len;
8209         if (SEQ_GEQ(c_end, rsm->r_end)) {
8210                 /*
8211                  * We retransmitted the whole piece or more than the whole
8212                  * slopping into the next rsm.
8213                  */
8214                 rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8215                 if (c_end == rsm->r_end) {
8216                         *lenp = 0;
8217                         return (0);
8218                 } else {
8219                         int32_t act_len;
8220
8221                         /* Hangs over the end return whats left */
8222                         act_len = rsm->r_end - rsm->r_start;
8223                         *lenp = (len - act_len);
8224                         return (rsm->r_end);
8225                 }
8226                 /* We don't get out of this block. */
8227         }
8228         /*
8229          * Here we retransmitted less than the whole thing which means we
8230          * have to split this into what was transmitted and what was not.
8231          */
8232         nrsm = rack_alloc_full_limit(rack);
8233         if (nrsm == NULL) {
8234                 /*
8235                  * We can't get memory, so lets not proceed.
8236                  */
8237                 *lenp = 0;
8238                 return (0);
8239         }
8240         /*
8241          * So here we are going to take the original rsm and make it what we
8242          * retransmitted. nrsm will be the tail portion we did not
8243          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8244          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8245          * 1, 6 and the new piece will be 6, 11.
8246          */
8247         rack_clone_rsm(rack, nrsm, rsm, c_end);
8248         nrsm->r_dupack = 0;
8249         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8250 #ifndef INVARIANTS
8251         (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8252 #else
8253         if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8254                 panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
8255                       nrsm, insret, rack, rsm);
8256         }
8257 #endif
8258         if (rsm->r_in_tmap) {
8259                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8260                 nrsm->r_in_tmap = 1;
8261         }
8262         rsm->r_flags &= (~RACK_HAS_FIN);
8263         rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8264         /* Log a split of rsm into rsm and nrsm */
8265         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8266         *lenp = 0;
8267         return (0);
8268 }
8269
8270 static void
8271 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8272                 uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8273                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb,
8274                 uint32_t s_moff, int hw_tls, int segsiz)
8275 {
8276         struct tcp_rack *rack;
8277         struct rack_sendmap *rsm, *nrsm;
8278         int insret __diagused;
8279
8280         register uint32_t snd_max, snd_una;
8281
8282         /*
8283          * Add to the RACK log of packets in flight or retransmitted. If
8284          * there is a TS option we will use the TS echoed, if not we will
8285          * grab a TS.
8286          *
8287          * Retransmissions will increment the count and move the ts to its
8288          * proper place. Note that if options do not include TS's then we
8289          * won't be able to effectively use the ACK for an RTT on a retran.
8290          *
8291          * Notes about r_start and r_end. Lets consider a send starting at
8292          * sequence 1 for 10 bytes. In such an example the r_start would be
8293          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8294          * This means that r_end is actually the first sequence for the next
8295          * slot (11).
8296          *
8297          */
8298         /*
8299          * If err is set what do we do XXXrrs? should we not add the thing?
8300          * -- i.e. return if err != 0 or should we pretend we sent it? --
8301          * i.e. proceed with add ** do this for now.
8302          */
8303         INP_WLOCK_ASSERT(tptoinpcb(tp));
8304         if (err)
8305                 /*
8306                  * We don't log errors -- we could but snd_max does not
8307                  * advance in this case either.
8308                  */
8309                 return;
8310
8311         if (th_flags & TH_RST) {
8312                 /*
8313                  * We don't log resets and we return immediately from
8314                  * sending
8315                  */
8316                 return;
8317         }
8318         rack = (struct tcp_rack *)tp->t_fb_ptr;
8319         snd_una = tp->snd_una;
8320         snd_max = tp->snd_max;
8321         if (th_flags & (TH_SYN | TH_FIN)) {
8322                 /*
8323                  * The call to rack_log_output is made before bumping
8324                  * snd_max. This means we can record one extra byte on a SYN
8325                  * or FIN if seq_out is adding more on and a FIN is present
8326                  * (and we are not resending).
8327                  */
8328                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
8329                         len++;
8330                 if (th_flags & TH_FIN)
8331                         len++;
8332                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
8333                         /*
8334                          * The add/update as not been done for the FIN/SYN
8335                          * yet.
8336                          */
8337                         snd_max = tp->snd_nxt;
8338                 }
8339         }
8340         if (SEQ_LEQ((seq_out + len), snd_una)) {
8341                 /* Are sending an old segment to induce an ack (keep-alive)? */
8342                 return;
8343         }
8344         if (SEQ_LT(seq_out, snd_una)) {
8345                 /* huh? should we panic? */
8346                 uint32_t end;
8347
8348                 end = seq_out + len;
8349                 seq_out = snd_una;
8350                 if (SEQ_GEQ(end, seq_out))
8351                         len = end - seq_out;
8352                 else
8353                         len = 0;
8354         }
8355         if (len == 0) {
8356                 /* We don't log zero window probes */
8357                 return;
8358         }
8359         if (IN_FASTRECOVERY(tp->t_flags)) {
8360                 rack->r_ctl.rc_prr_out += len;
8361         }
8362         /* First question is it a retransmission or new? */
8363         if (seq_out == snd_max) {
8364                 /* Its new */
8365                 rack_chk_http_and_hybrid_on_out(rack, seq_out, len, cts);
8366 again:
8367                 rsm = rack_alloc(rack);
8368                 if (rsm == NULL) {
8369                         /*
8370                          * Hmm out of memory and the tcb got destroyed while
8371                          * we tried to wait.
8372                          */
8373                         return;
8374                 }
8375                 if (th_flags & TH_FIN) {
8376                         rsm->r_flags = RACK_HAS_FIN|add_flag;
8377                 } else {
8378                         rsm->r_flags = add_flag;
8379                 }
8380                 if (hw_tls)
8381                         rsm->r_hw_tls = 1;
8382                 rsm->r_tim_lastsent[0] = cts;
8383                 rsm->r_rtr_cnt = 1;
8384                 rsm->r_rtr_bytes = 0;
8385                 if (th_flags & TH_SYN) {
8386                         /* The data space is one beyond snd_una */
8387                         rsm->r_flags |= RACK_HAS_SYN;
8388                 }
8389                 rsm->r_start = seq_out;
8390                 rsm->r_end = rsm->r_start + len;
8391                 rack_mark_in_gp_win(tp, rsm);
8392                 rsm->r_dupack = 0;
8393                 /*
8394                  * save off the mbuf location that
8395                  * sndmbuf_noadv returned (which is
8396                  * where we started copying from)..
8397                  */
8398                 rsm->m = s_mb;
8399                 rsm->soff = s_moff;
8400                 /*
8401                  * Here we do add in the len of send, since its not yet
8402                  * reflected in in snduna <->snd_max
8403                  */
8404                 rsm->r_fas = (ctf_flight_size(rack->rc_tp,
8405                                               rack->r_ctl.rc_sacked) +
8406                               (rsm->r_end - rsm->r_start));
8407                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
8408                 if (rsm->m) {
8409                         if (rsm->m->m_len <= rsm->soff) {
8410                                 /*
8411                                  * XXXrrs Question, will this happen?
8412                                  *
8413                                  * If sbsndptr is set at the correct place
8414                                  * then s_moff should always be somewhere
8415                                  * within rsm->m. But if the sbsndptr was
8416                                  * off then that won't be true. If it occurs
8417                                  * we need to walkout to the correct location.
8418                                  */
8419                                 struct mbuf *lm;
8420
8421                                 lm = rsm->m;
8422                                 while (lm->m_len <= rsm->soff) {
8423                                         rsm->soff -= lm->m_len;
8424                                         lm = lm->m_next;
8425                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
8426                                                              __func__, rack, s_moff, s_mb, rsm->soff));
8427                                 }
8428                                 rsm->m = lm;
8429                         }
8430                         rsm->orig_m_len = rsm->m->m_len;
8431                         rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
8432                 } else {
8433                         rsm->orig_m_len = 0;
8434                         rsm->orig_t_space = 0;
8435                 }
8436                 rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
8437                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8438                 /* Log a new rsm */
8439                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
8440 #ifndef INVARIANTS
8441                 (void)tqhash_insert(rack->r_ctl.tqh, rsm);
8442 #else
8443                 if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
8444                         panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
8445                               nrsm, insret, rack, rsm);
8446                 }
8447 #endif
8448                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8449                 rsm->r_in_tmap = 1;
8450                 /*
8451                  * Special case detection, is there just a single
8452                  * packet outstanding when we are not in recovery?
8453                  *
8454                  * If this is true mark it so.
8455                  */
8456                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
8457                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
8458                         struct rack_sendmap *prsm;
8459
8460                         prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
8461                         if (prsm)
8462                                 prsm->r_one_out_nr = 1;
8463                 }
8464                 return;
8465         }
8466         /*
8467          * If we reach here its a retransmission and we need to find it.
8468          */
8469 more:
8470         if (hintrsm && (hintrsm->r_start == seq_out)) {
8471                 rsm = hintrsm;
8472                 hintrsm = NULL;
8473         } else {
8474                 /* No hints sorry */
8475                 rsm = NULL;
8476         }
8477         if ((rsm) && (rsm->r_start == seq_out)) {
8478                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8479                 if (len == 0) {
8480                         return;
8481                 } else {
8482                         goto more;
8483                 }
8484         }
8485         /* Ok it was not the last pointer go through it the hard way. */
8486 refind:
8487         rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
8488         if (rsm) {
8489                 if (rsm->r_start == seq_out) {
8490                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8491                         if (len == 0) {
8492                                 return;
8493                         } else {
8494                                 goto refind;
8495                         }
8496                 }
8497                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
8498                         /* Transmitted within this piece */
8499                         /*
8500                          * Ok we must split off the front and then let the
8501                          * update do the rest
8502                          */
8503                         nrsm = rack_alloc_full_limit(rack);
8504                         if (nrsm == NULL) {
8505                                 rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
8506                                 return;
8507                         }
8508                         /*
8509                          * copy rsm to nrsm and then trim the front of rsm
8510                          * to not include this part.
8511                          */
8512                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
8513                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8514 #ifndef INVARIANTS
8515                         (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8516 #else
8517                         if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8518                                 panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
8519                                       nrsm, insret, rack, rsm);
8520                         }
8521 #endif
8522                         if (rsm->r_in_tmap) {
8523                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8524                                 nrsm->r_in_tmap = 1;
8525                         }
8526                         rsm->r_flags &= (~RACK_HAS_FIN);
8527                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
8528                         if (len == 0) {
8529                                 return;
8530                         } else if (len > 0)
8531                                 goto refind;
8532                 }
8533         }
8534         /*
8535          * Hmm not found in map did they retransmit both old and on into the
8536          * new?
8537          */
8538         if (seq_out == tp->snd_max) {
8539                 goto again;
8540         } else if (SEQ_LT(seq_out, tp->snd_max)) {
8541 #ifdef INVARIANTS
8542                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
8543                        seq_out, len, tp->snd_una, tp->snd_max);
8544                 printf("Starting Dump of all rack entries\n");
8545                 TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
8546                         printf("rsm:%p start:%u end:%u\n",
8547                                rsm, rsm->r_start, rsm->r_end);
8548                 }
8549                 printf("Dump complete\n");
8550                 panic("seq_out not found rack:%p tp:%p",
8551                       rack, tp);
8552 #endif
8553         } else {
8554 #ifdef INVARIANTS
8555                 /*
8556                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
8557                  * flag)
8558                  */
8559                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
8560                       seq_out, len, tp->snd_max, tp);
8561 #endif
8562         }
8563 }
8564
8565 /*
8566  * Record one of the RTT updates from an ack into
8567  * our sample structure.
8568  */
8569
8570 static void
8571 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
8572                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
8573 {
8574         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8575             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
8576                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
8577         }
8578         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8579             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
8580                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
8581         }
8582         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
8583             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
8584                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
8585             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
8586                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
8587         }
8588         if ((confidence == 1) &&
8589             ((rsm == NULL) ||
8590              (rsm->r_just_ret) ||
8591              (rsm->r_one_out_nr &&
8592               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
8593                 /*
8594                  * If the rsm had a just return
8595                  * hit it then we can't trust the
8596                  * rtt measurement for buffer deterimination
8597                  * Note that a confidence of 2, indicates
8598                  * SACK'd which overrides the r_just_ret or
8599                  * the r_one_out_nr. If it was a CUM-ACK and
8600                  * we had only two outstanding, but get an
8601                  * ack for only 1. Then that also lowers our
8602                  * confidence.
8603                  */
8604                 confidence = 0;
8605         }
8606         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8607             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
8608                 if (rack->r_ctl.rack_rs.confidence == 0) {
8609                         /*
8610                          * We take anything with no current confidence
8611                          * saved.
8612                          */
8613                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8614                         rack->r_ctl.rack_rs.confidence = confidence;
8615                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8616                 } else if (confidence != 0) {
8617                         /*
8618                          * Once we have a confident number,
8619                          * we can update it with a smaller
8620                          * value since this confident number
8621                          * may include the DSACK time until
8622                          * the next segment (the second one) arrived.
8623                          */
8624                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8625                         rack->r_ctl.rack_rs.confidence = confidence;
8626                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8627                 }
8628         }
8629         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
8630         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
8631         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
8632         rack->r_ctl.rack_rs.rs_rtt_cnt++;
8633 }
8634
8635 /*
8636  * Collect new round-trip time estimate
8637  * and update averages and current timeout.
8638  */
8639 static void
8640 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
8641 {
8642         int32_t delta;
8643         int32_t rtt;
8644
8645         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
8646                 /* No valid sample */
8647                 return;
8648         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
8649                 /* We are to use the lowest RTT seen in a single ack */
8650                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
8651         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
8652                 /* We are to use the highest RTT seen in a single ack */
8653                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
8654         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
8655                 /* We are to use the average RTT seen in a single ack */
8656                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
8657                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
8658         } else {
8659 #ifdef INVARIANTS
8660                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
8661 #endif
8662                 return;
8663         }
8664         if (rtt == 0)
8665                 rtt = 1;
8666         if (rack->rc_gp_rtt_set == 0) {
8667                 /*
8668                  * With no RTT we have to accept
8669                  * even one we are not confident of.
8670                  */
8671                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
8672                 rack->rc_gp_rtt_set = 1;
8673         } else if (rack->r_ctl.rack_rs.confidence) {
8674                 /* update the running gp srtt */
8675                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
8676                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
8677         }
8678         if (rack->r_ctl.rack_rs.confidence) {
8679                 /*
8680                  * record the low and high for highly buffered path computation,
8681                  * we only do this if we are confident (not a retransmission).
8682                  */
8683                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
8684                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8685                 }
8686                 if (rack->rc_highly_buffered == 0) {
8687                         /*
8688                          * Currently once we declare a path has
8689                          * highly buffered there is no going
8690                          * back, which may be a problem...
8691                          */
8692                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
8693                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
8694                                                      rack->r_ctl.rc_highest_us_rtt,
8695                                                      rack->r_ctl.rc_lowest_us_rtt,
8696                                                      RACK_RTTS_SEEHBP);
8697                                 rack->rc_highly_buffered = 1;
8698                         }
8699                 }
8700         }
8701         if ((rack->r_ctl.rack_rs.confidence) ||
8702             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
8703                 /*
8704                  * If we are highly confident of it <or> it was
8705                  * never retransmitted we accept it as the last us_rtt.
8706                  */
8707                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8708                 /* The lowest rtt can be set if its was not retransmited */
8709                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
8710                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8711                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
8712                                 rack->r_ctl.rc_lowest_us_rtt = 1;
8713                 }
8714         }
8715         rack = (struct tcp_rack *)tp->t_fb_ptr;
8716         if (tp->t_srtt != 0) {
8717                 /*
8718                  * We keep a simple srtt in microseconds, like our rtt
8719                  * measurement. We don't need to do any tricks with shifting
8720                  * etc. Instead we just add in 1/8th of the new measurement
8721                  * and subtract out 1/8 of the old srtt. We do the same with
8722                  * the variance after finding the absolute value of the
8723                  * difference between this sample and the current srtt.
8724                  */
8725                 delta = tp->t_srtt - rtt;
8726                 /* Take off 1/8th of the current sRTT */
8727                 tp->t_srtt -= (tp->t_srtt >> 3);
8728                 /* Add in 1/8th of the new RTT just measured */
8729                 tp->t_srtt += (rtt >> 3);
8730                 if (tp->t_srtt <= 0)
8731                         tp->t_srtt = 1;
8732                 /* Now lets make the absolute value of the variance */
8733                 if (delta < 0)
8734                         delta = -delta;
8735                 /* Subtract out 1/8th */
8736                 tp->t_rttvar -= (tp->t_rttvar >> 3);
8737                 /* Add in 1/8th of the new variance we just saw */
8738                 tp->t_rttvar += (delta >> 3);
8739                 if (tp->t_rttvar <= 0)
8740                         tp->t_rttvar = 1;
8741         } else {
8742                 /*
8743                  * No rtt measurement yet - use the unsmoothed rtt. Set the
8744                  * variance to half the rtt (so our first retransmit happens
8745                  * at 3*rtt).
8746                  */
8747                 tp->t_srtt = rtt;
8748                 tp->t_rttvar = rtt >> 1;
8749         }
8750         rack->rc_srtt_measure_made = 1;
8751         KMOD_TCPSTAT_INC(tcps_rttupdated);
8752         if (tp->t_rttupdated < UCHAR_MAX)
8753                 tp->t_rttupdated++;
8754 #ifdef STATS
8755         if (rack_stats_gets_ms_rtt == 0) {
8756                 /* Send in the microsecond rtt used for rxt timeout purposes */
8757                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
8758         } else if (rack_stats_gets_ms_rtt == 1) {
8759                 /* Send in the millisecond rtt used for rxt timeout purposes */
8760                 int32_t ms_rtt;
8761
8762                 /* Round up */
8763                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8764                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8765         } else if (rack_stats_gets_ms_rtt == 2) {
8766                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
8767                 int32_t ms_rtt;
8768
8769                 /* Round up */
8770                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8771                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8772         }  else {
8773                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
8774                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8775         }
8776         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8777 #endif
8778         /*
8779          * the retransmit should happen at rtt + 4 * rttvar. Because of the
8780          * way we do the smoothing, srtt and rttvar will each average +1/2
8781          * tick of bias.  When we compute the retransmit timer, we want 1/2
8782          * tick of rounding and 1 extra tick because of +-1/2 tick
8783          * uncertainty in the firing of the timer.  The bias will give us
8784          * exactly the 1.5 tick we need.  But, because the bias is
8785          * statistical, we have to test that we don't drop below the minimum
8786          * feasible timer (which is 2 ticks).
8787          */
8788         tp->t_rxtshift = 0;
8789         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8790                       max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
8791         rack_log_rtt_sample(rack, rtt);
8792         tp->t_softerror = 0;
8793 }
8794
8795
8796 static void
8797 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
8798 {
8799         /*
8800          * Apply to filter the inbound us-rtt at us_cts.
8801          */
8802         uint32_t old_rtt;
8803
8804         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
8805         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
8806                                us_rtt, us_cts);
8807         if (old_rtt > us_rtt) {
8808                 /* We just hit a new lower rtt time */
8809                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
8810                                      __LINE__, RACK_RTTS_NEWRTT);
8811                 /*
8812                  * Only count it if its lower than what we saw within our
8813                  * calculated range.
8814                  */
8815                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
8816                         if (rack_probertt_lower_within &&
8817                             rack->rc_gp_dyn_mul &&
8818                             (rack->use_fixed_rate == 0) &&
8819                             (rack->rc_always_pace)) {
8820                                 /*
8821                                  * We are seeing a new lower rtt very close
8822                                  * to the time that we would have entered probe-rtt.
8823                                  * This is probably due to the fact that a peer flow
8824                                  * has entered probe-rtt. Lets go in now too.
8825                                  */
8826                                 uint32_t val;
8827
8828                                 val = rack_probertt_lower_within * rack_time_between_probertt;
8829                                 val /= 100;
8830                                 if ((rack->in_probe_rtt == 0)  &&
8831                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
8832                                         rack_enter_probertt(rack, us_cts);
8833                                 }
8834                         }
8835                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
8836                 }
8837         }
8838 }
8839
8840 static int
8841 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
8842     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
8843 {
8844         uint32_t us_rtt;
8845         int32_t i, all;
8846         uint32_t t, len_acked;
8847
8848         if ((rsm->r_flags & RACK_ACKED) ||
8849             (rsm->r_flags & RACK_WAS_ACKED))
8850                 /* Already done */
8851                 return (0);
8852         if (rsm->r_no_rtt_allowed) {
8853                 /* Not allowed */
8854                 return (0);
8855         }
8856         if (ack_type == CUM_ACKED) {
8857                 if (SEQ_GT(th_ack, rsm->r_end)) {
8858                         len_acked = rsm->r_end - rsm->r_start;
8859                         all = 1;
8860                 } else {
8861                         len_acked = th_ack - rsm->r_start;
8862                         all = 0;
8863                 }
8864         } else {
8865                 len_acked = rsm->r_end - rsm->r_start;
8866                 all = 0;
8867         }
8868         if (rsm->r_rtr_cnt == 1) {
8869
8870                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8871                 if ((int)t <= 0)
8872                         t = 1;
8873                 if (!tp->t_rttlow || tp->t_rttlow > t)
8874                         tp->t_rttlow = t;
8875                 if (!rack->r_ctl.rc_rack_min_rtt ||
8876                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8877                         rack->r_ctl.rc_rack_min_rtt = t;
8878                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
8879                                 rack->r_ctl.rc_rack_min_rtt = 1;
8880                         }
8881                 }
8882                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8883                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8884                 else
8885                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8886                 if (us_rtt == 0)
8887                         us_rtt = 1;
8888                 if (CC_ALGO(tp)->rttsample != NULL) {
8889                         /* Kick the RTT to the CC */
8890                         CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8891                 }
8892                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
8893                 if (ack_type == SACKED) {
8894                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
8895                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
8896                 } else {
8897                         /*
8898                          * We need to setup what our confidence
8899                          * is in this ack.
8900                          *
8901                          * If the rsm was app limited and it is
8902                          * less than a mss in length (the end
8903                          * of the send) then we have a gap. If we
8904                          * were app limited but say we were sending
8905                          * multiple MSS's then we are more confident
8906                          * int it.
8907                          *
8908                          * When we are not app-limited then we see if
8909                          * the rsm is being included in the current
8910                          * measurement, we tell this by the app_limited_needs_set
8911                          * flag.
8912                          *
8913                          * Note that being cwnd blocked is not applimited
8914                          * as well as the pacing delay between packets which
8915                          * are sending only 1 or 2 MSS's also will show up
8916                          * in the RTT. We probably need to examine this algorithm
8917                          * a bit more and enhance it to account for the delay
8918                          * between rsm's. We could do that by saving off the
8919                          * pacing delay of each rsm (in an rsm) and then
8920                          * factoring that in somehow though for now I am
8921                          * not sure how :)
8922                          */
8923                         int calc_conf = 0;
8924
8925                         if (rsm->r_flags & RACK_APP_LIMITED) {
8926                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
8927                                         calc_conf = 0;
8928                                 else
8929                                         calc_conf = 1;
8930                         } else if (rack->app_limited_needs_set == 0) {
8931                                 calc_conf = 1;
8932                         } else {
8933                                 calc_conf = 0;
8934                         }
8935                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
8936                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
8937                                             calc_conf, rsm, rsm->r_rtr_cnt);
8938                 }
8939                 if ((rsm->r_flags & RACK_TLP) &&
8940                     (!IN_FASTRECOVERY(tp->t_flags))) {
8941                         /* Segment was a TLP and our retrans matched */
8942                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
8943                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
8944                         }
8945                 }
8946                 if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
8947                     (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
8948                             (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
8949                         /* New more recent rack_tmit_time */
8950                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8951                         if (rack->r_ctl.rc_rack_tmit_time == 0)
8952                                 rack->r_ctl.rc_rack_tmit_time = 1;
8953                         rack->rc_rack_rtt = t;
8954                 }
8955                 return (1);
8956         }
8957         /*
8958          * We clear the soft/rxtshift since we got an ack.
8959          * There is no assurance we will call the commit() function
8960          * so we need to clear these to avoid incorrect handling.
8961          */
8962         tp->t_rxtshift = 0;
8963         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8964                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
8965         tp->t_softerror = 0;
8966         if (to && (to->to_flags & TOF_TS) &&
8967             (ack_type == CUM_ACKED) &&
8968             (to->to_tsecr) &&
8969             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
8970                 /*
8971                  * Now which timestamp does it match? In this block the ACK
8972                  * must be coming from a previous transmission.
8973                  */
8974                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
8975                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
8976                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8977                                 if ((int)t <= 0)
8978                                         t = 1;
8979                                 if (CC_ALGO(tp)->rttsample != NULL) {
8980                                         /*
8981                                          * Kick the RTT to the CC, here
8982                                          * we lie a bit in that we know the
8983                                          * retransmission is correct even though
8984                                          * we retransmitted. This is because
8985                                          * we match the timestamps.
8986                                          */
8987                                         if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
8988                                                 us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
8989                                         else
8990                                                 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
8991                                         CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8992                                 }
8993                                 if ((i + 1) < rsm->r_rtr_cnt) {
8994                                         /*
8995                                          * The peer ack'd from our previous
8996                                          * transmission. We have a spurious
8997                                          * retransmission and thus we dont
8998                                          * want to update our rack_rtt.
8999                                          *
9000                                          * Hmm should there be a CC revert here?
9001                                          *
9002                                          */
9003                                         return (0);
9004                                 }
9005                                 if (!tp->t_rttlow || tp->t_rttlow > t)
9006                                         tp->t_rttlow = t;
9007                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9008                                         rack->r_ctl.rc_rack_min_rtt = t;
9009                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
9010                                                 rack->r_ctl.rc_rack_min_rtt = 1;
9011                                         }
9012                                 }
9013                                 if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9014                                     (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9015                                             (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9016                                         /* New more recent rack_tmit_time */
9017                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9018                                         if (rack->r_ctl.rc_rack_tmit_time == 0)
9019                                                 rack->r_ctl.rc_rack_tmit_time = 1;
9020                                         rack->rc_rack_rtt = t;
9021                                 }
9022                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9023                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9024                                                     rsm->r_rtr_cnt);
9025                                 return (1);
9026                         }
9027                 }
9028                 /* If we are logging log out the sendmap */
9029                 if (tcp_bblogging_on(rack->rc_tp)) {
9030                         for (i = 0; i < rsm->r_rtr_cnt; i++) {
9031                                 rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9032                         }
9033                 }
9034                 goto ts_not_found;
9035         } else {
9036                 /*
9037                  * Ok its a SACK block that we retransmitted. or a windows
9038                  * machine without timestamps. We can tell nothing from the
9039                  * time-stamp since its not there or the time the peer last
9040                  * recieved a segment that moved forward its cum-ack point.
9041                  */
9042 ts_not_found:
9043                 i = rsm->r_rtr_cnt - 1;
9044                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9045                 if ((int)t <= 0)
9046                         t = 1;
9047                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9048                         /*
9049                          * We retransmitted and the ack came back in less
9050                          * than the smallest rtt we have observed. We most
9051                          * likely did an improper retransmit as outlined in
9052                          * 6.2 Step 2 point 2 in the rack-draft so we
9053                          * don't want to update our rack_rtt. We in
9054                          * theory (in future) might want to think about reverting our
9055                          * cwnd state but we won't for now.
9056                          */
9057                         return (0);
9058                 } else if (rack->r_ctl.rc_rack_min_rtt) {
9059                         /*
9060                          * We retransmitted it and the retransmit did the
9061                          * job.
9062                          */
9063                         if (!rack->r_ctl.rc_rack_min_rtt ||
9064                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9065                                 rack->r_ctl.rc_rack_min_rtt = t;
9066                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
9067                                         rack->r_ctl.rc_rack_min_rtt = 1;
9068                                 }
9069                         }
9070                         if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9071                             (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9072                                     (uint32_t)rsm->r_tim_lastsent[i]))) {
9073                                 /* New more recent rack_tmit_time */
9074                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9075                                 if (rack->r_ctl.rc_rack_tmit_time == 0)
9076                                         rack->r_ctl.rc_rack_tmit_time = 1;
9077                                 rack->rc_rack_rtt = t;
9078                         }
9079                         return (1);
9080                 }
9081         }
9082         return (0);
9083 }
9084
9085 /*
9086  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9087  */
9088 static void
9089 rack_log_sack_passed(struct tcpcb *tp,
9090     struct tcp_rack *rack, struct rack_sendmap *rsm)
9091 {
9092         struct rack_sendmap *nrsm;
9093
9094         nrsm = rsm;
9095         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9096             rack_head, r_tnext) {
9097                 if (nrsm == rsm) {
9098                         /* Skip original segment he is acked */
9099                         continue;
9100                 }
9101                 if (nrsm->r_flags & RACK_ACKED) {
9102                         /*
9103                          * Skip ack'd segments, though we
9104                          * should not see these, since tmap
9105                          * should not have ack'd segments.
9106                          */
9107                         continue;
9108                 }
9109                 if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9110                         /*
9111                          * If the peer dropped the rwnd on
9112                          * these then we don't worry about them.
9113                          */
9114                         continue;
9115                 }
9116                 if (nrsm->r_flags & RACK_SACK_PASSED) {
9117                         /*
9118                          * We found one that is already marked
9119                          * passed, we have been here before and
9120                          * so all others below this are marked.
9121                          */
9122                         break;
9123                 }
9124                 nrsm->r_flags |= RACK_SACK_PASSED;
9125                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9126         }
9127 }
9128
9129 static void
9130 rack_need_set_test(struct tcpcb *tp,
9131                    struct tcp_rack *rack,
9132                    struct rack_sendmap *rsm,
9133                    tcp_seq th_ack,
9134                    int line,
9135                    int use_which)
9136 {
9137         struct rack_sendmap *s_rsm;
9138
9139         if ((tp->t_flags & TF_GPUTINPROG) &&
9140             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9141                 /*
9142                  * We were app limited, and this ack
9143                  * butts up or goes beyond the point where we want
9144                  * to start our next measurement. We need
9145                  * to record the new gput_ts as here and
9146                  * possibly update the start sequence.
9147                  */
9148                 uint32_t seq, ts;
9149
9150                 if (rsm->r_rtr_cnt > 1) {
9151                         /*
9152                          * This is a retransmit, can we
9153                          * really make any assessment at this
9154                          * point?  We are not really sure of
9155                          * the timestamp, is it this or the
9156                          * previous transmission?
9157                          *
9158                          * Lets wait for something better that
9159                          * is not retransmitted.
9160                          */
9161                         return;
9162                 }
9163                 seq = tp->gput_seq;
9164                 ts = tp->gput_ts;
9165                 rack->app_limited_needs_set = 0;
9166                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9167                 /* Do we start at a new end? */
9168                 if ((use_which == RACK_USE_BEG) &&
9169                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9170                         /*
9171                          * When we get an ACK that just eats
9172                          * up some of the rsm, we set RACK_USE_BEG
9173                          * since whats at r_start (i.e. th_ack)
9174                          * is left unacked and thats where the
9175                          * measurement now starts.
9176                          */
9177                         tp->gput_seq = rsm->r_start;
9178                 }
9179                 if ((use_which == RACK_USE_END) &&
9180                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9181                         /*
9182                          * We use the end when the cumack
9183                          * is moving forward and completely
9184                          * deleting the rsm passed so basically
9185                          * r_end holds th_ack.
9186                          *
9187                          * For SACK's we also want to use the end
9188                          * since this piece just got sacked and
9189                          * we want to target anything after that
9190                          * in our measurement.
9191                          */
9192                         tp->gput_seq = rsm->r_end;
9193                 }
9194                 if (use_which == RACK_USE_END_OR_THACK) {
9195                         /*
9196                          * special case for ack moving forward,
9197                          * not a sack, we need to move all the
9198                          * way up to where this ack cum-ack moves
9199                          * to.
9200                          */
9201                         if (SEQ_GT(th_ack, rsm->r_end))
9202                                 tp->gput_seq = th_ack;
9203                         else
9204                                 tp->gput_seq = rsm->r_end;
9205                 }
9206                 if (SEQ_LT(tp->gput_seq, tp->snd_max))
9207                         s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9208                 else
9209                         s_rsm = NULL;
9210                 /*
9211                  * Pick up the correct send time if we can the rsm passed in
9212                  * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9213                  * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9214                  * find a different seq i.e. the next send up.
9215                  *
9216                  * If that has not been sent, s_rsm will be NULL and we must
9217                  * arrange it so this function will get called again by setting
9218                  * app_limited_needs_set.
9219                  */
9220                 if (s_rsm)
9221                         rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9222                 else {
9223                         /* If we hit here we have to have *not* sent tp->gput_seq */
9224                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9225                         /* Set it up so we will go through here again */
9226                         rack->app_limited_needs_set = 1;
9227                 }
9228                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9229                         /*
9230                          * We moved beyond this guy's range, re-calculate
9231                          * the new end point.
9232                          */
9233                         if (rack->rc_gp_filled == 0) {
9234                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9235                         } else {
9236                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9237                         }
9238                 }
9239                 /*
9240                  * We are moving the goal post, we may be able to clear the
9241                  * measure_saw_probe_rtt flag.
9242                  */
9243                 if ((rack->in_probe_rtt == 0) &&
9244                     (rack->measure_saw_probe_rtt) &&
9245                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9246                         rack->measure_saw_probe_rtt = 0;
9247                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9248                                            seq, tp->gput_seq,
9249                                            (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9250                                             (uint64_t)rack->r_ctl.rc_gp_output_ts),
9251                                            5, line, NULL, 0);
9252                 if (rack->rc_gp_filled &&
9253                     ((tp->gput_ack - tp->gput_seq) <
9254                      max(rc_init_window(rack), (MIN_GP_WIN *
9255                                                 ctf_fixed_maxseg(tp))))) {
9256                         uint32_t ideal_amount;
9257
9258                         ideal_amount = rack_get_measure_window(tp, rack);
9259                         if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9260                                 /*
9261                                  * There is no sense of continuing this measurement
9262                                  * because its too small to gain us anything we
9263                                  * trust. Skip it and that way we can start a new
9264                                  * measurement quicker.
9265                                  */
9266                                 tp->t_flags &= ~TF_GPUTINPROG;
9267                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9268                                                            0, 0,
9269                                                            (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9270                                                             (uint64_t)rack->r_ctl.rc_gp_output_ts),
9271                                                            6, __LINE__, NULL, 0);
9272                         } else {
9273                                 /*
9274                                  * Reset the window further out.
9275                                  */
9276                                 tp->gput_ack = tp->gput_seq + ideal_amount;
9277                         }
9278                 }
9279                 rack_tend_gp_marks(tp, rack);
9280                 rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9281         }
9282 }
9283
9284 static inline int
9285 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9286 {
9287         if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
9288                 /* Behind our TLP definition or right at */
9289                 return (0);
9290         }
9291         if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
9292                 /* The start is beyond or right at our end of TLP definition */
9293                 return (0);
9294         }
9295         /* It has to be a sub-part of the original TLP recorded */
9296         return (1);
9297 }
9298
9299
9300
9301 static uint32_t
9302 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
9303                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
9304                    int *no_extra,
9305                    int *moved_two, uint32_t segsiz)
9306 {
9307         uint32_t start, end, changed = 0;
9308         struct rack_sendmap stack_map;
9309         struct rack_sendmap *rsm, *nrsm, *prev, *next;
9310         int insret __diagused;
9311         int32_t used_ref = 1;
9312         int moved = 0;
9313 #ifdef TCP_SAD_DETECTION
9314         int allow_segsiz;
9315         int first_time_through = 1;
9316 #endif
9317         int noextra = 0;
9318         int can_use_hookery = 0;
9319
9320         start = sack->start;
9321         end = sack->end;
9322         rsm = *prsm;
9323
9324 #ifdef TCP_SAD_DETECTION
9325         /*
9326          * There are a strange number of proxys and meddle boxes in the world
9327          * that seem to cut up segments on different boundaries. This gets us
9328          * smaller sacks that are still ok in terms of it being an attacker.
9329          * We use the base segsiz to calculate an allowable smallness but
9330          * also enforce a min on the segsiz in case it is an attacker playing
9331          * games with MSS. So basically if the sack arrives and it is
9332          * larger than a worse case 960 bytes, we don't classify the guy
9333          * as supicious.
9334          */
9335         allow_segsiz = max(segsiz, 1200) * sad_seg_size_per;
9336         allow_segsiz /= 1000;
9337 #endif
9338 do_rest_ofb:
9339         if ((rsm == NULL) ||
9340             (SEQ_LT(end, rsm->r_start)) ||
9341             (SEQ_GEQ(start, rsm->r_end)) ||
9342             (SEQ_LT(start, rsm->r_start))) {
9343                 /*
9344                  * We are not in the right spot,
9345                  * find the correct spot in the tree.
9346                  */
9347                 used_ref = 0;
9348                 rsm = tqhash_find(rack->r_ctl.tqh, start);
9349                 moved++;
9350         }
9351         if (rsm == NULL) {
9352                 /* TSNH */
9353                 goto out;
9354         }
9355 #ifdef TCP_SAD_DETECTION
9356         /* Now we must check for suspicous activity */
9357         if ((first_time_through == 1) &&
9358             ((end - start) < min((rsm->r_end - rsm->r_start), allow_segsiz)) &&
9359             ((rsm->r_flags & RACK_PMTU_CHG) == 0) &&
9360             ((rsm->r_flags & RACK_TLP) == 0)) {
9361                 /*
9362                  * Its less than a full MSS or the segment being acked
9363                  * this should only happen if the rsm in question had the
9364                  * r_just_ret flag set <and> the end matches the end of
9365                  * the rsm block.
9366                  *
9367                  * Note we do not look at segments that have had TLP's on
9368                  * them since we can get un-reported rwnd collapses that
9369                  * basically we TLP on and then we get back a sack block
9370                  * that goes from the start to only a small way.
9371                  *
9372                  */
9373                 int loss, ok;
9374
9375                 ok = 0;
9376                 if (SEQ_GEQ(end, rsm->r_end)) {
9377                         if (rsm->r_just_ret == 1) {
9378                                 /* This was at the end of a send which is ok */
9379                                 ok = 1;
9380                         } else {
9381                                 /* A bit harder was it the end of our segment */
9382                                 int segs, len;
9383
9384                                 len = (rsm->r_end - rsm->r_start);
9385                                 segs = len / segsiz;
9386                                 segs *= segsiz;
9387                                 if ((segs + (rsm->r_end - start)) == len) {
9388                                         /*
9389                                          * So this last bit was the
9390                                          * end of our send if we cut it
9391                                          * up into segsiz pieces so its ok.
9392                                          */
9393                                         ok = 1;
9394                                 }
9395                         }
9396                 }
9397                 if (ok == 0) {
9398                         /*
9399                          * This guy is doing something suspicious
9400                          * lets start detection.
9401                          */
9402                         if (rack->rc_suspicious == 0) {
9403                                 tcp_trace_point(rack->rc_tp, TCP_TP_SAD_SUSPECT);
9404                                 counter_u64_add(rack_sack_attacks_suspect, 1);
9405                                 rack->rc_suspicious = 1;
9406                                 rack_log_sad(rack, 4);
9407                                 if (tcp_bblogging_on(rack->rc_tp)) {
9408                                         union tcp_log_stackspecific log;
9409                                         struct timeval tv;
9410
9411                                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
9412                                         log.u_bbr.flex1 = end;
9413                                         log.u_bbr.flex2 = start;
9414                                         log.u_bbr.flex3 = rsm->r_end;
9415                                         log.u_bbr.flex4 = rsm->r_start;
9416                                         log.u_bbr.flex5 = segsiz;
9417                                         log.u_bbr.flex6 = rsm->r_fas;
9418                                         log.u_bbr.flex7 = rsm->r_bas;
9419                                         log.u_bbr.flex8 = 5;
9420                                         log.u_bbr.pkts_out = rsm->r_flags;
9421                                         log.u_bbr.bbr_state = rack->rc_suspicious;
9422                                         log.u_bbr.bbr_substate = rsm->r_just_ret;
9423                                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
9424                                         log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9425                                         TCP_LOG_EVENTP(rack->rc_tp, NULL,
9426                                                        &rack->rc_inp->inp_socket->so_rcv,
9427                                                        &rack->rc_inp->inp_socket->so_snd,
9428                                                        TCP_SAD_DETECTION, 0,
9429                                                        0, &log, false, &tv);
9430                                 }
9431                         }
9432                         /* You loose some ack count every time you sack
9433                          * a small bit that is not butting to the end of
9434                          * what we have sent. This is because we never
9435                          * send small bits unless its the end of the sb.
9436                          * Anyone sending a sack that is not at the end
9437                          * is thus very very suspicious.
9438                          */
9439                         loss = (segsiz/2) / (end - start);
9440                         if (loss < rack->r_ctl.ack_count)
9441                                 rack->r_ctl.ack_count -= loss;
9442                         else
9443                                 rack->r_ctl.ack_count = 0;
9444                 }
9445         }
9446         first_time_through = 0;
9447 #endif
9448         /* Ok we have an ACK for some piece of this rsm */
9449         if (rsm->r_start != start) {
9450                 if ((rsm->r_flags & RACK_ACKED) == 0) {
9451                         /*
9452                          * Before any splitting or hookery is
9453                          * done is it a TLP of interest i.e. rxt?
9454                          */
9455                         if ((rsm->r_flags & RACK_TLP) &&
9456                             (rsm->r_rtr_cnt > 1)) {
9457                                 /*
9458                                  * We are splitting a rxt TLP, check
9459                                  * if we need to save off the start/end
9460                                  */
9461                                 if (rack->rc_last_tlp_acked_set &&
9462                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9463                                         /*
9464                                          * We already turned this on since we are inside
9465                                          * the previous one was a partially sack now we
9466                                          * are getting another one (maybe all of it).
9467                                          *
9468                                          */
9469                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9470                                         /*
9471                                          * Lets make sure we have all of it though.
9472                                          */
9473                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9474                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9475                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9476                                                                      rack->r_ctl.last_tlp_acked_end);
9477                                         }
9478                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9479                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9480                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9481                                                                      rack->r_ctl.last_tlp_acked_end);
9482                                         }
9483                                 } else {
9484                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9485                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9486                                         rack->rc_last_tlp_past_cumack = 0;
9487                                         rack->rc_last_tlp_acked_set = 1;
9488                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9489                                 }
9490                         }
9491                         /**
9492                          * Need to split this in two pieces the before and after,
9493                          * the before remains in the map, the after must be
9494                          * added. In other words we have:
9495                          * rsm        |--------------|
9496                          * sackblk        |------->
9497                          * rsm will become
9498                          *     rsm    |---|
9499                          * and nrsm will be  the sacked piece
9500                          *     nrsm       |----------|
9501                          *
9502                          * But before we start down that path lets
9503                          * see if the sack spans over on top of
9504                          * the next guy and it is already sacked.
9505                          *
9506                          */
9507                         /*
9508                          * Hookery can only be used if the two entries
9509                          * are in the same bucket and neither one of
9510                          * them staddle the bucket line.
9511                          */
9512                         next = tqhash_next(rack->r_ctl.tqh, rsm);
9513                         if (next &&
9514                             (rsm->bindex == next->bindex) &&
9515                             ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9516                             ((next->r_flags & RACK_STRADDLE) == 0) &&
9517                             (rsm->r_flags & RACK_IN_GP_WIN) &&
9518                             (next->r_flags & RACK_IN_GP_WIN))
9519                                 can_use_hookery = 1;
9520                         else if (next &&
9521                                  (rsm->bindex == next->bindex) &&
9522                                  ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9523                                  ((next->r_flags & RACK_STRADDLE) == 0) &&
9524                                  ((rsm->r_flags & RACK_IN_GP_WIN) == 0) &&
9525                                  ((next->r_flags & RACK_IN_GP_WIN) == 0))
9526                                 can_use_hookery = 1;
9527                         else
9528                                 can_use_hookery = 0;
9529                         if (next && can_use_hookery &&
9530                             (next->r_flags & RACK_ACKED) &&
9531                             SEQ_GEQ(end, next->r_start)) {
9532                                 /**
9533                                  * So the next one is already acked, and
9534                                  * we can thus by hookery use our stack_map
9535                                  * to reflect the piece being sacked and
9536                                  * then adjust the two tree entries moving
9537                                  * the start and ends around. So we start like:
9538                                  *  rsm     |------------|             (not-acked)
9539                                  *  next                 |-----------| (acked)
9540                                  *  sackblk        |-------->
9541                                  *  We want to end like so:
9542                                  *  rsm     |------|                   (not-acked)
9543                                  *  next           |-----------------| (acked)
9544                                  *  nrsm           |-----|
9545                                  * Where nrsm is a temporary stack piece we
9546                                  * use to update all the gizmos.
9547                                  */
9548                                 /* Copy up our fudge block */
9549                                 noextra++;
9550                                 nrsm = &stack_map;
9551                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9552                                 /* Now adjust our tree blocks */
9553                                 rsm->r_end = start;
9554                                 next->r_start = start;
9555                                 rsm->r_flags |= RACK_SHUFFLED;
9556                                 next->r_flags |= RACK_SHUFFLED;
9557                                 /* Now we must adjust back where next->m is */
9558                                 rack_setup_offset_for_rsm(rack, rsm, next);
9559                                 /*
9560                                  * Which timestamp do we keep? It is rather
9561                                  * important in GP measurements to have the
9562                                  * accurate end of the send window.
9563                                  *
9564                                  * We keep the largest value, which is the newest
9565                                  * send. We do this in case a segment that is
9566                                  * joined together and not part of a GP estimate
9567                                  * later gets expanded into the GP estimate.
9568                                  *
9569                                  * We prohibit the merging of unlike kinds i.e.
9570                                  * all pieces that are in the GP estimate can be
9571                                  * merged and all pieces that are not in a GP estimate
9572                                  * can be merged, but not disimilar pieces. Combine
9573                                  * this with taking the highest here and we should
9574                                  * be ok unless of course the client reneges. Then
9575                                  * all bets are off.
9576                                  */
9577                                 if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
9578                                     nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
9579                                         next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
9580                                 /*
9581                                  * And we must keep the newest ack arrival time.
9582                                  */
9583                                 if (next->r_ack_arrival <
9584                                     rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9585                                         next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9586
9587
9588                                 /* We don't need to adjust rsm, it did not change */
9589                                 /* Clear out the dup ack count of the remainder */
9590                                 rsm->r_dupack = 0;
9591                                 rsm->r_just_ret = 0;
9592                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9593                                 /* Now lets make sure our fudge block is right */
9594                                 nrsm->r_start = start;
9595                                 /* Now lets update all the stats and such */
9596                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9597                                 if (rack->app_limited_needs_set)
9598                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9599                                 changed += (nrsm->r_end - nrsm->r_start);
9600                                 /* You get a count for acking a whole segment or more */
9601                                 if ((nrsm->r_end - nrsm->r_start) >= segsiz)
9602                                         rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
9603                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9604                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
9605                                         rack->r_ctl.rc_reorder_ts = cts;
9606                                         if (rack->r_ctl.rc_reorder_ts == 0)
9607                                                 rack->r_ctl.rc_reorder_ts = 1;
9608                                 }
9609                                 /*
9610                                  * Now we want to go up from rsm (the
9611                                  * one left un-acked) to the next one
9612                                  * in the tmap. We do this so when
9613                                  * we walk backwards we include marking
9614                                  * sack-passed on rsm (The one passed in
9615                                  * is skipped since it is generally called
9616                                  * on something sacked before removing it
9617                                  * from the tmap).
9618                                  */
9619                                 if (rsm->r_in_tmap) {
9620                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
9621                                         /*
9622                                          * Now that we have the next
9623                                          * one walk backwards from there.
9624                                          */
9625                                         if (nrsm && nrsm->r_in_tmap)
9626                                                 rack_log_sack_passed(tp, rack, nrsm);
9627                                 }
9628                                 /* Now are we done? */
9629                                 if (SEQ_LT(end, next->r_end) ||
9630                                     (end == next->r_end)) {
9631                                         /* Done with block */
9632                                         goto out;
9633                                 }
9634                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
9635                                 counter_u64_add(rack_sack_used_next_merge, 1);
9636                                 /* Postion for the next block */
9637                                 start = next->r_end;
9638                                 rsm = tqhash_next(rack->r_ctl.tqh, next);
9639                                 if (rsm == NULL)
9640                                         goto out;
9641                         } else {
9642                                 /**
9643                                  * We can't use any hookery here, so we
9644                                  * need to split the map. We enter like
9645                                  * so:
9646                                  *  rsm      |--------|
9647                                  *  sackblk       |----->
9648                                  * We will add the new block nrsm and
9649                                  * that will be the new portion, and then
9650                                  * fall through after reseting rsm. So we
9651                                  * split and look like this:
9652                                  *  rsm      |----|
9653                                  *  sackblk       |----->
9654                                  *  nrsm          |---|
9655                                  * We then fall through reseting
9656                                  * rsm to nrsm, so the next block
9657                                  * picks it up.
9658                                  */
9659                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9660                                 if (nrsm == NULL) {
9661                                         /*
9662                                          * failed XXXrrs what can we do but loose the sack
9663                                          * info?
9664                                          */
9665                                         goto out;
9666                                 }
9667                                 counter_u64_add(rack_sack_splits, 1);
9668                                 rack_clone_rsm(rack, nrsm, rsm, start);
9669                                 moved++;
9670                                 rsm->r_just_ret = 0;
9671 #ifndef INVARIANTS
9672                                 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9673 #else
9674                                 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9675                                         panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
9676                                               nrsm, insret, rack, rsm);
9677                                 }
9678 #endif
9679                                 if (rsm->r_in_tmap) {
9680                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9681                                         nrsm->r_in_tmap = 1;
9682                                 }
9683                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
9684                                 rsm->r_flags &= (~RACK_HAS_FIN);
9685                                 /* Position us to point to the new nrsm that starts the sack blk */
9686                                 rsm = nrsm;
9687                         }
9688                 } else {
9689                         /* Already sacked this piece */
9690                         counter_u64_add(rack_sack_skipped_acked, 1);
9691                         moved++;
9692                         if (end == rsm->r_end) {
9693                                 /* Done with block */
9694                                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9695                                 goto out;
9696                         } else if (SEQ_LT(end, rsm->r_end)) {
9697                                 /* A partial sack to a already sacked block */
9698                                 moved++;
9699                                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9700                                 goto out;
9701                         } else {
9702                                 /*
9703                                  * The end goes beyond this guy
9704                                  * reposition the start to the
9705                                  * next block.
9706                                  */
9707                                 start = rsm->r_end;
9708                                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9709                                 if (rsm == NULL)
9710                                         goto out;
9711                         }
9712                 }
9713         }
9714         if (SEQ_GEQ(end, rsm->r_end)) {
9715                 /**
9716                  * The end of this block is either beyond this guy or right
9717                  * at this guy. I.e.:
9718                  *  rsm ---                 |-----|
9719                  *  end                     |-----|
9720                  *  <or>
9721                  *  end                     |---------|
9722                  */
9723                 if ((rsm->r_flags & RACK_ACKED) == 0) {
9724                         /*
9725                          * Is it a TLP of interest?
9726                          */
9727                         if ((rsm->r_flags & RACK_TLP) &&
9728                             (rsm->r_rtr_cnt > 1)) {
9729                                 /*
9730                                  * We are splitting a rxt TLP, check
9731                                  * if we need to save off the start/end
9732                                  */
9733                                 if (rack->rc_last_tlp_acked_set &&
9734                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9735                                         /*
9736                                          * We already turned this on since we are inside
9737                                          * the previous one was a partially sack now we
9738                                          * are getting another one (maybe all of it).
9739                                          */
9740                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9741                                         /*
9742                                          * Lets make sure we have all of it though.
9743                                          */
9744                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9745                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9746                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9747                                                                      rack->r_ctl.last_tlp_acked_end);
9748                                         }
9749                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9750                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9751                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9752                                                                      rack->r_ctl.last_tlp_acked_end);
9753                                         }
9754                                 } else {
9755                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9756                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9757                                         rack->rc_last_tlp_past_cumack = 0;
9758                                         rack->rc_last_tlp_acked_set = 1;
9759                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9760                                 }
9761                         }
9762                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9763                         changed += (rsm->r_end - rsm->r_start);
9764                         /* You get a count for acking a whole segment or more */
9765                         if ((rsm->r_end - rsm->r_start) >= segsiz)
9766                                 rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
9767                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9768                         if (rsm->r_in_tmap) /* should be true */
9769                                 rack_log_sack_passed(tp, rack, rsm);
9770                         /* Is Reordering occuring? */
9771                         if (rsm->r_flags & RACK_SACK_PASSED) {
9772                                 rsm->r_flags &= ~RACK_SACK_PASSED;
9773                                 rack->r_ctl.rc_reorder_ts = cts;
9774                                 if (rack->r_ctl.rc_reorder_ts == 0)
9775                                         rack->r_ctl.rc_reorder_ts = 1;
9776                         }
9777                         if (rack->app_limited_needs_set)
9778                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9779                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9780                         rsm->r_flags |= RACK_ACKED;
9781                         if (rsm->r_in_tmap) {
9782                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9783                                 rsm->r_in_tmap = 0;
9784                         }
9785                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
9786                 } else {
9787                         counter_u64_add(rack_sack_skipped_acked, 1);
9788                         moved++;
9789                 }
9790                 if (end == rsm->r_end) {
9791                         /* This block only - done, setup for next */
9792                         goto out;
9793                 }
9794                 /*
9795                  * There is more not coverend by this rsm move on
9796                  * to the next block in the RB tree.
9797                  */
9798                 nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
9799                 start = rsm->r_end;
9800                 rsm = nrsm;
9801                 if (rsm == NULL)
9802                         goto out;
9803                 goto do_rest_ofb;
9804         }
9805         /**
9806          * The end of this sack block is smaller than
9807          * our rsm i.e.:
9808          *  rsm ---                 |-----|
9809          *  end                     |--|
9810          */
9811         if ((rsm->r_flags & RACK_ACKED) == 0) {
9812                 /*
9813                  * Is it a TLP of interest?
9814                  */
9815                 if ((rsm->r_flags & RACK_TLP) &&
9816                     (rsm->r_rtr_cnt > 1)) {
9817                         /*
9818                          * We are splitting a rxt TLP, check
9819                          * if we need to save off the start/end
9820                          */
9821                         if (rack->rc_last_tlp_acked_set &&
9822                             (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9823                                 /*
9824                                  * We already turned this on since we are inside
9825                                  * the previous one was a partially sack now we
9826                                  * are getting another one (maybe all of it).
9827                                  */
9828                                 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9829                                 /*
9830                                  * Lets make sure we have all of it though.
9831                                  */
9832                                 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9833                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9834                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9835                                                              rack->r_ctl.last_tlp_acked_end);
9836                                 }
9837                                 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9838                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9839                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9840                                                              rack->r_ctl.last_tlp_acked_end);
9841                                 }
9842                         } else {
9843                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9844                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9845                                 rack->rc_last_tlp_past_cumack = 0;
9846                                 rack->rc_last_tlp_acked_set = 1;
9847                                 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9848                         }
9849                 }
9850                 /*
9851                  * Hookery can only be used if the two entries
9852                  * are in the same bucket and neither one of
9853                  * them staddle the bucket line.
9854                  */
9855                 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
9856                 if (prev &&
9857                     (rsm->bindex == prev->bindex) &&
9858                     ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9859                     ((prev->r_flags & RACK_STRADDLE) == 0) &&
9860                     (rsm->r_flags & RACK_IN_GP_WIN) &&
9861                     (prev->r_flags & RACK_IN_GP_WIN))
9862                         can_use_hookery = 1;
9863                 else if (prev &&
9864                          (rsm->bindex == prev->bindex) &&
9865                          ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9866                          ((prev->r_flags & RACK_STRADDLE) == 0) &&
9867                          ((rsm->r_flags & RACK_IN_GP_WIN) == 0) &&
9868                          ((prev->r_flags & RACK_IN_GP_WIN) == 0))
9869                         can_use_hookery = 1;
9870                 else
9871                         can_use_hookery = 0;
9872
9873                 if (prev && can_use_hookery &&
9874                     (prev->r_flags & RACK_ACKED)) {
9875                         /**
9876                          * Goal, we want the right remainder of rsm to shrink
9877                          * in place and span from (rsm->r_start = end) to rsm->r_end.
9878                          * We want to expand prev to go all the way
9879                          * to prev->r_end <- end.
9880                          * so in the tree we have before:
9881                          *   prev     |--------|         (acked)
9882                          *   rsm               |-------| (non-acked)
9883                          *   sackblk           |-|
9884                          * We churn it so we end up with
9885                          *   prev     |----------|       (acked)
9886                          *   rsm                 |-----| (non-acked)
9887                          *   nrsm              |-| (temporary)
9888                          *
9889                          * Note if either prev/rsm is a TLP we don't
9890                          * do this.
9891                          */
9892                         noextra++;
9893                         nrsm = &stack_map;
9894                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9895                         prev->r_end = end;
9896                         rsm->r_start = end;
9897                         rsm->r_flags |= RACK_SHUFFLED;
9898                         prev->r_flags |= RACK_SHUFFLED;
9899                         /* Now adjust nrsm (stack copy) to be
9900                          * the one that is the small
9901                          * piece that was "sacked".
9902                          */
9903                         nrsm->r_end = end;
9904                         rsm->r_dupack = 0;
9905                         /*
9906                          * Which timestamp do we keep? It is rather
9907                          * important in GP measurements to have the
9908                          * accurate end of the send window.
9909                          *
9910                          * We keep the largest value, which is the newest
9911                          * send. We do this in case a segment that is
9912                          * joined together and not part of a GP estimate
9913                          * later gets expanded into the GP estimate.
9914                          *
9915                          * We prohibit the merging of unlike kinds i.e.
9916                          * all pieces that are in the GP estimate can be
9917                          * merged and all pieces that are not in a GP estimate
9918                          * can be merged, but not disimilar pieces. Combine
9919                          * this with taking the highest here and we should
9920                          * be ok unless of course the client reneges. Then
9921                          * all bets are off.
9922                          */
9923                         if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
9924                            nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
9925                                 prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9926                         }
9927                         /*
9928                          * And we must keep the newest ack arrival time.
9929                          */
9930
9931                         if(prev->r_ack_arrival <
9932                            rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9933                                 prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9934
9935                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9936                         /*
9937                          * Now that the rsm has had its start moved forward
9938                          * lets go ahead and get its new place in the world.
9939                          */
9940                         rack_setup_offset_for_rsm(rack, prev, rsm);
9941                         /*
9942                          * Now nrsm is our new little piece
9943                          * that is acked (which was merged
9944                          * to prev). Update the rtt and changed
9945                          * based on that. Also check for reordering.
9946                          */
9947                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9948                         if (rack->app_limited_needs_set)
9949                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9950                         changed += (nrsm->r_end - nrsm->r_start);
9951                         /* You get a count for acking a whole segment or more */
9952                         if ((nrsm->r_end - nrsm->r_start) >= segsiz)
9953                                 rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
9954
9955                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9956                         if (nrsm->r_flags & RACK_SACK_PASSED) {
9957                                 rack->r_ctl.rc_reorder_ts = cts;
9958                                 if (rack->r_ctl.rc_reorder_ts == 0)
9959                                         rack->r_ctl.rc_reorder_ts = 1;
9960                         }
9961                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
9962                         rsm = prev;
9963                         counter_u64_add(rack_sack_used_prev_merge, 1);
9964                 } else {
9965                         /**
9966                          * This is the case where our previous
9967                          * block is not acked either, so we must
9968                          * split the block in two.
9969                          */
9970                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9971                         if (nrsm == NULL) {
9972                                 /* failed rrs what can we do but loose the sack info? */
9973                                 goto out;
9974                         }
9975                         if ((rsm->r_flags & RACK_TLP) &&
9976                             (rsm->r_rtr_cnt > 1)) {
9977                                 /*
9978                                  * We are splitting a rxt TLP, check
9979                                  * if we need to save off the start/end
9980                                  */
9981                                 if (rack->rc_last_tlp_acked_set &&
9982                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9983                                         /*
9984                                          * We already turned this on since this block is inside
9985                                          * the previous one was a partially sack now we
9986                                          * are getting another one (maybe all of it).
9987                                          */
9988                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9989                                         /*
9990                                          * Lets make sure we have all of it though.
9991                                          */
9992                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9993                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9994                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9995                                                                      rack->r_ctl.last_tlp_acked_end);
9996                                         }
9997                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9998                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9999                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10000                                                                      rack->r_ctl.last_tlp_acked_end);
10001                                         }
10002                                 } else {
10003                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10004                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10005                                         rack->rc_last_tlp_acked_set = 1;
10006                                         rack->rc_last_tlp_past_cumack = 0;
10007                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10008                                 }
10009                         }
10010                         /**
10011                          * In this case nrsm becomes
10012                          * nrsm->r_start = end;
10013                          * nrsm->r_end = rsm->r_end;
10014                          * which is un-acked.
10015                          * <and>
10016                          * rsm->r_end = nrsm->r_start;
10017                          * i.e. the remaining un-acked
10018                          * piece is left on the left
10019                          * hand side.
10020                          *
10021                          * So we start like this
10022                          * rsm      |----------| (not acked)
10023                          * sackblk  |---|
10024                          * build it so we have
10025                          * rsm      |---|         (acked)
10026                          * nrsm         |------|  (not acked)
10027                          */
10028                         counter_u64_add(rack_sack_splits, 1);
10029                         rack_clone_rsm(rack, nrsm, rsm, end);
10030                         moved++;
10031                         rsm->r_flags &= (~RACK_HAS_FIN);
10032                         rsm->r_just_ret = 0;
10033 #ifndef INVARIANTS
10034                         (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
10035 #else
10036                         if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
10037                                 panic("Insert in rb tree of %p fails ret:% rack:%p rsm:%p",
10038                                       nrsm, insret, rack, rsm);
10039                         }
10040 #endif
10041                         if (rsm->r_in_tmap) {
10042                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10043                                 nrsm->r_in_tmap = 1;
10044                         }
10045                         nrsm->r_dupack = 0;
10046                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
10047                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
10048                         changed += (rsm->r_end - rsm->r_start);
10049                         /* You get a count for acking a whole segment or more */
10050                         if ((rsm->r_end - rsm->r_start) >= segsiz)
10051                                 rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
10052
10053                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
10054                         if (rsm->r_in_tmap) /* should be true */
10055                                 rack_log_sack_passed(tp, rack, rsm);
10056                         /* Is Reordering occuring? */
10057                         if (rsm->r_flags & RACK_SACK_PASSED) {
10058                                 rsm->r_flags &= ~RACK_SACK_PASSED;
10059                                 rack->r_ctl.rc_reorder_ts = cts;
10060                                 if (rack->r_ctl.rc_reorder_ts == 0)
10061                                         rack->r_ctl.rc_reorder_ts = 1;
10062                         }
10063                         if (rack->app_limited_needs_set)
10064                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10065                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10066                         rsm->r_flags |= RACK_ACKED;
10067                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
10068                         if (rsm->r_in_tmap) {
10069                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10070                                 rsm->r_in_tmap = 0;
10071                         }
10072                 }
10073         } else if (start != end){
10074                 /*
10075                  * The block was already acked.
10076                  */
10077                 counter_u64_add(rack_sack_skipped_acked, 1);
10078                 moved++;
10079         }
10080 out:
10081         if (rsm &&
10082             ((rsm->r_flags & RACK_TLP) == 0) &&
10083             (rsm->r_flags & RACK_ACKED)) {
10084                 /*
10085                  * Now can we merge where we worked
10086                  * with either the previous or
10087                  * next block?
10088                  */
10089                 next = tqhash_next(rack->r_ctl.tqh, rsm);
10090                 while (next) {
10091                         if (next->r_flags & RACK_TLP)
10092                                 break;
10093                         /* Only allow merges between ones in or out of GP window */
10094                         if ((next->r_flags & RACK_IN_GP_WIN) &&
10095                             ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10096                                 break;
10097                         }
10098                         if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10099                             ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10100                                 break;
10101                         }
10102                         if (rsm->bindex != next->bindex)
10103                                 break;
10104                         if (rsm->r_flags & RACK_STRADDLE)
10105                                 break;
10106                         if (next->r_flags & RACK_STRADDLE)
10107                                 break;
10108                         if (next->r_flags & RACK_ACKED) {
10109                                 /* yep this and next can be merged */
10110                                 rsm = rack_merge_rsm(rack, rsm, next);
10111                                 noextra++;
10112                                 next = tqhash_next(rack->r_ctl.tqh, rsm);
10113                         } else
10114                                 break;
10115                 }
10116                 /* Now what about the previous? */
10117                 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10118                 while (prev) {
10119                         if (prev->r_flags & RACK_TLP)
10120                                 break;
10121                         /* Only allow merges between ones in or out of GP window */
10122                         if ((prev->r_flags & RACK_IN_GP_WIN) &&
10123                             ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10124                                 break;
10125                         }
10126                         if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10127                             ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10128                                 break;
10129                         }
10130                         if (rsm->bindex != prev->bindex)
10131                                 break;
10132                         if (rsm->r_flags & RACK_STRADDLE)
10133                                 break;
10134                         if (prev->r_flags & RACK_STRADDLE)
10135                                 break;
10136                         if (prev->r_flags & RACK_ACKED) {
10137                                 /* yep the previous and this can be merged */
10138                                 rsm = rack_merge_rsm(rack, prev, rsm);
10139                                 noextra++;
10140                                 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10141                         } else
10142                                 break;
10143                 }
10144         }
10145         if (used_ref == 0) {
10146                 counter_u64_add(rack_sack_proc_all, 1);
10147         } else {
10148                 counter_u64_add(rack_sack_proc_short, 1);
10149         }
10150         /* Save off the next one for quick reference. */
10151         nrsm = tqhash_find(rack->r_ctl.tqh, end);
10152         *prsm = rack->r_ctl.rc_sacklast = nrsm;
10153         /* Pass back the moved. */
10154         *moved_two = moved;
10155         *no_extra = noextra;
10156         return (changed);
10157 }
10158
10159 static void inline
10160 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10161 {
10162         struct rack_sendmap *tmap;
10163
10164         tmap = NULL;
10165         while (rsm && (rsm->r_flags & RACK_ACKED)) {
10166                 /* Its no longer sacked, mark it so */
10167                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10168 #ifdef INVARIANTS
10169                 if (rsm->r_in_tmap) {
10170                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
10171                               rack, rsm, rsm->r_flags);
10172                 }
10173 #endif
10174                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10175                 /* Rebuild it into our tmap */
10176                 if (tmap == NULL) {
10177                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10178                         tmap = rsm;
10179                 } else {
10180                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10181                         tmap = rsm;
10182                 }
10183                 tmap->r_in_tmap = 1;
10184                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10185         }
10186         /*
10187          * Now lets possibly clear the sack filter so we start
10188          * recognizing sacks that cover this area.
10189          */
10190         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10191
10192 }
10193
10194 static void
10195 rack_do_decay(struct tcp_rack *rack)
10196 {
10197         struct timeval res;
10198
10199 #define timersub(tvp, uvp, vvp)                                         \
10200         do {                                                            \
10201                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
10202                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
10203                 if ((vvp)->tv_usec < 0) {                               \
10204                         (vvp)->tv_sec--;                                \
10205                         (vvp)->tv_usec += 1000000;                      \
10206                 }                                                       \
10207         } while (0)
10208
10209         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
10210 #undef timersub
10211
10212         rack->r_ctl.input_pkt++;
10213         if ((rack->rc_in_persist) ||
10214             (res.tv_sec >= 1) ||
10215             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
10216                 /*
10217                  * Check for decay of non-SAD,
10218                  * we want all SAD detection metrics to
10219                  * decay 1/4 per second (or more) passed.
10220                  * Current default is 800 so it decays
10221                  * 80% every second.
10222                  */
10223 #ifdef TCP_SAD_DETECTION
10224                 uint32_t pkt_delta;
10225
10226                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
10227 #endif
10228                 /* Update our saved tracking values */
10229                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
10230                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
10231                 /* Now do we escape without decay? */
10232 #ifdef TCP_SAD_DETECTION
10233                 if (rack->rc_in_persist ||
10234                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
10235                     (pkt_delta < tcp_sad_low_pps)){
10236                         /*
10237                          * We don't decay idle connections
10238                          * or ones that have a low input pps.
10239                          */
10240                         return;
10241                 }
10242                 /* Decay the counters */
10243                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
10244                                                         tcp_sad_decay_val);
10245                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
10246                                                          tcp_sad_decay_val);
10247                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
10248                                                                tcp_sad_decay_val);
10249                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
10250                                                                 tcp_sad_decay_val);
10251 #endif
10252         }
10253 }
10254
10255 static void inline
10256 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10257 {
10258         /*
10259          * We look at advancing the end send time for our GP
10260          * measurement tracking only as the cumulative acknowledgment
10261          * moves forward. You might wonder about this, why not
10262          * at every transmission or retransmission within the
10263          * GP window update the rc_gp_cumack_ts? Well its rather
10264          * nuanced but basically the GP window *may* expand (as
10265          * it does below) or worse and harder to track it may shrink.
10266          *
10267          * This last makes it impossible to track at the time of
10268          * the send, since you may set forward your rc_gp_cumack_ts
10269          * when you send, because that send *is* in your currently
10270          * "guessed" window, but then it shrinks. Now which was
10271          * the send time of the last bytes in the window, by the
10272          * time you ask that question that part of the sendmap
10273          * is freed. So you don't know and you will have too
10274          * long of send window. Instead by updating the time
10275          * marker only when the cumack advances this assures us
10276          * that we will have only the sends in the window of our
10277          * GP measurement.
10278          *
10279          * Another complication from this is the
10280          * merging of sendmap entries. During SACK processing this
10281          * can happen to conserve the sendmap size. That breaks
10282          * everything down in tracking the send window of the GP
10283          * estimate. So to prevent that and keep it working with
10284          * a tiny bit more limited merging, we only allow like
10285          * types to be merged. I.e. if two sends are in the GP window
10286          * then its ok to merge them together. If two sends are not
10287          * in the GP window its ok to merge them together too. Though
10288          * one send in and one send out cannot be merged. We combine
10289          * this with never allowing the shrinking of the GP window when
10290          * we are in recovery so that we can properly calculate the
10291          * sending times.
10292          *
10293          * This all of course seems complicated, because it is.. :)
10294          *
10295          * The cum-ack is being advanced upon the sendmap.
10296          * If we are not doing a GP estimate don't
10297          * proceed.
10298          */
10299         uint64_t ts;
10300
10301         if ((tp->t_flags & TF_GPUTINPROG) == 0)
10302                 return;
10303         /*
10304          * If this sendmap entry is going
10305          * beyond the measurement window we had picked,
10306          * expand the measurement window by that much.
10307          */
10308         if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10309                 tp->gput_ack = rsm->r_end;
10310         }
10311         /*
10312          * If we have not setup a ack, then we
10313          * have no idea if the newly acked pieces
10314          * will be "in our seq measurement range". If
10315          * it is when we clear the app_limited_needs_set
10316          * flag the timestamp will be updated.
10317          */
10318         if (rack->app_limited_needs_set)
10319                 return;
10320         /*
10321          * Finally, we grab out the latest timestamp
10322          * that this packet was sent and then see
10323          * if:
10324          *  a) The packet touches are newly defined GP range.
10325          *  b) The time is greater than (newer) than the
10326          *     one we currently have. If so we update
10327          *     our sending end time window.
10328          *
10329          * Note we *do not* do this at send time. The reason
10330          * is that if you do you *may* pick up a newer timestamp
10331          * for a range you are not going to measure. We project
10332          * out how far and then sometimes modify that to be
10333          * smaller. If that occurs then you will have a send
10334          * that does not belong to the range included.
10335          */
10336         if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10337             rack->r_ctl.rc_gp_cumack_ts)
10338                 return;
10339         if (rack_in_gp_window(tp, rsm)) {
10340                 rack->r_ctl.rc_gp_cumack_ts = ts;
10341                 rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10342                                __LINE__, from, rsm);
10343         }
10344 }
10345
10346 static void
10347 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to, uint64_t acktime)
10348 {
10349         struct rack_sendmap *rsm;
10350         /*
10351          * The ACK point is advancing to th_ack, we must drop off
10352          * the packets in the rack log and calculate any eligble
10353          * RTT's.
10354          */
10355
10356         rack->r_wanted_output = 1;
10357         if (SEQ_GT(th_ack, tp->snd_una))
10358             rack->r_ctl.last_cumack_advance = acktime;
10359
10360         /* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
10361         if ((rack->rc_last_tlp_acked_set == 1)&&
10362             (rack->rc_last_tlp_past_cumack == 1) &&
10363             (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10364                 /*
10365                  * We have reached the point where our last rack
10366                  * tlp retransmit sequence is ahead of the cum-ack.
10367                  * This can only happen when the cum-ack moves all
10368                  * the way around (its been a full 2^^31+1 bytes
10369                  * or more since we sent a retransmitted TLP). Lets
10370                  * turn off the valid flag since its not really valid.
10371                  *
10372                  * Note since sack's also turn on this event we have
10373                  * a complication, we have to wait to age it out until
10374                  * the cum-ack is by the TLP before checking which is
10375                  * what the next else clause does.
10376                  */
10377                 rack_log_dsack_event(rack, 9, __LINE__,
10378                                      rack->r_ctl.last_tlp_acked_start,
10379                                      rack->r_ctl.last_tlp_acked_end);
10380                 rack->rc_last_tlp_acked_set = 0;
10381                 rack->rc_last_tlp_past_cumack = 0;
10382         } else if ((rack->rc_last_tlp_acked_set == 1) &&
10383                    (rack->rc_last_tlp_past_cumack == 0) &&
10384                    (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10385                 /*
10386                  * It is safe to start aging TLP's out.
10387                  */
10388                 rack->rc_last_tlp_past_cumack = 1;
10389         }
10390         /* We do the same for the tlp send seq as well */
10391         if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10392             (rack->rc_last_sent_tlp_past_cumack == 1) &&
10393             (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
10394                 rack_log_dsack_event(rack, 9, __LINE__,
10395                                      rack->r_ctl.last_sent_tlp_seq,
10396                                      (rack->r_ctl.last_sent_tlp_seq +
10397                                       rack->r_ctl.last_sent_tlp_len));
10398                 rack->rc_last_sent_tlp_seq_valid = 0;
10399                 rack->rc_last_sent_tlp_past_cumack = 0;
10400         } else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10401                    (rack->rc_last_sent_tlp_past_cumack == 0) &&
10402                    (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10403                 /*
10404                  * It is safe to start aging TLP's send.
10405                  */
10406                 rack->rc_last_sent_tlp_past_cumack = 1;
10407         }
10408 more:
10409         rsm = tqhash_min(rack->r_ctl.tqh);
10410         if (rsm == NULL) {
10411                 if ((th_ack - 1) == tp->iss) {
10412                         /*
10413                          * For the SYN incoming case we will not
10414                          * have called tcp_output for the sending of
10415                          * the SYN, so there will be no map. All
10416                          * other cases should probably be a panic.
10417                          */
10418                         return;
10419                 }
10420                 if (tp->t_flags & TF_SENTFIN) {
10421                         /* if we sent a FIN we often will not have map */
10422                         return;
10423                 }
10424 #ifdef INVARIANTS
10425                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
10426                       tp,
10427                       tp->t_state, th_ack, rack,
10428                       tp->snd_una, tp->snd_max, tp->snd_nxt);
10429 #endif
10430                 return;
10431         }
10432         if (SEQ_LT(th_ack, rsm->r_start)) {
10433                 /* Huh map is missing this */
10434 #ifdef INVARIANTS
10435                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
10436                        rsm->r_start,
10437                        th_ack, tp->t_state, rack->r_state);
10438 #endif
10439                 return;
10440         }
10441         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
10442
10443         /* Now was it a retransmitted TLP? */
10444         if ((rsm->r_flags & RACK_TLP) &&
10445             (rsm->r_rtr_cnt > 1)) {
10446                 /*
10447                  * Yes, this rsm was a TLP and retransmitted, remember that
10448                  * since if a DSACK comes back on this we don't want
10449                  * to think of it as a reordered segment. This may
10450                  * get updated again with possibly even other TLPs
10451                  * in flight, but thats ok. Only when we don't send
10452                  * a retransmitted TLP for 1/2 the sequences space
10453                  * will it get turned off (above).
10454                  */
10455                 if (rack->rc_last_tlp_acked_set &&
10456                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10457                         /*
10458                          * We already turned this on since the end matches,
10459                          * the previous one was a partially ack now we
10460                          * are getting another one (maybe all of it).
10461                          */
10462                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10463                         /*
10464                          * Lets make sure we have all of it though.
10465                          */
10466                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10467                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10468                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10469                                                      rack->r_ctl.last_tlp_acked_end);
10470                         }
10471                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10472                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10473                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10474                                                      rack->r_ctl.last_tlp_acked_end);
10475                         }
10476                 } else {
10477                         rack->rc_last_tlp_past_cumack = 1;
10478                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10479                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10480                         rack->rc_last_tlp_acked_set = 1;
10481                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10482                 }
10483         }
10484         /* Now do we consume the whole thing? */
10485         if (SEQ_GEQ(th_ack, rsm->r_end)) {
10486                 /* Its all consumed. */
10487                 uint32_t left;
10488                 uint8_t newly_acked;
10489
10490                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
10491                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
10492                 rsm->r_rtr_bytes = 0;
10493                 /*
10494                  * Record the time of highest cumack sent if its in our measurement
10495                  * window and possibly bump out the end.
10496                  */
10497                 rack_rsm_sender_update(rack, tp, rsm, 4);
10498                 tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
10499                 if (rsm->r_in_tmap) {
10500                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10501                         rsm->r_in_tmap = 0;
10502                 }
10503                 newly_acked = 1;
10504                 if (rsm->r_flags & RACK_ACKED) {
10505                         /*
10506                          * It was acked on the scoreboard -- remove
10507                          * it from total
10508                          */
10509                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10510                         newly_acked = 0;
10511                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
10512                         /*
10513                          * There are segments ACKED on the
10514                          * scoreboard further up. We are seeing
10515                          * reordering.
10516                          */
10517                         rsm->r_flags &= ~RACK_SACK_PASSED;
10518                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10519                         rsm->r_flags |= RACK_ACKED;
10520                         rack->r_ctl.rc_reorder_ts = cts;
10521                         if (rack->r_ctl.rc_reorder_ts == 0)
10522                                 rack->r_ctl.rc_reorder_ts = 1;
10523                         if (rack->r_ent_rec_ns) {
10524                                 /*
10525                                  * We have sent no more, and we saw an sack
10526                                  * then ack arrive.
10527                                  */
10528                                 rack->r_might_revert = 1;
10529                         }
10530                 }
10531                 if ((rsm->r_flags & RACK_TO_REXT) &&
10532                     (tp->t_flags & TF_RCVD_TSTMP) &&
10533                     (to->to_flags & TOF_TS) &&
10534                     (to->to_tsecr != 0) &&
10535                     (tp->t_flags & TF_PREVVALID)) {
10536                         /*
10537                          * We can use the timestamp to see
10538                          * if this retransmission was from the
10539                          * first transmit. If so we made a mistake.
10540                          */
10541                         tp->t_flags &= ~TF_PREVVALID;
10542                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
10543                                 /* The first transmit is what this ack is for */
10544                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
10545                         }
10546                 }
10547                 left = th_ack - rsm->r_end;
10548                 if (rack->app_limited_needs_set && newly_acked)
10549                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
10550                 /* Free back to zone */
10551                 rack_free(rack, rsm);
10552                 if (left) {
10553                         goto more;
10554                 }
10555                 /* Check for reneging */
10556                 rsm = tqhash_min(rack->r_ctl.tqh);
10557                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
10558                         /*
10559                          * The peer has moved snd_una up to
10560                          * the edge of this send, i.e. one
10561                          * that it had previously acked. The only
10562                          * way that can be true if the peer threw
10563                          * away data (space issues) that it had
10564                          * previously sacked (else it would have
10565                          * given us snd_una up to (rsm->r_end).
10566                          * We need to undo the acked markings here.
10567                          *
10568                          * Note we have to look to make sure th_ack is
10569                          * our rsm->r_start in case we get an old ack
10570                          * where th_ack is behind snd_una.
10571                          */
10572                         rack_peer_reneges(rack, rsm, th_ack);
10573                 }
10574                 return;
10575         }
10576         if (rsm->r_flags & RACK_ACKED) {
10577                 /*
10578                  * It was acked on the scoreboard -- remove it from
10579                  * total for the part being cum-acked.
10580                  */
10581                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
10582         }
10583         /*
10584          * Clear the dup ack count for
10585          * the piece that remains.
10586          */
10587         rsm->r_dupack = 0;
10588         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10589         if (rsm->r_rtr_bytes) {
10590                 /*
10591                  * It was retransmitted adjust the
10592                  * sack holes for what was acked.
10593                  */
10594                 int ack_am;
10595
10596                 ack_am = (th_ack - rsm->r_start);
10597                 if (ack_am >= rsm->r_rtr_bytes) {
10598                         rack->r_ctl.rc_holes_rxt -= ack_am;
10599                         rsm->r_rtr_bytes -= ack_am;
10600                 }
10601         }
10602         /*
10603          * Update where the piece starts and record
10604          * the time of send of highest cumack sent if
10605          * its in our GP range.
10606          */
10607         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
10608         /* Now we need to move our offset forward too */
10609         if (rsm->m &&
10610             ((rsm->orig_m_len != rsm->m->m_len) ||
10611              (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
10612                 /* Fix up the orig_m_len and possibly the mbuf offset */
10613                 rack_adjust_orig_mlen(rsm);
10614         }
10615         rsm->soff += (th_ack - rsm->r_start);
10616         rack_rsm_sender_update(rack, tp, rsm, 5);
10617         /* The trim will move th_ack into r_start for us */
10618         tqhash_trim(rack->r_ctl.tqh, th_ack);
10619         /* Now do we need to move the mbuf fwd too? */
10620         if (rsm->m) {
10621                 while (rsm->soff >= rsm->m->m_len) {
10622                         rsm->soff -= rsm->m->m_len;
10623                         rsm->m = rsm->m->m_next;
10624                         KASSERT((rsm->m != NULL),
10625                                 (" nrsm:%p hit at soff:%u null m",
10626                                  rsm, rsm->soff));
10627                 }
10628                 rsm->orig_m_len = rsm->m->m_len;
10629                 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
10630         }
10631         if (rack->app_limited_needs_set &&
10632             SEQ_GEQ(th_ack, tp->gput_seq))
10633                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
10634 }
10635
10636 static void
10637 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
10638 {
10639         struct rack_sendmap *rsm;
10640         int sack_pass_fnd = 0;
10641
10642         if (rack->r_might_revert) {
10643                 /*
10644                  * Ok we have reordering, have not sent anything, we
10645                  * might want to revert the congestion state if nothing
10646                  * further has SACK_PASSED on it. Lets check.
10647                  *
10648                  * We also get here when we have DSACKs come in for
10649                  * all the data that we FR'd. Note that a rxt or tlp
10650                  * timer clears this from happening.
10651                  */
10652
10653                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
10654                         if (rsm->r_flags & RACK_SACK_PASSED) {
10655                                 sack_pass_fnd = 1;
10656                                 break;
10657                         }
10658                 }
10659                 if (sack_pass_fnd == 0) {
10660                         /*
10661                          * We went into recovery
10662                          * incorrectly due to reordering!
10663                          */
10664                         int orig_cwnd;
10665
10666                         rack->r_ent_rec_ns = 0;
10667                         orig_cwnd = tp->snd_cwnd;
10668                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
10669                         tp->snd_recover = tp->snd_una;
10670                         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
10671                         EXIT_RECOVERY(tp->t_flags);
10672                 }
10673                 rack->r_might_revert = 0;
10674         }
10675 }
10676
10677 #ifdef TCP_SAD_DETECTION
10678
10679 static void
10680 rack_merge_out_sacks(struct tcp_rack *rack)
10681 {
10682         struct rack_sendmap *cur, *next, *rsm, *trsm = NULL;
10683
10684         cur = tqhash_min(rack->r_ctl.tqh);
10685         while(cur) {
10686                 next = tqhash_next(rack->r_ctl.tqh, cur);
10687                 /*
10688                  * The idea is to go through all and merge back
10689                  * together the pieces sent together,
10690                  */
10691                 if ((next != NULL) &&
10692                     (cur->r_tim_lastsent[0] == next->r_tim_lastsent[0])) {
10693                         rack_merge_rsm(rack, cur, next);
10694                 } else {
10695                         cur = next;
10696                 }
10697         }
10698         /*
10699          * now treat it like a rxt event, everything is outstanding
10700          * and sent nothing acvked and dupacks are all zero. If this
10701          * is not an attacker it will have to dupack its way through
10702          * it all.
10703          */
10704         TAILQ_INIT(&rack->r_ctl.rc_tmap);
10705         TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
10706                 rsm->r_dupack = 0;
10707                 /* We must re-add it back to the tlist */
10708                 if (trsm == NULL) {
10709                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10710                 } else {
10711                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
10712                 }
10713                 rsm->r_in_tmap = 1;
10714                 trsm = rsm;
10715                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
10716         }
10717         sack_filter_clear(&rack->r_ctl.rack_sf, rack->rc_tp->snd_una);
10718 }
10719
10720 static void
10721 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
10722 {
10723         int do_detection = 0;
10724
10725         if (rack->sack_attack_disable || rack->rc_suspicious) {
10726                 /*
10727                  * If we have been disabled we must detect
10728                  * to possibly reverse it. Or if the guy has
10729                  * sent in suspicious sacks we want to do detection too.
10730                  */
10731                 do_detection = 1;
10732
10733         } else if  ((rack->do_detection || tcp_force_detection) &&
10734                     (tcp_sack_to_ack_thresh > 0) &&
10735                     (tcp_sack_to_move_thresh > 0) &&
10736                     (rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum)) {
10737                 /*
10738                  * We only detect here if:
10739                  * 1) System wide forcing is on <or> do_detection is on
10740                  *   <and>
10741                  * 2) We have thresholds for move and ack (set one to 0 and we are off)
10742                  *   <and>
10743                  * 3) We have maps allocated larger than our min (500).
10744                  */
10745                 do_detection = 1;
10746         }
10747         if (do_detection > 0) {
10748                 /*
10749                  * We have thresholds set to find
10750                  * possible attackers and disable sack.
10751                  * Check them.
10752                  */
10753                 uint64_t ackratio, moveratio, movetotal;
10754
10755                 /* Log detecting */
10756                 rack_log_sad(rack, 1);
10757                 /* Do we establish a ack ratio */
10758                 if ((rack->r_ctl.sack_count > tcp_map_minimum)  ||
10759                     (rack->rc_suspicious == 1) ||
10760                     (rack->sack_attack_disable > 0)) {
10761                         ackratio = (uint64_t)(rack->r_ctl.sack_count);
10762                         ackratio *= (uint64_t)(1000);
10763                         if (rack->r_ctl.ack_count)
10764                                 ackratio /= (uint64_t)(rack->r_ctl.ack_count);
10765                         else {
10766                                 /* We can hit this due to ack totals degregation (via small sacks) */
10767                                 ackratio = 1000;
10768                         }
10769                 } else {
10770                         /*
10771                          * No ack ratio needed if we have not
10772                          * seen more sacks then the number of map entries.
10773                          * The exception to that is if we have disabled sack then
10774                          * we need to find a ratio.
10775                          */
10776                         ackratio = 0;
10777                 }
10778
10779                 if ((rack->sack_attack_disable == 0) &&
10780                     (ackratio > rack_highest_sack_thresh_seen))
10781                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
10782                 /* Do we establish a move ratio? */
10783                 if ((rack->r_ctl.sack_moved_extra > tcp_map_minimum) ||
10784                     (rack->rc_suspicious == 1) ||
10785                     (rack->sack_attack_disable > 0)) {
10786                         /*
10787                          * We need to have more sack moves than maps
10788                          * allocated to have a move ratio considered.
10789                          */
10790                         movetotal = rack->r_ctl.sack_moved_extra;
10791                         movetotal += rack->r_ctl.sack_noextra_move;
10792                         moveratio = rack->r_ctl.sack_moved_extra;
10793                         moveratio *= (uint64_t)1000;
10794                         if (movetotal)
10795                                 moveratio /= movetotal;
10796                         else {
10797                                 /* No moves, thats pretty good */
10798                                 moveratio = 0;
10799                         }
10800                 } else {
10801                         /*
10802                          * Not enough moves have occured to consider
10803                          * if we are out of whack in that ratio.
10804                          * The exception to that is if we have disabled sack then
10805                          * we need to find a ratio.
10806                          */
10807                         moveratio = 0;
10808                 }
10809                 if ((rack->sack_attack_disable == 0) &&
10810                     (moveratio > rack_highest_move_thresh_seen))
10811                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
10812                 /* Now the tests */
10813                 if (rack->sack_attack_disable == 0) {
10814                         /* Not disabled, do we need to disable? */
10815                         if ((ackratio > tcp_sack_to_ack_thresh) &&
10816                             (moveratio > tcp_sack_to_move_thresh)) {
10817                                 /* Disable sack processing */
10818                                 tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
10819                                 rack->sack_attack_disable = 1;
10820                                 /* set it so we have the built in delay */
10821                                 rack->r_ctl.ack_during_sd = 1;
10822                                 if (rack_merge_out_sacks_on_attack)
10823                                         rack_merge_out_sacks(rack);
10824                                 counter_u64_add(rack_sack_attacks_detected, 1);
10825                                 tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
10826                                 /* Clamp the cwnd at flight size */
10827                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
10828                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10829                                 rack_log_sad(rack, 2);
10830                         }
10831                 } else {
10832                         /* We are sack-disabled check for false positives */
10833                         if ((ackratio <= tcp_restoral_thresh) ||
10834                             ((rack_merge_out_sacks_on_attack == 0) &&
10835                              (rack->rc_suspicious == 0) &&
10836                              (rack->r_ctl.rc_num_maps_alloced <= (tcp_map_minimum/2)))) {
10837                                 rack->sack_attack_disable = 0;
10838                                 rack_log_sad(rack, 3);
10839                                 /* Restart counting */
10840                                 rack->r_ctl.sack_count = 0;
10841                                 rack->r_ctl.sack_moved_extra = 0;
10842                                 rack->r_ctl.sack_noextra_move = 1;
10843                                 rack->rc_suspicious = 0;
10844                                 rack->r_ctl.ack_count = max(1,
10845                                                             (bytes_this_ack / segsiz));
10846
10847                                 counter_u64_add(rack_sack_attacks_reversed, 1);
10848                                 /* Restore the cwnd */
10849                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
10850                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
10851                         }
10852                 }
10853         }
10854 }
10855 #endif
10856
10857 static int
10858 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
10859 {
10860
10861         uint32_t am, l_end;
10862         int was_tlp = 0;
10863
10864         if (SEQ_GT(end, start))
10865                 am = end - start;
10866         else
10867                 am = 0;
10868         if ((rack->rc_last_tlp_acked_set ) &&
10869             (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
10870             (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
10871                 /*
10872                  * The DSACK is because of a TLP which we don't
10873                  * do anything with the reordering window over since
10874                  * it was not reordering that caused the DSACK but
10875                  * our previous retransmit TLP.
10876                  */
10877                 rack_log_dsack_event(rack, 7, __LINE__, start, end);
10878                 was_tlp = 1;
10879                 goto skip_dsack_round;
10880         }
10881         if (rack->rc_last_sent_tlp_seq_valid) {
10882                 l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
10883                 if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
10884                     (SEQ_LEQ(end, l_end))) {
10885                         /*
10886                          * This dsack is from the last sent TLP, ignore it
10887                          * for reordering purposes.
10888                          */
10889                         rack_log_dsack_event(rack, 7, __LINE__, start, end);
10890                         was_tlp = 1;
10891                         goto skip_dsack_round;
10892                 }
10893         }
10894         if (rack->rc_dsack_round_seen == 0) {
10895                 rack->rc_dsack_round_seen = 1;
10896                 rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
10897                 rack->r_ctl.num_dsack++;
10898                 rack->r_ctl.dsack_persist = 16; /* 16 is from the standard */
10899                 rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
10900         }
10901 skip_dsack_round:
10902         /*
10903          * We keep track of how many DSACK blocks we get
10904          * after a recovery incident.
10905          */
10906         rack->r_ctl.dsack_byte_cnt += am;
10907         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
10908             rack->r_ctl.retran_during_recovery &&
10909             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
10910                 /*
10911                  * False recovery most likely culprit is reordering. If
10912                  * nothing else is missing we need to revert.
10913                  */
10914                 rack->r_might_revert = 1;
10915                 rack_handle_might_revert(rack->rc_tp, rack);
10916                 rack->r_might_revert = 0;
10917                 rack->r_ctl.retran_during_recovery = 0;
10918                 rack->r_ctl.dsack_byte_cnt = 0;
10919         }
10920         return (was_tlp);
10921 }
10922
10923 static uint32_t
10924 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
10925 {
10926         return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
10927 }
10928
10929 static int32_t
10930 rack_compute_pipe(struct tcpcb *tp)
10931 {
10932         return ((int32_t)do_rack_compute_pipe(tp,
10933                                               (struct tcp_rack *)tp->t_fb_ptr,
10934                                               tp->snd_una));
10935 }
10936
10937 static void
10938 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
10939 {
10940         /* Deal with changed and PRR here (in recovery only) */
10941         uint32_t pipe, snd_una;
10942
10943         rack->r_ctl.rc_prr_delivered += changed;
10944
10945         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
10946                 /*
10947                  * It is all outstanding, we are application limited
10948                  * and thus we don't need more room to send anything.
10949                  * Note we use tp->snd_una here and not th_ack because
10950                  * the data as yet not been cut from the sb.
10951                  */
10952                 rack->r_ctl.rc_prr_sndcnt = 0;
10953                 return;
10954         }
10955         /* Compute prr_sndcnt */
10956         if (SEQ_GT(tp->snd_una, th_ack)) {
10957                 snd_una = tp->snd_una;
10958         } else {
10959                 snd_una = th_ack;
10960         }
10961         pipe = do_rack_compute_pipe(tp, rack, snd_una);
10962         if (pipe > tp->snd_ssthresh) {
10963                 long sndcnt;
10964
10965                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
10966                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
10967                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
10968                 else {
10969                         rack->r_ctl.rc_prr_sndcnt = 0;
10970                         rack_log_to_prr(rack, 9, 0, __LINE__);
10971                         sndcnt = 0;
10972                 }
10973                 sndcnt++;
10974                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
10975                         sndcnt -= rack->r_ctl.rc_prr_out;
10976                 else
10977                         sndcnt = 0;
10978                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
10979                 rack_log_to_prr(rack, 10, 0, __LINE__);
10980         } else {
10981                 uint32_t limit;
10982
10983                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
10984                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
10985                 else
10986                         limit = 0;
10987                 if (changed > limit)
10988                         limit = changed;
10989                 limit += ctf_fixed_maxseg(tp);
10990                 if (tp->snd_ssthresh > pipe) {
10991                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
10992                         rack_log_to_prr(rack, 11, 0, __LINE__);
10993                 } else {
10994                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
10995                         rack_log_to_prr(rack, 12, 0, __LINE__);
10996                 }
10997         }
10998 }
10999
11000 static void
11001 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
11002              int *dsack_seen, int *sacks_seen)
11003 {
11004         uint32_t changed;
11005         struct tcp_rack *rack;
11006         struct rack_sendmap *rsm;
11007         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
11008         register uint32_t th_ack;
11009         int32_t i, j, k, num_sack_blks = 0;
11010         uint32_t cts, acked, ack_point;
11011         int loop_start = 0, moved_two = 0, no_extra = 0;
11012         uint32_t tsused;
11013         uint32_t segsiz, o_cnt;
11014
11015
11016         INP_WLOCK_ASSERT(tptoinpcb(tp));
11017         if (tcp_get_flags(th) & TH_RST) {
11018                 /* We don't log resets */
11019                 return;
11020         }
11021         rack = (struct tcp_rack *)tp->t_fb_ptr;
11022         cts = tcp_get_usecs(NULL);
11023         rsm = tqhash_min(rack->r_ctl.tqh);
11024         changed = 0;
11025         th_ack = th->th_ack;
11026         if (rack->sack_attack_disable == 0)
11027                 rack_do_decay(rack);
11028         segsiz = ctf_fixed_maxseg(rack->rc_tp);
11029         if (BYTES_THIS_ACK(tp, th) >=  segsiz) {
11030                 /*
11031                  * You only get credit for
11032                  * MSS and greater (and you get extra
11033                  * credit for larger cum-ack moves).
11034                  */
11035                 int ac;
11036
11037                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
11038                 rack->r_ctl.ack_count += ac;
11039                 counter_u64_add(rack_ack_total, ac);
11040         }
11041         if (rack->r_ctl.ack_count > 0xfff00000) {
11042                 /*
11043                  * reduce the number to keep us under
11044                  * a uint32_t.
11045                  */
11046                 rack->r_ctl.ack_count /= 2;
11047                 rack->r_ctl.sack_count /= 2;
11048         }
11049         if (SEQ_GT(th_ack, tp->snd_una)) {
11050                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
11051                 tp->t_acktime = ticks;
11052         }
11053         if (rsm && SEQ_GT(th_ack, rsm->r_start))
11054                 changed = th_ack - rsm->r_start;
11055         if (changed) {
11056                 rack_process_to_cumack(tp, rack, th_ack, cts, to,
11057                                        tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
11058         }
11059         if ((to->to_flags & TOF_SACK) == 0) {
11060                 /* We are done nothing left and no sack. */
11061                 rack_handle_might_revert(tp, rack);
11062                 /*
11063                  * For cases where we struck a dup-ack
11064                  * with no SACK, add to the changes so
11065                  * PRR will work right.
11066                  */
11067                 if (dup_ack_struck && (changed == 0)) {
11068                         changed += ctf_fixed_maxseg(rack->rc_tp);
11069                 }
11070                 goto out;
11071         }
11072         /* Sack block processing */
11073         if (SEQ_GT(th_ack, tp->snd_una))
11074                 ack_point = th_ack;
11075         else
11076                 ack_point = tp->snd_una;
11077         for (i = 0; i < to->to_nsacks; i++) {
11078                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
11079                       &sack, sizeof(sack));
11080                 sack.start = ntohl(sack.start);
11081                 sack.end = ntohl(sack.end);
11082                 if (SEQ_GT(sack.end, sack.start) &&
11083                     SEQ_GT(sack.start, ack_point) &&
11084                     SEQ_LT(sack.start, tp->snd_max) &&
11085                     SEQ_GT(sack.end, ack_point) &&
11086                     SEQ_LEQ(sack.end, tp->snd_max)) {
11087                         sack_blocks[num_sack_blks] = sack;
11088                         num_sack_blks++;
11089                 } else if (SEQ_LEQ(sack.start, th_ack) &&
11090                            SEQ_LEQ(sack.end, th_ack)) {
11091                         int was_tlp;
11092
11093                         if (dsack_seen != NULL)
11094                                 *dsack_seen = 1;
11095                         was_tlp = rack_note_dsack(rack, sack.start, sack.end);
11096                         /*
11097                          * Its a D-SACK block.
11098                          */
11099                         tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
11100                 }
11101         }
11102         if (rack->rc_dsack_round_seen) {
11103                 /* Is the dsack roound over? */
11104                 if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
11105                         /* Yes it is */
11106                         rack->rc_dsack_round_seen = 0;
11107                         rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
11108                 }
11109         }
11110         /*
11111          * Sort the SACK blocks so we can update the rack scoreboard with
11112          * just one pass.
11113          */
11114         o_cnt = num_sack_blks;
11115         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
11116                                          num_sack_blks, th->th_ack);
11117         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
11118         if (sacks_seen != NULL)
11119                 *sacks_seen = num_sack_blks;
11120         if (num_sack_blks == 0) {
11121                 /* Nothing to sack, but we need to update counts */
11122                 if ((o_cnt == 1) &&
11123                     (*dsack_seen != 1))
11124                         rack->r_ctl.sack_count++;
11125                 else if (o_cnt > 1)
11126                         rack->r_ctl.sack_count++;
11127                 goto out_with_totals;
11128         }
11129         if (rack->sack_attack_disable) {
11130                 /*
11131                  * An attacker disablement is in place, for
11132                  * every sack block that is not at least a full MSS
11133                  * count up sack_count.
11134                  */
11135                 for (i = 0; i < num_sack_blks; i++) {
11136                         if ((sack_blocks[i].end - sack_blocks[i].start) < segsiz) {
11137                                 rack->r_ctl.sack_count++;
11138                         }
11139                         if (rack->r_ctl.sack_count > 0xfff00000) {
11140                                 /*
11141                                  * reduce the number to keep us under
11142                                  * a uint32_t.
11143                                  */
11144                                 rack->r_ctl.ack_count /= 2;
11145                                 rack->r_ctl.sack_count /= 2;
11146                         }
11147                 }
11148                 goto out;
11149         }
11150         /* Its a sack of some sort */
11151         rack->r_ctl.sack_count += num_sack_blks;
11152         if (rack->r_ctl.sack_count > 0xfff00000) {
11153                 /*
11154                  * reduce the number to keep us under
11155                  * a uint32_t.
11156                  */
11157                 rack->r_ctl.ack_count /= 2;
11158                 rack->r_ctl.sack_count /= 2;
11159         }
11160         if (num_sack_blks < 2) {
11161                 /* Only one, we don't need to sort */
11162                 goto do_sack_work;
11163         }
11164         /* Sort the sacks */
11165         for (i = 0; i < num_sack_blks; i++) {
11166                 for (j = i + 1; j < num_sack_blks; j++) {
11167                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
11168                                 sack = sack_blocks[i];
11169                                 sack_blocks[i] = sack_blocks[j];
11170                                 sack_blocks[j] = sack;
11171                         }
11172                 }
11173         }
11174         /*
11175          * Now are any of the sack block ends the same (yes some
11176          * implementations send these)?
11177          */
11178 again:
11179         if (num_sack_blks == 0)
11180                 goto out_with_totals;
11181         if (num_sack_blks > 1) {
11182                 for (i = 0; i < num_sack_blks; i++) {
11183                         for (j = i + 1; j < num_sack_blks; j++) {
11184                                 if (sack_blocks[i].end == sack_blocks[j].end) {
11185                                         /*
11186                                          * Ok these two have the same end we
11187                                          * want the smallest end and then
11188                                          * throw away the larger and start
11189                                          * again.
11190                                          */
11191                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
11192                                                 /*
11193                                                  * The second block covers
11194                                                  * more area use that
11195                                                  */
11196                                                 sack_blocks[i].start = sack_blocks[j].start;
11197                                         }
11198                                         /*
11199                                          * Now collapse out the dup-sack and
11200                                          * lower the count
11201                                          */
11202                                         for (k = (j + 1); k < num_sack_blks; k++) {
11203                                                 sack_blocks[j].start = sack_blocks[k].start;
11204                                                 sack_blocks[j].end = sack_blocks[k].end;
11205                                                 j++;
11206                                         }
11207                                         num_sack_blks--;
11208                                         goto again;
11209                                 }
11210                         }
11211                 }
11212         }
11213 do_sack_work:
11214         /*
11215          * First lets look to see if
11216          * we have retransmitted and
11217          * can use the transmit next?
11218          */
11219         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11220         if (rsm &&
11221             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
11222             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
11223                 /*
11224                  * We probably did the FR and the next
11225                  * SACK in continues as we would expect.
11226                  */
11227                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &no_extra, &moved_two, segsiz);
11228                 if (acked) {
11229                         rack->r_wanted_output = 1;
11230                         changed += acked;
11231                 }
11232                 if (num_sack_blks == 1) {
11233                         /*
11234                          * This is what we would expect from
11235                          * a normal implementation to happen
11236                          * after we have retransmitted the FR,
11237                          * i.e the sack-filter pushes down
11238                          * to 1 block and the next to be retransmitted
11239                          * is the sequence in the sack block (has more
11240                          * are acked). Count this as ACK'd data to boost
11241                          * up the chances of recovering any false positives.
11242                          */
11243                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
11244                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
11245                         counter_u64_add(rack_express_sack, 1);
11246                         if (rack->r_ctl.ack_count > 0xfff00000) {
11247                                 /*
11248                                  * reduce the number to keep us under
11249                                  * a uint32_t.
11250                                  */
11251                                 rack->r_ctl.ack_count /= 2;
11252                                 rack->r_ctl.sack_count /= 2;
11253                         }
11254                         if (moved_two) {
11255                                 /*
11256                                  * If we did not get a SACK for at least a MSS and
11257                                  * had to move at all, or if we moved more than our
11258                                  * threshold, it counts against the "extra" move.
11259                                  */
11260                                 rack->r_ctl.sack_moved_extra += moved_two;
11261                                 rack->r_ctl.sack_noextra_move += no_extra;
11262                                 counter_u64_add(rack_move_some, 1);
11263                         } else {
11264                                 /*
11265                                  * else we did not have to move
11266                                  * any more than we would expect.
11267                                  */
11268                                 rack->r_ctl.sack_noextra_move += no_extra;
11269                                 rack->r_ctl.sack_noextra_move++;
11270                                 counter_u64_add(rack_move_none, 1);
11271                         }
11272                         if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
11273                             (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
11274                                 rack->r_ctl.sack_moved_extra /= 2;
11275                                 rack->r_ctl.sack_noextra_move /= 2;
11276                         }
11277                         goto out_with_totals;
11278                 } else {
11279                         /*
11280                          * Start the loop through the
11281                          * rest of blocks, past the first block.
11282                          */
11283                         loop_start = 1;
11284                 }
11285         }
11286         counter_u64_add(rack_sack_total, 1);
11287         rsm = rack->r_ctl.rc_sacklast;
11288         for (i = loop_start; i < num_sack_blks; i++) {
11289                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &no_extra, &moved_two, segsiz);
11290                 if (acked) {
11291                         rack->r_wanted_output = 1;
11292                         changed += acked;
11293                 }
11294                 if (moved_two) {
11295                         /*
11296                          * If we did not get a SACK for at least a MSS and
11297                          * had to move at all, or if we moved more than our
11298                          * threshold, it counts against the "extra" move.
11299                          */
11300                         rack->r_ctl.sack_moved_extra += moved_two;
11301                         rack->r_ctl.sack_noextra_move += no_extra;
11302                         counter_u64_add(rack_move_some, 1);
11303                 } else {
11304                         /*
11305                          * else we did not have to move
11306                          * any more than we would expect.
11307                          */
11308                         rack->r_ctl.sack_noextra_move += no_extra;
11309                         rack->r_ctl.sack_noextra_move++;
11310                         counter_u64_add(rack_move_none, 1);
11311                 }
11312                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
11313                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
11314                         rack->r_ctl.sack_moved_extra /= 2;
11315                         rack->r_ctl.sack_noextra_move /= 2;
11316                 }
11317                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
11318                         /*
11319                          * If the SACK was not a full MSS then
11320                          * we add to sack_count the number of
11321                          * MSS's (or possibly more than
11322                          * a MSS if its a TSO send) we had to skip by.
11323                          */
11324                         rack->r_ctl.sack_count += moved_two;
11325                         if (rack->r_ctl.sack_count > 0xfff00000) {
11326                                 rack->r_ctl.ack_count /= 2;
11327                                 rack->r_ctl.sack_count /= 2;
11328                         }
11329                         counter_u64_add(rack_sack_total, moved_two);
11330                 }
11331                 /*
11332                  * Now we need to setup for the next
11333                  * round. First we make sure we won't
11334                  * exceed the size of our uint32_t on
11335                  * the various counts, and then clear out
11336                  * moved_two.
11337                  */
11338                 moved_two = 0;
11339                 no_extra = 0;
11340         }
11341 out_with_totals:
11342         if (num_sack_blks > 1) {
11343                 /*
11344                  * You get an extra stroke if
11345                  * you have more than one sack-blk, this
11346                  * could be where we are skipping forward
11347                  * and the sack-filter is still working, or
11348                  * it could be an attacker constantly
11349                  * moving us.
11350                  */
11351                 rack->r_ctl.sack_moved_extra++;
11352                 counter_u64_add(rack_move_some, 1);
11353         }
11354 out:
11355 #ifdef TCP_SAD_DETECTION
11356         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
11357 #endif
11358         if (changed) {
11359                 /* Something changed cancel the rack timer */
11360                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11361         }
11362         tsused = tcp_get_usecs(NULL);
11363         rsm = tcp_rack_output(tp, rack, tsused);
11364         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11365             rsm &&
11366             ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11367                 /* Enter recovery */
11368                 entered_recovery = 1;
11369                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
11370                 /*
11371                  * When we enter recovery we need to assure we send
11372                  * one packet.
11373                  */
11374                 if (rack->rack_no_prr == 0) {
11375                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11376                         rack_log_to_prr(rack, 8, 0, __LINE__);
11377                 }
11378                 rack->r_timer_override = 1;
11379                 rack->r_early = 0;
11380                 rack->r_ctl.rc_agg_early = 0;
11381         } else if (IN_FASTRECOVERY(tp->t_flags) &&
11382                    rsm &&
11383                    (rack->r_rr_config == 3)) {
11384                 /*
11385                  * Assure we can output and we get no
11386                  * remembered pace time except the retransmit.
11387                  */
11388                 rack->r_timer_override = 1;
11389                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11390                 rack->r_ctl.rc_resend = rsm;
11391         }
11392         if (IN_FASTRECOVERY(tp->t_flags) &&
11393             (rack->rack_no_prr == 0) &&
11394             (entered_recovery == 0)) {
11395                 rack_update_prr(tp, rack, changed, th_ack);
11396                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11397                      ((tcp_in_hpts(rack->rc_inp) == 0) &&
11398                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11399                         /*
11400                          * If you are pacing output you don't want
11401                          * to override.
11402                          */
11403                         rack->r_early = 0;
11404                         rack->r_ctl.rc_agg_early = 0;
11405                         rack->r_timer_override = 1;
11406                 }
11407         }
11408 }
11409
11410 static void
11411 rack_strike_dupack(struct tcp_rack *rack)
11412 {
11413         struct rack_sendmap *rsm;
11414
11415         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11416         while (rsm) {
11417                 /*
11418                  * We need to skip anything already set
11419                  * to be retransmitted.
11420                  */
11421                 if ((rsm->r_dupack >= DUP_ACK_THRESHOLD)  ||
11422                     (rsm->r_flags & RACK_MUST_RXT)) {
11423                         rsm = TAILQ_NEXT(rsm, r_tnext);
11424                         continue;
11425                 }
11426                 break;
11427         }
11428         if (rsm && (rsm->r_dupack < 0xff)) {
11429                 rsm->r_dupack++;
11430                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11431                         struct timeval tv;
11432                         uint32_t cts;
11433                         /*
11434                          * Here we see if we need to retransmit. For
11435                          * a SACK type connection if enough time has passed
11436                          * we will get a return of the rsm. For a non-sack
11437                          * connection we will get the rsm returned if the
11438                          * dupack value is 3 or more.
11439                          */
11440                         cts = tcp_get_usecs(&tv);
11441                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11442                         if (rack->r_ctl.rc_resend != NULL) {
11443                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11444                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11445                                                          rack->rc_tp->snd_una, __LINE__);
11446                                 }
11447                                 rack->r_wanted_output = 1;
11448                                 rack->r_timer_override = 1;
11449                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11450                         }
11451                 } else {
11452                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11453                 }
11454         }
11455 }
11456
11457 static void
11458 rack_check_bottom_drag(struct tcpcb *tp,
11459                        struct tcp_rack *rack,
11460                        struct socket *so)
11461 {
11462         uint32_t segsiz, minseg;
11463
11464         segsiz = ctf_fixed_maxseg(tp);
11465         minseg = segsiz;
11466         if (tp->snd_max == tp->snd_una) {
11467                 /*
11468                  * We are doing dynamic pacing and we are way
11469                  * under. Basically everything got acked while
11470                  * we were still waiting on the pacer to expire.
11471                  *
11472                  * This means we need to boost the b/w in
11473                  * addition to any earlier boosting of
11474                  * the multiplier.
11475                  */
11476                 uint64_t lt_bw;
11477
11478                 lt_bw = rack_get_lt_bw(rack);
11479                 rack->rc_dragged_bottom = 1;
11480                 rack_validate_multipliers_at_or_above100(rack);
11481                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11482                     (lt_bw > 0)) {
11483                         /*
11484                          * Lets use the long-term b/w we have
11485                          * been getting as a base.
11486                          */
11487                         if (rack->rc_gp_filled == 0) {
11488                                 if (lt_bw > ONE_POINT_TWO_MEG) {
11489                                         /*
11490                                          * If we have no measurement
11491                                          * don't let us set in more than
11492                                          * 1.2Mbps. If we are still too
11493                                          * low after pacing with this we
11494                                          * will hopefully have a max b/w
11495                                          * available to sanity check things.
11496                                          */
11497                                         lt_bw = ONE_POINT_TWO_MEG;
11498                                 }
11499                                 rack->r_ctl.rc_rtt_diff = 0;
11500                                 rack->r_ctl.gp_bw = lt_bw;
11501                                 rack->rc_gp_filled = 1;
11502                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11503                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
11504                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11505                         } else if (lt_bw > rack->r_ctl.gp_bw) {
11506                                 rack->r_ctl.rc_rtt_diff = 0;
11507                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11508                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
11509                                 rack->r_ctl.gp_bw = lt_bw;
11510                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11511                         } else
11512                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
11513                         if ((rack->gp_ready == 0) &&
11514                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11515                                 /* We have enough measurements now */
11516                                 rack->gp_ready = 1;
11517                                 if ((rack->rc_always_pace && (rack->use_fixed_rate == 0)) ||
11518                                     rack->rack_hibeta)
11519                                         rack_set_cc_pacing(rack);
11520                                 if (rack->defer_options)
11521                                         rack_apply_deferred_options(rack);
11522                         }
11523                 } else {
11524                         /*
11525                          * zero rtt possibly?, settle for just an old increase.
11526                          */
11527                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
11528                 }
11529         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11530                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11531                                                minseg)) &&
11532                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11533                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11534                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11535                     (segsiz * rack_req_segs))) {
11536                 /*
11537                  * We are doing dynamic GP pacing and
11538                  * we have everything except 1MSS or less
11539                  * bytes left out. We are still pacing away.
11540                  * And there is data that could be sent, This
11541                  * means we are inserting delayed ack time in
11542                  * our measurements because we are pacing too slow.
11543                  */
11544                 rack_validate_multipliers_at_or_above100(rack);
11545                 rack->rc_dragged_bottom = 1;
11546                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
11547         }
11548 }
11549
11550 #ifdef TCP_REQUEST_TRK
11551 static void
11552 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11553                 struct http_sendfile_track *cur, uint8_t mod, int line, int err)
11554 {
11555         int do_log;
11556
11557         do_log = tcp_bblogging_on(rack->rc_tp);
11558         if (do_log == 0) {
11559                 if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11560                         return;
11561                 /* We only allow the three below with point logging on */
11562                 if ((mod != HYBRID_LOG_RULES_APP) &&
11563                     (mod != HYBRID_LOG_RULES_SET) &&
11564                     (mod != HYBRID_LOG_REQ_COMP))
11565                         return;
11566
11567         }
11568         if (do_log) {
11569                 union tcp_log_stackspecific log;
11570                 struct timeval tv;
11571
11572                 /* Convert our ms to a microsecond */
11573                 memset(&log, 0, sizeof(log));
11574                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11575                 log.u_bbr.flex1 = seq;
11576                 log.u_bbr.cwnd_gain = line;
11577                 if (cur != NULL) {
11578                         uint64_t off;
11579
11580                         log.u_bbr.flex2 = cur->start_seq;
11581                         log.u_bbr.flex3 = cur->end_seq;
11582                         log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11583                         log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11584                         log.u_bbr.flex6 = cur->flags;
11585                         log.u_bbr.pkts_out = cur->hybrid_flags;
11586                         log.u_bbr.rttProp = cur->timestamp;
11587                         log.u_bbr.cur_del_rate = cur->cspr;
11588                         log.u_bbr.bw_inuse = cur->start;
11589                         log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11590                         log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11591                         log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11592                         log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11593                         log.u_bbr.bbr_state = 1;
11594                         off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_http_info[0]);
11595                         log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct http_sendfile_track));
11596                 } else {
11597                         log.u_bbr.flex2 = err;
11598                 }
11599                 /*
11600                  * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11601                  */
11602                 log.u_bbr.flex7 = rack->rc_catch_up;
11603                 log.u_bbr.flex7 <<= 1;
11604                 log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11605                 log.u_bbr.flex7 <<= 1;
11606                 log.u_bbr.flex7 |= rack->dgp_on;
11607                 log.u_bbr.flex8 = mod;
11608                 log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11609                 log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11610                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11611                 log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
11612                 log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
11613                 log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
11614                 tcp_log_event(rack->rc_tp, NULL,
11615                     &rack->rc_inp->inp_socket->so_rcv,
11616                     &rack->rc_inp->inp_socket->so_snd,
11617                     TCP_HYBRID_PACING_LOG, 0,
11618                     0, &log, false, NULL, __func__, __LINE__, &tv);
11619         }
11620 }
11621 #endif
11622
11623 #ifdef TCP_REQUEST_TRK
11624 static void
11625 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len)
11626 {
11627         struct http_sendfile_track *rc_cur;
11628         struct tcpcb *tp;
11629         int err = 0;
11630
11631         rc_cur = tcp_http_find_req_for_seq(rack->rc_tp, seq);
11632         if (rc_cur == NULL) {
11633                 /* If not in the beginning what about the end piece */
11634                 rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11635                 rc_cur = tcp_http_find_req_for_seq(rack->rc_tp, (seq + len - 1));
11636         } else {
11637                 err = 12345;
11638         }
11639         /* If we find no parameters we are in straight DGP mode */
11640         if(rc_cur == NULL) {
11641                 /* None found for this seq, just DGP for now */
11642                 rack->r_ctl.client_suggested_maxseg = 0;
11643                 rack->rc_catch_up = 0;
11644                 rack->r_ctl.bw_rate_cap = 0;
11645                 rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11646                 if (rack->r_ctl.rc_last_sft) {
11647                         rack->r_ctl.rc_last_sft = NULL;
11648                 }
11649                 return;
11650         }
11651         /*
11652          * Ok if we have a new entry *or* have never
11653          * set up an entry we need to proceed. If
11654          * we have already set it up this entry we
11655          * just continue along with what we already
11656          * setup.
11657          */
11658         tp = rack->rc_tp;
11659         if ((rack->r_ctl.rc_last_sft != NULL) &&
11660             (rack->r_ctl.rc_last_sft == rc_cur)) {
11661                 /* Its already in place */
11662                 rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
11663                 return;
11664         }
11665         if (rack->rc_hybrid_mode == 0) {
11666                 rack->r_ctl.rc_last_sft = rc_cur;
11667                 rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11668                 return;
11669         }
11670         if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
11671                 /* Compensate for all the header overhead's */
11672                 rack->r_ctl.bw_rate_cap = rack_compensate_for_linerate(rack, rc_cur->cspr);
11673         } else
11674                 rack->r_ctl.bw_rate_cap = 0;
11675         if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
11676                 rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
11677         else
11678                 rack->r_ctl.client_suggested_maxseg = 0;
11679         if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
11680             (rc_cur->cspr > 0)) {
11681                 uint64_t len;
11682
11683                 rack->rc_catch_up = 1;
11684                 /*
11685                  * Calculate the deadline time, first set the
11686                  * time to when the request arrived.
11687                  */
11688                 rc_cur->deadline = rc_cur->localtime;
11689                 /*
11690                  * Next calculate the length and compensate for
11691                  * TLS if need be.
11692                  */
11693                 len = rc_cur->end - rc_cur->start;
11694                 if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
11695                         /*
11696                          * This session is doing TLS. Take a swag guess
11697                          * at the overhead.
11698                          */
11699                         len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
11700                 }
11701                 /*
11702                  * Now considering the size, and the cspr, what is the time that
11703                  * would be required at the cspr rate. Here we use the raw
11704                  * cspr value since the client only looks at the raw data. We
11705                  * do use len which includes TLS overhead, but not the TCP/IP etc.
11706                  * That will get made up for in the CU pacing rate set.
11707                  */
11708                 len *= HPTS_USEC_IN_SEC;
11709                 len /= rc_cur->cspr;
11710                 rc_cur->deadline += len;
11711         } else {
11712                 rack->rc_catch_up = 0;
11713                 rc_cur->deadline = 0;
11714         }
11715         if (rack->r_ctl.client_suggested_maxseg != 0) {
11716                 /*
11717                  * We need to reset the max pace segs if we have a
11718                  * client_suggested_maxseg.
11719                  */
11720                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
11721         }
11722         rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11723         /* Remember it for next time and for CU mode */
11724         rack->r_ctl.rc_last_sft = rc_cur;
11725 }
11726 #endif
11727
11728 static void
11729 rack_chk_http_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11730 {
11731 #ifdef TCP_REQUEST_TRK
11732         struct http_sendfile_track *ent;
11733
11734         ent = rack->r_ctl.rc_last_sft;
11735         if ((ent == NULL) ||
11736             (ent->flags == TCP_HTTP_TRACK_FLG_EMPTY) ||
11737             (SEQ_GEQ(seq, ent->end_seq))) {
11738                 /* Time to update the track. */
11739                 rack_set_dgp_hybrid_mode(rack, seq, len);
11740                 ent = rack->r_ctl.rc_last_sft;
11741         }
11742         /* Out of all */
11743         if (ent == NULL) {
11744                 return;
11745         }
11746         if (SEQ_LT(ent->end_seq, (seq + len))) {
11747                 /*
11748                  * This is the case where our end_seq guess
11749                  * was wrong. This is usually due to TLS having
11750                  * more bytes then our guess. It could also be the
11751                  * case that the client sent in two requests closely
11752                  * and the SB is full of both so we are sending part
11753                  * of each (end|beg). In such a case lets move this
11754                  * guys end to match the end of this send. That
11755                  * way it will complete when all of it is acked.
11756                  */
11757                 ent->end_seq = (seq + len);
11758                 rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent);
11759         }
11760         /* Now validate we have set the send time of this one */
11761         if ((ent->flags & TCP_HTTP_TRACK_FLG_FSND) == 0) {
11762                 ent->flags |= TCP_HTTP_TRACK_FLG_FSND;
11763                 ent->first_send = cts;
11764                 ent->sent_at_fs = rack->rc_tp->t_sndbytes;
11765                 ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11766         }
11767 #endif
11768 }
11769
11770 static void
11771 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
11772 {
11773         /*
11774          * The fast output path is enabled and we
11775          * have moved the cumack forward. Lets see if
11776          * we can expand forward the fast path length by
11777          * that amount. What we would ideally like to
11778          * do is increase the number of bytes in the
11779          * fast path block (left_to_send) by the
11780          * acked amount. However we have to gate that
11781          * by two factors:
11782          * 1) The amount outstanding and the rwnd of the peer
11783          *    (i.e. we don't want to exceed the rwnd of the peer).
11784          *    <and>
11785          * 2) The amount of data left in the socket buffer (i.e.
11786          *    we can't send beyond what is in the buffer).
11787          *
11788          * Note that this does not take into account any increase
11789          * in the cwnd. We will only extend the fast path by
11790          * what was acked.
11791          */
11792         uint32_t new_total, gating_val;
11793
11794         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
11795         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
11796                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
11797         if (new_total <= gating_val) {
11798                 /* We can increase left_to_send by the acked amount */
11799                 counter_u64_add(rack_extended_rfo, 1);
11800                 rack->r_ctl.fsb.left_to_send = new_total;
11801                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
11802                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
11803                          rack, rack->r_ctl.fsb.left_to_send,
11804                          sbavail(&rack->rc_inp->inp_socket->so_snd),
11805                          (tp->snd_max - tp->snd_una)));
11806
11807         }
11808 }
11809
11810 static void
11811 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
11812 {
11813         /*
11814          * Here any sendmap entry that points to the
11815          * beginning mbuf must be adjusted to the correct
11816          * offset. This must be called with:
11817          * 1) The socket buffer locked
11818          * 2) snd_una adjusted to its new position.
11819          *
11820          * Note that (2) implies rack_ack_received has also
11821          * been called and all the sbcut's have been done.
11822          *
11823          * We grab the first mbuf in the socket buffer and
11824          * then go through the front of the sendmap, recalculating
11825          * the stored offset for any sendmap entry that has
11826          * that mbuf. We must use the sb functions to do this
11827          * since its possible an add was done has well as
11828          * the subtraction we may have just completed. This should
11829          * not be a penalty though, since we just referenced the sb
11830          * to go in and trim off the mbufs that we freed (of course
11831          * there will be a penalty for the sendmap references though).
11832          *
11833          * Note also with INVARIANT on, we validate with a KASSERT
11834          * that the first sendmap entry has a soff of 0.
11835          *
11836          */
11837         struct mbuf *m;
11838         struct rack_sendmap *rsm;
11839         tcp_seq snd_una;
11840 #ifdef INVARIANTS
11841         int first_processed = 0;
11842 #endif
11843
11844         snd_una = rack->rc_tp->snd_una;
11845         SOCKBUF_LOCK_ASSERT(sb);
11846         m = sb->sb_mb;
11847         rsm = tqhash_min(rack->r_ctl.tqh);
11848         if ((rsm == NULL) || (m == NULL)) {
11849                 /* Nothing outstanding */
11850                 return;
11851         }
11852         /* The very first RSM's mbuf must point to the head mbuf in the sb */
11853         KASSERT((rsm->m == m),
11854                 ("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
11855                  rack, sb, rsm));
11856         while (rsm->m && (rsm->m == m)) {
11857                 /* one to adjust */
11858 #ifdef INVARIANTS
11859                 struct mbuf *tm;
11860                 uint32_t soff;
11861
11862                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
11863                 if ((rsm->orig_m_len != m->m_len) ||
11864                     (rsm->orig_t_space != M_TRAILINGROOM(m))){
11865                         rack_adjust_orig_mlen(rsm);
11866                 }
11867                 if (first_processed == 0) {
11868                         KASSERT((rsm->soff == 0),
11869                                 ("Rack:%p rsm:%p -- rsm at head but soff not zero",
11870                                  rack, rsm));
11871                         first_processed = 1;
11872                 }
11873                 if ((rsm->soff != soff) || (rsm->m != tm)) {
11874                         /*
11875                          * This is not a fatal error, we anticipate it
11876                          * might happen (the else code), so we count it here
11877                          * so that under invariant we can see that it really
11878                          * does happen.
11879                          */
11880                         counter_u64_add(rack_adjust_map_bw, 1);
11881                 }
11882                 rsm->m = tm;
11883                 rsm->soff = soff;
11884                 if (tm) {
11885                         rsm->orig_m_len = rsm->m->m_len;
11886                         rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11887                 } else {
11888                         rsm->orig_m_len = 0;
11889                         rsm->orig_t_space = 0;
11890                 }
11891 #else
11892                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
11893                 if (rsm->m) {
11894                         rsm->orig_m_len = rsm->m->m_len;
11895                         rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11896                 } else {
11897                         rsm->orig_m_len = 0;
11898                         rsm->orig_t_space = 0;
11899                 }
11900 #endif
11901                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
11902                 if (rsm == NULL)
11903                         break;
11904         }
11905 }
11906
11907 #ifdef TCP_REQUEST_TRK
11908 static inline void
11909 rack_http_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
11910 {
11911         struct http_sendfile_track *ent;
11912         int i;
11913
11914         if ((rack->rc_hybrid_mode == 0) &&
11915             (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
11916                 /*
11917                  * Just do normal completions hybrid pacing is not on
11918                  * and CLDL is off as well.
11919                  */
11920                 tcp_http_check_for_comp(rack->rc_tp, th_ack);
11921                 return;
11922         }
11923         /*
11924          * Originally I was just going to find the th_ack associated
11925          * with an entry. But then I realized a large strech ack could
11926          * in theory ack two or more requests at once. So instead we
11927          * need to find all entries that are completed by th_ack not
11928          * just a single entry and do our logging.
11929          */
11930         ent = tcp_http_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11931         while (ent != NULL) {
11932                 /*
11933                  * We may be doing hybrid pacing or CLDL and need more details possibly
11934                  * so we do it manually instead of calling
11935                  * tcp_http_check_for_comp()
11936                  */
11937                 uint64_t laa, tim, data, cbw, ftim;
11938
11939                 /* Ok this ack frees it */
11940                 rack_log_hybrid(rack, th_ack,
11941                                 ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
11942                 /* calculate the time based on the ack arrival */
11943                 data = ent->end - ent->start;
11944                 laa = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
11945                 if (ent->flags & TCP_HTTP_TRACK_FLG_FSND) {
11946                         if (ent->first_send > ent->localtime)
11947                                 ftim = ent->first_send;
11948                         else
11949                                 ftim = ent->localtime;
11950                 } else {
11951                         /* TSNH */
11952                         ftim = ent->localtime;
11953                 }
11954                 if (laa > ent->localtime)
11955                         tim = laa - ftim;
11956                 else
11957                         tim = 0;
11958                 cbw = data * HPTS_USEC_IN_SEC;
11959                 if (tim > 0)
11960                         cbw /= tim;
11961                 else
11962                         cbw = 0;
11963                 rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent);
11964                 /*
11965                  * Check to see if we are freeing what we are pointing to send wise
11966                  * if so be sure to NULL the pointer so we know we are no longer
11967                  * set to anything.
11968                  */
11969                 if (ent == rack->r_ctl.rc_last_sft)
11970                         rack->r_ctl.rc_last_sft = NULL;
11971                 /* Generate the log that the tcp_netflix call would have */
11972                 tcp_http_log_req_info(rack->rc_tp, ent,
11973                                       i, TCP_HTTP_REQ_LOG_FREED, 0, 0);
11974                 /* Free it and see if there is another one */
11975                 tcp_http_free_a_slot(rack->rc_tp, ent);
11976                 ent = tcp_http_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11977         }
11978 }
11979 #endif
11980
11981
11982 /*
11983  * Return value of 1, we do not need to call rack_process_data().
11984  * return value of 0, rack_process_data can be called.
11985  * For ret_val if its 0 the TCP is locked, if its non-zero
11986  * its unlocked and probably unsafe to touch the TCB.
11987  */
11988 static int
11989 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11990     struct tcpcb *tp, struct tcpopt *to,
11991     uint32_t tiwin, int32_t tlen,
11992     int32_t * ofia, int32_t thflags, int32_t *ret_val)
11993 {
11994         int32_t ourfinisacked = 0;
11995         int32_t nsegs, acked_amount;
11996         int32_t acked;
11997         struct mbuf *mfree;
11998         struct tcp_rack *rack;
11999         int32_t under_pacing = 0;
12000         int32_t recovery = 0;
12001
12002         INP_WLOCK_ASSERT(tptoinpcb(tp));
12003
12004         rack = (struct tcp_rack *)tp->t_fb_ptr;
12005         if (SEQ_GT(th->th_ack, tp->snd_max)) {
12006                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
12007                                       &rack->r_ctl.challenge_ack_ts,
12008                                       &rack->r_ctl.challenge_ack_cnt);
12009                 rack->r_wanted_output = 1;
12010                 return (1);
12011         }
12012         if (rack->gp_ready &&
12013             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12014                 under_pacing = 1;
12015         }
12016         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
12017                 int in_rec, dup_ack_struck = 0;
12018                 int dsack_seen = 0, sacks_seen = 0;
12019
12020                 in_rec = IN_FASTRECOVERY(tp->t_flags);
12021                 if (rack->rc_in_persist) {
12022                         tp->t_rxtshift = 0;
12023                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12024                                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12025                 }
12026
12027                 if ((th->th_ack == tp->snd_una) &&
12028                     (tiwin == tp->snd_wnd) &&
12029                     ((to->to_flags & TOF_SACK) == 0)) {
12030                         rack_strike_dupack(rack);
12031                         dup_ack_struck = 1;
12032                 }
12033                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
12034                              dup_ack_struck, &dsack_seen, &sacks_seen);
12035                 if ((rack->sack_attack_disable > 0) &&
12036                     (th->th_ack == tp->snd_una) &&
12037                     (tiwin == tp->snd_wnd) &&
12038                     (dsack_seen == 0) &&
12039                     (sacks_seen > 0)) {
12040                         /*
12041                          * If sacks have been disabled we may
12042                          * want to strike a dup-ack "ignoring" the
12043                          * sack as long as the sack was not a "dsack". Note
12044                          * that if no sack is sent (TOF_SACK is off) then the
12045                          * normal dsack code above rack_log_ack() would have
12046                          * already struck. So this is just to catch the case
12047                          * were we are ignoring sacks from this guy due to
12048                          * it being a suspected attacker.
12049                          */
12050                         rack_strike_dupack(rack);
12051                 }
12052
12053         }
12054         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12055                 /*
12056                  * Old ack, behind (or duplicate to) the last one rcv'd
12057                  * Note: We mark reordering is occuring if its
12058                  * less than and we have not closed our window.
12059                  */
12060                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
12061                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12062                         if (rack->r_ctl.rc_reorder_ts == 0)
12063                                 rack->r_ctl.rc_reorder_ts = 1;
12064                 }
12065                 return (0);
12066         }
12067         /*
12068          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
12069          * something we sent.
12070          */
12071         if (tp->t_flags & TF_NEEDSYN) {
12072                 /*
12073                  * T/TCP: Connection was half-synchronized, and our SYN has
12074                  * been ACK'd (so connection is now fully synchronized).  Go
12075                  * to non-starred state, increment snd_una for ACK of SYN,
12076                  * and check if we can do window scaling.
12077                  */
12078                 tp->t_flags &= ~TF_NEEDSYN;
12079                 tp->snd_una++;
12080                 /* Do window scaling? */
12081                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12082                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12083                         tp->rcv_scale = tp->request_r_scale;
12084                         /* Send window already scaled. */
12085                 }
12086         }
12087         nsegs = max(1, m->m_pkthdr.lro_nsegs);
12088
12089         acked = BYTES_THIS_ACK(tp, th);
12090         if (acked) {
12091                 /*
12092                  * Any time we move the cum-ack forward clear
12093                  * keep-alive tied probe-not-answered. The
12094                  * persists clears its own on entry.
12095                  */
12096                 rack->probe_not_answered = 0;
12097         }
12098         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12099         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12100         /*
12101          * If we just performed our first retransmit, and the ACK arrives
12102          * within our recovery window, then it was a mistake to do the
12103          * retransmit in the first place.  Recover our original cwnd and
12104          * ssthresh, and proceed to transmit where we left off.
12105          */
12106         if ((tp->t_flags & TF_PREVVALID) &&
12107             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12108                 tp->t_flags &= ~TF_PREVVALID;
12109                 if (tp->t_rxtshift == 1 &&
12110                     (int)(ticks - tp->t_badrxtwin) < 0)
12111                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12112         }
12113         if (acked) {
12114                 /* assure we are not backed off */
12115                 tp->t_rxtshift = 0;
12116                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12117                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12118                 rack->rc_tlp_in_progress = 0;
12119                 rack->r_ctl.rc_tlp_cnt_out = 0;
12120                 /*
12121                  * If it is the RXT timer we want to
12122                  * stop it, so we can restart a TLP.
12123                  */
12124                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12125                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12126 #ifdef TCP_REQUEST_TRK
12127                 rack_http_check_for_comp(rack, th->th_ack);
12128 #endif
12129         }
12130         /*
12131          * If we have a timestamp reply, update smoothed round trip time. If
12132          * no timestamp is present but transmit timer is running and timed
12133          * sequence number was acked, update smoothed round trip time. Since
12134          * we now have an rtt measurement, cancel the timer backoff (cf.,
12135          * Phil Karn's retransmit alg.). Recompute the initial retransmit
12136          * timer.
12137          *
12138          * Some boxes send broken timestamp replies during the SYN+ACK
12139          * phase, ignore timestamps of 0 or we could calculate a huge RTT
12140          * and blow up the retransmit timer.
12141          */
12142         /*
12143          * If all outstanding data is acked, stop retransmit timer and
12144          * remember to restart (more output or persist). If there is more
12145          * data to be acked, restart retransmit timer, using current
12146          * (possibly backed-off) value.
12147          */
12148         if (acked == 0) {
12149                 if (ofia)
12150                         *ofia = ourfinisacked;
12151                 return (0);
12152         }
12153         if (IN_RECOVERY(tp->t_flags)) {
12154                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
12155                     (SEQ_LT(th->th_ack, tp->snd_max))) {
12156                         tcp_rack_partialack(tp);
12157                 } else {
12158                         rack_post_recovery(tp, th->th_ack);
12159                         recovery = 1;
12160                 }
12161         }
12162         /*
12163          * Let the congestion control algorithm update congestion control
12164          * related information. This typically means increasing the
12165          * congestion window.
12166          */
12167         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
12168         SOCKBUF_LOCK(&so->so_snd);
12169         acked_amount = min(acked, (int)sbavail(&so->so_snd));
12170         tp->snd_wnd -= acked_amount;
12171         mfree = sbcut_locked(&so->so_snd, acked_amount);
12172         if ((sbused(&so->so_snd) == 0) &&
12173             (acked > acked_amount) &&
12174             (tp->t_state >= TCPS_FIN_WAIT_1) &&
12175             (tp->t_flags & TF_SENTFIN)) {
12176                 /*
12177                  * We must be sure our fin
12178                  * was sent and acked (we can be
12179                  * in FIN_WAIT_1 without having
12180                  * sent the fin).
12181                  */
12182                 ourfinisacked = 1;
12183         }
12184         tp->snd_una = th->th_ack;
12185         /* wakeups? */
12186         if (acked_amount && sbavail(&so->so_snd))
12187                 rack_adjust_sendmap_head(rack, &so->so_snd);
12188         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12189         /* NB: sowwakeup_locked() does an implicit unlock. */
12190         sowwakeup_locked(so);
12191         /* now check the rxt clamps */
12192         if ((recovery == 1) &&
12193             (rack->excess_rxt_on) &&
12194             (rack->r_cwnd_was_clamped == 0))  {
12195                 do_rack_excess_rxt(tp, rack);
12196         } else if (rack->r_cwnd_was_clamped)
12197                 do_rack_check_for_unclamp(tp, rack);
12198         m_freem(mfree);
12199         if (SEQ_GT(tp->snd_una, tp->snd_recover))
12200                 tp->snd_recover = tp->snd_una;
12201
12202         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
12203                 tp->snd_nxt = tp->snd_una;
12204         }
12205         if (under_pacing &&
12206             (rack->use_fixed_rate == 0) &&
12207             (rack->in_probe_rtt == 0) &&
12208             rack->rc_gp_dyn_mul &&
12209             rack->rc_always_pace) {
12210                 /* Check if we are dragging bottom */
12211                 rack_check_bottom_drag(tp, rack, so);
12212         }
12213         if (tp->snd_una == tp->snd_max) {
12214                 /* Nothing left outstanding */
12215                 tp->t_flags &= ~TF_PREVVALID;
12216                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12217                 rack->r_ctl.retran_during_recovery = 0;
12218                 rack->r_ctl.dsack_byte_cnt = 0;
12219                 if (rack->r_ctl.rc_went_idle_time == 0)
12220                         rack->r_ctl.rc_went_idle_time = 1;
12221                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12222                 if (sbavail(&tptosocket(tp)->so_snd) == 0)
12223                         tp->t_acktime = 0;
12224                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12225                 rack->rc_suspicious = 0;
12226                 /* Set need output so persist might get set */
12227                 rack->r_wanted_output = 1;
12228                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12229                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12230                     (sbavail(&so->so_snd) == 0) &&
12231                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12232                         /*
12233                          * The socket was gone and the
12234                          * peer sent data (now or in the past), time to
12235                          * reset him.
12236                          */
12237                         *ret_val = 1;
12238                         /* tcp_close will kill the inp pre-log the Reset */
12239                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12240                         tp = tcp_close(tp);
12241                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
12242                         return (1);
12243                 }
12244         }
12245         if (ofia)
12246                 *ofia = ourfinisacked;
12247         return (0);
12248 }
12249
12250
12251 static void
12252 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12253                   int dir, uint32_t flags, struct rack_sendmap *rsm)
12254 {
12255         if (tcp_bblogging_on(rack->rc_tp)) {
12256                 union tcp_log_stackspecific log;
12257                 struct timeval tv;
12258
12259                 memset(&log, 0, sizeof(log));
12260                 log.u_bbr.flex1 = cnt;
12261                 log.u_bbr.flex2 = split;
12262                 log.u_bbr.flex3 = out;
12263                 log.u_bbr.flex4 = line;
12264                 log.u_bbr.flex5 = rack->r_must_retran;
12265                 log.u_bbr.flex6 = flags;
12266                 log.u_bbr.flex7 = rack->rc_has_collapsed;
12267                 log.u_bbr.flex8 = dir;  /*
12268                                          * 1 is collapsed, 0 is uncollapsed,
12269                                          * 2 is log of a rsm being marked, 3 is a split.
12270                                          */
12271                 if (rsm == NULL)
12272                         log.u_bbr.rttProp = 0;
12273                 else
12274                         log.u_bbr.rttProp = (uint64_t)rsm;
12275                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12276                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12277                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
12278                     &rack->rc_inp->inp_socket->so_rcv,
12279                     &rack->rc_inp->inp_socket->so_snd,
12280                     TCP_RACK_LOG_COLLAPSE, 0,
12281                     0, &log, false, &tv);
12282         }
12283 }
12284
12285 static void
12286 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12287 {
12288         /*
12289          * Here all we do is mark the collapsed point and set the flag.
12290          * This may happen again and again, but there is no
12291          * sense splitting our map until we know where the
12292          * peer finally lands in the collapse.
12293          */
12294         tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12295         if ((rack->rc_has_collapsed == 0) ||
12296             (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12297                 counter_u64_add(rack_collapsed_win_seen, 1);
12298         rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12299         rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12300         rack->rc_has_collapsed = 1;
12301         rack->r_collapse_point_valid = 1;
12302         rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12303 }
12304
12305 static void
12306 rack_un_collapse_window(struct tcp_rack *rack, int line)
12307 {
12308         struct rack_sendmap *nrsm, *rsm;
12309         int cnt = 0, split = 0;
12310         int insret __diagused;
12311
12312
12313         tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12314         rack->rc_has_collapsed = 0;
12315         rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12316         if (rsm == NULL) {
12317                 /* Nothing to do maybe the peer ack'ed it all */
12318                 rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12319                 return;
12320         }
12321         /* Now do we need to split this one? */
12322         if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12323                 rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12324                                   rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12325                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12326                 if (nrsm == NULL) {
12327                         /* We can't get a rsm, mark all? */
12328                         nrsm = rsm;
12329                         goto no_split;
12330                 }
12331                 /* Clone it */
12332                 split = 1;
12333                 rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12334 #ifndef INVARIANTS
12335                 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12336 #else
12337                 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12338                         panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
12339                               nrsm, insret, rack, rsm);
12340                 }
12341 #endif
12342                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12343                                  rack->r_ctl.last_collapse_point, __LINE__);
12344                 if (rsm->r_in_tmap) {
12345                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12346                         nrsm->r_in_tmap = 1;
12347                 }
12348                 /*
12349                  * Set in the new RSM as the
12350                  * collapsed starting point
12351                  */
12352                 rsm = nrsm;
12353         }
12354
12355 no_split:
12356         TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm)  {
12357                 cnt++;
12358                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
12359                 rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12360                 cnt++;
12361         }
12362         if (cnt) {
12363                 counter_u64_add(rack_collapsed_win, 1);
12364         }
12365         rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12366 }
12367
12368 static void
12369 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12370                         int32_t tlen, int32_t tfo_syn)
12371 {
12372         if (DELAY_ACK(tp, tlen) || tfo_syn) {
12373                 rack_timer_cancel(tp, rack,
12374                                   rack->r_ctl.rc_rcvtime, __LINE__);
12375                 tp->t_flags |= TF_DELACK;
12376         } else {
12377                 rack->r_wanted_output = 1;
12378                 tp->t_flags |= TF_ACKNOW;
12379         }
12380 }
12381
12382 static void
12383 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12384 {
12385         /*
12386          * If fast output is in progress, lets validate that
12387          * the new window did not shrink on us and make it
12388          * so fast output should end.
12389          */
12390         if (rack->r_fast_output) {
12391                 uint32_t out;
12392
12393                 /*
12394                  * Calculate what we will send if left as is
12395                  * and compare that to our send window.
12396                  */
12397                 out = ctf_outstanding(tp);
12398                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12399                         /* ok we have an issue */
12400                         if (out >= tp->snd_wnd) {
12401                                 /* Turn off fast output the window is met or collapsed */
12402                                 rack->r_fast_output = 0;
12403                         } else {
12404                                 /* we have some room left */
12405                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12406                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12407                                         /* If not at least 1 full segment never mind */
12408                                         rack->r_fast_output = 0;
12409                                 }
12410                         }
12411                 }
12412         }
12413 }
12414
12415
12416 /*
12417  * Return value of 1, the TCB is unlocked and most
12418  * likely gone, return value of 0, the TCP is still
12419  * locked.
12420  */
12421 static int
12422 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12423     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12424     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12425 {
12426         /*
12427          * Update window information. Don't look at window if no ACK: TAC's
12428          * send garbage on first SYN.
12429          */
12430         int32_t nsegs;
12431         int32_t tfo_syn;
12432         struct tcp_rack *rack;
12433
12434         INP_WLOCK_ASSERT(tptoinpcb(tp));
12435
12436         rack = (struct tcp_rack *)tp->t_fb_ptr;
12437         nsegs = max(1, m->m_pkthdr.lro_nsegs);
12438         if ((thflags & TH_ACK) &&
12439             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12440             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12441             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12442                 /* keep track of pure window updates */
12443                 if (tlen == 0 &&
12444                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12445                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12446                 tp->snd_wnd = tiwin;
12447                 rack_validate_fo_sendwin_up(tp, rack);
12448                 tp->snd_wl1 = th->th_seq;
12449                 tp->snd_wl2 = th->th_ack;
12450                 if (tp->snd_wnd > tp->max_sndwnd)
12451                         tp->max_sndwnd = tp->snd_wnd;
12452                 rack->r_wanted_output = 1;
12453         } else if (thflags & TH_ACK) {
12454                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12455                         tp->snd_wnd = tiwin;
12456                         rack_validate_fo_sendwin_up(tp, rack);
12457                         tp->snd_wl1 = th->th_seq;
12458                         tp->snd_wl2 = th->th_ack;
12459                 }
12460         }
12461         if (tp->snd_wnd < ctf_outstanding(tp))
12462                 /* The peer collapsed the window */
12463                 rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12464         else if (rack->rc_has_collapsed)
12465                 rack_un_collapse_window(rack, __LINE__);
12466         if ((rack->r_collapse_point_valid) &&
12467             (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12468                 rack->r_collapse_point_valid = 0;
12469         /* Was persist timer active and now we have window space? */
12470         if ((rack->rc_in_persist != 0) &&
12471             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12472                                 rack->r_ctl.rc_pace_min_segs))) {
12473                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12474                 tp->snd_nxt = tp->snd_max;
12475                 /* Make sure we output to start the timer */
12476                 rack->r_wanted_output = 1;
12477         }
12478         /* Do we enter persists? */
12479         if ((rack->rc_in_persist == 0) &&
12480             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12481             TCPS_HAVEESTABLISHED(tp->t_state) &&
12482             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12483             sbavail(&tptosocket(tp)->so_snd) &&
12484             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12485                 /*
12486                  * Here the rwnd is less than
12487                  * the pacing size, we are established,
12488                  * nothing is outstanding, and there is
12489                  * data to send. Enter persists.
12490                  */
12491                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12492         }
12493         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12494                 m_freem(m);
12495                 return (0);
12496         }
12497         /*
12498          * don't process the URG bit, ignore them drag
12499          * along the up.
12500          */
12501         tp->rcv_up = tp->rcv_nxt;
12502
12503         /*
12504          * Process the segment text, merging it into the TCP sequencing
12505          * queue, and arranging for acknowledgment of receipt if necessary.
12506          * This process logically involves adjusting tp->rcv_wnd as data is
12507          * presented to the user (this happens in tcp_usrreq.c, case
12508          * PRU_RCVD).  If a FIN has already been received on this connection
12509          * then we just ignore the text.
12510          */
12511         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12512                    IS_FASTOPEN(tp->t_flags));
12513         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12514             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12515                 tcp_seq save_start = th->th_seq;
12516                 tcp_seq save_rnxt  = tp->rcv_nxt;
12517                 int     save_tlen  = tlen;
12518
12519                 m_adj(m, drop_hdrlen);  /* delayed header drop */
12520                 /*
12521                  * Insert segment which includes th into TCP reassembly
12522                  * queue with control block tp.  Set thflags to whether
12523                  * reassembly now includes a segment with FIN.  This handles
12524                  * the common case inline (segment is the next to be
12525                  * received on an established connection, and the queue is
12526                  * empty), avoiding linkage into and removal from the queue
12527                  * and repetition of various conversions. Set DELACK for
12528                  * segments received in order, but ack immediately when
12529                  * segments are out of order (so fast retransmit can work).
12530                  */
12531                 if (th->th_seq == tp->rcv_nxt &&
12532                     SEGQ_EMPTY(tp) &&
12533                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
12534                     tfo_syn)) {
12535 #ifdef NETFLIX_SB_LIMITS
12536                         u_int mcnt, appended;
12537
12538                         if (so->so_rcv.sb_shlim) {
12539                                 mcnt = m_memcnt(m);
12540                                 appended = 0;
12541                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12542                                     CFO_NOSLEEP, NULL) == false) {
12543                                         counter_u64_add(tcp_sb_shlim_fails, 1);
12544                                         m_freem(m);
12545                                         return (0);
12546                                 }
12547                         }
12548 #endif
12549                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
12550                         tp->rcv_nxt += tlen;
12551                         if (tlen &&
12552                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12553                             (tp->t_fbyte_in == 0)) {
12554                                 tp->t_fbyte_in = ticks;
12555                                 if (tp->t_fbyte_in == 0)
12556                                         tp->t_fbyte_in = 1;
12557                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
12558                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12559                         }
12560                         thflags = tcp_get_flags(th) & TH_FIN;
12561                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12562                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12563                         SOCKBUF_LOCK(&so->so_rcv);
12564                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12565                                 m_freem(m);
12566                         } else
12567 #ifdef NETFLIX_SB_LIMITS
12568                                 appended =
12569 #endif
12570                                         sbappendstream_locked(&so->so_rcv, m, 0);
12571
12572                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12573                         /* NB: sorwakeup_locked() does an implicit unlock. */
12574                         sorwakeup_locked(so);
12575 #ifdef NETFLIX_SB_LIMITS
12576                         if (so->so_rcv.sb_shlim && appended != mcnt)
12577                                 counter_fo_release(so->so_rcv.sb_shlim,
12578                                     mcnt - appended);
12579 #endif
12580                 } else {
12581                         /*
12582                          * XXX: Due to the header drop above "th" is
12583                          * theoretically invalid by now.  Fortunately
12584                          * m_adj() doesn't actually frees any mbufs when
12585                          * trimming from the head.
12586                          */
12587                         tcp_seq temp = save_start;
12588
12589                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
12590                         tp->t_flags |= TF_ACKNOW;
12591                         if (tp->t_flags & TF_WAKESOR) {
12592                                 tp->t_flags &= ~TF_WAKESOR;
12593                                 /* NB: sorwakeup_locked() does an implicit unlock. */
12594                                 sorwakeup_locked(so);
12595                         }
12596                 }
12597                 if ((tp->t_flags & TF_SACK_PERMIT) &&
12598                     (save_tlen > 0) &&
12599                     TCPS_HAVEESTABLISHED(tp->t_state)) {
12600                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
12601                                 /*
12602                                  * DSACK actually handled in the fastpath
12603                                  * above.
12604                                  */
12605                                 tcp_update_sack_list(tp, save_start,
12606                                     save_start + save_tlen);
12607                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
12608                                 if ((tp->rcv_numsacks >= 1) &&
12609                                     (tp->sackblks[0].end == save_start)) {
12610                                         /*
12611                                          * Partial overlap, recorded at todrop
12612                                          * above.
12613                                          */
12614                                         tcp_update_sack_list(tp,
12615                                             tp->sackblks[0].start,
12616                                             tp->sackblks[0].end);
12617                                 } else {
12618                                         tcp_update_dsack_list(tp, save_start,
12619                                             save_start + save_tlen);
12620                                 }
12621                         } else if (tlen >= save_tlen) {
12622                                 /* Update of sackblks. */
12623                                 tcp_update_dsack_list(tp, save_start,
12624                                     save_start + save_tlen);
12625                         } else if (tlen > 0) {
12626                                 tcp_update_dsack_list(tp, save_start,
12627                                     save_start + tlen);
12628                         }
12629                 }
12630         } else {
12631                 m_freem(m);
12632                 thflags &= ~TH_FIN;
12633         }
12634
12635         /*
12636          * If FIN is received ACK the FIN and let the user know that the
12637          * connection is closing.
12638          */
12639         if (thflags & TH_FIN) {
12640                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12641                         /* The socket upcall is handled by socantrcvmore. */
12642                         socantrcvmore(so);
12643                         /*
12644                          * If connection is half-synchronized (ie NEEDSYN
12645                          * flag on) then delay ACK, so it may be piggybacked
12646                          * when SYN is sent. Otherwise, since we received a
12647                          * FIN then no more input can be expected, send ACK
12648                          * now.
12649                          */
12650                         if (tp->t_flags & TF_NEEDSYN) {
12651                                 rack_timer_cancel(tp, rack,
12652                                     rack->r_ctl.rc_rcvtime, __LINE__);
12653                                 tp->t_flags |= TF_DELACK;
12654                         } else {
12655                                 tp->t_flags |= TF_ACKNOW;
12656                         }
12657                         tp->rcv_nxt++;
12658                 }
12659                 switch (tp->t_state) {
12660                         /*
12661                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
12662                          * CLOSE_WAIT state.
12663                          */
12664                 case TCPS_SYN_RECEIVED:
12665                         tp->t_starttime = ticks;
12666                         /* FALLTHROUGH */
12667                 case TCPS_ESTABLISHED:
12668                         rack_timer_cancel(tp, rack,
12669                             rack->r_ctl.rc_rcvtime, __LINE__);
12670                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
12671                         break;
12672
12673                         /*
12674                          * If still in FIN_WAIT_1 STATE FIN has not been
12675                          * acked so enter the CLOSING state.
12676                          */
12677                 case TCPS_FIN_WAIT_1:
12678                         rack_timer_cancel(tp, rack,
12679                             rack->r_ctl.rc_rcvtime, __LINE__);
12680                         tcp_state_change(tp, TCPS_CLOSING);
12681                         break;
12682
12683                         /*
12684                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
12685                          * starting the time-wait timer, turning off the
12686                          * other standard timers.
12687                          */
12688                 case TCPS_FIN_WAIT_2:
12689                         rack_timer_cancel(tp, rack,
12690                             rack->r_ctl.rc_rcvtime, __LINE__);
12691                         tcp_twstart(tp);
12692                         return (1);
12693                 }
12694         }
12695         /*
12696          * Return any desired output.
12697          */
12698         if ((tp->t_flags & TF_ACKNOW) ||
12699             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
12700                 rack->r_wanted_output = 1;
12701         }
12702         return (0);
12703 }
12704
12705 /*
12706  * Here nothing is really faster, its just that we
12707  * have broken out the fast-data path also just like
12708  * the fast-ack.
12709  */
12710 static int
12711 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
12712     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12713     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
12714 {
12715         int32_t nsegs;
12716         int32_t newsize = 0;    /* automatic sockbuf scaling */
12717         struct tcp_rack *rack;
12718 #ifdef NETFLIX_SB_LIMITS
12719         u_int mcnt, appended;
12720 #endif
12721
12722         /*
12723          * If last ACK falls within this segment's sequence numbers, record
12724          * the timestamp. NOTE that the test is modified according to the
12725          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12726          */
12727         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
12728                 return (0);
12729         }
12730         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
12731                 return (0);
12732         }
12733         if (tiwin && tiwin != tp->snd_wnd) {
12734                 return (0);
12735         }
12736         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
12737                 return (0);
12738         }
12739         if (__predict_false((to->to_flags & TOF_TS) &&
12740             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
12741                 return (0);
12742         }
12743         if (__predict_false((th->th_ack != tp->snd_una))) {
12744                 return (0);
12745         }
12746         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
12747                 return (0);
12748         }
12749         if ((to->to_flags & TOF_TS) != 0 &&
12750             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12751                 tp->ts_recent_age = tcp_ts_getticks();
12752                 tp->ts_recent = to->to_tsval;
12753         }
12754         rack = (struct tcp_rack *)tp->t_fb_ptr;
12755         /*
12756          * This is a pure, in-sequence data packet with nothing on the
12757          * reassembly queue and we have enough buffer space to take it.
12758          */
12759         nsegs = max(1, m->m_pkthdr.lro_nsegs);
12760
12761 #ifdef NETFLIX_SB_LIMITS
12762         if (so->so_rcv.sb_shlim) {
12763                 mcnt = m_memcnt(m);
12764                 appended = 0;
12765                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12766                     CFO_NOSLEEP, NULL) == false) {
12767                         counter_u64_add(tcp_sb_shlim_fails, 1);
12768                         m_freem(m);
12769                         return (1);
12770                 }
12771         }
12772 #endif
12773         /* Clean receiver SACK report if present */
12774         if (tp->rcv_numsacks)
12775                 tcp_clean_sackreport(tp);
12776         KMOD_TCPSTAT_INC(tcps_preddat);
12777         tp->rcv_nxt += tlen;
12778         if (tlen &&
12779             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12780             (tp->t_fbyte_in == 0)) {
12781                 tp->t_fbyte_in = ticks;
12782                 if (tp->t_fbyte_in == 0)
12783                         tp->t_fbyte_in = 1;
12784                 if (tp->t_fbyte_out && tp->t_fbyte_in)
12785                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12786         }
12787         /*
12788          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
12789          */
12790         tp->snd_wl1 = th->th_seq;
12791         /*
12792          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
12793          */
12794         tp->rcv_up = tp->rcv_nxt;
12795         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12796         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12797         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12798
12799         /* Add data to socket buffer. */
12800         SOCKBUF_LOCK(&so->so_rcv);
12801         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12802                 m_freem(m);
12803         } else {
12804                 /*
12805                  * Set new socket buffer size. Give up when limit is
12806                  * reached.
12807                  */
12808                 if (newsize)
12809                         if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12810                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12811                 m_adj(m, drop_hdrlen);  /* delayed header drop */
12812 #ifdef NETFLIX_SB_LIMITS
12813                 appended =
12814 #endif
12815                         sbappendstream_locked(&so->so_rcv, m, 0);
12816                 ctf_calc_rwin(so, tp);
12817         }
12818         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12819         /* NB: sorwakeup_locked() does an implicit unlock. */
12820         sorwakeup_locked(so);
12821 #ifdef NETFLIX_SB_LIMITS
12822         if (so->so_rcv.sb_shlim && mcnt != appended)
12823                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
12824 #endif
12825         rack_handle_delayed_ack(tp, rack, tlen, 0);
12826         if (tp->snd_una == tp->snd_max)
12827                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12828         return (1);
12829 }
12830
12831 /*
12832  * This subfunction is used to try to highly optimize the
12833  * fast path. We again allow window updates that are
12834  * in sequence to remain in the fast-path. We also add
12835  * in the __predict's to attempt to help the compiler.
12836  * Note that if we return a 0, then we can *not* process
12837  * it and the caller should push the packet into the
12838  * slow-path.
12839  */
12840 static int
12841 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12842     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12843     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
12844 {
12845         int32_t acked;
12846         int32_t nsegs;
12847         int32_t under_pacing = 0;
12848         struct tcp_rack *rack;
12849
12850         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12851                 /* Old ack, behind (or duplicate to) the last one rcv'd */
12852                 return (0);
12853         }
12854         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
12855                 /* Above what we have sent? */
12856                 return (0);
12857         }
12858         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
12859                 /* We are retransmitting */
12860                 return (0);
12861         }
12862         if (__predict_false(tiwin == 0)) {
12863                 /* zero window */
12864                 return (0);
12865         }
12866         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
12867                 /* We need a SYN or a FIN, unlikely.. */
12868                 return (0);
12869         }
12870         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
12871                 /* Timestamp is behind .. old ack with seq wrap? */
12872                 return (0);
12873         }
12874         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
12875                 /* Still recovering */
12876                 return (0);
12877         }
12878         rack = (struct tcp_rack *)tp->t_fb_ptr;
12879         if (rack->r_ctl.rc_sacked) {
12880                 /* We have sack holes on our scoreboard */
12881                 return (0);
12882         }
12883         /* Ok if we reach here, we can process a fast-ack */
12884         if (rack->gp_ready &&
12885             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12886                 under_pacing = 1;
12887         }
12888         nsegs = max(1, m->m_pkthdr.lro_nsegs);
12889         rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
12890         /* Did the window get updated? */
12891         if (tiwin != tp->snd_wnd) {
12892                 tp->snd_wnd = tiwin;
12893                 rack_validate_fo_sendwin_up(tp, rack);
12894                 tp->snd_wl1 = th->th_seq;
12895                 if (tp->snd_wnd > tp->max_sndwnd)
12896                         tp->max_sndwnd = tp->snd_wnd;
12897         }
12898         /* Do we exit persists? */
12899         if ((rack->rc_in_persist != 0) &&
12900             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12901                                rack->r_ctl.rc_pace_min_segs))) {
12902                 rack_exit_persist(tp, rack, cts);
12903         }
12904         /* Do we enter persists? */
12905         if ((rack->rc_in_persist == 0) &&
12906             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12907             TCPS_HAVEESTABLISHED(tp->t_state) &&
12908             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12909             sbavail(&tptosocket(tp)->so_snd) &&
12910             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12911                 /*
12912                  * Here the rwnd is less than
12913                  * the pacing size, we are established,
12914                  * nothing is outstanding, and there is
12915                  * data to send. Enter persists.
12916                  */
12917                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
12918         }
12919         /*
12920          * If last ACK falls within this segment's sequence numbers, record
12921          * the timestamp. NOTE that the test is modified according to the
12922          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12923          */
12924         if ((to->to_flags & TOF_TS) != 0 &&
12925             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12926                 tp->ts_recent_age = tcp_ts_getticks();
12927                 tp->ts_recent = to->to_tsval;
12928         }
12929         /*
12930          * This is a pure ack for outstanding data.
12931          */
12932         KMOD_TCPSTAT_INC(tcps_predack);
12933
12934         /*
12935          * "bad retransmit" recovery.
12936          */
12937         if ((tp->t_flags & TF_PREVVALID) &&
12938             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12939                 tp->t_flags &= ~TF_PREVVALID;
12940                 if (tp->t_rxtshift == 1 &&
12941                     (int)(ticks - tp->t_badrxtwin) < 0)
12942                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12943         }
12944         /*
12945          * Recalculate the transmit timer / rtt.
12946          *
12947          * Some boxes send broken timestamp replies during the SYN+ACK
12948          * phase, ignore timestamps of 0 or we could calculate a huge RTT
12949          * and blow up the retransmit timer.
12950          */
12951         acked = BYTES_THIS_ACK(tp, th);
12952
12953 #ifdef TCP_HHOOK
12954         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
12955         hhook_run_tcp_est_in(tp, th, to);
12956 #endif
12957         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12958         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12959         if (acked) {
12960                 struct mbuf *mfree;
12961
12962                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
12963                 SOCKBUF_LOCK(&so->so_snd);
12964                 mfree = sbcut_locked(&so->so_snd, acked);
12965                 tp->snd_una = th->th_ack;
12966                 /* Note we want to hold the sb lock through the sendmap adjust */
12967                 rack_adjust_sendmap_head(rack, &so->so_snd);
12968                 /* Wake up the socket if we have room to write more */
12969                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12970                 sowwakeup_locked(so);
12971                 m_freem(mfree);
12972                 tp->t_rxtshift = 0;
12973                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12974                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12975                 rack->rc_tlp_in_progress = 0;
12976                 rack->r_ctl.rc_tlp_cnt_out = 0;
12977                 /*
12978                  * If it is the RXT timer we want to
12979                  * stop it, so we can restart a TLP.
12980                  */
12981                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12982                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12983
12984 #ifdef TCP_REQUEST_TRK
12985                 rack_http_check_for_comp(rack, th->th_ack);
12986 #endif
12987         }
12988         /*
12989          * Let the congestion control algorithm update congestion control
12990          * related information. This typically means increasing the
12991          * congestion window.
12992          */
12993         if (tp->snd_wnd < ctf_outstanding(tp)) {
12994                 /* The peer collapsed the window */
12995                 rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12996         } else if (rack->rc_has_collapsed)
12997                 rack_un_collapse_window(rack, __LINE__);
12998         if ((rack->r_collapse_point_valid) &&
12999             (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
13000                 rack->r_collapse_point_valid = 0;
13001         /*
13002          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
13003          */
13004         tp->snd_wl2 = th->th_ack;
13005         tp->t_dupacks = 0;
13006         m_freem(m);
13007         /* ND6_HINT(tp);         *//* Some progress has been made. */
13008
13009         /*
13010          * If all outstanding data are acked, stop retransmit timer,
13011          * otherwise restart timer using current (possibly backed-off)
13012          * value. If process is waiting for space, wakeup/selwakeup/signal.
13013          * If data are ready to send, let tcp_output decide between more
13014          * output or persist.
13015          */
13016         if (under_pacing &&
13017             (rack->use_fixed_rate == 0) &&
13018             (rack->in_probe_rtt == 0) &&
13019             rack->rc_gp_dyn_mul &&
13020             rack->rc_always_pace) {
13021                 /* Check if we are dragging bottom */
13022                 rack_check_bottom_drag(tp, rack, so);
13023         }
13024         if (tp->snd_una == tp->snd_max) {
13025                 tp->t_flags &= ~TF_PREVVALID;
13026                 rack->r_ctl.retran_during_recovery = 0;
13027                 rack->rc_suspicious = 0;
13028                 rack->r_ctl.dsack_byte_cnt = 0;
13029                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13030                 if (rack->r_ctl.rc_went_idle_time == 0)
13031                         rack->r_ctl.rc_went_idle_time = 1;
13032                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13033                 if (sbavail(&tptosocket(tp)->so_snd) == 0)
13034                         tp->t_acktime = 0;
13035                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13036         }
13037         if (acked && rack->r_fast_output)
13038                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
13039         if (sbavail(&so->so_snd)) {
13040                 rack->r_wanted_output = 1;
13041         }
13042         return (1);
13043 }
13044
13045 /*
13046  * Return value of 1, the TCB is unlocked and most
13047  * likely gone, return value of 0, the TCP is still
13048  * locked.
13049  */
13050 static int
13051 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
13052     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13053     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13054 {
13055         int32_t ret_val = 0;
13056         int32_t todrop;
13057         int32_t ourfinisacked = 0;
13058         struct tcp_rack *rack;
13059
13060         INP_WLOCK_ASSERT(tptoinpcb(tp));
13061
13062         ctf_calc_rwin(so, tp);
13063         /*
13064          * If the state is SYN_SENT: if seg contains an ACK, but not for our
13065          * SYN, drop the input. if seg contains a RST, then drop the
13066          * connection. if seg does not contain SYN, then drop it. Otherwise
13067          * this is an acceptable SYN segment initialize tp->rcv_nxt and
13068          * tp->irs if seg contains ack then advance tp->snd_una if seg
13069          * contains an ECE and ECN support is enabled, the stream is ECN
13070          * capable. if SYN has been acked change to ESTABLISHED else
13071          * SYN_RCVD state arrange for segment to be acked (eventually)
13072          * continue processing rest of data/controls.
13073          */
13074         if ((thflags & TH_ACK) &&
13075             (SEQ_LEQ(th->th_ack, tp->iss) ||
13076             SEQ_GT(th->th_ack, tp->snd_max))) {
13077                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13078                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13079                 return (1);
13080         }
13081         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
13082                 TCP_PROBE5(connect__refused, NULL, tp,
13083                     mtod(m, const char *), tp, th);
13084                 tp = tcp_drop(tp, ECONNREFUSED);
13085                 ctf_do_drop(m, tp);
13086                 return (1);
13087         }
13088         if (thflags & TH_RST) {
13089                 ctf_do_drop(m, tp);
13090                 return (1);
13091         }
13092         if (!(thflags & TH_SYN)) {
13093                 ctf_do_drop(m, tp);
13094                 return (1);
13095         }
13096         tp->irs = th->th_seq;
13097         tcp_rcvseqinit(tp);
13098         rack = (struct tcp_rack *)tp->t_fb_ptr;
13099         if (thflags & TH_ACK) {
13100                 int tfo_partial = 0;
13101
13102                 KMOD_TCPSTAT_INC(tcps_connects);
13103                 soisconnected(so);
13104 #ifdef MAC
13105                 mac_socketpeer_set_from_mbuf(m, so);
13106 #endif
13107                 /* Do window scaling on this connection? */
13108                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13109                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13110                         tp->rcv_scale = tp->request_r_scale;
13111                 }
13112                 tp->rcv_adv += min(tp->rcv_wnd,
13113                     TCP_MAXWIN << tp->rcv_scale);
13114                 /*
13115                  * If not all the data that was sent in the TFO SYN
13116                  * has been acked, resend the remainder right away.
13117                  */
13118                 if (IS_FASTOPEN(tp->t_flags) &&
13119                     (tp->snd_una != tp->snd_max)) {
13120                         tp->snd_nxt = th->th_ack;
13121                         tfo_partial = 1;
13122                 }
13123                 /*
13124                  * If there's data, delay ACK; if there's also a FIN ACKNOW
13125                  * will be turned on later.
13126                  */
13127                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
13128                         rack_timer_cancel(tp, rack,
13129                                           rack->r_ctl.rc_rcvtime, __LINE__);
13130                         tp->t_flags |= TF_DELACK;
13131                 } else {
13132                         rack->r_wanted_output = 1;
13133                         tp->t_flags |= TF_ACKNOW;
13134                 }
13135
13136                 tcp_ecn_input_syn_sent(tp, thflags, iptos);
13137
13138                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
13139                         /*
13140                          * We advance snd_una for the
13141                          * fast open case. If th_ack is
13142                          * acknowledging data beyond
13143                          * snd_una we can't just call
13144                          * ack-processing since the
13145                          * data stream in our send-map
13146                          * will start at snd_una + 1 (one
13147                          * beyond the SYN). If its just
13148                          * equal we don't need to do that
13149                          * and there is no send_map.
13150                          */
13151                         tp->snd_una++;
13152                 }
13153                 /*
13154                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
13155                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
13156                  */
13157                 tp->t_starttime = ticks;
13158                 if (tp->t_flags & TF_NEEDFIN) {
13159                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
13160                         tp->t_flags &= ~TF_NEEDFIN;
13161                         thflags &= ~TH_SYN;
13162                 } else {
13163                         tcp_state_change(tp, TCPS_ESTABLISHED);
13164                         TCP_PROBE5(connect__established, NULL, tp,
13165                             mtod(m, const char *), tp, th);
13166                         rack_cc_conn_init(tp);
13167                 }
13168         } else {
13169                 /*
13170                  * Received initial SYN in SYN-SENT[*] state => simultaneous
13171                  * open.  If segment contains CC option and there is a
13172                  * cached CC, apply TAO test. If it succeeds, connection is *
13173                  * half-synchronized. Otherwise, do 3-way handshake:
13174                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
13175                  * there was no CC option, clear cached CC value.
13176                  */
13177                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
13178                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
13179         }
13180         /*
13181          * Advance th->th_seq to correspond to first data byte. If data,
13182          * trim to stay within window, dropping FIN if necessary.
13183          */
13184         th->th_seq++;
13185         if (tlen > tp->rcv_wnd) {
13186                 todrop = tlen - tp->rcv_wnd;
13187                 m_adj(m, -todrop);
13188                 tlen = tp->rcv_wnd;
13189                 thflags &= ~TH_FIN;
13190                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13191                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13192         }
13193         tp->snd_wl1 = th->th_seq - 1;
13194         tp->rcv_up = th->th_seq;
13195         /*
13196          * Client side of transaction: already sent SYN and data. If the
13197          * remote host used T/TCP to validate the SYN, our data will be
13198          * ACK'd; if so, enter normal data segment processing in the middle
13199          * of step 5, ack processing. Otherwise, goto step 6.
13200          */
13201         if (thflags & TH_ACK) {
13202                 /* For syn-sent we need to possibly update the rtt */
13203                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13204                         uint32_t t, mcts;
13205
13206                         mcts = tcp_ts_getticks();
13207                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13208                         if (!tp->t_rttlow || tp->t_rttlow > t)
13209                                 tp->t_rttlow = t;
13210                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13211                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13212                         tcp_rack_xmit_timer_commit(rack, tp);
13213                 }
13214                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
13215                         return (ret_val);
13216                 /* We may have changed to FIN_WAIT_1 above */
13217                 if (tp->t_state == TCPS_FIN_WAIT_1) {
13218                         /*
13219                          * In FIN_WAIT_1 STATE in addition to the processing
13220                          * for the ESTABLISHED state if our FIN is now
13221                          * acknowledged then enter FIN_WAIT_2.
13222                          */
13223                         if (ourfinisacked) {
13224                                 /*
13225                                  * If we can't receive any more data, then
13226                                  * closing user can proceed. Starting the
13227                                  * timer is contrary to the specification,
13228                                  * but if we don't get a FIN we'll hang
13229                                  * forever.
13230                                  *
13231                                  * XXXjl: we should release the tp also, and
13232                                  * use a compressed state.
13233                                  */
13234                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13235                                         soisdisconnected(so);
13236                                         tcp_timer_activate(tp, TT_2MSL,
13237                                             (tcp_fast_finwait2_recycle ?
13238                                             tcp_finwait2_timeout :
13239                                             TP_MAXIDLE(tp)));
13240                                 }
13241                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13242                         }
13243                 }
13244         }
13245         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13246            tiwin, thflags, nxt_pkt));
13247 }
13248
13249 /*
13250  * Return value of 1, the TCB is unlocked and most
13251  * likely gone, return value of 0, the TCP is still
13252  * locked.
13253  */
13254 static int
13255 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13256     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13257     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13258 {
13259         struct tcp_rack *rack;
13260         int32_t ret_val = 0;
13261         int32_t ourfinisacked = 0;
13262
13263         ctf_calc_rwin(so, tp);
13264         if ((thflags & TH_ACK) &&
13265             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13266             SEQ_GT(th->th_ack, tp->snd_max))) {
13267                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13268                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13269                 return (1);
13270         }
13271         rack = (struct tcp_rack *)tp->t_fb_ptr;
13272         if (IS_FASTOPEN(tp->t_flags)) {
13273                 /*
13274                  * When a TFO connection is in SYN_RECEIVED, the
13275                  * only valid packets are the initial SYN, a
13276                  * retransmit/copy of the initial SYN (possibly with
13277                  * a subset of the original data), a valid ACK, a
13278                  * FIN, or a RST.
13279                  */
13280                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13281                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13282                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13283                         return (1);
13284                 } else if (thflags & TH_SYN) {
13285                         /* non-initial SYN is ignored */
13286                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13287                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13288                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13289                                 ctf_do_drop(m, NULL);
13290                                 return (0);
13291                         }
13292                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13293                         ctf_do_drop(m, NULL);
13294                         return (0);
13295                 }
13296         }
13297
13298         if ((thflags & TH_RST) ||
13299             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13300                 return (__ctf_process_rst(m, th, so, tp,
13301                                           &rack->r_ctl.challenge_ack_ts,
13302                                           &rack->r_ctl.challenge_ack_cnt));
13303         /*
13304          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13305          * it's less than ts_recent, drop it.
13306          */
13307         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13308             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13309                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13310                         return (ret_val);
13311         }
13312         /*
13313          * In the SYN-RECEIVED state, validate that the packet belongs to
13314          * this connection before trimming the data to fit the receive
13315          * window.  Check the sequence number versus IRS since we know the
13316          * sequence numbers haven't wrapped.  This is a partial fix for the
13317          * "LAND" DoS attack.
13318          */
13319         if (SEQ_LT(th->th_seq, tp->irs)) {
13320                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13321                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13322                 return (1);
13323         }
13324         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13325                               &rack->r_ctl.challenge_ack_ts,
13326                               &rack->r_ctl.challenge_ack_cnt)) {
13327                 return (ret_val);
13328         }
13329         /*
13330          * If last ACK falls within this segment's sequence numbers, record
13331          * its timestamp. NOTE: 1) That the test incorporates suggestions
13332          * from the latest proposal of the tcplw@cray.com list (Braden
13333          * 1993/04/26). 2) That updating only on newer timestamps interferes
13334          * with our earlier PAWS tests, so this check should be solely
13335          * predicated on the sequence space of this segment. 3) That we
13336          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13337          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13338          * SEG.Len, This modified check allows us to overcome RFC1323's
13339          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13340          * p.869. In such cases, we can still calculate the RTT correctly
13341          * when RCV.NXT == Last.ACK.Sent.
13342          */
13343         if ((to->to_flags & TOF_TS) != 0 &&
13344             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13345             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13346             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13347                 tp->ts_recent_age = tcp_ts_getticks();
13348                 tp->ts_recent = to->to_tsval;
13349         }
13350         tp->snd_wnd = tiwin;
13351         rack_validate_fo_sendwin_up(tp, rack);
13352         /*
13353          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13354          * is on (half-synchronized state), then queue data for later
13355          * processing; else drop segment and return.
13356          */
13357         if ((thflags & TH_ACK) == 0) {
13358                 if (IS_FASTOPEN(tp->t_flags)) {
13359                         rack_cc_conn_init(tp);
13360                 }
13361                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13362                     tiwin, thflags, nxt_pkt));
13363         }
13364         KMOD_TCPSTAT_INC(tcps_connects);
13365         if (tp->t_flags & TF_SONOTCONN) {
13366                 tp->t_flags &= ~TF_SONOTCONN;
13367                 soisconnected(so);
13368         }
13369         /* Do window scaling? */
13370         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13371             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13372                 tp->rcv_scale = tp->request_r_scale;
13373         }
13374         /*
13375          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
13376          * FIN-WAIT-1
13377          */
13378         tp->t_starttime = ticks;
13379         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
13380                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13381                 tp->t_tfo_pending = NULL;
13382         }
13383         if (tp->t_flags & TF_NEEDFIN) {
13384                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
13385                 tp->t_flags &= ~TF_NEEDFIN;
13386         } else {
13387                 tcp_state_change(tp, TCPS_ESTABLISHED);
13388                 TCP_PROBE5(accept__established, NULL, tp,
13389                     mtod(m, const char *), tp, th);
13390                 /*
13391                  * TFO connections call cc_conn_init() during SYN
13392                  * processing.  Calling it again here for such connections
13393                  * is not harmless as it would undo the snd_cwnd reduction
13394                  * that occurs when a TFO SYN|ACK is retransmitted.
13395                  */
13396                 if (!IS_FASTOPEN(tp->t_flags))
13397                         rack_cc_conn_init(tp);
13398         }
13399         /*
13400          * Account for the ACK of our SYN prior to
13401          * regular ACK processing below, except for
13402          * simultaneous SYN, which is handled later.
13403          */
13404         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13405                 tp->snd_una++;
13406         /*
13407          * If segment contains data or ACK, will call tcp_reass() later; if
13408          * not, do so now to pass queued data to user.
13409          */
13410         if (tlen == 0 && (thflags & TH_FIN) == 0) {
13411                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13412                     (struct mbuf *)0);
13413                 if (tp->t_flags & TF_WAKESOR) {
13414                         tp->t_flags &= ~TF_WAKESOR;
13415                         /* NB: sorwakeup_locked() does an implicit unlock. */
13416                         sorwakeup_locked(so);
13417                 }
13418         }
13419         tp->snd_wl1 = th->th_seq - 1;
13420         /* For syn-recv we need to possibly update the rtt */
13421         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13422                 uint32_t t, mcts;
13423
13424                 mcts = tcp_ts_getticks();
13425                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13426                 if (!tp->t_rttlow || tp->t_rttlow > t)
13427                         tp->t_rttlow = t;
13428                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13429                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13430                 tcp_rack_xmit_timer_commit(rack, tp);
13431         }
13432         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13433                 return (ret_val);
13434         }
13435         if (tp->t_state == TCPS_FIN_WAIT_1) {
13436                 /* We could have went to FIN_WAIT_1 (or EST) above */
13437                 /*
13438                  * In FIN_WAIT_1 STATE in addition to the processing for the
13439                  * ESTABLISHED state if our FIN is now acknowledged then
13440                  * enter FIN_WAIT_2.
13441                  */
13442                 if (ourfinisacked) {
13443                         /*
13444                          * If we can't receive any more data, then closing
13445                          * user can proceed. Starting the timer is contrary
13446                          * to the specification, but if we don't get a FIN
13447                          * we'll hang forever.
13448                          *
13449                          * XXXjl: we should release the tp also, and use a
13450                          * compressed state.
13451                          */
13452                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13453                                 soisdisconnected(so);
13454                                 tcp_timer_activate(tp, TT_2MSL,
13455                                     (tcp_fast_finwait2_recycle ?
13456                                     tcp_finwait2_timeout :
13457                                     TP_MAXIDLE(tp)));
13458                         }
13459                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13460                 }
13461         }
13462         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13463             tiwin, thflags, nxt_pkt));
13464 }
13465
13466 /*
13467  * Return value of 1, the TCB is unlocked and most
13468  * likely gone, return value of 0, the TCP is still
13469  * locked.
13470  */
13471 static int
13472 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13473     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13474     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13475 {
13476         int32_t ret_val = 0;
13477         struct tcp_rack *rack;
13478
13479         /*
13480          * Header prediction: check for the two common cases of a
13481          * uni-directional data xfer.  If the packet has no control flags,
13482          * is in-sequence, the window didn't change and we're not
13483          * retransmitting, it's a candidate.  If the length is zero and the
13484          * ack moved forward, we're the sender side of the xfer.  Just free
13485          * the data acked & wake any higher level process that was blocked
13486          * waiting for space.  If the length is non-zero and the ack didn't
13487          * move, we're the receiver side.  If we're getting packets in-order
13488          * (the reassembly queue is empty), add the data toc The socket
13489          * buffer and note that we need a delayed ack. Make sure that the
13490          * hidden state-flags are also off. Since we check for
13491          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13492          */
13493         rack = (struct tcp_rack *)tp->t_fb_ptr;
13494         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13495             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13496             __predict_true(SEGQ_EMPTY(tp)) &&
13497             __predict_true(th->th_seq == tp->rcv_nxt)) {
13498                 if (tlen == 0) {
13499                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13500                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13501                                 return (0);
13502                         }
13503                 } else {
13504                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13505                             tiwin, nxt_pkt, iptos)) {
13506                                 return (0);
13507                         }
13508                 }
13509         }
13510         ctf_calc_rwin(so, tp);
13511
13512         if ((thflags & TH_RST) ||
13513             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13514                 return (__ctf_process_rst(m, th, so, tp,
13515                                           &rack->r_ctl.challenge_ack_ts,
13516                                           &rack->r_ctl.challenge_ack_cnt));
13517
13518         /*
13519          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13520          * synchronized state.
13521          */
13522         if (thflags & TH_SYN) {
13523                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13524                 return (ret_val);
13525         }
13526         /*
13527          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13528          * it's less than ts_recent, drop it.
13529          */
13530         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13531             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13532                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13533                         return (ret_val);
13534         }
13535         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13536                               &rack->r_ctl.challenge_ack_ts,
13537                               &rack->r_ctl.challenge_ack_cnt)) {
13538                 return (ret_val);
13539         }
13540         /*
13541          * If last ACK falls within this segment's sequence numbers, record
13542          * its timestamp. NOTE: 1) That the test incorporates suggestions
13543          * from the latest proposal of the tcplw@cray.com list (Braden
13544          * 1993/04/26). 2) That updating only on newer timestamps interferes
13545          * with our earlier PAWS tests, so this check should be solely
13546          * predicated on the sequence space of this segment. 3) That we
13547          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13548          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13549          * SEG.Len, This modified check allows us to overcome RFC1323's
13550          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13551          * p.869. In such cases, we can still calculate the RTT correctly
13552          * when RCV.NXT == Last.ACK.Sent.
13553          */
13554         if ((to->to_flags & TOF_TS) != 0 &&
13555             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13556             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13557             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13558                 tp->ts_recent_age = tcp_ts_getticks();
13559                 tp->ts_recent = to->to_tsval;
13560         }
13561         /*
13562          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13563          * is on (half-synchronized state), then queue data for later
13564          * processing; else drop segment and return.
13565          */
13566         if ((thflags & TH_ACK) == 0) {
13567                 if (tp->t_flags & TF_NEEDSYN) {
13568                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13569                             tiwin, thflags, nxt_pkt));
13570
13571                 } else if (tp->t_flags & TF_ACKNOW) {
13572                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13573                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13574                         return (ret_val);
13575                 } else {
13576                         ctf_do_drop(m, NULL);
13577                         return (0);
13578                 }
13579         }
13580         /*
13581          * Ack processing.
13582          */
13583         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
13584                 return (ret_val);
13585         }
13586         if (sbavail(&so->so_snd)) {
13587                 if (ctf_progress_timeout_check(tp, true)) {
13588                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
13589                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13590                         return (1);
13591                 }
13592         }
13593         /* State changes only happen in rack_process_data() */
13594         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13595             tiwin, thflags, nxt_pkt));
13596 }
13597
13598 /*
13599  * Return value of 1, the TCB is unlocked and most
13600  * likely gone, return value of 0, the TCP is still
13601  * locked.
13602  */
13603 static int
13604 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
13605     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13606     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13607 {
13608         int32_t ret_val = 0;
13609         struct tcp_rack *rack;
13610
13611         rack = (struct tcp_rack *)tp->t_fb_ptr;
13612         ctf_calc_rwin(so, tp);
13613         if ((thflags & TH_RST) ||
13614             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13615                 return (__ctf_process_rst(m, th, so, tp,
13616                                           &rack->r_ctl.challenge_ack_ts,
13617                                           &rack->r_ctl.challenge_ack_cnt));
13618         /*
13619          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13620          * synchronized state.
13621          */
13622         if (thflags & TH_SYN) {
13623                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13624                 return (ret_val);
13625         }
13626         /*
13627          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13628          * it's less than ts_recent, drop it.
13629          */
13630         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13631             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13632                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13633                         return (ret_val);
13634         }
13635         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13636                               &rack->r_ctl.challenge_ack_ts,
13637                               &rack->r_ctl.challenge_ack_cnt)) {
13638                 return (ret_val);
13639         }
13640         /*
13641          * If last ACK falls within this segment's sequence numbers, record
13642          * its timestamp. NOTE: 1) That the test incorporates suggestions
13643          * from the latest proposal of the tcplw@cray.com list (Braden
13644          * 1993/04/26). 2) That updating only on newer timestamps interferes
13645          * with our earlier PAWS tests, so this check should be solely
13646          * predicated on the sequence space of this segment. 3) That we
13647          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13648          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13649          * SEG.Len, This modified check allows us to overcome RFC1323's
13650          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13651          * p.869. In such cases, we can still calculate the RTT correctly
13652          * when RCV.NXT == Last.ACK.Sent.
13653          */
13654         if ((to->to_flags & TOF_TS) != 0 &&
13655             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13656             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13657             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13658                 tp->ts_recent_age = tcp_ts_getticks();
13659                 tp->ts_recent = to->to_tsval;
13660         }
13661         /*
13662          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13663          * is on (half-synchronized state), then queue data for later
13664          * processing; else drop segment and return.
13665          */
13666         if ((thflags & TH_ACK) == 0) {
13667                 if (tp->t_flags & TF_NEEDSYN) {
13668                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13669                             tiwin, thflags, nxt_pkt));
13670
13671                 } else if (tp->t_flags & TF_ACKNOW) {
13672                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13673                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13674                         return (ret_val);
13675                 } else {
13676                         ctf_do_drop(m, NULL);
13677                         return (0);
13678                 }
13679         }
13680         /*
13681          * Ack processing.
13682          */
13683         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
13684                 return (ret_val);
13685         }
13686         if (sbavail(&so->so_snd)) {
13687                 if (ctf_progress_timeout_check(tp, true)) {
13688                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13689                                                 tp, tick, PROGRESS_DROP, __LINE__);
13690                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13691                         return (1);
13692                 }
13693         }
13694         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13695             tiwin, thflags, nxt_pkt));
13696 }
13697
13698 static int
13699 rack_check_data_after_close(struct mbuf *m,
13700     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
13701 {
13702         struct tcp_rack *rack;
13703
13704         rack = (struct tcp_rack *)tp->t_fb_ptr;
13705         if (rack->rc_allow_data_af_clo == 0) {
13706         close_now:
13707                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13708                 /* tcp_close will kill the inp pre-log the Reset */
13709                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13710                 tp = tcp_close(tp);
13711                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
13712                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
13713                 return (1);
13714         }
13715         if (sbavail(&so->so_snd) == 0)
13716                 goto close_now;
13717         /* Ok we allow data that is ignored and a followup reset */
13718         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13719         tp->rcv_nxt = th->th_seq + *tlen;
13720         tp->t_flags2 |= TF2_DROP_AF_DATA;
13721         rack->r_wanted_output = 1;
13722         *tlen = 0;
13723         return (0);
13724 }
13725
13726 /*
13727  * Return value of 1, the TCB is unlocked and most
13728  * likely gone, return value of 0, the TCP is still
13729  * locked.
13730  */
13731 static int
13732 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
13733     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13734     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13735 {
13736         int32_t ret_val = 0;
13737         int32_t ourfinisacked = 0;
13738         struct tcp_rack *rack;
13739
13740         rack = (struct tcp_rack *)tp->t_fb_ptr;
13741         ctf_calc_rwin(so, tp);
13742
13743         if ((thflags & TH_RST) ||
13744             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13745                 return (__ctf_process_rst(m, th, so, tp,
13746                                           &rack->r_ctl.challenge_ack_ts,
13747                                           &rack->r_ctl.challenge_ack_cnt));
13748         /*
13749          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13750          * synchronized state.
13751          */
13752         if (thflags & TH_SYN) {
13753                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13754                 return (ret_val);
13755         }
13756         /*
13757          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13758          * it's less than ts_recent, drop it.
13759          */
13760         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13761             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13762                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13763                         return (ret_val);
13764         }
13765         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13766                               &rack->r_ctl.challenge_ack_ts,
13767                               &rack->r_ctl.challenge_ack_cnt)) {
13768                 return (ret_val);
13769         }
13770         /*
13771          * If new data are received on a connection after the user processes
13772          * are gone, then RST the other end.
13773          */
13774         if ((tp->t_flags & TF_CLOSED) && tlen &&
13775             rack_check_data_after_close(m, tp, &tlen, th, so))
13776                 return (1);
13777         /*
13778          * If last ACK falls within this segment's sequence numbers, record
13779          * its timestamp. NOTE: 1) That the test incorporates suggestions
13780          * from the latest proposal of the tcplw@cray.com list (Braden
13781          * 1993/04/26). 2) That updating only on newer timestamps interferes
13782          * with our earlier PAWS tests, so this check should be solely
13783          * predicated on the sequence space of this segment. 3) That we
13784          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13785          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13786          * SEG.Len, This modified check allows us to overcome RFC1323's
13787          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13788          * p.869. In such cases, we can still calculate the RTT correctly
13789          * when RCV.NXT == Last.ACK.Sent.
13790          */
13791         if ((to->to_flags & TOF_TS) != 0 &&
13792             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13793             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13794             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13795                 tp->ts_recent_age = tcp_ts_getticks();
13796                 tp->ts_recent = to->to_tsval;
13797         }
13798         /*
13799          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13800          * is on (half-synchronized state), then queue data for later
13801          * processing; else drop segment and return.
13802          */
13803         if ((thflags & TH_ACK) == 0) {
13804                 if (tp->t_flags & TF_NEEDSYN) {
13805                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13806                             tiwin, thflags, nxt_pkt));
13807                 } else if (tp->t_flags & TF_ACKNOW) {
13808                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13809                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13810                         return (ret_val);
13811                 } else {
13812                         ctf_do_drop(m, NULL);
13813                         return (0);
13814                 }
13815         }
13816         /*
13817          * Ack processing.
13818          */
13819         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13820                 return (ret_val);
13821         }
13822         if (ourfinisacked) {
13823                 /*
13824                  * If we can't receive any more data, then closing user can
13825                  * proceed. Starting the timer is contrary to the
13826                  * specification, but if we don't get a FIN we'll hang
13827                  * forever.
13828                  *
13829                  * XXXjl: we should release the tp also, and use a
13830                  * compressed state.
13831                  */
13832                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13833                         soisdisconnected(so);
13834                         tcp_timer_activate(tp, TT_2MSL,
13835                             (tcp_fast_finwait2_recycle ?
13836                             tcp_finwait2_timeout :
13837                             TP_MAXIDLE(tp)));
13838                 }
13839                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13840         }
13841         if (sbavail(&so->so_snd)) {
13842                 if (ctf_progress_timeout_check(tp, true)) {
13843                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13844                                                 tp, tick, PROGRESS_DROP, __LINE__);
13845                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13846                         return (1);
13847                 }
13848         }
13849         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13850             tiwin, thflags, nxt_pkt));
13851 }
13852
13853 /*
13854  * Return value of 1, the TCB is unlocked and most
13855  * likely gone, return value of 0, the TCP is still
13856  * locked.
13857  */
13858 static int
13859 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
13860     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13861     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13862 {
13863         int32_t ret_val = 0;
13864         int32_t ourfinisacked = 0;
13865         struct tcp_rack *rack;
13866
13867         rack = (struct tcp_rack *)tp->t_fb_ptr;
13868         ctf_calc_rwin(so, tp);
13869
13870         if ((thflags & TH_RST) ||
13871             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13872                 return (__ctf_process_rst(m, th, so, tp,
13873                                           &rack->r_ctl.challenge_ack_ts,
13874                                           &rack->r_ctl.challenge_ack_cnt));
13875         /*
13876          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13877          * synchronized state.
13878          */
13879         if (thflags & TH_SYN) {
13880                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13881                 return (ret_val);
13882         }
13883         /*
13884          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13885          * it's less than ts_recent, drop it.
13886          */
13887         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13888             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13889                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13890                         return (ret_val);
13891         }
13892         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13893                               &rack->r_ctl.challenge_ack_ts,
13894                               &rack->r_ctl.challenge_ack_cnt)) {
13895                 return (ret_val);
13896         }
13897         /*
13898          * If new data are received on a connection after the user processes
13899          * are gone, then RST the other end.
13900          */
13901         if ((tp->t_flags & TF_CLOSED) && tlen &&
13902             rack_check_data_after_close(m, tp, &tlen, th, so))
13903                 return (1);
13904         /*
13905          * If last ACK falls within this segment's sequence numbers, record
13906          * its timestamp. NOTE: 1) That the test incorporates suggestions
13907          * from the latest proposal of the tcplw@cray.com list (Braden
13908          * 1993/04/26). 2) That updating only on newer timestamps interferes
13909          * with our earlier PAWS tests, so this check should be solely
13910          * predicated on the sequence space of this segment. 3) That we
13911          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13912          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13913          * SEG.Len, This modified check allows us to overcome RFC1323's
13914          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13915          * p.869. In such cases, we can still calculate the RTT correctly
13916          * when RCV.NXT == Last.ACK.Sent.
13917          */
13918         if ((to->to_flags & TOF_TS) != 0 &&
13919             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13920             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13921             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13922                 tp->ts_recent_age = tcp_ts_getticks();
13923                 tp->ts_recent = to->to_tsval;
13924         }
13925         /*
13926          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13927          * is on (half-synchronized state), then queue data for later
13928          * processing; else drop segment and return.
13929          */
13930         if ((thflags & TH_ACK) == 0) {
13931                 if (tp->t_flags & TF_NEEDSYN) {
13932                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13933                             tiwin, thflags, nxt_pkt));
13934                 } else if (tp->t_flags & TF_ACKNOW) {
13935                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13936                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13937                         return (ret_val);
13938                 } else {
13939                         ctf_do_drop(m, NULL);
13940                         return (0);
13941                 }
13942         }
13943         /*
13944          * Ack processing.
13945          */
13946         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13947                 return (ret_val);
13948         }
13949         if (ourfinisacked) {
13950                 tcp_twstart(tp);
13951                 m_freem(m);
13952                 return (1);
13953         }
13954         if (sbavail(&so->so_snd)) {
13955                 if (ctf_progress_timeout_check(tp, true)) {
13956                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13957                                                 tp, tick, PROGRESS_DROP, __LINE__);
13958                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13959                         return (1);
13960                 }
13961         }
13962         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13963             tiwin, thflags, nxt_pkt));
13964 }
13965
13966 /*
13967  * Return value of 1, the TCB is unlocked and most
13968  * likely gone, return value of 0, the TCP is still
13969  * locked.
13970  */
13971 static int
13972 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
13973     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13974     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13975 {
13976         int32_t ret_val = 0;
13977         int32_t ourfinisacked = 0;
13978         struct tcp_rack *rack;
13979
13980         rack = (struct tcp_rack *)tp->t_fb_ptr;
13981         ctf_calc_rwin(so, tp);
13982
13983         if ((thflags & TH_RST) ||
13984             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13985                 return (__ctf_process_rst(m, th, so, tp,
13986                                           &rack->r_ctl.challenge_ack_ts,
13987                                           &rack->r_ctl.challenge_ack_cnt));
13988         /*
13989          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13990          * synchronized state.
13991          */
13992         if (thflags & TH_SYN) {
13993                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13994                 return (ret_val);
13995         }
13996         /*
13997          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13998          * it's less than ts_recent, drop it.
13999          */
14000         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14001             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14002                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14003                         return (ret_val);
14004         }
14005         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14006                               &rack->r_ctl.challenge_ack_ts,
14007                               &rack->r_ctl.challenge_ack_cnt)) {
14008                 return (ret_val);
14009         }
14010         /*
14011          * If new data are received on a connection after the user processes
14012          * are gone, then RST the other end.
14013          */
14014         if ((tp->t_flags & TF_CLOSED) && tlen &&
14015             rack_check_data_after_close(m, tp, &tlen, th, so))
14016                 return (1);
14017         /*
14018          * If last ACK falls within this segment's sequence numbers, record
14019          * its timestamp. NOTE: 1) That the test incorporates suggestions
14020          * from the latest proposal of the tcplw@cray.com list (Braden
14021          * 1993/04/26). 2) That updating only on newer timestamps interferes
14022          * with our earlier PAWS tests, so this check should be solely
14023          * predicated on the sequence space of this segment. 3) That we
14024          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14025          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14026          * SEG.Len, This modified check allows us to overcome RFC1323's
14027          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14028          * p.869. In such cases, we can still calculate the RTT correctly
14029          * when RCV.NXT == Last.ACK.Sent.
14030          */
14031         if ((to->to_flags & TOF_TS) != 0 &&
14032             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14033             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14034             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14035                 tp->ts_recent_age = tcp_ts_getticks();
14036                 tp->ts_recent = to->to_tsval;
14037         }
14038         /*
14039          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14040          * is on (half-synchronized state), then queue data for later
14041          * processing; else drop segment and return.
14042          */
14043         if ((thflags & TH_ACK) == 0) {
14044                 if (tp->t_flags & TF_NEEDSYN) {
14045                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14046                             tiwin, thflags, nxt_pkt));
14047                 } else if (tp->t_flags & TF_ACKNOW) {
14048                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14049                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14050                         return (ret_val);
14051                 } else {
14052                         ctf_do_drop(m, NULL);
14053                         return (0);
14054                 }
14055         }
14056         /*
14057          * case TCPS_LAST_ACK: Ack processing.
14058          */
14059         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14060                 return (ret_val);
14061         }
14062         if (ourfinisacked) {
14063                 tp = tcp_close(tp);
14064                 ctf_do_drop(m, tp);
14065                 return (1);
14066         }
14067         if (sbavail(&so->so_snd)) {
14068                 if (ctf_progress_timeout_check(tp, true)) {
14069                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14070                                                 tp, tick, PROGRESS_DROP, __LINE__);
14071                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14072                         return (1);
14073                 }
14074         }
14075         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14076             tiwin, thflags, nxt_pkt));
14077 }
14078
14079 /*
14080  * Return value of 1, the TCB is unlocked and most
14081  * likely gone, return value of 0, the TCP is still
14082  * locked.
14083  */
14084 static int
14085 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
14086     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14087     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14088 {
14089         int32_t ret_val = 0;
14090         int32_t ourfinisacked = 0;
14091         struct tcp_rack *rack;
14092
14093         rack = (struct tcp_rack *)tp->t_fb_ptr;
14094         ctf_calc_rwin(so, tp);
14095
14096         /* Reset receive buffer auto scaling when not in bulk receive mode. */
14097         if ((thflags & TH_RST) ||
14098             (tp->t_fin_is_rst && (thflags & TH_FIN)))
14099                 return (__ctf_process_rst(m, th, so, tp,
14100                                           &rack->r_ctl.challenge_ack_ts,
14101                                           &rack->r_ctl.challenge_ack_cnt));
14102         /*
14103          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14104          * synchronized state.
14105          */
14106         if (thflags & TH_SYN) {
14107                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14108                 return (ret_val);
14109         }
14110         /*
14111          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14112          * it's less than ts_recent, drop it.
14113          */
14114         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14115             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14116                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14117                         return (ret_val);
14118         }
14119         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14120                               &rack->r_ctl.challenge_ack_ts,
14121                               &rack->r_ctl.challenge_ack_cnt)) {
14122                 return (ret_val);
14123         }
14124         /*
14125          * If new data are received on a connection after the user processes
14126          * are gone, then RST the other end.
14127          */
14128         if ((tp->t_flags & TF_CLOSED) && tlen &&
14129             rack_check_data_after_close(m, tp, &tlen, th, so))
14130                 return (1);
14131         /*
14132          * If last ACK falls within this segment's sequence numbers, record
14133          * its timestamp. NOTE: 1) That the test incorporates suggestions
14134          * from the latest proposal of the tcplw@cray.com list (Braden
14135          * 1993/04/26). 2) That updating only on newer timestamps interferes
14136          * with our earlier PAWS tests, so this check should be solely
14137          * predicated on the sequence space of this segment. 3) That we
14138          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14139          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14140          * SEG.Len, This modified check allows us to overcome RFC1323's
14141          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14142          * p.869. In such cases, we can still calculate the RTT correctly
14143          * when RCV.NXT == Last.ACK.Sent.
14144          */
14145         if ((to->to_flags & TOF_TS) != 0 &&
14146             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14147             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14148             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14149                 tp->ts_recent_age = tcp_ts_getticks();
14150                 tp->ts_recent = to->to_tsval;
14151         }
14152         /*
14153          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14154          * is on (half-synchronized state), then queue data for later
14155          * processing; else drop segment and return.
14156          */
14157         if ((thflags & TH_ACK) == 0) {
14158                 if (tp->t_flags & TF_NEEDSYN) {
14159                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14160                             tiwin, thflags, nxt_pkt));
14161                 } else if (tp->t_flags & TF_ACKNOW) {
14162                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14163                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14164                         return (ret_val);
14165                 } else {
14166                         ctf_do_drop(m, NULL);
14167                         return (0);
14168                 }
14169         }
14170         /*
14171          * Ack processing.
14172          */
14173         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14174                 return (ret_val);
14175         }
14176         if (sbavail(&so->so_snd)) {
14177                 if (ctf_progress_timeout_check(tp, true)) {
14178                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14179                                                 tp, tick, PROGRESS_DROP, __LINE__);
14180                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14181                         return (1);
14182                 }
14183         }
14184         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14185             tiwin, thflags, nxt_pkt));
14186 }
14187
14188 static void inline
14189 rack_clear_rate_sample(struct tcp_rack *rack)
14190 {
14191         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
14192         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
14193         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
14194 }
14195
14196 static void
14197 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
14198 {
14199         uint64_t bw_est, rate_wanted;
14200         int chged = 0;
14201         uint32_t user_max, orig_min, orig_max;
14202
14203 #ifdef TCP_REQUEST_TRK
14204         if (rack->rc_hybrid_mode &&
14205             (rack->r_ctl.rc_pace_max_segs != 0) &&
14206             (rack_hybrid_allow_set_maxseg == 1) &&
14207             (rack->r_ctl.rc_last_sft != NULL)) {
14208                 rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
14209                 return;
14210         }
14211 #endif
14212         orig_min = rack->r_ctl.rc_pace_min_segs;
14213         orig_max = rack->r_ctl.rc_pace_max_segs;
14214         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
14215         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
14216                 chged = 1;
14217         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
14218         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
14219                 if (user_max != rack->r_ctl.rc_pace_max_segs)
14220                         chged = 1;
14221         }
14222         if (rack->rc_force_max_seg) {
14223                 rack->r_ctl.rc_pace_max_segs = user_max;
14224         } else if (rack->use_fixed_rate) {
14225                 bw_est = rack_get_bw(rack);
14226                 if ((rack->r_ctl.crte == NULL) ||
14227                     (bw_est != rack->r_ctl.crte->rate)) {
14228                         rack->r_ctl.rc_pace_max_segs = user_max;
14229                 } else {
14230                         /* We are pacing right at the hardware rate */
14231                         uint32_t segsiz, pace_one;
14232
14233                         if (rack_pace_one_seg ||
14234                             (rack->r_ctl.rc_user_set_min_segs == 1))
14235                                 pace_one = 1;
14236                         else
14237                                 pace_one = 0;
14238                         segsiz = min(ctf_fixed_maxseg(tp),
14239                                      rack->r_ctl.rc_pace_min_segs);
14240                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14241                                 tp, bw_est, segsiz, pace_one,
14242                                 rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14243                 }
14244         } else if (rack->rc_always_pace) {
14245                 if (rack->r_ctl.gp_bw ||
14246                     rack->r_ctl.init_rate) {
14247                         /* We have a rate of some sort set */
14248                         uint32_t  orig;
14249
14250                         bw_est = rack_get_bw(rack);
14251                         orig = rack->r_ctl.rc_pace_max_segs;
14252                         if (fill_override)
14253                                 rate_wanted = *fill_override;
14254                         else
14255                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
14256                         if (rate_wanted) {
14257                                 /* We have something */
14258                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14259                                                                                    rate_wanted,
14260                                                                                    ctf_fixed_maxseg(rack->rc_tp));
14261                         } else
14262                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14263                         if (orig != rack->r_ctl.rc_pace_max_segs)
14264                                 chged = 1;
14265                 } else if ((rack->r_ctl.gp_bw == 0) &&
14266                            (rack->r_ctl.rc_pace_max_segs == 0)) {
14267                         /*
14268                          * If we have nothing limit us to bursting
14269                          * out IW sized pieces.
14270                          */
14271                         chged = 1;
14272                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14273                 }
14274         }
14275         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14276                 chged = 1;
14277                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14278         }
14279         if (chged)
14280                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14281 }
14282
14283
14284 static void
14285 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14286 {
14287 #ifdef INET6
14288         struct ip6_hdr *ip6 = NULL;
14289 #endif
14290 #ifdef INET
14291         struct ip *ip = NULL;
14292 #endif
14293         struct udphdr *udp = NULL;
14294
14295         /* Ok lets fill in the fast block, it can only be used with no IP options! */
14296 #ifdef INET6
14297         if (rack->r_is_v6) {
14298                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14299                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14300                 if (tp->t_port) {
14301                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14302                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14303                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14304                         udp->uh_dport = tp->t_port;
14305                         rack->r_ctl.fsb.udp = udp;
14306                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14307                 } else
14308                 {
14309                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14310                         rack->r_ctl.fsb.udp = NULL;
14311                 }
14312                 tcpip_fillheaders(rack->rc_inp,
14313                                   tp->t_port,
14314                                   ip6, rack->r_ctl.fsb.th);
14315                 rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14316         } else
14317 #endif                          /* INET6 */
14318 #ifdef INET
14319         {
14320                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14321                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14322                 if (tp->t_port) {
14323                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14324                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14325                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14326                         udp->uh_dport = tp->t_port;
14327                         rack->r_ctl.fsb.udp = udp;
14328                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14329                 } else
14330                 {
14331                         rack->r_ctl.fsb.udp = NULL;
14332                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14333                 }
14334                 tcpip_fillheaders(rack->rc_inp,
14335                                   tp->t_port,
14336                                   ip, rack->r_ctl.fsb.th);
14337                 rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14338         }
14339 #endif
14340         rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14341             (long)TCP_MAXWIN << tp->rcv_scale);
14342         rack->r_fsb_inited = 1;
14343 }
14344
14345 static int
14346 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14347 {
14348         /*
14349          * Allocate the larger of spaces V6 if available else just
14350          * V4 and include udphdr (overbook)
14351          */
14352 #ifdef INET6
14353         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14354 #else
14355         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14356 #endif
14357         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14358                                             M_TCPFSB, M_NOWAIT|M_ZERO);
14359         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14360                 return (ENOMEM);
14361         }
14362         rack->r_fsb_inited = 0;
14363         return (0);
14364 }
14365
14366 static void
14367 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14368 {
14369         /*
14370          * Types of logs (mod value)
14371          * 20 - Initial round setup
14372          * 21 - Rack declares a new round.
14373          */
14374         struct tcpcb *tp;
14375
14376         tp = rack->rc_tp;
14377         if (tcp_bblogging_on(tp)) {
14378                 union tcp_log_stackspecific log;
14379                 struct timeval tv;
14380
14381                 memset(&log, 0, sizeof(log));
14382                 log.u_bbr.flex1 = rack->r_ctl.current_round;
14383                 log.u_bbr.flex2 = rack->r_ctl.roundends;
14384                 log.u_bbr.flex3 = high_seq;
14385                 log.u_bbr.flex4 = tp->snd_max;
14386                 log.u_bbr.flex8 = mod;
14387                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14388                 log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14389                 log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14390                 TCP_LOG_EVENTP(tp, NULL,
14391                     &tptosocket(tp)->so_rcv,
14392                     &tptosocket(tp)->so_snd,
14393                     TCP_HYSTART, 0,
14394                     0, &log, false, &tv);
14395         }
14396 }
14397
14398 static void
14399 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14400 {
14401         rack->rack_deferred_inited = 1;
14402         rack->r_ctl.roundends = tp->snd_max;
14403         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14404         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14405 }
14406
14407 static void
14408 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14409 {
14410         /* Retransmit bit controls.
14411          *
14412          * The setting of these values control one of
14413          * three settings you can have and dictate
14414          * how rack does retransmissions. Note this
14415          * is in *any* mode i.e. pacing on or off DGP
14416          * fixed rate pacing, or just bursting rack.
14417          *
14418          * 1 - Use full sized retransmits i.e. limit
14419          *     the size to whatever the pace_max_segments
14420          *     size is.
14421          *
14422          * 2 - Use pacer min granularity as a guide to
14423          *     the size combined with the current calculated
14424          *     goodput b/w measurement. So for example if
14425          *     the goodput is measured at 20Mbps we would
14426          *     calculate 8125 (pacer minimum 250usec in
14427          *     that b/w) and then round it up to the next
14428          *     MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14429          *
14430          * 0 - The rack default 1 MSS (anything not 0/1/2
14431          *     fall here too if we are setting via rack_init()).
14432          *
14433          */
14434         if (ctl == 1) {
14435                 rack->full_size_rxt = 1;
14436                 rack->shape_rxt_to_pacing_min  = 0;
14437         } else if (ctl == 2) {
14438                 rack->full_size_rxt = 0;
14439                 rack->shape_rxt_to_pacing_min  = 1;
14440         } else {
14441                 rack->full_size_rxt = 0;
14442                 rack->shape_rxt_to_pacing_min  = 0;
14443         }
14444 }
14445
14446 static void
14447 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14448                   uint32_t flex1,
14449                   uint32_t flex2,
14450                   uint32_t flex3)
14451 {
14452         if (tcp_bblogging_on(rack->rc_tp)) {
14453                 union tcp_log_stackspecific log;
14454                 struct timeval tv;
14455
14456                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14457                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14458                 log.u_bbr.flex8 = mod;
14459                 log.u_bbr.flex1 = flex1;
14460                 log.u_bbr.flex2 = flex2;
14461                 log.u_bbr.flex3 = flex3;
14462                 tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14463                                0, &log, false, NULL, __func__, __LINE__, &tv);
14464         }
14465 }
14466
14467 static int
14468 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14469 {
14470         struct tcp_rack *rack;
14471         struct rack_sendmap *rsm;
14472         int i;
14473
14474
14475         rack = (struct tcp_rack *)tp->t_fb_ptr;
14476         switch (reqr->req) {
14477         case TCP_QUERY_SENDMAP:
14478                 if ((reqr->req_param == tp->snd_max) ||
14479                     (tp->snd_max == tp->snd_una)){
14480                         /* Unlikely */
14481                         return (0);
14482                 }
14483                 rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14484                 if (rsm == NULL) {
14485                         /* Can't find that seq -- unlikely */
14486                         return (0);
14487                 }
14488                 reqr->sendmap_start = rsm->r_start;
14489                 reqr->sendmap_end = rsm->r_end;
14490                 reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14491                 reqr->sendmap_fas = rsm->r_fas;
14492                 if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14493                         reqr->sendmap_send_cnt = SNDMAP_NRTX;
14494                 for(i=0; i<reqr->sendmap_send_cnt; i++)
14495                         reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14496                 reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14497                 reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14498                 reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14499                 reqr->sendmap_dupacks = rsm->r_dupack;
14500                 rack_log_chg_info(tp, rack, 1,
14501                                   rsm->r_start,
14502                                   rsm->r_end,
14503                                   rsm->r_flags);
14504                 return(1);
14505                 break;
14506         case TCP_QUERY_TIMERS_UP:
14507                 if (rack->r_ctl.rc_hpts_flags == 0) {
14508                         /* no timers up */
14509                         return (0);
14510                 }
14511                 reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14512                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14513                         reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14514                 }
14515                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14516                         reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
14517                 }
14518                 rack_log_chg_info(tp, rack, 2,
14519                                   rack->r_ctl.rc_hpts_flags,
14520                                   rack->r_ctl.rc_last_output_to,
14521                                   rack->r_ctl.rc_timer_exp);
14522                 return (1);
14523                 break;
14524         case TCP_QUERY_RACK_TIMES:
14525                 /* Reordering items */
14526                 reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
14527                 reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
14528                 /* Timerstamps and timers */
14529                 reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
14530                 reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
14531                 reqr->rack_rtt = rack->rc_rack_rtt;
14532                 reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
14533                 reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
14534                 /* PRR data */
14535                 reqr->rack_sacked = rack->r_ctl.rc_sacked;
14536                 reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
14537                 reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
14538                 reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
14539                 reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
14540                 reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
14541                 /* TLP and persists info */
14542                 reqr->rack_tlp_out = rack->rc_tlp_in_progress;
14543                 reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
14544                 if (rack->rc_in_persist) {
14545                         reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
14546                         reqr->rack_in_persist = 1;
14547                 } else {
14548                         reqr->rack_time_went_idle = 0;
14549                         reqr->rack_in_persist = 0;
14550                 }
14551                 if (rack->r_wanted_output)
14552                         reqr->rack_wanted_output = 1;
14553                 else
14554                         reqr->rack_wanted_output = 0;
14555                 return (1);
14556                 break;
14557         default:
14558                 return (-EINVAL);
14559         }
14560 }
14561
14562 static void
14563 rack_switch_failed(struct tcpcb *tp)
14564 {
14565         /*
14566          * This method gets called if a stack switch was
14567          * attempted and it failed. We are left
14568          * but our hpts timers were stopped and we
14569          * need to validate time units and inp_flags2.
14570          */
14571         struct inpcb *inp = tptoinpcb(tp);
14572         struct tcp_rack *rack;
14573         struct timeval tv;
14574         uint32_t cts;
14575         uint32_t toval;
14576         struct hpts_diag diag;
14577
14578         rack = (struct tcp_rack *)tp->t_fb_ptr;
14579         tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
14580         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14581                 inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
14582         else
14583                 inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
14584         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14585                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
14586         if (inp->inp_in_hpts) {
14587                 /* Strange */
14588                 return;
14589         }
14590         cts = tcp_get_usecs(&tv);
14591         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14592                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
14593                         toval = rack->r_ctl.rc_last_output_to - cts;
14594                 } else {
14595                         /* one slot please */
14596                         toval = HPTS_TICKS_PER_SLOT;
14597                 }
14598         } else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14599                 if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
14600                         toval = rack->r_ctl.rc_timer_exp - cts;
14601                 } else {
14602                         /* one slot please */
14603                         toval = HPTS_TICKS_PER_SLOT;
14604                 }
14605         } else
14606                 toval = HPTS_TICKS_PER_SLOT;
14607         (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(toval),
14608                                    __LINE__, &diag);
14609         rack_log_hpts_diag(rack, cts, &diag, &tv);
14610 }
14611
14612 static int
14613 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
14614 {
14615         struct rack_sendmap *rsm, *ersm;
14616         int insret __diagused;
14617         /*
14618          * When initing outstanding, we must be quite careful
14619          * to not refer to tp->t_fb_ptr. This has the old rack
14620          * pointer in it, not the "new" one (when we are doing
14621          * a stack switch).
14622          */
14623
14624
14625         if (tp->t_fb->tfb_chg_query == NULL) {
14626                 /* Create a send map for the current outstanding data */
14627
14628                 rsm = rack_alloc(rack);
14629                 if (rsm == NULL) {
14630                         uma_zfree(rack_pcb_zone, ptr);
14631                         return (ENOMEM);
14632                 }
14633                 rsm->r_no_rtt_allowed = 1;
14634                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
14635                 rsm->r_rtr_cnt = 1;
14636                 rsm->r_rtr_bytes = 0;
14637                 if (tp->t_flags & TF_SENTFIN)
14638                         rsm->r_flags |= RACK_HAS_FIN;
14639                 rsm->r_end = tp->snd_max;
14640                 if (tp->snd_una == tp->iss) {
14641                         /* The data space is one beyond snd_una */
14642                         rsm->r_flags |= RACK_HAS_SYN;
14643                         rsm->r_start = tp->iss;
14644                         rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
14645                 } else
14646                         rsm->r_start = tp->snd_una;
14647                 rsm->r_dupack = 0;
14648                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
14649                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
14650                         if (rsm->m) {
14651                                 rsm->orig_m_len = rsm->m->m_len;
14652                                 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14653                         } else {
14654                                 rsm->orig_m_len = 0;
14655                                 rsm->orig_t_space = 0;
14656                         }
14657                 } else {
14658                         /*
14659                          * This can happen if we have a stand-alone FIN or
14660                          *  SYN.
14661                          */
14662                         rsm->m = NULL;
14663                         rsm->orig_m_len = 0;
14664                         rsm->orig_t_space = 0;
14665                         rsm->soff = 0;
14666                 }
14667 #ifdef INVARIANTS
14668                 if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14669                         panic("Insert in rb tree fails ret:%d rack:%p rsm:%p",
14670                               insret, rack, rsm);
14671                 }
14672 #else
14673                 (void)tqhash_insert(rack->r_ctl.tqh, rsm);
14674 #endif
14675                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14676                 rsm->r_in_tmap = 1;
14677         } else {
14678                 /* We have a query mechanism, lets use it */
14679                 struct tcp_query_resp qr;
14680                 int i;
14681                 tcp_seq at;
14682
14683                 at = tp->snd_una;
14684                 while (at != tp->snd_max) {
14685                         memset(&qr, 0, sizeof(qr));
14686                         qr.req = TCP_QUERY_SENDMAP;
14687                         qr.req_param = at;
14688                         if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
14689                                 break;
14690                         /* Move forward */
14691                         at = qr.sendmap_end;
14692                         /* Now lets build the entry for this one */
14693                         rsm = rack_alloc(rack);
14694                         if (rsm == NULL) {
14695                                 uma_zfree(rack_pcb_zone, ptr);
14696                                 return (ENOMEM);
14697                         }
14698                         memset(rsm, 0, sizeof(struct rack_sendmap));
14699                         /* Now configure the rsm and insert it */
14700                         rsm->r_dupack = qr.sendmap_dupacks;
14701                         rsm->r_start = qr.sendmap_start;
14702                         rsm->r_end = qr.sendmap_end;
14703                         if (qr.sendmap_fas)
14704                                 rsm->r_fas = qr.sendmap_end;
14705                         else
14706                                 rsm->r_fas = rsm->r_start - tp->snd_una;
14707                         /*
14708                          * We have carefully aligned the bits
14709                          * so that all we have to do is copy over
14710                          * the bits with the mask.
14711                          */
14712                         rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
14713                         rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
14714                         rsm->r_rtr_cnt = qr.sendmap_send_cnt;
14715                         rsm->r_ack_arrival = qr.sendmap_ack_arrival;
14716                         for (i=0 ; i<rsm->r_rtr_cnt; i++)
14717                                 rsm->r_tim_lastsent[i]  = qr.sendmap_time[i];
14718                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
14719                                            (rsm->r_start - tp->snd_una), &rsm->soff);
14720                         if (rsm->m) {
14721                                 rsm->orig_m_len = rsm->m->m_len;
14722                                 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14723                         } else {
14724                                 rsm->orig_m_len = 0;
14725                                 rsm->orig_t_space = 0;
14726                         }
14727 #ifdef INVARIANTS
14728                         if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14729                                 panic("Insert in rb tree fails ret:%d rack:%p rsm:%p",
14730                                       insret, rack, rsm);
14731                         }
14732 #else
14733                         (void)tqhash_insert(rack->r_ctl.tqh, rsm);
14734 #endif
14735                         if ((rsm->r_flags & RACK_ACKED) == 0)  {
14736                                 TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
14737                                         if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
14738                                             rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
14739                                                 /*
14740                                                  * If the existing ersm was sent at
14741                                                  * a later time than the new one, then
14742                                                  * the new one should appear ahead of this
14743                                                  * ersm.
14744                                                  */
14745                                                 rsm->r_in_tmap = 1;
14746                                                 TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
14747                                                 break;
14748                                         }
14749                                 }
14750                                 if (rsm->r_in_tmap == 0) {
14751                                         /*
14752                                          * Not found so shove it on the tail.
14753                                          */
14754                                         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14755                                         rsm->r_in_tmap = 1;
14756                                 }
14757                         } else {
14758                                 if ((rack->r_ctl.rc_sacklast == NULL) ||
14759                                     (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
14760                                         rack->r_ctl.rc_sacklast = rsm;
14761                                 }
14762                         }
14763                         rack_log_chg_info(tp, rack, 3,
14764                                           rsm->r_start,
14765                                           rsm->r_end,
14766                                           rsm->r_flags);
14767                 }
14768         }
14769         return (0);
14770 }
14771
14772 static void
14773 rack_translate_clamp_value(struct tcp_rack *rack, uint32_t optval)
14774 {
14775         /*
14776          * P = percent bits
14777          * F = fill cw bit -- Toggle fillcw if this bit is set.
14778          * S = Segment bits
14779          * M = set max segment bit
14780          * U = Unclamined
14781          * C = If set to non-zero override the max number of clamps.
14782          * L = Bit to indicate if clamped gets lower.
14783          *
14784          * CCCC CCCCC UUUU UULF PPPP PPPP PPPP PPPP
14785          *
14786          * The lowest 3 nibbles is the perentage .1 - 6553.5%
14787          * where 10.1 = 101, max 6553.5
14788          * The upper 16 bits  holds some options.
14789          * The F bit will turn on fill-cw on if you are
14790          * not pacing, it will turn it off if dgp is on.
14791          * The L bit will change it so when clamped we get
14792          * the min(gp, lt-bw) for dgp.
14793          */
14794         uint16_t per;
14795
14796         rack->r_ctl.saved_rxt_clamp_val = optval;
14797         per = optval & 0x0000ffff;
14798         rack->r_ctl.rxt_threshold = (uint64_t)(per & 0xffff);
14799         if (optval > 0) {
14800                 uint16_t clamp_opt;
14801
14802                 rack->excess_rxt_on = 1;
14803                 clamp_opt = ((optval & 0xffff0000) >> 16);
14804                 rack->r_ctl.clamp_options = clamp_opt & 0x00ff;
14805                 if (clamp_opt & 0xff00) {
14806                         /* A max clamps is also present */
14807                         rack->r_ctl.max_clamps = (clamp_opt >> 8);
14808                 } else {
14809                         /* No specified clamps means no limit */
14810                         rack->r_ctl.max_clamps = 0;
14811                 }
14812                 if (rack->r_ctl.clamp_options & 0x0002) {
14813                         rack->r_clamped_gets_lower  = 1;
14814                 } else {
14815                         rack->r_clamped_gets_lower  = 0;
14816                 }
14817         } else {
14818                 /* Turn it off back to default */
14819                 rack->excess_rxt_on = 0;
14820                 rack->r_clamped_gets_lower  = 0;
14821         }
14822
14823 }
14824
14825
14826 static int32_t
14827 rack_init(struct tcpcb *tp, void **ptr)
14828 {
14829         struct inpcb *inp = tptoinpcb(tp);
14830         struct tcp_rack *rack = NULL;
14831         uint32_t iwin, snt, us_cts;
14832         int err, no_query;
14833
14834         /*
14835          * First are we the initial or are we a switched stack?
14836          * If we are initing via tcp_newtcppcb the ptr passed
14837          * will be tp->t_fb_ptr. If its a stack switch that
14838          * has a previous stack we can query it will be a local
14839          * var that will in the end be set into t_fb_ptr.
14840          */
14841         if (ptr == &tp->t_fb_ptr)
14842                 no_query = 1;
14843         else
14844                 no_query = 0;
14845         *ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
14846         if (*ptr == NULL) {
14847                 /*
14848                  * We need to allocate memory but cant. The INP and INP_INFO
14849                  * locks and they are recursive (happens during setup. So a
14850                  * scheme to drop the locks fails :(
14851                  *
14852                  */
14853                 return(ENOMEM);
14854         }
14855         memset(*ptr, 0, sizeof(struct tcp_rack));
14856         rack = (struct tcp_rack *)*ptr;
14857         rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
14858         if (rack->r_ctl.tqh == NULL) {
14859                 uma_zfree(rack_pcb_zone, rack);
14860                 return(ENOMEM);
14861         }
14862         tqhash_init(rack->r_ctl.tqh);
14863         TAILQ_INIT(&rack->r_ctl.rc_free);
14864         TAILQ_INIT(&rack->r_ctl.rc_tmap);
14865         rack->rc_tp = tp;
14866         rack->rc_inp = inp;
14867         /* Set the flag */
14868         rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
14869         /* Probably not needed but lets be sure */
14870         rack_clear_rate_sample(rack);
14871         /*
14872          * Save off the default values, socket options will poke
14873          * at these if pacing is not on or we have not yet
14874          * reached where pacing is on (gp_ready/fixed enabled).
14875          * When they get set into the CC module (when gp_ready
14876          * is enabled or we enable fixed) then we will set these
14877          * values into the CC and place in here the old values
14878          * so we have a restoral. Then we will set the flag
14879          * rc_pacing_cc_set. That way whenever we turn off pacing
14880          * or switch off this stack, we will know to go restore
14881          * the saved values.
14882          *
14883          * We specifically put into the beta the ecn value for pacing.
14884          */
14885         rack->rc_new_rnd_needed = 1;
14886         rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
14887         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
14888         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
14889         /* We want abe like behavior as well */
14890         rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
14891         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
14892         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
14893         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
14894         if (rack_rxt_clamp_thresh) {
14895                 rack_translate_clamp_value(rack, rack_rxt_clamp_thresh);
14896                 rack->excess_rxt_on = 1;
14897         }
14898         if (rack_uses_full_dgp_in_rec)
14899                 rack->r_ctl.full_dgp_in_rec = 1;
14900         if (rack_fill_cw_state)
14901                 rack->rc_pace_to_cwnd = 1;
14902         if (rack_pacing_min_seg)
14903                 rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
14904         if (use_rack_rr)
14905                 rack->use_rack_rr = 1;
14906         if (rack_dnd_default) {
14907                 rack->rc_pace_dnd = 1;
14908         }
14909         if (V_tcp_delack_enabled)
14910                 tp->t_delayed_ack = 1;
14911         else
14912                 tp->t_delayed_ack = 0;
14913 #ifdef TCP_ACCOUNTING
14914         if (rack_tcp_accounting) {
14915                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
14916         }
14917 #endif
14918         rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
14919         rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
14920         if (rack_enable_shared_cwnd)
14921                 rack->rack_enable_scwnd = 1;
14922         rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
14923         rack->rc_user_set_max_segs = rack_hptsi_segments;
14924         rack->rc_force_max_seg = 0;
14925         TAILQ_INIT(&rack->r_ctl.opt_list);
14926         if (rack_hibeta_setting)
14927                 rack->rack_hibeta = 1;
14928         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
14929         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
14930         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
14931         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
14932         rack->r_ctl.rc_highest_us_rtt = 0;
14933         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
14934         rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
14935         if (rack_use_cmp_acks)
14936                 rack->r_use_cmp_ack = 1;
14937         if (rack_disable_prr)
14938                 rack->rack_no_prr = 1;
14939         if (rack_gp_no_rec_chg)
14940                 rack->rc_gp_no_rec_chg = 1;
14941         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
14942                 rack->rc_always_pace = 1;
14943                 if ((rack->gp_ready) && (rack->rc_always_pace && (rack->use_fixed_rate == 0)))
14944                         rack_set_cc_pacing(rack);
14945         } else
14946                 rack->rc_always_pace = 0;
14947         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
14948                 rack->r_mbuf_queue = 1;
14949         else
14950                 rack->r_mbuf_queue = 0;
14951         rack_set_pace_segments(tp, rack, __LINE__, NULL);
14952         if (rack_limits_scwnd)
14953                 rack->r_limit_scw = 1;
14954         else
14955                 rack->r_limit_scw = 0;
14956         rack_init_retransmit_value(rack, rack_rxt_controls);
14957         rack->rc_labc = V_tcp_abc_l_var;
14958         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
14959         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
14960         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
14961         rack->r_ctl.rc_min_to = rack_min_to;
14962         microuptime(&rack->r_ctl.act_rcv_time);
14963         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
14964         rack->rc_init_win = rack_default_init_window;
14965         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
14966         if (rack_hw_up_only)
14967                 rack->r_up_only = 1;
14968         if (rack_do_dyn_mul) {
14969                 /* When dynamic adjustment is on CA needs to start at 100% */
14970                 rack->rc_gp_dyn_mul = 1;
14971                 if (rack_do_dyn_mul >= 100)
14972                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
14973         } else
14974                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
14975         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
14976         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
14977         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
14978         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
14979                                 rack_probertt_filter_life);
14980         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14981         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
14982         rack->r_ctl.rc_time_of_last_probertt = us_cts;
14983         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
14984         rack->r_ctl.rc_time_probertt_starts = 0;
14985         if (rack_dsack_std_based & 0x1) {
14986                 /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
14987                 rack->rc_rack_tmr_std_based = 1;
14988         }
14989         if (rack_dsack_std_based & 0x2) {
14990                 /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
14991                 rack->rc_rack_use_dsack = 1;
14992         }
14993         /* We require at least one measurement, even if the sysctl is 0 */
14994         if (rack_req_measurements)
14995                 rack->r_ctl.req_measurements = rack_req_measurements;
14996         else
14997                 rack->r_ctl.req_measurements = 1;
14998         if (rack_enable_hw_pacing)
14999                 rack->rack_hdw_pace_ena = 1;
15000         if (rack_hw_rate_caps)
15001                 rack->r_rack_hw_rate_caps = 1;
15002 #ifdef TCP_SAD_DETECTION
15003         rack->do_detection = 1;
15004 #else
15005         rack->do_detection = 0;
15006 #endif
15007         if (rack_non_rxt_use_cr)
15008                 rack->rack_rec_nonrxt_use_cr = 1;
15009         /* Lets setup the fsb block */
15010         err = rack_init_fsb(tp, rack);
15011         if (err) {
15012                 uma_zfree(rack_pcb_zone, *ptr);
15013                 *ptr = NULL;
15014                 return (err);
15015         }
15016         if (rack_do_hystart) {
15017                 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
15018                 if (rack_do_hystart > 1)
15019                         tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
15020                 if (rack_do_hystart > 2)
15021                         tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
15022         }
15023         /* Log what we will do with queries */
15024         rack_log_chg_info(tp, rack, 7,
15025                           no_query, 0, 0);
15026         if (rack_def_profile)
15027                 rack_set_profile(rack, rack_def_profile);
15028         /* Cancel the GP measurement in progress */
15029         tp->t_flags &= ~TF_GPUTINPROG;
15030         if ((tp->t_state != TCPS_CLOSED) &&
15031             (tp->t_state != TCPS_TIME_WAIT)) {
15032                 /*
15033                  * We are already open, we may
15034                  * need to adjust a few things.
15035                  */
15036                 if (SEQ_GT(tp->snd_max, tp->iss))
15037                         snt = tp->snd_max - tp->iss;
15038                 else
15039                         snt = 0;
15040                 iwin = rc_init_window(rack);
15041                 if ((snt < iwin) &&
15042                     (no_query == 1)) {
15043                         /* We are not past the initial window
15044                          * on the first init (i.e. a stack switch
15045                          * has not yet occured) so we need to make
15046                          * sure cwnd and ssthresh is correct.
15047                          */
15048                         if (tp->snd_cwnd < iwin)
15049                                 tp->snd_cwnd = iwin;
15050                         /*
15051                          * If we are within the initial window
15052                          * we want ssthresh to be unlimited. Setting
15053                          * it to the rwnd (which the default stack does
15054                          * and older racks) is not really a good idea
15055                          * since we want to be in SS and grow both the
15056                          * cwnd and the rwnd (via dynamic rwnd growth). If
15057                          * we set it to the rwnd then as the peer grows its
15058                          * rwnd we will be stuck in CA and never hit SS.
15059                          *
15060                          * Its far better to raise it up high (this takes the
15061                          * risk that there as been a loss already, probably
15062                          * we should have an indicator in all stacks of loss
15063                          * but we don't), but considering the normal use this
15064                          * is a risk worth taking. The consequences of not
15065                          * hitting SS are far worse than going one more time
15066                          * into it early on (before we have sent even a IW).
15067                          * It is highly unlikely that we will have had a loss
15068                          * before getting the IW out.
15069                          */
15070                         tp->snd_ssthresh = 0xffffffff;
15071                 }
15072                 /*
15073                  * Any init based on sequence numbers
15074                  * should be done in the deferred init path
15075                  * since we can be CLOSED and not have them
15076                  * inited when rack_init() is called. We
15077                  * are not closed so lets call it.
15078                  */
15079                 rack_deferred_init(tp, rack);
15080         }
15081         if ((tp->t_state != TCPS_CLOSED) &&
15082             (tp->t_state != TCPS_TIME_WAIT) &&
15083             (no_query == 0) &&
15084             (tp->snd_una != tp->snd_max))  {
15085                 err = rack_init_outstanding(tp, rack, us_cts, *ptr);
15086                 if (err) {
15087                         *ptr = NULL;
15088                         return(err);
15089                 }
15090         }
15091         rack_stop_all_timers(tp, rack);
15092         /* Setup all the inp_flags2 */
15093         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
15094                 tptoinpcb(tp)->inp_flags2 |= INP_SUPPORTS_MBUFQ;
15095         else
15096                 tptoinpcb(tp)->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
15097         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15098                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
15099         /*
15100          * Timers in Rack are kept in microseconds so lets
15101          * convert any initial incoming variables
15102          * from ticks into usecs. Note that we
15103          * also change the values of t_srtt and t_rttvar, if
15104          * they are non-zero. They are kept with a 5
15105          * bit decimal so we have to carefully convert
15106          * these to get the full precision.
15107          */
15108         rack_convert_rtts(tp);
15109         rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
15110         if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
15111                 /* We do not start any timers on DROPPED connections */
15112                 if (tp->t_fb->tfb_chg_query == NULL) {
15113                         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15114                 } else {
15115                         struct tcp_query_resp qr;
15116                         int ret;
15117
15118                         memset(&qr, 0, sizeof(qr));
15119
15120                         /* Get the misc time stamps and such for rack */
15121                         qr.req = TCP_QUERY_RACK_TIMES;
15122                         ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15123                         if (ret == 1) {
15124                                 rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
15125                                 rack->r_ctl.num_dsack  = qr.rack_num_dsacks;
15126                                 rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
15127                                 rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
15128                                 rack->rc_rack_rtt = qr.rack_rtt;
15129                                 rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
15130                                 rack->r_ctl.rc_sacked = qr.rack_sacked;
15131                                 rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
15132                                 rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
15133                                 rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
15134                                 rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
15135                                 rack->r_ctl.rc_prr_out = qr.rack_prr_out;
15136                                 if (qr.rack_tlp_out) {
15137                                         rack->rc_tlp_in_progress = 1;
15138                                         rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
15139                                 } else {
15140                                         rack->rc_tlp_in_progress = 0;
15141                                         rack->r_ctl.rc_tlp_cnt_out = 0;
15142                                 }
15143                                 if (qr.rack_srtt_measured)
15144                                         rack->rc_srtt_measure_made = 1;
15145                                 if (qr.rack_in_persist == 1) {
15146                                         rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
15147 #ifdef NETFLIX_SHARED_CWND
15148                                         if (rack->r_ctl.rc_scw) {
15149                                                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
15150                                                 rack->rack_scwnd_is_idle = 1;
15151                                         }
15152 #endif
15153                                         rack->r_ctl.persist_lost_ends = 0;
15154                                         rack->probe_not_answered = 0;
15155                                         rack->forced_ack = 0;
15156                                         tp->t_rxtshift = 0;
15157                                         rack->rc_in_persist = 1;
15158                                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
15159                                                            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
15160                                 }
15161                                 if (qr.rack_wanted_output)
15162                                         rack->r_wanted_output = 1;
15163                                 rack_log_chg_info(tp, rack, 6,
15164                                                   qr.rack_min_rtt,
15165                                                   qr.rack_rtt,
15166                                                   qr.rack_reorder_ts);
15167                         }
15168                         /* Get the old stack timers */
15169                         qr.req_param = 0;
15170                         qr.req = TCP_QUERY_TIMERS_UP;
15171                         ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15172                         if (ret) {
15173                                 /*
15174                                  * non-zero return means we have a timer('s)
15175                                  * to start. Zero means no timer (no keepalive
15176                                  * I suppose).
15177                                  */
15178                                 uint32_t tov = 0;
15179
15180                                 rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
15181                                 if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
15182                                         rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
15183                                         if (TSTMP_GT(qr.timer_pacing_to, us_cts))
15184                                                 tov = qr.timer_pacing_to - us_cts;
15185                                         else
15186                                                 tov = HPTS_TICKS_PER_SLOT;
15187                                 }
15188                                 if (qr.timer_hpts_flags & PACE_TMR_MASK) {
15189                                         rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
15190                                         if (tov == 0) {
15191                                                 if (TSTMP_GT(qr.timer_timer_exp, us_cts))
15192                                                         tov = qr.timer_timer_exp - us_cts;
15193                                                 else
15194                                                         tov = HPTS_TICKS_PER_SLOT;
15195                                         }
15196                                 }
15197                                 rack_log_chg_info(tp, rack, 4,
15198                                                   rack->r_ctl.rc_hpts_flags,
15199                                                   rack->r_ctl.rc_last_output_to,
15200                                                   rack->r_ctl.rc_timer_exp);
15201                                 if (tov) {
15202                                         struct hpts_diag diag;
15203
15204                                         (void)tcp_hpts_insert_diag(rack->rc_inp, HPTS_USEC_TO_SLOTS(tov),
15205                                                                    __LINE__, &diag);
15206                                         rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
15207                                 }
15208                         }
15209                 }
15210                 rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
15211                                      __LINE__, RACK_RTTS_INIT);
15212         }
15213         return (0);
15214 }
15215
15216 static int
15217 rack_handoff_ok(struct tcpcb *tp)
15218 {
15219         if ((tp->t_state == TCPS_CLOSED) ||
15220             (tp->t_state == TCPS_LISTEN)) {
15221                 /* Sure no problem though it may not stick */
15222                 return (0);
15223         }
15224         if ((tp->t_state == TCPS_SYN_SENT) ||
15225             (tp->t_state == TCPS_SYN_RECEIVED)) {
15226                 /*
15227                  * We really don't know if you support sack,
15228                  * you have to get to ESTAB or beyond to tell.
15229                  */
15230                 return (EAGAIN);
15231         }
15232         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15233                 /*
15234                  * Rack will only send a FIN after all data is acknowledged.
15235                  * So in this case we have more data outstanding. We can't
15236                  * switch stacks until either all data and only the FIN
15237                  * is left (in which case rack_init() now knows how
15238                  * to deal with that) <or> all is acknowledged and we
15239                  * are only left with incoming data, though why you
15240                  * would want to switch to rack after all data is acknowledged
15241                  * I have no idea (rrs)!
15242                  */
15243                 return (EAGAIN);
15244         }
15245         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15246                 return (0);
15247         }
15248         /*
15249          * If we reach here we don't do SACK on this connection so we can
15250          * never do rack.
15251          */
15252         return (EINVAL);
15253 }
15254
15255 static void
15256 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15257 {
15258
15259         if (tp->t_fb_ptr) {
15260                 uint32_t cnt_free = 0;
15261                 struct tcp_rack *rack;
15262                 struct rack_sendmap *rsm;
15263
15264                 tcp_handle_orphaned_packets(tp);
15265                 tp->t_flags &= ~TF_FORCEDATA;
15266                 rack = (struct tcp_rack *)tp->t_fb_ptr;
15267                 rack_log_pacing_delay_calc(rack,
15268                                            0,
15269                                            0,
15270                                            0,
15271                                            rack_get_gp_est(rack), /* delRate */
15272                                            rack_get_lt_bw(rack), /* rttProp */
15273                                            20, __LINE__, NULL, 0);
15274 #ifdef NETFLIX_SHARED_CWND
15275                 if (rack->r_ctl.rc_scw) {
15276                         uint32_t limit;
15277
15278                         if (rack->r_limit_scw)
15279                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15280                         else
15281                                 limit = 0;
15282                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15283                                                   rack->r_ctl.rc_scw_index,
15284                                                   limit);
15285                         rack->r_ctl.rc_scw = NULL;
15286                 }
15287 #endif
15288                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
15289                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15290                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15291                         rack->r_ctl.fsb.th = NULL;
15292                 }
15293                 if (rack->rc_always_pace) {
15294                         tcp_decrement_paced_conn();
15295                         rack_undo_cc_pacing(rack);
15296                         rack->rc_always_pace = 0;
15297                 }
15298                 /* Clean up any options if they were not applied */
15299                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15300                         struct deferred_opt_list *dol;
15301
15302                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15303                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15304                         free(dol, M_TCPDO);
15305                 }
15306                 /* rack does not use force data but other stacks may clear it */
15307                 if (rack->r_ctl.crte != NULL) {
15308                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15309                         rack->rack_hdrw_pacing = 0;
15310                         rack->r_ctl.crte = NULL;
15311                 }
15312 #ifdef TCP_BLACKBOX
15313                 tcp_log_flowend(tp);
15314 #endif
15315                 /*
15316                  * Lets take a different approach to purging just
15317                  * get each one and free it like a cum-ack would and
15318                  * not use a foreach loop.
15319                  */
15320                 rsm = tqhash_min(rack->r_ctl.tqh);
15321                 while (rsm) {
15322                         tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15323                         rack->r_ctl.rc_num_maps_alloced--;
15324                         uma_zfree(rack_zone, rsm);
15325                         rsm = tqhash_min(rack->r_ctl.tqh);
15326                 }
15327                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15328                 while (rsm) {
15329                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15330                         rack->r_ctl.rc_num_maps_alloced--;
15331                         rack->rc_free_cnt--;
15332                         cnt_free++;
15333                         uma_zfree(rack_zone, rsm);
15334                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15335                 }
15336                 if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15337                     (tcp_bblogging_on(tp))) {
15338                         union tcp_log_stackspecific log;
15339                         struct timeval tv;
15340
15341                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15342                         log.u_bbr.flex8 = 10;
15343                         log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15344                         log.u_bbr.flex2 = rack->rc_free_cnt;
15345                         log.u_bbr.flex3 = cnt_free;
15346                         log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15347                         rsm = tqhash_min(rack->r_ctl.tqh);
15348                         log.u_bbr.delRate = (uint64_t)rsm;
15349                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15350                         log.u_bbr.cur_del_rate = (uint64_t)rsm;
15351                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15352                         log.u_bbr.pkt_epoch = __LINE__;
15353                         (void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15354                                              0, &log, false, NULL, NULL, 0, &tv);
15355                 }
15356                 KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15357                         ("rack:%p num_aloc:%u after freeing all?",
15358                          rack,
15359                          rack->r_ctl.rc_num_maps_alloced));
15360                 rack->rc_free_cnt = 0;
15361                 free(rack->r_ctl.tqh, M_TCPFSB);
15362                 rack->r_ctl.tqh = NULL;
15363                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15364                 tp->t_fb_ptr = NULL;
15365         }
15366         /* Make sure snd_nxt is correctly set */
15367         tp->snd_nxt = tp->snd_max;
15368 }
15369
15370 static void
15371 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15372 {
15373         if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15374                 rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15375         }
15376         switch (tp->t_state) {
15377         case TCPS_SYN_SENT:
15378                 rack->r_state = TCPS_SYN_SENT;
15379                 rack->r_substate = rack_do_syn_sent;
15380                 break;
15381         case TCPS_SYN_RECEIVED:
15382                 rack->r_state = TCPS_SYN_RECEIVED;
15383                 rack->r_substate = rack_do_syn_recv;
15384                 break;
15385         case TCPS_ESTABLISHED:
15386                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15387                 rack->r_state = TCPS_ESTABLISHED;
15388                 rack->r_substate = rack_do_established;
15389                 break;
15390         case TCPS_CLOSE_WAIT:
15391                 rack->r_state = TCPS_CLOSE_WAIT;
15392                 rack->r_substate = rack_do_close_wait;
15393                 break;
15394         case TCPS_FIN_WAIT_1:
15395                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15396                 rack->r_state = TCPS_FIN_WAIT_1;
15397                 rack->r_substate = rack_do_fin_wait_1;
15398                 break;
15399         case TCPS_CLOSING:
15400                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15401                 rack->r_state = TCPS_CLOSING;
15402                 rack->r_substate = rack_do_closing;
15403                 break;
15404         case TCPS_LAST_ACK:
15405                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15406                 rack->r_state = TCPS_LAST_ACK;
15407                 rack->r_substate = rack_do_lastack;
15408                 break;
15409         case TCPS_FIN_WAIT_2:
15410                 rack->r_state = TCPS_FIN_WAIT_2;
15411                 rack->r_substate = rack_do_fin_wait_2;
15412                 break;
15413         case TCPS_LISTEN:
15414         case TCPS_CLOSED:
15415         case TCPS_TIME_WAIT:
15416         default:
15417                 break;
15418         };
15419         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15420                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
15421
15422 }
15423
15424 static void
15425 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15426 {
15427         /*
15428          * We received an ack, and then did not
15429          * call send or were bounced out due to the
15430          * hpts was running. Now a timer is up as well, is
15431          * it the right timer?
15432          */
15433         struct rack_sendmap *rsm;
15434         int tmr_up;
15435
15436         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15437         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15438                 return;
15439         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15440         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15441             (tmr_up == PACE_TMR_RXT)) {
15442                 /* Should be an RXT */
15443                 return;
15444         }
15445         if (rsm == NULL) {
15446                 /* Nothing outstanding? */
15447                 if (tp->t_flags & TF_DELACK) {
15448                         if (tmr_up == PACE_TMR_DELACK)
15449                                 /* We are supposed to have delayed ack up and we do */
15450                                 return;
15451                 } else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
15452                         /*
15453                          * if we hit enobufs then we would expect the possibility
15454                          * of nothing outstanding and the RXT up (and the hptsi timer).
15455                          */
15456                         return;
15457                 } else if (((V_tcp_always_keepalive ||
15458                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
15459                             (tp->t_state <= TCPS_CLOSING)) &&
15460                            (tmr_up == PACE_TMR_KEEP) &&
15461                            (tp->snd_max == tp->snd_una)) {
15462                         /* We should have keep alive up and we do */
15463                         return;
15464                 }
15465         }
15466         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
15467                    ((tmr_up == PACE_TMR_TLP) ||
15468                     (tmr_up == PACE_TMR_RACK) ||
15469                     (tmr_up == PACE_TMR_RXT))) {
15470                 /*
15471                  * Either a Rack, TLP or RXT is fine if  we
15472                  * have outstanding data.
15473                  */
15474                 return;
15475         } else if (tmr_up == PACE_TMR_DELACK) {
15476                 /*
15477                  * If the delayed ack was going to go off
15478                  * before the rtx/tlp/rack timer were going to
15479                  * expire, then that would be the timer in control.
15480                  * Note we don't check the time here trusting the
15481                  * code is correct.
15482                  */
15483                 return;
15484         }
15485         /*
15486          * Ok the timer originally started is not what we want now.
15487          * We will force the hpts to be stopped if any, and restart
15488          * with the slot set to what was in the saved slot.
15489          */
15490         if (tcp_in_hpts(rack->rc_inp)) {
15491                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15492                         uint32_t us_cts;
15493
15494                         us_cts = tcp_get_usecs(NULL);
15495                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
15496                                 rack->r_early = 1;
15497                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
15498                         }
15499                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
15500                 }
15501                 tcp_hpts_remove(rack->rc_inp);
15502         }
15503         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15504         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15505 }
15506
15507
15508 static void
15509 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
15510 {
15511         if ((SEQ_LT(tp->snd_wl1, seq) ||
15512             (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
15513             (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
15514                 /* keep track of pure window updates */
15515                 if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
15516                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
15517                 tp->snd_wnd = tiwin;
15518                 rack_validate_fo_sendwin_up(tp, rack);
15519                 tp->snd_wl1 = seq;
15520                 tp->snd_wl2 = ack;
15521                 if (tp->snd_wnd > tp->max_sndwnd)
15522                         tp->max_sndwnd = tp->snd_wnd;
15523             rack->r_wanted_output = 1;
15524         } else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
15525                 tp->snd_wnd = tiwin;
15526                 rack_validate_fo_sendwin_up(tp, rack);
15527                 tp->snd_wl1 = seq;
15528                 tp->snd_wl2 = ack;
15529         } else {
15530                 /* Not a valid win update */
15531                 return;
15532         }
15533         if (tp->snd_wnd > tp->max_sndwnd)
15534                 tp->max_sndwnd = tp->snd_wnd;
15535         /* Do we exit persists? */
15536         if ((rack->rc_in_persist != 0) &&
15537             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
15538                                 rack->r_ctl.rc_pace_min_segs))) {
15539                 rack_exit_persist(tp, rack, cts);
15540         }
15541         /* Do we enter persists? */
15542         if ((rack->rc_in_persist == 0) &&
15543             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
15544             TCPS_HAVEESTABLISHED(tp->t_state) &&
15545             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
15546             sbavail(&tptosocket(tp)->so_snd) &&
15547             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
15548                 /*
15549                  * Here the rwnd is less than
15550                  * the pacing size, we are established,
15551                  * nothing is outstanding, and there is
15552                  * data to send. Enter persists.
15553                  */
15554                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
15555         }
15556 }
15557
15558 static void
15559 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
15560 {
15561
15562         if (tcp_bblogging_on(rack->rc_tp)) {
15563                 struct inpcb *inp = tptoinpcb(tp);
15564                 union tcp_log_stackspecific log;
15565                 struct timeval ltv;
15566                 char tcp_hdr_buf[60];
15567                 struct tcphdr *th;
15568                 struct timespec ts;
15569                 uint32_t orig_snd_una;
15570                 uint8_t xx = 0;
15571
15572 #ifdef TCP_REQUEST_TRK
15573                 struct http_sendfile_track *http_req;
15574
15575                 if (SEQ_GT(ae->ack, tp->snd_una)) {
15576                         http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
15577                 } else {
15578                         http_req = tcp_http_find_req_for_seq(tp, ae->ack);
15579                 }
15580 #endif
15581                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15582                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15583                 if (rack->rack_no_prr == 0)
15584                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15585                 else
15586                         log.u_bbr.flex1 = 0;
15587                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
15588                 log.u_bbr.use_lt_bw <<= 1;
15589                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
15590                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
15591                 log.u_bbr.bbr_state = rack->rc_free_cnt;
15592                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15593                 log.u_bbr.pkts_out = tp->t_maxseg;
15594                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
15595                 log.u_bbr.flex7 = 1;
15596                 log.u_bbr.lost = ae->flags;
15597                 log.u_bbr.cwnd_gain = ackval;
15598                 log.u_bbr.pacing_gain = 0x2;
15599                 if (ae->flags & TSTMP_HDWR) {
15600                         /* Record the hardware timestamp if present */
15601                         log.u_bbr.flex3 = M_TSTMP;
15602                         ts.tv_sec = ae->timestamp / 1000000000;
15603                         ts.tv_nsec = ae->timestamp % 1000000000;
15604                         ltv.tv_sec = ts.tv_sec;
15605                         ltv.tv_usec = ts.tv_nsec / 1000;
15606                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
15607                 } else if (ae->flags & TSTMP_LRO) {
15608                         /* Record the LRO the arrival timestamp */
15609                         log.u_bbr.flex3 = M_TSTMP_LRO;
15610                         ts.tv_sec = ae->timestamp / 1000000000;
15611                         ts.tv_nsec = ae->timestamp % 1000000000;
15612                         ltv.tv_sec = ts.tv_sec;
15613                         ltv.tv_usec = ts.tv_nsec / 1000;
15614                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
15615                 }
15616                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
15617                 /* Log the rcv time */
15618                 log.u_bbr.delRate = ae->timestamp;
15619 #ifdef TCP_REQUEST_TRK
15620                 log.u_bbr.applimited = tp->t_http_closed;
15621                 log.u_bbr.applimited <<= 8;
15622                 log.u_bbr.applimited |= tp->t_http_open;
15623                 log.u_bbr.applimited <<= 8;
15624                 log.u_bbr.applimited |= tp->t_http_req;
15625                 if (http_req) {
15626                         /* Copy out any client req info */
15627                         /* seconds */
15628                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
15629                         /* useconds */
15630                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
15631                         log.u_bbr.rttProp = http_req->timestamp;
15632                         log.u_bbr.cur_del_rate = http_req->start;
15633                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
15634                                 log.u_bbr.flex8 |= 1;
15635                         } else {
15636                                 log.u_bbr.flex8 |= 2;
15637                                 log.u_bbr.bw_inuse = http_req->end;
15638                         }
15639                         log.u_bbr.flex6 = http_req->start_seq;
15640                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
15641                                 log.u_bbr.flex8 |= 4;
15642                                 log.u_bbr.epoch = http_req->end_seq;
15643                         }
15644                 }
15645 #endif
15646                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
15647                 th = (struct tcphdr *)tcp_hdr_buf;
15648                 th->th_seq = ae->seq;
15649                 th->th_ack = ae->ack;
15650                 th->th_win = ae->win;
15651                 /* Now fill in the ports */
15652                 th->th_sport = inp->inp_fport;
15653                 th->th_dport = inp->inp_lport;
15654                 tcp_set_flags(th, ae->flags);
15655                 /* Now do we have a timestamp option? */
15656                 if (ae->flags & HAS_TSTMP) {
15657                         u_char *cp;
15658                         uint32_t val;
15659
15660                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
15661                         cp = (u_char *)(th + 1);
15662                         *cp = TCPOPT_NOP;
15663                         cp++;
15664                         *cp = TCPOPT_NOP;
15665                         cp++;
15666                         *cp = TCPOPT_TIMESTAMP;
15667                         cp++;
15668                         *cp = TCPOLEN_TIMESTAMP;
15669                         cp++;
15670                         val = htonl(ae->ts_value);
15671                         bcopy((char *)&val,
15672                               (char *)cp, sizeof(uint32_t));
15673                         val = htonl(ae->ts_echo);
15674                         bcopy((char *)&val,
15675                               (char *)(cp + 4), sizeof(uint32_t));
15676                 } else
15677                         th->th_off = (sizeof(struct tcphdr) >> 2);
15678
15679                 /*
15680                  * For sane logging we need to play a little trick.
15681                  * If the ack were fully processed we would have moved
15682                  * snd_una to high_seq, but since compressed acks are
15683                  * processed in two phases, at this point (logging) snd_una
15684                  * won't be advanced. So we would see multiple acks showing
15685                  * the advancement. We can prevent that by "pretending" that
15686                  * snd_una was advanced and then un-advancing it so that the
15687                  * logging code has the right value for tlb_snd_una.
15688                  */
15689                 if (tp->snd_una != high_seq) {
15690                         orig_snd_una = tp->snd_una;
15691                         tp->snd_una = high_seq;
15692                         xx = 1;
15693                 } else
15694                         xx = 0;
15695                 TCP_LOG_EVENTP(tp, th,
15696                                &tptosocket(tp)->so_rcv,
15697                                &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
15698                                0, &log, true, &ltv);
15699                 if (xx) {
15700                         tp->snd_una = orig_snd_una;
15701                 }
15702         }
15703
15704 }
15705
15706 static void
15707 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
15708 {
15709         uint32_t us_rtt;
15710         /*
15711          * A persist or keep-alive was forced out, update our
15712          * min rtt time. Note now worry about lost responses.
15713          * When a subsequent keep-alive or persist times out
15714          * and forced_ack is still on, then the last probe
15715          * was not responded to. In such cases we have a
15716          * sysctl that controls the behavior. Either we apply
15717          * the rtt but with reduced confidence (0). Or we just
15718          * plain don't apply the rtt estimate. Having data flow
15719          * will clear the probe_not_answered flag i.e. cum-ack
15720          * move forward <or> exiting and reentering persists.
15721          */
15722
15723         rack->forced_ack = 0;
15724         rack->rc_tp->t_rxtshift = 0;
15725         if ((rack->rc_in_persist &&
15726              (tiwin == rack->rc_tp->snd_wnd)) ||
15727             (rack->rc_in_persist == 0)) {
15728                 /*
15729                  * In persists only apply the RTT update if this is
15730                  * a response to our window probe. And that
15731                  * means the rwnd sent must match the current
15732                  * snd_wnd. If it does not, then we got a
15733                  * window update ack instead. For keepalive
15734                  * we allow the answer no matter what the window.
15735                  *
15736                  * Note that if the probe_not_answered is set then
15737                  * the forced_ack_ts is the oldest one i.e. the first
15738                  * probe sent that might have been lost. This assures
15739                  * us that if we do calculate an RTT it is longer not
15740                  * some short thing.
15741                  */
15742                 if (rack->rc_in_persist)
15743                         counter_u64_add(rack_persists_acks, 1);
15744                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
15745                 if (us_rtt == 0)
15746                         us_rtt = 1;
15747                 if (rack->probe_not_answered == 0) {
15748                         rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15749                         tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
15750                 } else {
15751                         /* We have a retransmitted probe here too */
15752                         if (rack_apply_rtt_with_reduced_conf) {
15753                                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15754                                 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
15755                         }
15756                 }
15757         }
15758 }
15759
15760 static int
15761 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
15762 {
15763         /*
15764          * Handle a "special" compressed ack mbuf. Each incoming
15765          * ack has only four possible dispositions:
15766          *
15767          * A) It moves the cum-ack forward
15768          * B) It is behind the cum-ack.
15769          * C) It is a window-update ack.
15770          * D) It is a dup-ack.
15771          *
15772          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
15773          * in the incoming mbuf. We also need to still pay attention
15774          * to nxt_pkt since there may be another packet after this
15775          * one.
15776          */
15777 #ifdef TCP_ACCOUNTING
15778         uint64_t ts_val;
15779         uint64_t rdstc;
15780 #endif
15781         int segsiz;
15782         struct timespec ts;
15783         struct tcp_rack *rack;
15784         struct tcp_ackent *ae;
15785         uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
15786         int cnt, i, did_out, ourfinisacked = 0;
15787         struct tcpopt to_holder, *to = NULL;
15788 #ifdef TCP_ACCOUNTING
15789         int win_up_req = 0;
15790 #endif
15791         int nsegs = 0;
15792         int under_pacing = 0;
15793         int recovery = 0;
15794 #ifdef TCP_ACCOUNTING
15795         sched_pin();
15796 #endif
15797         rack = (struct tcp_rack *)tp->t_fb_ptr;
15798         if (rack->gp_ready &&
15799             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
15800                 under_pacing = 1;
15801
15802         if (rack->r_state != tp->t_state)
15803                 rack_set_state(tp, rack);
15804         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
15805             (tp->t_flags & TF_GPUTINPROG)) {
15806                 /*
15807                  * We have a goodput in progress
15808                  * and we have entered a late state.
15809                  * Do we have enough data in the sb
15810                  * to handle the GPUT request?
15811                  */
15812                 uint32_t bytes;
15813
15814                 bytes = tp->gput_ack - tp->gput_seq;
15815                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
15816                         bytes += tp->gput_seq - tp->snd_una;
15817                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
15818                         /*
15819                          * There are not enough bytes in the socket
15820                          * buffer that have been sent to cover this
15821                          * measurement. Cancel it.
15822                          */
15823                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
15824                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
15825                                                    tp->gput_seq,
15826                                                    0, 0, 18, __LINE__, NULL, 0);
15827                         tp->t_flags &= ~TF_GPUTINPROG;
15828                 }
15829         }
15830         to = &to_holder;
15831         to->to_flags = 0;
15832         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
15833                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
15834         cnt = m->m_len / sizeof(struct tcp_ackent);
15835         counter_u64_add(rack_multi_single_eq, cnt);
15836         high_seq = tp->snd_una;
15837         the_win = tp->snd_wnd;
15838         win_seq = tp->snd_wl1;
15839         win_upd_ack = tp->snd_wl2;
15840         cts = tcp_tv_to_usectick(tv);
15841         ms_cts = tcp_tv_to_mssectick(tv);
15842         rack->r_ctl.rc_rcvtime = cts;
15843         segsiz = ctf_fixed_maxseg(tp);
15844         if ((rack->rc_gp_dyn_mul) &&
15845             (rack->use_fixed_rate == 0) &&
15846             (rack->rc_always_pace)) {
15847                 /* Check in on probertt */
15848                 rack_check_probe_rtt(rack, cts);
15849         }
15850         for (i = 0; i < cnt; i++) {
15851 #ifdef TCP_ACCOUNTING
15852                 ts_val = get_cyclecount();
15853 #endif
15854                 rack_clear_rate_sample(rack);
15855                 ae = ((mtod(m, struct tcp_ackent *)) + i);
15856                 if (ae->flags & TH_FIN)
15857                         rack_log_pacing_delay_calc(rack,
15858                                                    0,
15859                                                    0,
15860                                                    0,
15861                                                    rack_get_gp_est(rack), /* delRate */
15862                                                    rack_get_lt_bw(rack), /* rttProp */
15863                                                    20, __LINE__, NULL, 0);
15864                 /* Setup the window */
15865                 tiwin = ae->win << tp->snd_scale;
15866                 if (tiwin > rack->r_ctl.rc_high_rwnd)
15867                         rack->r_ctl.rc_high_rwnd = tiwin;
15868                 /* figure out the type of ack */
15869                 if (SEQ_LT(ae->ack, high_seq)) {
15870                         /* Case B*/
15871                         ae->ack_val_set = ACK_BEHIND;
15872                 } else if (SEQ_GT(ae->ack, high_seq)) {
15873                         /* Case A */
15874                         ae->ack_val_set = ACK_CUMACK;
15875                 } else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
15876                         /* Case D */
15877                         ae->ack_val_set = ACK_DUPACK;
15878                 } else {
15879                         /* Case C */
15880                         ae->ack_val_set = ACK_RWND;
15881                 }
15882                 if (rack->sack_attack_disable > 0) {
15883                         rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
15884                         rack->r_ctl.ack_during_sd++;
15885                 }
15886                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
15887                 /* Validate timestamp */
15888                 if (ae->flags & HAS_TSTMP) {
15889                         /* Setup for a timestamp */
15890                         to->to_flags = TOF_TS;
15891                         ae->ts_echo -= tp->ts_offset;
15892                         to->to_tsecr = ae->ts_echo;
15893                         to->to_tsval = ae->ts_value;
15894                         /*
15895                          * If echoed timestamp is later than the current time, fall back to
15896                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
15897                          * were used when this connection was established.
15898                          */
15899                         if (TSTMP_GT(ae->ts_echo, ms_cts))
15900                                 to->to_tsecr = 0;
15901                         if (tp->ts_recent &&
15902                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
15903                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
15904 #ifdef TCP_ACCOUNTING
15905                                         rdstc = get_cyclecount();
15906                                         if (rdstc > ts_val) {
15907                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15908                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15909                                                 }
15910                                         }
15911 #endif
15912                                         continue;
15913                                 }
15914                         }
15915                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
15916                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
15917                                 tp->ts_recent_age = tcp_ts_getticks();
15918                                 tp->ts_recent = ae->ts_value;
15919                         }
15920                 } else {
15921                         /* Setup for a no options */
15922                         to->to_flags = 0;
15923                 }
15924                 /* Update the rcv time and perform idle reduction possibly */
15925                 if  (tp->t_idle_reduce &&
15926                      (tp->snd_max == tp->snd_una) &&
15927                      (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
15928                         counter_u64_add(rack_input_idle_reduces, 1);
15929                         rack_cc_after_idle(rack, tp);
15930                 }
15931                 tp->t_rcvtime = ticks;
15932                 /* Now what about ECN of a chain of pure ACKs? */
15933                 if (tcp_ecn_input_segment(tp, ae->flags, 0,
15934                         tcp_packets_this_ack(tp, ae->ack),
15935                         ae->codepoint))
15936                         rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
15937 #ifdef TCP_ACCOUNTING
15938                 /* Count for the specific type of ack in */
15939                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15940                         tp->tcp_cnt_counters[ae->ack_val_set]++;
15941                 }
15942 #endif
15943                 /*
15944                  * Note how we could move up these in the determination
15945                  * above, but we don't so that way the timestamp checks (and ECN)
15946                  * is done first before we do any processing on the ACK.
15947                  * The non-compressed path through the code has this
15948                  * weakness (noted by @jtl) that it actually does some
15949                  * processing before verifying the timestamp information.
15950                  * We don't take that path here which is why we set
15951                  * the ack_val_set first, do the timestamp and ecn
15952                  * processing, and then look at what we have setup.
15953                  */
15954                 if (ae->ack_val_set == ACK_BEHIND) {
15955                         /*
15956                          * Case B flag reordering, if window is not closed
15957                          * or it could be a keep-alive or persists
15958                          */
15959                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
15960                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15961                                 if (rack->r_ctl.rc_reorder_ts == 0)
15962                                         rack->r_ctl.rc_reorder_ts = 1;
15963                         }
15964                 } else if (ae->ack_val_set == ACK_DUPACK) {
15965                         /* Case D */
15966                         rack_strike_dupack(rack);
15967                 } else if (ae->ack_val_set == ACK_RWND) {
15968                         /* Case C */
15969                         if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15970                                 ts.tv_sec = ae->timestamp / 1000000000;
15971                                 ts.tv_nsec = ae->timestamp % 1000000000;
15972                                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15973                                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15974                         } else {
15975                                 rack->r_ctl.act_rcv_time = *tv;
15976                         }
15977                         if (rack->forced_ack) {
15978                                 rack_handle_probe_response(rack, tiwin,
15979                                                            tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
15980                         }
15981 #ifdef TCP_ACCOUNTING
15982                         win_up_req = 1;
15983 #endif
15984                         win_upd_ack = ae->ack;
15985                         win_seq = ae->seq;
15986                         the_win = tiwin;
15987                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15988                 } else {
15989                         /* Case A */
15990                         if (SEQ_GT(ae->ack, tp->snd_max)) {
15991                                 /*
15992                                  * We just send an ack since the incoming
15993                                  * ack is beyond the largest seq we sent.
15994                                  */
15995                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
15996                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
15997                                         if (tp->t_flags && TF_ACKNOW)
15998                                                 rack->r_wanted_output = 1;
15999                                 }
16000                         } else {
16001                                 nsegs++;
16002                                 /* If the window changed setup to update */
16003                                 if (tiwin != tp->snd_wnd) {
16004                                         win_upd_ack = ae->ack;
16005                                         win_seq = ae->seq;
16006                                         the_win = tiwin;
16007                                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
16008                                 }
16009 #ifdef TCP_ACCOUNTING
16010                                 /* Account for the acks */
16011                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16012                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
16013                                 }
16014 #endif
16015                                 high_seq = ae->ack;
16016                                 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp))
16017                                         rack_log_hystart_event(rack, high_seq, 8);
16018                                 /* Setup our act_rcv_time */
16019                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
16020                                         ts.tv_sec = ae->timestamp / 1000000000;
16021                                         ts.tv_nsec = ae->timestamp % 1000000000;
16022                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16023                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16024                                 } else {
16025                                         rack->r_ctl.act_rcv_time = *tv;
16026                                 }
16027                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to,
16028                                                        tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
16029 #ifdef TCP_REQUEST_TRK
16030                                 rack_http_check_for_comp(rack, high_seq);
16031 #endif
16032                                 if (rack->rc_dsack_round_seen) {
16033                                         /* Is the dsack round over? */
16034                                         if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
16035                                                 /* Yes it is */
16036                                                 rack->rc_dsack_round_seen = 0;
16037                                                 rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
16038                                         }
16039                                 }
16040                         }
16041                 }
16042                 /* And lets be sure to commit the rtt measurements for this ack */
16043                 tcp_rack_xmit_timer_commit(rack, tp);
16044 #ifdef TCP_ACCOUNTING
16045                 rdstc = get_cyclecount();
16046                 if (rdstc > ts_val) {
16047                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16048                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
16049                                 if (ae->ack_val_set == ACK_CUMACK)
16050                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
16051                         }
16052                 }
16053 #endif
16054         }
16055 #ifdef TCP_ACCOUNTING
16056         ts_val = get_cyclecount();
16057 #endif
16058         /* Tend to any collapsed window */
16059         if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
16060                 /* The peer collapsed the window */
16061                 rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
16062         } else if (rack->rc_has_collapsed)
16063                 rack_un_collapse_window(rack, __LINE__);
16064         if ((rack->r_collapse_point_valid) &&
16065             (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
16066                 rack->r_collapse_point_valid = 0;
16067         acked_amount = acked = (high_seq - tp->snd_una);
16068         if (acked) {
16069                 /*
16070                  * The draft (v3) calls for us to use SEQ_GEQ, but that
16071                  * causes issues when we are just going app limited. Lets
16072                  * instead use SEQ_GT <or> where its equal but more data
16073                  * is outstanding.
16074                  *
16075                  * Also make sure we are on the last ack of a series. We
16076                  * have to have all the ack's processed in queue to know
16077                  * if there is something left outstanding.
16078                  *
16079                  */
16080                 if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
16081                     (rack->rc_new_rnd_needed == 0) &&
16082                     (nxt_pkt == 0)) {
16083                         rack_log_hystart_event(rack, high_seq, 21);
16084                         rack->r_ctl.current_round++;
16085                         /* Force the next send to setup the next round */
16086                         rack->rc_new_rnd_needed = 1;
16087                         if (CC_ALGO(tp)->newround != NULL) {
16088                                 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
16089                         }
16090                 }
16091                 /*
16092                  * Clear the probe not answered flag
16093                  * since cum-ack moved forward.
16094                  */
16095                 rack->probe_not_answered = 0;
16096                 if (rack->sack_attack_disable == 0)
16097                         rack_do_decay(rack);
16098                 if (acked >= segsiz) {
16099                         /*
16100                          * You only get credit for
16101                          * MSS and greater (and you get extra
16102                          * credit for larger cum-ack moves).
16103                          */
16104                         int ac;
16105
16106                         ac = acked / segsiz;
16107                         rack->r_ctl.ack_count += ac;
16108                         counter_u64_add(rack_ack_total, ac);
16109                 }
16110                 if (rack->r_ctl.ack_count > 0xfff00000) {
16111                         /*
16112                          * reduce the number to keep us under
16113                          * a uint32_t.
16114                          */
16115                         rack->r_ctl.ack_count /= 2;
16116                         rack->r_ctl.sack_count /= 2;
16117                 }
16118                 if (tp->t_flags & TF_NEEDSYN) {
16119                         /*
16120                          * T/TCP: Connection was half-synchronized, and our SYN has
16121                          * been ACK'd (so connection is now fully synchronized).  Go
16122                          * to non-starred state, increment snd_una for ACK of SYN,
16123                          * and check if we can do window scaling.
16124                          */
16125                         tp->t_flags &= ~TF_NEEDSYN;
16126                         tp->snd_una++;
16127                         acked_amount = acked = (high_seq - tp->snd_una);
16128                 }
16129                 if (acked > sbavail(&so->so_snd))
16130                         acked_amount = sbavail(&so->so_snd);
16131 #ifdef TCP_SAD_DETECTION
16132                 /*
16133                  * We only care on a cum-ack move if we are in a sack-disabled
16134                  * state. We have already added in to the ack_count, and we never
16135                  * would disable on a cum-ack move, so we only care to do the
16136                  * detection if it may "undo" it, i.e. we were in disabled already.
16137                  */
16138                 if (rack->sack_attack_disable)
16139                         rack_do_detection(tp, rack, acked_amount, segsiz);
16140 #endif
16141                 if (IN_FASTRECOVERY(tp->t_flags) &&
16142                     (rack->rack_no_prr == 0))
16143                         rack_update_prr(tp, rack, acked_amount, high_seq);
16144                 if (IN_RECOVERY(tp->t_flags)) {
16145                         if (SEQ_LT(high_seq, tp->snd_recover) &&
16146                             (SEQ_LT(high_seq, tp->snd_max))) {
16147                                 tcp_rack_partialack(tp);
16148                         } else {
16149                                 rack_post_recovery(tp, high_seq);
16150                                 recovery = 1;
16151                         }
16152                 }
16153                 /* Handle the rack-log-ack part (sendmap) */
16154                 if ((sbused(&so->so_snd) == 0) &&
16155                     (acked > acked_amount) &&
16156                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
16157                     (tp->t_flags & TF_SENTFIN)) {
16158                         /*
16159                          * We must be sure our fin
16160                          * was sent and acked (we can be
16161                          * in FIN_WAIT_1 without having
16162                          * sent the fin).
16163                          */
16164                         ourfinisacked = 1;
16165                         /*
16166                          * Lets make sure snd_una is updated
16167                          * since most likely acked_amount = 0 (it
16168                          * should be).
16169                          */
16170                         tp->snd_una = high_seq;
16171                 }
16172                 /* Did we make a RTO error? */
16173                 if ((tp->t_flags & TF_PREVVALID) &&
16174                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16175                         tp->t_flags &= ~TF_PREVVALID;
16176                         if (tp->t_rxtshift == 1 &&
16177                             (int)(ticks - tp->t_badrxtwin) < 0)
16178                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16179                 }
16180                 /* Handle the data in the socket buffer */
16181                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16182                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16183                 if (acked_amount > 0) {
16184                         struct mbuf *mfree;
16185
16186                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
16187                         SOCKBUF_LOCK(&so->so_snd);
16188                         mfree = sbcut_locked(&so->so_snd, acked_amount);
16189                         tp->snd_una = high_seq;
16190                         /* Note we want to hold the sb lock through the sendmap adjust */
16191                         rack_adjust_sendmap_head(rack, &so->so_snd);
16192                         /* Wake up the socket if we have room to write more */
16193                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16194                         sowwakeup_locked(so);
16195                         if ((recovery == 1) &&
16196                             (rack->excess_rxt_on) &&
16197                             (rack->r_cwnd_was_clamped == 0)) {
16198                                 do_rack_excess_rxt(tp, rack);
16199                         } else if (rack->r_cwnd_was_clamped)
16200                                 do_rack_check_for_unclamp(tp, rack);
16201                         m_freem(mfree);
16202                 }
16203                 /* update progress */
16204                 tp->t_acktime = ticks;
16205                 rack_log_progress_event(rack, tp, tp->t_acktime,
16206                                         PROGRESS_UPDATE, __LINE__);
16207                 /* Clear out shifts and such */
16208                 tp->t_rxtshift = 0;
16209                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16210                                    rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16211                 rack->rc_tlp_in_progress = 0;
16212                 rack->r_ctl.rc_tlp_cnt_out = 0;
16213                 /* Send recover and snd_nxt must be dragged along */
16214                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
16215                         tp->snd_recover = tp->snd_una;
16216                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
16217                         tp->snd_nxt = tp->snd_una;
16218                 /*
16219                  * If the RXT timer is running we want to
16220                  * stop it, so we can restart a TLP (or new RXT).
16221                  */
16222                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16223                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16224                 tp->snd_wl2 = high_seq;
16225                 tp->t_dupacks = 0;
16226                 if (under_pacing &&
16227                     (rack->use_fixed_rate == 0) &&
16228                     (rack->in_probe_rtt == 0) &&
16229                     rack->rc_gp_dyn_mul &&
16230                     rack->rc_always_pace) {
16231                         /* Check if we are dragging bottom */
16232                         rack_check_bottom_drag(tp, rack, so);
16233                 }
16234                 if (tp->snd_una == tp->snd_max) {
16235                         tp->t_flags &= ~TF_PREVVALID;
16236                         rack->r_ctl.retran_during_recovery = 0;
16237                         rack->rc_suspicious = 0;
16238                         rack->r_ctl.dsack_byte_cnt = 0;
16239                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16240                         if (rack->r_ctl.rc_went_idle_time == 0)
16241                                 rack->r_ctl.rc_went_idle_time = 1;
16242                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16243                         if (sbavail(&tptosocket(tp)->so_snd) == 0)
16244                                 tp->t_acktime = 0;
16245                         /* Set so we might enter persists... */
16246                         rack->r_wanted_output = 1;
16247                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16248                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16249                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16250                             (sbavail(&so->so_snd) == 0) &&
16251                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16252                                 /*
16253                                  * The socket was gone and the
16254                                  * peer sent data (not now in the past), time to
16255                                  * reset him.
16256                                  */
16257                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16258                                 /* tcp_close will kill the inp pre-log the Reset */
16259                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16260 #ifdef TCP_ACCOUNTING
16261                                 rdstc = get_cyclecount();
16262                                 if (rdstc > ts_val) {
16263                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16264                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16265                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16266                                         }
16267                                 }
16268 #endif
16269                                 m_freem(m);
16270                                 tp = tcp_close(tp);
16271                                 if (tp == NULL) {
16272 #ifdef TCP_ACCOUNTING
16273                                         sched_unpin();
16274 #endif
16275                                         return (1);
16276                                 }
16277                                 /*
16278                                  * We would normally do drop-with-reset which would
16279                                  * send back a reset. We can't since we don't have
16280                                  * all the needed bits. Instead lets arrange for
16281                                  * a call to tcp_output(). That way since we
16282                                  * are in the closed state we will generate a reset.
16283                                  *
16284                                  * Note if tcp_accounting is on we don't unpin since
16285                                  * we do that after the goto label.
16286                                  */
16287                                 goto send_out_a_rst;
16288                         }
16289                         if ((sbused(&so->so_snd) == 0) &&
16290                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
16291                             (tp->t_flags & TF_SENTFIN)) {
16292                                 /*
16293                                  * If we can't receive any more data, then closing user can
16294                                  * proceed. Starting the timer is contrary to the
16295                                  * specification, but if we don't get a FIN we'll hang
16296                                  * forever.
16297                                  *
16298                                  */
16299                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16300                                         soisdisconnected(so);
16301                                         tcp_timer_activate(tp, TT_2MSL,
16302                                                            (tcp_fast_finwait2_recycle ?
16303                                                             tcp_finwait2_timeout :
16304                                                             TP_MAXIDLE(tp)));
16305                                 }
16306                                 if (ourfinisacked == 0) {
16307                                         /*
16308                                          * We don't change to fin-wait-2 if we have our fin acked
16309                                          * which means we are probably in TCPS_CLOSING.
16310                                          */
16311                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
16312                                 }
16313                         }
16314                 }
16315                 /* Wake up the socket if we have room to write more */
16316                 if (sbavail(&so->so_snd)) {
16317                         rack->r_wanted_output = 1;
16318                         if (ctf_progress_timeout_check(tp, true)) {
16319                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16320                                                         tp, tick, PROGRESS_DROP, __LINE__);
16321                                 /*
16322                                  * We cheat here and don't send a RST, we should send one
16323                                  * when the pacer drops the connection.
16324                                  */
16325 #ifdef TCP_ACCOUNTING
16326                                 rdstc = get_cyclecount();
16327                                 if (rdstc > ts_val) {
16328                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16329                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16330                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16331                                         }
16332                                 }
16333                                 sched_unpin();
16334 #endif
16335                                 (void)tcp_drop(tp, ETIMEDOUT);
16336                                 m_freem(m);
16337                                 return (1);
16338                         }
16339                 }
16340                 if (ourfinisacked) {
16341                         switch(tp->t_state) {
16342                         case TCPS_CLOSING:
16343 #ifdef TCP_ACCOUNTING
16344                                 rdstc = get_cyclecount();
16345                                 if (rdstc > ts_val) {
16346                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16347                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16348                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16349                                         }
16350                                 }
16351                                 sched_unpin();
16352 #endif
16353                                 tcp_twstart(tp);
16354                                 m_freem(m);
16355                                 return (1);
16356                                 break;
16357                         case TCPS_LAST_ACK:
16358 #ifdef TCP_ACCOUNTING
16359                                 rdstc = get_cyclecount();
16360                                 if (rdstc > ts_val) {
16361                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16362                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16363                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16364                                         }
16365                                 }
16366                                 sched_unpin();
16367 #endif
16368                                 tp = tcp_close(tp);
16369                                 ctf_do_drop(m, tp);
16370                                 return (1);
16371                                 break;
16372                         case TCPS_FIN_WAIT_1:
16373 #ifdef TCP_ACCOUNTING
16374                                 rdstc = get_cyclecount();
16375                                 if (rdstc > ts_val) {
16376                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16377                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16378                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16379                                         }
16380                                 }
16381 #endif
16382                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16383                                         soisdisconnected(so);
16384                                         tcp_timer_activate(tp, TT_2MSL,
16385                                                            (tcp_fast_finwait2_recycle ?
16386                                                             tcp_finwait2_timeout :
16387                                                             TP_MAXIDLE(tp)));
16388                                 }
16389                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
16390                                 break;
16391                         default:
16392                                 break;
16393                         }
16394                 }
16395                 if (rack->r_fast_output) {
16396                         /*
16397                          * We re doing fast output.. can we expand that?
16398                          */
16399                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
16400                 }
16401 #ifdef TCP_ACCOUNTING
16402                 rdstc = get_cyclecount();
16403                 if (rdstc > ts_val) {
16404                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16405                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16406                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16407                         }
16408                 }
16409
16410         } else if (win_up_req) {
16411                 rdstc = get_cyclecount();
16412                 if (rdstc > ts_val) {
16413                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16414                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
16415                         }
16416                 }
16417 #endif
16418         }
16419         /* Now is there a next packet, if so we are done */
16420         m_freem(m);
16421         did_out = 0;
16422         if (nxt_pkt) {
16423 #ifdef TCP_ACCOUNTING
16424                 sched_unpin();
16425 #endif
16426                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
16427                 return (0);
16428         }
16429         rack_handle_might_revert(tp, rack);
16430         ctf_calc_rwin(so, tp);
16431         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
16432         send_out_a_rst:
16433                 if (tcp_output(tp) < 0) {
16434 #ifdef TCP_ACCOUNTING
16435                         sched_unpin();
16436 #endif
16437                         return (1);
16438                 }
16439                 did_out = 1;
16440         }
16441         if (rack->rc_inp->inp_hpts_calls)
16442                 rack->rc_inp->inp_hpts_calls = 0;
16443         rack_free_trim(rack);
16444 #ifdef TCP_ACCOUNTING
16445         sched_unpin();
16446 #endif
16447         rack_timer_audit(tp, rack, &so->so_snd);
16448         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
16449         return (0);
16450 }
16451
16452 #define TCP_LRO_TS_OPTION \
16453     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
16454           (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
16455
16456 static int
16457 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
16458     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
16459     struct timeval *tv)
16460 {
16461         struct inpcb *inp = tptoinpcb(tp);
16462         struct socket *so = tptosocket(tp);
16463 #ifdef TCP_ACCOUNTING
16464         uint64_t ts_val;
16465 #endif
16466         int32_t thflags, retval, did_out = 0;
16467         int32_t way_out = 0;
16468         /*
16469          * cts - is the current time from tv (caller gets ts) in microseconds.
16470          * ms_cts - is the current time from tv in milliseconds.
16471          * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
16472          */
16473         uint32_t cts, us_cts, ms_cts;
16474         uint32_t tiwin, high_seq;
16475         struct timespec ts;
16476         struct tcpopt to;
16477         struct tcp_rack *rack;
16478         struct rack_sendmap *rsm;
16479         int32_t prev_state = 0;
16480         int no_output = 0;
16481         int slot_remaining = 0;
16482 #ifdef TCP_ACCOUNTING
16483         int ack_val_set = 0xf;
16484 #endif
16485         int nsegs;
16486
16487         NET_EPOCH_ASSERT();
16488         INP_WLOCK_ASSERT(inp);
16489
16490         /*
16491          * tv passed from common code is from either M_TSTMP_LRO or
16492          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
16493          */
16494         rack = (struct tcp_rack *)tp->t_fb_ptr;
16495         if (rack->rack_deferred_inited == 0) {
16496                 /*
16497                  * If we are the connecting socket we will
16498                  * hit rack_init() when no sequence numbers
16499                  * are setup. This makes it so we must defer
16500                  * some initialization. Call that now.
16501                  */
16502                 rack_deferred_init(tp, rack);
16503         }
16504         /*
16505          * Check to see if we need to skip any output plans. This
16506          * can happen in the non-LRO path where we are pacing and
16507          * must process the ack coming in but need to defer sending
16508          * anything becase a pacing timer is running.
16509          */
16510         us_cts = tcp_tv_to_usectick(tv);
16511         if (m->m_flags & M_ACKCMP) {
16512                 /*
16513                  * All compressed ack's are ack's by definition so
16514                  * remove any ack required flag and then do the processing.
16515                  */
16516                 rack->rc_ack_required = 0;
16517                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
16518         }
16519         thflags = tcp_get_flags(th);
16520         if ((rack->rc_always_pace == 1) &&
16521             (rack->rc_ack_can_sendout_data == 0) &&
16522             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16523             (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
16524                 /*
16525                  * Ok conditions are right for queuing the packets
16526                  * but we do have to check the flags in the inp, it
16527                  * could be, if a sack is present, we want to be awoken and
16528                  * so should process the packets.
16529                  */
16530                 slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
16531                 if (rack->rc_inp->inp_flags2 & INP_DONT_SACK_QUEUE) {
16532                         no_output = 1;
16533                 } else {
16534                         /*
16535                          * If there is no options, or just a
16536                          * timestamp option, we will want to queue
16537                          * the packets. This is the same that LRO does
16538                          * and will need to change with accurate ECN.
16539                          */
16540                         uint32_t *ts_ptr;
16541                         int optlen;
16542
16543                         optlen = (th->th_off << 2) - sizeof(struct tcphdr);
16544                         ts_ptr = (uint32_t *)(th + 1);
16545                         if ((optlen == 0) ||
16546                             ((optlen == TCPOLEN_TSTAMP_APPA) &&
16547                              (*ts_ptr == TCP_LRO_TS_OPTION)))
16548                                 no_output = 1;
16549                 }
16550                 if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
16551                         /*
16552                          * It is unrealistic to think we can pace in less than
16553                          * the minimum granularity of the pacer (def:250usec). So
16554                          * if we have less than that time remaining we should go
16555                          * ahead and allow output to be "early". We will attempt to
16556                          * make up for it in any pacing time we try to apply on
16557                          * the outbound packet.
16558                          */
16559                         no_output = 0;
16560                 }
16561         }
16562         /*
16563          * If there is a RST or FIN lets dump out the bw
16564          * with a FIN the connection may go on but we
16565          * may not.
16566          */
16567         if ((thflags & TH_FIN) || (thflags & TH_RST))
16568                 rack_log_pacing_delay_calc(rack,
16569                                            rack->r_ctl.gp_bw,
16570                                            0,
16571                                            0,
16572                                            rack_get_gp_est(rack), /* delRate */
16573                                            rack_get_lt_bw(rack), /* rttProp */
16574                                            20, __LINE__, NULL, 0);
16575         if (m->m_flags & M_ACKCMP) {
16576                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
16577         }
16578         cts = tcp_tv_to_usectick(tv);
16579         ms_cts =  tcp_tv_to_mssectick(tv);
16580         nsegs = m->m_pkthdr.lro_nsegs;
16581         counter_u64_add(rack_proc_non_comp_ack, 1);
16582 #ifdef TCP_ACCOUNTING
16583         sched_pin();
16584         if (thflags & TH_ACK)
16585                 ts_val = get_cyclecount();
16586 #endif
16587         if ((m->m_flags & M_TSTMP) ||
16588             (m->m_flags & M_TSTMP_LRO)) {
16589                 mbuf_tstmp2timespec(m, &ts);
16590                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16591                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16592         } else
16593                 rack->r_ctl.act_rcv_time = *tv;
16594         kern_prefetch(rack, &prev_state);
16595         prev_state = 0;
16596         /*
16597          * Unscale the window into a 32-bit value. For the SYN_SENT state
16598          * the scale is zero.
16599          */
16600         tiwin = th->th_win << tp->snd_scale;
16601 #ifdef TCP_ACCOUNTING
16602         if (thflags & TH_ACK) {
16603                 /*
16604                  * We have a tradeoff here. We can either do what we are
16605                  * doing i.e. pinning to this CPU and then doing the accounting
16606                  * <or> we could do a critical enter, setup the rdtsc and cpu
16607                  * as in below, and then validate we are on the same CPU on
16608                  * exit. I have choosen to not do the critical enter since
16609                  * that often will gain you a context switch, and instead lock
16610                  * us (line above this if) to the same CPU with sched_pin(). This
16611                  * means we may be context switched out for a higher priority
16612                  * interupt but we won't be moved to another CPU.
16613                  *
16614                  * If this occurs (which it won't very often since we most likely
16615                  * are running this code in interupt context and only a higher
16616                  * priority will bump us ... clock?) we will falsely add in
16617                  * to the time the interupt processing time plus the ack processing
16618                  * time. This is ok since its a rare event.
16619                  */
16620                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
16621                                                     ctf_fixed_maxseg(tp));
16622         }
16623 #endif
16624         /*
16625          * Parse options on any incoming segment.
16626          */
16627         memset(&to, 0, sizeof(to));
16628         tcp_dooptions(&to, (u_char *)(th + 1),
16629             (th->th_off << 2) - sizeof(struct tcphdr),
16630             (thflags & TH_SYN) ? TO_SYN : 0);
16631         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
16632             __func__));
16633         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
16634             __func__));
16635
16636         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16637             (tp->t_flags & TF_GPUTINPROG)) {
16638                 /*
16639                  * We have a goodput in progress
16640                  * and we have entered a late state.
16641                  * Do we have enough data in the sb
16642                  * to handle the GPUT request?
16643                  */
16644                 uint32_t bytes;
16645
16646                 bytes = tp->gput_ack - tp->gput_seq;
16647                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
16648                         bytes += tp->gput_seq - tp->snd_una;
16649                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16650                         /*
16651                          * There are not enough bytes in the socket
16652                          * buffer that have been sent to cover this
16653                          * measurement. Cancel it.
16654                          */
16655                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16656                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
16657                                                    tp->gput_seq,
16658                                                    0, 0, 18, __LINE__, NULL, 0);
16659                         tp->t_flags &= ~TF_GPUTINPROG;
16660                 }
16661         }
16662         high_seq = th->th_ack;
16663         if (tcp_bblogging_on(rack->rc_tp)) {
16664                 union tcp_log_stackspecific log;
16665                 struct timeval ltv;
16666 #ifdef TCP_REQUEST_TRK
16667                 struct http_sendfile_track *http_req;
16668
16669                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
16670                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
16671                 } else {
16672                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
16673                 }
16674 #endif
16675                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16676                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16677                 if (rack->rack_no_prr == 0)
16678                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16679                 else
16680                         log.u_bbr.flex1 = 0;
16681                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16682                 log.u_bbr.use_lt_bw <<= 1;
16683                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
16684                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16685                 log.u_bbr.bbr_state = rack->rc_free_cnt;
16686                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16687                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
16688                 log.u_bbr.flex3 = m->m_flags;
16689                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16690                 log.u_bbr.lost = thflags;
16691                 log.u_bbr.pacing_gain = 0x1;
16692 #ifdef TCP_ACCOUNTING
16693                 log.u_bbr.cwnd_gain = ack_val_set;
16694 #endif
16695                 log.u_bbr.flex7 = 2;
16696                 if (m->m_flags & M_TSTMP) {
16697                         /* Record the hardware timestamp if present */
16698                         mbuf_tstmp2timespec(m, &ts);
16699                         ltv.tv_sec = ts.tv_sec;
16700                         ltv.tv_usec = ts.tv_nsec / 1000;
16701                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
16702                 } else if (m->m_flags & M_TSTMP_LRO) {
16703                         /* Record the LRO the arrival timestamp */
16704                         mbuf_tstmp2timespec(m, &ts);
16705                         ltv.tv_sec = ts.tv_sec;
16706                         ltv.tv_usec = ts.tv_nsec / 1000;
16707                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
16708                 }
16709                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
16710                 /* Log the rcv time */
16711                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
16712 #ifdef TCP_REQUEST_TRK
16713                 log.u_bbr.applimited = tp->t_http_closed;
16714                 log.u_bbr.applimited <<= 8;
16715                 log.u_bbr.applimited |= tp->t_http_open;
16716                 log.u_bbr.applimited <<= 8;
16717                 log.u_bbr.applimited |= tp->t_http_req;
16718                 if (http_req) {
16719                         /* Copy out any client req info */
16720                         /* seconds */
16721                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
16722                         /* useconds */
16723                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
16724                         log.u_bbr.rttProp = http_req->timestamp;
16725                         log.u_bbr.cur_del_rate = http_req->start;
16726                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
16727                                 log.u_bbr.flex8 |= 1;
16728                         } else {
16729                                 log.u_bbr.flex8 |= 2;
16730                                 log.u_bbr.bw_inuse = http_req->end;
16731                         }
16732                         log.u_bbr.flex6 = http_req->start_seq;
16733                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
16734                                 log.u_bbr.flex8 |= 4;
16735                                 log.u_bbr.epoch = http_req->end_seq;
16736                         }
16737                 }
16738 #endif
16739                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
16740                     tlen, &log, true, &ltv);
16741         }
16742         /* Remove ack required flag if set, we have one  */
16743         if (thflags & TH_ACK)
16744                 rack->rc_ack_required = 0;
16745         if (rack->sack_attack_disable > 0) {
16746                 rack->r_ctl.ack_during_sd++;
16747                 rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16748         }
16749         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
16750                 way_out = 4;
16751                 retval = 0;
16752                 m_freem(m);
16753                 goto done_with_input;
16754         }
16755         /*
16756          * If a segment with the ACK-bit set arrives in the SYN-SENT state
16757          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
16758          */
16759         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
16760             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
16761                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
16762                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
16763 #ifdef TCP_ACCOUNTING
16764                 sched_unpin();
16765 #endif
16766                 return (1);
16767         }
16768         /*
16769          * If timestamps were negotiated during SYN/ACK and a
16770          * segment without a timestamp is received, silently drop
16771          * the segment, unless it is a RST segment or missing timestamps are
16772          * tolerated.
16773          * See section 3.2 of RFC 7323.
16774          */
16775         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
16776             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
16777                 way_out = 5;
16778                 retval = 0;
16779                 m_freem(m);
16780                 goto done_with_input;
16781         }
16782
16783         /*
16784          * Segment received on connection. Reset idle time and keep-alive
16785          * timer. XXX: This should be done after segment validation to
16786          * ignore broken/spoofed segs.
16787          */
16788         if  (tp->t_idle_reduce &&
16789              (tp->snd_max == tp->snd_una) &&
16790              (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16791                 counter_u64_add(rack_input_idle_reduces, 1);
16792                 rack_cc_after_idle(rack, tp);
16793         }
16794         tp->t_rcvtime = ticks;
16795 #ifdef STATS
16796         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
16797 #endif
16798         if (tiwin > rack->r_ctl.rc_high_rwnd)
16799                 rack->r_ctl.rc_high_rwnd = tiwin;
16800         /*
16801          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
16802          * this to occur after we've validated the segment.
16803          */
16804         if (tcp_ecn_input_segment(tp, thflags, tlen,
16805             tcp_packets_this_ack(tp, th->th_ack),
16806             iptos))
16807                 rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
16808
16809         /*
16810          * If echoed timestamp is later than the current time, fall back to
16811          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16812          * were used when this connection was established.
16813          */
16814         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
16815                 to.to_tsecr -= tp->ts_offset;
16816                 if (TSTMP_GT(to.to_tsecr, ms_cts))
16817                         to.to_tsecr = 0;
16818         }
16819
16820         /*
16821          * If its the first time in we need to take care of options and
16822          * verify we can do SACK for rack!
16823          */
16824         if (rack->r_state == 0) {
16825                 /* Should be init'd by rack_init() */
16826                 KASSERT(rack->rc_inp != NULL,
16827                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
16828                 if (rack->rc_inp == NULL) {
16829                         rack->rc_inp = inp;
16830                 }
16831
16832                 /*
16833                  * Process options only when we get SYN/ACK back. The SYN
16834                  * case for incoming connections is handled in tcp_syncache.
16835                  * According to RFC1323 the window field in a SYN (i.e., a
16836                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
16837                  * this is traditional behavior, may need to be cleaned up.
16838                  */
16839                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
16840                         /* Handle parallel SYN for ECN */
16841                         tcp_ecn_input_parallel_syn(tp, thflags, iptos);
16842                         if ((to.to_flags & TOF_SCALE) &&
16843                             (tp->t_flags & TF_REQ_SCALE)) {
16844                                 tp->t_flags |= TF_RCVD_SCALE;
16845                                 tp->snd_scale = to.to_wscale;
16846                         } else
16847                                 tp->t_flags &= ~TF_REQ_SCALE;
16848                         /*
16849                          * Initial send window.  It will be updated with the
16850                          * next incoming segment to the scaled value.
16851                          */
16852                         tp->snd_wnd = th->th_win;
16853                         rack_validate_fo_sendwin_up(tp, rack);
16854                         if ((to.to_flags & TOF_TS) &&
16855                             (tp->t_flags & TF_REQ_TSTMP)) {
16856                                 tp->t_flags |= TF_RCVD_TSTMP;
16857                                 tp->ts_recent = to.to_tsval;
16858                                 tp->ts_recent_age = cts;
16859                         } else
16860                                 tp->t_flags &= ~TF_REQ_TSTMP;
16861                         if (to.to_flags & TOF_MSS) {
16862                                 tcp_mss(tp, to.to_mss);
16863                         }
16864                         if ((tp->t_flags & TF_SACK_PERMIT) &&
16865                             (to.to_flags & TOF_SACKPERM) == 0)
16866                                 tp->t_flags &= ~TF_SACK_PERMIT;
16867                         if (IS_FASTOPEN(tp->t_flags)) {
16868                                 if (to.to_flags & TOF_FASTOPEN) {
16869                                         uint16_t mss;
16870
16871                                         if (to.to_flags & TOF_MSS)
16872                                                 mss = to.to_mss;
16873                                         else
16874                                                 if ((inp->inp_vflag & INP_IPV6) != 0)
16875                                                         mss = TCP6_MSS;
16876                                                 else
16877                                                         mss = TCP_MSS;
16878                                         tcp_fastopen_update_cache(tp, mss,
16879                                             to.to_tfo_len, to.to_tfo_cookie);
16880                                 } else
16881                                         tcp_fastopen_disable_path(tp);
16882                         }
16883                 }
16884                 /*
16885                  * At this point we are at the initial call. Here we decide
16886                  * if we are doing RACK or not. We do this by seeing if
16887                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
16888                  * The code now does do dup-ack counting so if you don't
16889                  * switch back you won't get rack & TLP, but you will still
16890                  * get this stack.
16891                  */
16892
16893                 if ((rack_sack_not_required == 0) &&
16894                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
16895                         tcp_switch_back_to_default(tp);
16896                         (*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
16897                             tlen, iptos);
16898 #ifdef TCP_ACCOUNTING
16899                         sched_unpin();
16900 #endif
16901                         return (1);
16902                 }
16903                 tcp_set_hpts(inp);
16904                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
16905         }
16906         if (thflags & TH_FIN)
16907                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
16908         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
16909         if ((rack->rc_gp_dyn_mul) &&
16910             (rack->use_fixed_rate == 0) &&
16911             (rack->rc_always_pace)) {
16912                 /* Check in on probertt */
16913                 rack_check_probe_rtt(rack, us_cts);
16914         }
16915         rack_clear_rate_sample(rack);
16916         if ((rack->forced_ack) &&
16917             ((tcp_get_flags(th) & TH_RST) == 0)) {
16918                 rack_handle_probe_response(rack, tiwin, us_cts);
16919         }
16920         /*
16921          * This is the one exception case where we set the rack state
16922          * always. All other times (timers etc) we must have a rack-state
16923          * set (so we assure we have done the checks above for SACK).
16924          */
16925         rack->r_ctl.rc_rcvtime = cts;
16926         if (rack->r_state != tp->t_state)
16927                 rack_set_state(tp, rack);
16928         if (SEQ_GT(th->th_ack, tp->snd_una) &&
16929             (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
16930                 kern_prefetch(rsm, &prev_state);
16931         prev_state = rack->r_state;
16932         if ((thflags & TH_RST) &&
16933             ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
16934               SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
16935              (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
16936                 /* The connection will be killed by a reset check the tracepoint */
16937                 tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
16938         }
16939         retval = (*rack->r_substate) (m, th, so,
16940             tp, &to, drop_hdrlen,
16941             tlen, tiwin, thflags, nxt_pkt, iptos);
16942         if (retval == 0) {
16943                 /*
16944                  * If retval is 1 the tcb is unlocked and most likely the tp
16945                  * is gone.
16946                  */
16947                 INP_WLOCK_ASSERT(inp);
16948                 if ((rack->rc_gp_dyn_mul) &&
16949                     (rack->rc_always_pace) &&
16950                     (rack->use_fixed_rate == 0) &&
16951                     rack->in_probe_rtt &&
16952                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
16953                         /*
16954                          * If we are going for target, lets recheck before
16955                          * we output.
16956                          */
16957                         rack_check_probe_rtt(rack, us_cts);
16958                 }
16959                 if (rack->set_pacing_done_a_iw == 0) {
16960                         /* How much has been acked? */
16961                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
16962                                 /* We have enough to set in the pacing segment size */
16963                                 rack->set_pacing_done_a_iw = 1;
16964                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
16965                         }
16966                 }
16967                 tcp_rack_xmit_timer_commit(rack, tp);
16968 #ifdef TCP_ACCOUNTING
16969                 /*
16970                  * If we set the ack_val_se to what ack processing we are doing
16971                  * we also want to track how many cycles we burned. Note
16972                  * the bits after tcp_output we let be "free". This is because
16973                  * we are also tracking the tcp_output times as well. Note the
16974                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
16975                  * 0xf cannot be returned and is what we initialize it too to
16976                  * indicate we are not doing the tabulations.
16977                  */
16978                 if (ack_val_set != 0xf) {
16979                         uint64_t crtsc;
16980
16981                         crtsc = get_cyclecount();
16982                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16983                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
16984                         }
16985                 }
16986 #endif
16987                 if ((nxt_pkt == 0) && (no_output == 0)) {
16988                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
16989 do_output_now:
16990                                 if (tcp_output(tp) < 0) {
16991 #ifdef TCP_ACCOUNTING
16992                                         sched_unpin();
16993 #endif
16994                                         return (1);
16995                                 }
16996                                 did_out = 1;
16997                         }
16998                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16999                         rack_free_trim(rack);
17000                 } else if ((no_output == 1) &&
17001                            (nxt_pkt == 0)  &&
17002                            (tcp_in_hpts(rack->rc_inp) == 0)) {
17003                         /*
17004                          * We are not in hpts and we had a pacing timer up. Use
17005                          * the remaining time (slot_remaining) to restart the timer.
17006                          */
17007                         KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
17008                         rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
17009                         rack_free_trim(rack);
17010                 }
17011                 /* Clear the flag, it may have been cleared by output but we may not have  */
17012                 if ((nxt_pkt == 0) && (inp->inp_hpts_calls))
17013                         inp->inp_hpts_calls = 0;
17014                 /* Update any rounds needed */
17015                 if (rack_verbose_logging &&  tcp_bblogging_on(rack->rc_tp))
17016                         rack_log_hystart_event(rack, high_seq, 8);
17017                 /*
17018                  * The draft (v3) calls for us to use SEQ_GEQ, but that
17019                  * causes issues when we are just going app limited. Lets
17020                  * instead use SEQ_GT <or> where its equal but more data
17021                  * is outstanding.
17022                  *
17023                  * Also make sure we are on the last ack of a series. We
17024                  * have to have all the ack's processed in queue to know
17025                  * if there is something left outstanding.
17026                  */
17027                 if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
17028                     (rack->rc_new_rnd_needed == 0) &&
17029                     (nxt_pkt == 0)) {
17030                         rack_log_hystart_event(rack, tp->snd_una, 21);
17031                         rack->r_ctl.current_round++;
17032                         /* Force the next send to setup the next round */
17033                         rack->rc_new_rnd_needed = 1;
17034                         if (CC_ALGO(tp)->newround != NULL) {
17035                                 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
17036                         }
17037                 }
17038                 if ((nxt_pkt == 0) &&
17039                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
17040                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
17041                      (tp->t_flags & TF_DELACK) ||
17042                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
17043                       (tp->t_state <= TCPS_CLOSING)))) {
17044                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
17045                         if ((tp->snd_max == tp->snd_una) &&
17046                             ((tp->t_flags & TF_DELACK) == 0) &&
17047                             (tcp_in_hpts(rack->rc_inp)) &&
17048                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
17049                                 /* keep alive not needed if we are hptsi output yet */
17050                                 ;
17051                         } else {
17052                                 int late = 0;
17053                                 if (tcp_in_hpts(inp)) {
17054                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
17055                                                 us_cts = tcp_get_usecs(NULL);
17056                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
17057                                                         rack->r_early = 1;
17058                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
17059                                                 } else
17060                                                         late = 1;
17061                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
17062                                         }
17063                                         tcp_hpts_remove(inp);
17064                                 }
17065                                 if (late && (did_out == 0)) {
17066                                         /*
17067                                          * We are late in the sending
17068                                          * and we did not call the output
17069                                          * (this probably should not happen).
17070                                          */
17071                                         goto do_output_now;
17072                                 }
17073                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
17074                         }
17075                         way_out = 1;
17076                 } else if (nxt_pkt == 0) {
17077                         /* Do we have the correct timer running? */
17078                         rack_timer_audit(tp, rack, &so->so_snd);
17079                         way_out = 2;
17080                 }
17081         done_with_input:
17082                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
17083                 if (did_out)
17084                         rack->r_wanted_output = 0;
17085         }
17086 #ifdef TCP_ACCOUNTING
17087         sched_unpin();
17088 #endif
17089         return (retval);
17090 }
17091
17092 static void
17093 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17094     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17095 {
17096         struct timeval tv;
17097
17098         /* First lets see if we have old packets */
17099         if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17100                 if (ctf_do_queued_segments(tp, 1)) {
17101                         m_freem(m);
17102                         return;
17103                 }
17104         }
17105         if (m->m_flags & M_TSTMP_LRO) {
17106                 mbuf_tstmp2timeval(m, &tv);
17107         } else {
17108                 /* Should not be should we kassert instead? */
17109                 tcp_get_usecs(&tv);
17110         }
17111         if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17112             &tv) == 0) {
17113                 INP_WUNLOCK(tptoinpcb(tp));
17114         }
17115 }
17116
17117 struct rack_sendmap *
17118 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17119 {
17120         struct rack_sendmap *rsm = NULL;
17121         int32_t idx;
17122         uint32_t srtt = 0, thresh = 0, ts_low = 0;
17123         int no_sack = 0;
17124
17125         /* Return the next guy to be re-transmitted */
17126         if (tqhash_empty(rack->r_ctl.tqh)) {
17127                 return (NULL);
17128         }
17129         if (tp->t_flags & TF_SENTFIN) {
17130                 /* retran the end FIN? */
17131                 return (NULL);
17132         }
17133         /* ok lets look at this one */
17134         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17135         if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17136                 return (rsm);
17137         }
17138         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17139                 goto check_it;
17140         }
17141         rsm = rack_find_lowest_rsm(rack);
17142         if (rsm == NULL) {
17143                 return (NULL);
17144         }
17145 check_it:
17146         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) ||
17147             (rack->sack_attack_disable > 0)) {
17148                 no_sack = 1;
17149         }
17150         if ((no_sack > 0) &&
17151             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17152                 /*
17153                  * No sack so we automatically do the 3 strikes and
17154                  * retransmit (no rack timer would be started).
17155                  */
17156                 return (rsm);
17157         }
17158         if (rsm->r_flags & RACK_ACKED) {
17159                 return (NULL);
17160         }
17161         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17162             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17163                 /* Its not yet ready */
17164                 return (NULL);
17165         }
17166         srtt = rack_grab_rtt(tp, rack);
17167         idx = rsm->r_rtr_cnt - 1;
17168         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17169         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
17170         if ((tsused == ts_low) ||
17171             (TSTMP_LT(tsused, ts_low))) {
17172                 /* No time since sending */
17173                 return (NULL);
17174         }
17175         if ((tsused - ts_low) < thresh) {
17176                 /* It has not been long enough yet */
17177                 return (NULL);
17178         }
17179         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17180             ((rsm->r_flags & RACK_SACK_PASSED) &&
17181              (rack->sack_attack_disable == 0))) {
17182                 /*
17183                  * We have passed the dup-ack threshold <or>
17184                  * a SACK has indicated this is missing.
17185                  * Note that if you are a declared attacker
17186                  * it is only the dup-ack threshold that
17187                  * will cause retransmits.
17188                  */
17189                 /* log retransmit reason */
17190                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17191                 rack->r_fast_output = 0;
17192                 return (rsm);
17193         }
17194         return (NULL);
17195 }
17196
17197 static void
17198 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
17199                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17200                            int line, struct rack_sendmap *rsm, uint8_t quality)
17201 {
17202         if (tcp_bblogging_on(rack->rc_tp)) {
17203                 union tcp_log_stackspecific log;
17204                 struct timeval tv;
17205
17206                 memset(&log, 0, sizeof(log));
17207                 log.u_bbr.flex1 = slot;
17208                 log.u_bbr.flex2 = len;
17209                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17210                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17211                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17212                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17213                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17214                 log.u_bbr.use_lt_bw <<= 1;
17215                 log.u_bbr.use_lt_bw |= rack->r_late;
17216                 log.u_bbr.use_lt_bw <<= 1;
17217                 log.u_bbr.use_lt_bw |= rack->r_early;
17218                 log.u_bbr.use_lt_bw <<= 1;
17219                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17220                 log.u_bbr.use_lt_bw <<= 1;
17221                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17222                 log.u_bbr.use_lt_bw <<= 1;
17223                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17224                 log.u_bbr.use_lt_bw <<= 1;
17225                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17226                 log.u_bbr.use_lt_bw <<= 1;
17227                 log.u_bbr.use_lt_bw |= rack->gp_ready;
17228                 log.u_bbr.pkt_epoch = line;
17229                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17230                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17231                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17232                 log.u_bbr.bw_inuse = bw_est;
17233                 log.u_bbr.delRate = bw;
17234                 if (rack->r_ctl.gp_bw == 0)
17235                         log.u_bbr.cur_del_rate = 0;
17236                 else
17237                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
17238                 log.u_bbr.rttProp = len_time;
17239                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17240                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17241                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17242                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17243                         /* We are in slow start */
17244                         log.u_bbr.flex7 = 1;
17245                 } else {
17246                         /* we are on congestion avoidance */
17247                         log.u_bbr.flex7 = 0;
17248                 }
17249                 log.u_bbr.flex8 = method;
17250                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17251                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17252                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17253                 log.u_bbr.cwnd_gain <<= 1;
17254                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17255                 log.u_bbr.cwnd_gain <<= 1;
17256                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17257                 log.u_bbr.bbr_substate = quality;
17258                 log.u_bbr.bbr_state = rack->dgp_on;
17259                 log.u_bbr.bbr_state <<= 1;
17260                 log.u_bbr.bbr_state |= rack->r_fill_less_agg;
17261                 log.u_bbr.bbr_state <<= 1;
17262                 log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17263                 log.u_bbr.bbr_state <<= 2;
17264                 log.u_bbr.bbr_state |= rack->r_pacing_discount;
17265                 log.u_bbr.flex7 = ((rack->r_ctl.pacing_discount_amm << 1) | log.u_bbr.flex7);
17266                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
17267                     &rack->rc_inp->inp_socket->so_rcv,
17268                     &rack->rc_inp->inp_socket->so_snd,
17269                     BBR_LOG_HPTSI_CALC, 0,
17270                     0, &log, false, &tv);
17271         }
17272 }
17273
17274 static uint32_t
17275 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17276 {
17277         uint32_t new_tso, user_max, pace_one;
17278
17279         user_max = rack->rc_user_set_max_segs * mss;
17280         if (rack->rc_force_max_seg) {
17281                 return (user_max);
17282         }
17283         if (rack->use_fixed_rate &&
17284             ((rack->r_ctl.crte == NULL) ||
17285              (bw != rack->r_ctl.crte->rate))) {
17286                 /* Use the user mss since we are not exactly matched */
17287                 return (user_max);
17288         }
17289         if (rack_pace_one_seg ||
17290             (rack->r_ctl.rc_user_set_min_segs == 1))
17291                 pace_one = 1;
17292         else
17293                 pace_one = 0;
17294
17295         new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17296                      pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17297         if (new_tso > user_max)
17298                 new_tso = user_max;
17299         if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17300                 if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17301                         new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17302         }
17303         if (rack->r_ctl.rc_user_set_min_segs &&
17304             ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17305             new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17306         return (new_tso);
17307 }
17308
17309 static int32_t
17310 pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz, int *capped, uint64_t *rate_wanted, uint8_t non_paced)
17311 {
17312         uint64_t lentim, fill_bw;
17313
17314         /* Lets first see if we are full, if so continue with normal rate */
17315         rack->r_via_fill_cw = 0;
17316         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
17317                 return (slot);
17318         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
17319                 return (slot);
17320         if (rack->r_ctl.rc_last_us_rtt == 0)
17321                 return (slot);
17322         if (rack->rc_pace_fill_if_rttin_range &&
17323             (rack->r_ctl.rc_last_us_rtt >=
17324              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
17325                 /* The rtt is huge, N * smallest, lets not fill */
17326                 return (slot);
17327         }
17328         /*
17329          * first lets calculate the b/w based on the last us-rtt
17330          * and the the smallest send window.
17331          */
17332         fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17333         /* Take the rwnd if its smaller */
17334         if (fill_bw > rack->rc_tp->snd_wnd)
17335                 fill_bw = rack->rc_tp->snd_wnd;
17336         /* Now lets make it into a b/w */
17337         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
17338         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17339         if (rack->r_fill_less_agg) {
17340                 /*
17341                  * We want the average of the rate_wanted
17342                  * and our fill-cw calculated bw. We also want
17343                  * to cap any increase to be no more than
17344                  * X times the lt_bw (where X is the rack_bw_multipler).
17345                  */
17346                 uint64_t lt_bw, rate;
17347
17348                 lt_bw = rack_get_lt_bw(rack);
17349                 if (lt_bw > *rate_wanted)
17350                         rate = lt_bw;
17351                 else
17352                         rate = *rate_wanted;
17353                 fill_bw += rate;
17354                 fill_bw /= 2;
17355                 if (rack_bw_multipler && (fill_bw > (rate * rack_bw_multipler))) {
17356                         fill_bw = rate * rack_bw_multipler;
17357                 }
17358         }
17359         /* We are below the min b/w */
17360         if (non_paced)
17361                 *rate_wanted = fill_bw;
17362         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
17363                 return (slot);
17364         rack->r_via_fill_cw = 1;
17365         if (rack->r_rack_hw_rate_caps &&
17366             (rack->r_ctl.crte != NULL)) {
17367                 uint64_t high_rate;
17368
17369                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
17370                 if (fill_bw > high_rate) {
17371                         /* We are capping bw at the highest rate table entry */
17372                         if (*rate_wanted > high_rate) {
17373                                 /* The original rate was also capped */
17374                                 rack->r_via_fill_cw = 0;
17375                         }
17376                         rack_log_hdwr_pacing(rack,
17377                                              fill_bw, high_rate, __LINE__,
17378                                              0, 3);
17379                         fill_bw = high_rate;
17380                         if (capped)
17381                                 *capped = 1;
17382                 }
17383         } else if ((rack->r_ctl.crte == NULL) &&
17384                    (rack->rack_hdrw_pacing == 0) &&
17385                    (rack->rack_hdw_pace_ena) &&
17386                    rack->r_rack_hw_rate_caps &&
17387                    (rack->rack_attempt_hdwr_pace == 0) &&
17388                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
17389                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17390                 /*
17391                  * Ok we may have a first attempt that is greater than our top rate
17392                  * lets check.
17393                  */
17394                 uint64_t high_rate;
17395
17396                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
17397                 if (high_rate) {
17398                         if (fill_bw > high_rate) {
17399                                 fill_bw = high_rate;
17400                                 if (capped)
17401                                         *capped = 1;
17402                         }
17403                 }
17404         }
17405         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
17406                 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
17407                                    fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL);
17408                 fill_bw = rack->r_ctl.bw_rate_cap;
17409         }
17410         /*
17411          * Ok fill_bw holds our mythical b/w to fill the cwnd
17412          * in an rtt (unless it was capped), what does that
17413          * time wise equate too?
17414          */
17415         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
17416         lentim /= fill_bw;
17417         *rate_wanted = fill_bw;
17418         if (non_paced || (lentim < slot)) {
17419                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
17420                                            0, lentim, 12, __LINE__, NULL, 0);
17421                 return ((int32_t)lentim);
17422         } else
17423                 return (slot);
17424 }
17425
17426 static int32_t
17427 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
17428 {
17429         uint64_t srtt;
17430         int32_t slot = 0;
17431         int32_t minslot = 0;
17432         int can_start_hw_pacing = 1;
17433         int err;
17434         int pace_one;
17435
17436         if (rack_pace_one_seg ||
17437             (rack->r_ctl.rc_user_set_min_segs == 1))
17438                 pace_one = 1;
17439         else
17440                 pace_one = 0;
17441         if (rack->rc_always_pace == 0) {
17442                 /*
17443                  * We use the most optimistic possible cwnd/srtt for
17444                  * sending calculations. This will make our
17445                  * calculation anticipate getting more through
17446                  * quicker then possible. But thats ok we don't want
17447                  * the peer to have a gap in data sending.
17448                  */
17449                 uint64_t cwnd, tr_perms = 0;
17450                 int32_t reduce = 0;
17451
17452         old_method:
17453                 /*
17454                  * We keep no precise pacing with the old method
17455                  * instead we use the pacer to mitigate bursts.
17456                  */
17457                 if (rack->r_ctl.rc_rack_min_rtt)
17458                         srtt = rack->r_ctl.rc_rack_min_rtt;
17459                 else
17460                         srtt = max(tp->t_srtt, 1);
17461                 if (rack->r_ctl.rc_rack_largest_cwnd)
17462                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
17463                 else
17464                         cwnd = rack->r_ctl.cwnd_to_use;
17465                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
17466                 tr_perms = (cwnd * 1000) / srtt;
17467                 if (tr_perms == 0) {
17468                         tr_perms = ctf_fixed_maxseg(tp);
17469                 }
17470                 /*
17471                  * Calculate how long this will take to drain, if
17472                  * the calculation comes out to zero, thats ok we
17473                  * will use send_a_lot to possibly spin around for
17474                  * more increasing tot_len_this_send to the point
17475                  * that its going to require a pace, or we hit the
17476                  * cwnd. Which in that case we are just waiting for
17477                  * a ACK.
17478                  */
17479                 slot = len / tr_perms;
17480                 /* Now do we reduce the time so we don't run dry? */
17481                 if (slot && rack_slot_reduction) {
17482                         reduce = (slot / rack_slot_reduction);
17483                         if (reduce < slot) {
17484                                 slot -= reduce;
17485                         } else
17486                                 slot = 0;
17487                 }
17488                 slot *= HPTS_USEC_IN_MSEC;
17489                 if (rack->rc_pace_to_cwnd) {
17490                         uint64_t rate_wanted = 0;
17491
17492                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
17493                         rack->rc_ack_can_sendout_data = 1;
17494                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
17495                 } else
17496                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
17497                 /*******************************************************/
17498                 /* RRS: We insert non-paced call to stats here for len */
17499                 /*******************************************************/
17500         } else {
17501                 uint64_t bw_est, res, lentim, rate_wanted;
17502                 uint32_t segs, oh;
17503                 int capped = 0;
17504                 int prev_fill;
17505
17506                 if ((rack->r_rr_config == 1) && rsm) {
17507                         return (rack->r_ctl.rc_min_to);
17508                 }
17509                 if (rack->use_fixed_rate) {
17510                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
17511                 } else if ((rack->r_ctl.init_rate == 0) &&
17512                            (rack->r_ctl.gp_bw == 0)) {
17513                         /* no way to yet do an estimate */
17514                         bw_est = rate_wanted = 0;
17515                 } else {
17516                         bw_est = rack_get_bw(rack);
17517                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
17518                 }
17519                 if ((bw_est == 0) || (rate_wanted == 0) ||
17520                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
17521                         /*
17522                          * No way yet to make a b/w estimate or
17523                          * our raise is set incorrectly.
17524                          */
17525                         goto old_method;
17526                 }
17527                 rack_rate_cap_bw(rack, &rate_wanted, &capped);
17528                 /* We need to account for all the overheads */
17529                 segs = (len + segsiz - 1) / segsiz;
17530                 /*
17531                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
17532                  * and how much data we put in each packet. Yes this
17533                  * means we may be off if we are larger than 1500 bytes
17534                  * or smaller. But this just makes us more conservative.
17535                  */
17536
17537                 oh =  (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
17538                 if (rack->r_is_v6) {
17539 #ifdef INET6
17540                         oh += sizeof(struct ip6_hdr);
17541 #endif
17542                 } else {
17543 #ifdef INET
17544                         oh += sizeof(struct ip);
17545 #endif
17546                 }
17547                 /* We add a fixed 14 for the ethernet header */
17548                 oh += 14;
17549                 segs *= oh;
17550                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
17551                 res = lentim / rate_wanted;
17552                 slot = (uint32_t)res;
17553                 if (rack_hw_rate_min &&
17554                     (rate_wanted < rack_hw_rate_min)) {
17555                         can_start_hw_pacing = 0;
17556                         if (rack->r_ctl.crte) {
17557                                 /*
17558                                  * Ok we need to release it, we
17559                                  * have fallen too low.
17560                                  */
17561                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17562                                 rack->r_ctl.crte = NULL;
17563                                 rack->rack_attempt_hdwr_pace = 0;
17564                                 rack->rack_hdrw_pacing = 0;
17565                         }
17566                 }
17567                 if (rack->r_ctl.crte &&
17568                     (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17569                         /*
17570                          * We want more than the hardware can give us,
17571                          * don't start any hw pacing.
17572                          */
17573                         can_start_hw_pacing = 0;
17574                         if (rack->r_rack_hw_rate_caps == 0) {
17575                                 /*
17576                                  * Ok we need to release it, we
17577                                  * want more than the card can give us and
17578                                  * no rate cap is in place. Set it up so
17579                                  * when we want less we can retry.
17580                                  */
17581                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17582                                 rack->r_ctl.crte = NULL;
17583                                 rack->rack_attempt_hdwr_pace = 0;
17584                                 rack->rack_hdrw_pacing = 0;
17585                         }
17586                 }
17587                 if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
17588                         /*
17589                          * We lost our rate somehow, this can happen
17590                          * if the interface changed underneath us.
17591                          */
17592                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17593                         rack->r_ctl.crte = NULL;
17594                         /* Lets re-allow attempting to setup pacing */
17595                         rack->rack_hdrw_pacing = 0;
17596                         rack->rack_attempt_hdwr_pace = 0;
17597                         rack_log_hdwr_pacing(rack,
17598                                              rate_wanted, bw_est, __LINE__,
17599                                              0, 6);
17600                 }
17601                 prev_fill = rack->r_via_fill_cw;
17602                 if ((rack->rc_pace_to_cwnd) &&
17603                     (capped == 0) &&
17604                     (rack->use_fixed_rate == 0) &&
17605                     (rack->in_probe_rtt == 0) &&
17606                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
17607                         /*
17608                          * We want to pace at our rate *or* faster to
17609                          * fill the cwnd to the max if its not full.
17610                          */
17611                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
17612                         /* Re-check to make sure we are not exceeding our max b/w */
17613                         if ((rack->r_ctl.crte != NULL) &&
17614                             (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17615                                 /*
17616                                  * We want more than the hardware can give us,
17617                                  * don't start any hw pacing.
17618                                  */
17619                                 can_start_hw_pacing = 0;
17620                                 if (rack->r_rack_hw_rate_caps == 0) {
17621                                         /*
17622                                          * Ok we need to release it, we
17623                                          * want more than the card can give us and
17624                                          * no rate cap is in place. Set it up so
17625                                          * when we want less we can retry.
17626                                          */
17627                                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17628                                         rack->r_ctl.crte = NULL;
17629                                         rack->rack_attempt_hdwr_pace = 0;
17630                                         rack->rack_hdrw_pacing = 0;
17631                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
17632                                 }
17633                         }
17634                 }
17635                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
17636                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17637                         if ((rack->rack_hdw_pace_ena) &&
17638                             (can_start_hw_pacing > 0) &&
17639                             (rack->rack_hdrw_pacing == 0) &&
17640                             (rack->rack_attempt_hdwr_pace == 0)) {
17641                                 /*
17642                                  * Lets attempt to turn on hardware pacing
17643                                  * if we can.
17644                                  */
17645                                 rack->rack_attempt_hdwr_pace = 1;
17646                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
17647                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
17648                                                                        rate_wanted,
17649                                                                        RS_PACING_GEQ,
17650                                                                        &err, &rack->r_ctl.crte_prev_rate);
17651                                 if (rack->r_ctl.crte) {
17652                                         rack->rack_hdrw_pacing = 1;
17653                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
17654                                                                           pace_one, rack->r_ctl.crte,
17655                                                                           NULL, rack->r_ctl.pace_len_divisor);
17656                                         rack_log_hdwr_pacing(rack,
17657                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17658                                                              err, 0);
17659                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
17660                                 } else {
17661                                         counter_u64_add(rack_hw_pace_init_fail, 1);
17662                                 }
17663                         } else if (rack->rack_hdrw_pacing &&
17664                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
17665                                 /* Do we need to adjust our rate? */
17666                                 const struct tcp_hwrate_limit_table *nrte;
17667
17668                                 if (rack->r_up_only &&
17669                                     (rate_wanted < rack->r_ctl.crte->rate)) {
17670                                         /**
17671                                          * We have four possible states here
17672                                          * having to do with the previous time
17673                                          * and this time.
17674                                          *   previous  |  this-time
17675                                          * A)     0      |     0   -- fill_cw not in the picture
17676                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
17677                                          * C)     1      |     1   -- all rates from fill_cw
17678                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
17679                                          *
17680                                          * For case A, C and D we don't allow a drop. But for
17681                                          * case B where we now our on our steady rate we do
17682                                          * allow a drop.
17683                                          *
17684                                          */
17685                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
17686                                                 goto done_w_hdwr;
17687                                 }
17688                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
17689                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
17690                                         if (rack_hw_rate_to_low &&
17691                                             (bw_est < rack_hw_rate_to_low)) {
17692                                                 /*
17693                                                  * The pacing rate is too low for hardware, but
17694                                                  * do allow hardware pacing to be restarted.
17695                                                  */
17696                                                 rack_log_hdwr_pacing(rack,
17697                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
17698                                                              0, 5);
17699                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17700                                                 rack->r_ctl.crte = NULL;
17701                                                 rack->rack_attempt_hdwr_pace = 0;
17702                                                 rack->rack_hdrw_pacing = 0;
17703                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17704                                                 goto done_w_hdwr;
17705                                         }
17706                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
17707                                                                             rack->rc_tp,
17708                                                                             rack->rc_inp->inp_route.ro_nh->nh_ifp,
17709                                                                             rate_wanted,
17710                                                                             RS_PACING_GEQ,
17711                                                                             &err, &rack->r_ctl.crte_prev_rate);
17712                                         if (nrte == NULL) {
17713                                                 /*
17714                                                  * Lost the rate, lets drop hardware pacing
17715                                                  * period.
17716                                                  */
17717                                                 rack->rack_hdrw_pacing = 0;
17718                                                 rack->r_ctl.crte = NULL;
17719                                                 rack_log_hdwr_pacing(rack,
17720                                                                      rate_wanted, 0, __LINE__,
17721                                                                      err, 1);
17722                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17723                                                 counter_u64_add(rack_hw_pace_lost, 1);
17724                                         } else if (nrte != rack->r_ctl.crte) {
17725                                                 rack->r_ctl.crte = nrte;
17726                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
17727                                                                                  segsiz, pace_one, rack->r_ctl.crte,
17728                                                                                  NULL, rack->r_ctl.pace_len_divisor);
17729                                                 rack_log_hdwr_pacing(rack,
17730                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17731                                                                      err, 2);
17732                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
17733                                         }
17734                                 } else {
17735                                         /* We just need to adjust the segment size */
17736                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17737                                         rack_log_hdwr_pacing(rack,
17738                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17739                                                              0, 4);
17740                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
17741                                 }
17742                         }
17743                 }
17744                 if (minslot && (minslot > slot)) {
17745                         rack_log_pacing_delay_calc(rack, minslot, slot, rack->r_ctl.crte->rate, bw_est, lentim,
17746                                                    98, __LINE__, NULL, 0);
17747                         slot = minslot;
17748                 }
17749 done_w_hdwr:
17750                 if (rack_limit_time_with_srtt &&
17751                     (rack->use_fixed_rate == 0) &&
17752                     (rack->rack_hdrw_pacing == 0)) {
17753                         /*
17754                          * Sanity check, we do not allow the pacing delay
17755                          * to be longer than the SRTT of the path. If it is
17756                          * a slow path, then adding a packet should increase
17757                          * the RTT and compensate for this i.e. the srtt will
17758                          * be greater so the allowed pacing time will be greater.
17759                          *
17760                          * Note this restriction is not for where a peak rate
17761                          * is set, we are doing fixed pacing or hardware pacing.
17762                          */
17763                         if (rack->rc_tp->t_srtt)
17764                                 srtt = rack->rc_tp->t_srtt;
17765                         else
17766                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
17767                         if (srtt < (uint64_t)slot) {
17768                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
17769                                 slot = srtt;
17770                         }
17771                 }
17772                 /*******************************************************************/
17773                 /* RRS: We insert paced call to stats here for len and rate_wanted */
17774                 /*******************************************************************/
17775                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
17776         }
17777         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
17778                 /*
17779                  * If this rate is seeing enobufs when it
17780                  * goes to send then either the nic is out
17781                  * of gas or we are mis-estimating the time
17782                  * somehow and not letting the queue empty
17783                  * completely. Lets add to the pacing time.
17784                  */
17785                 int hw_boost_delay;
17786
17787                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
17788                 if (hw_boost_delay > rack_enobuf_hw_max)
17789                         hw_boost_delay = rack_enobuf_hw_max;
17790                 else if (hw_boost_delay < rack_enobuf_hw_min)
17791                         hw_boost_delay = rack_enobuf_hw_min;
17792                 slot += hw_boost_delay;
17793         }
17794         return (slot);
17795 }
17796
17797 static void
17798 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
17799     tcp_seq startseq, uint32_t sb_offset)
17800 {
17801         struct rack_sendmap *my_rsm = NULL;
17802
17803         if (tp->t_state < TCPS_ESTABLISHED) {
17804                 /*
17805                  * We don't start any measurements if we are
17806                  * not at least established.
17807                  */
17808                 return;
17809         }
17810         if (tp->t_state >= TCPS_FIN_WAIT_1) {
17811                 /*
17812                  * We will get no more data into the SB
17813                  * this means we need to have the data available
17814                  * before we start a measurement.
17815                  */
17816
17817                 if (sbavail(&tptosocket(tp)->so_snd) <
17818                     max(rc_init_window(rack),
17819                         (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
17820                         /* Nope not enough data */
17821                         return;
17822                 }
17823         }
17824         tp->t_flags |= TF_GPUTINPROG;
17825         rack->r_ctl.rc_gp_cumack_ts = 0;
17826         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
17827         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
17828         tp->gput_seq = startseq;
17829         rack->app_limited_needs_set = 0;
17830         if (rack->in_probe_rtt)
17831                 rack->measure_saw_probe_rtt = 1;
17832         else if ((rack->measure_saw_probe_rtt) &&
17833                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
17834                 rack->measure_saw_probe_rtt = 0;
17835         if (rack->rc_gp_filled)
17836                 tp->gput_ts = rack->r_ctl.last_cumack_advance;
17837         else {
17838                 /* Special case initial measurement */
17839                 struct timeval tv;
17840
17841                 tp->gput_ts = tcp_get_usecs(&tv);
17842                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
17843         }
17844         /*
17845          * We take a guess out into the future,
17846          * if we have no measurement and no
17847          * initial rate, we measure the first
17848          * initial-windows worth of data to
17849          * speed up getting some GP measurement and
17850          * thus start pacing.
17851          */
17852         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
17853                 rack->app_limited_needs_set = 1;
17854                 tp->gput_ack = startseq + max(rc_init_window(rack),
17855                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
17856                 rack_log_pacing_delay_calc(rack,
17857                                            tp->gput_seq,
17858                                            tp->gput_ack,
17859                                            0,
17860                                            tp->gput_ts,
17861                                            (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
17862                                            9,
17863                                            __LINE__, NULL, 0);
17864                 rack_tend_gp_marks(tp, rack);
17865                 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
17866                 return;
17867         }
17868         if (sb_offset) {
17869                 /*
17870                  * We are out somewhere in the sb
17871                  * can we use the already outstanding data?
17872                  */
17873
17874                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
17875                         /*
17876                          * Yes first one is good and in this case
17877                          * the tp->gput_ts is correctly set based on
17878                          * the last ack that arrived (no need to
17879                          * set things up when an ack comes in).
17880                          */
17881                         my_rsm = tqhash_min(rack->r_ctl.tqh);
17882                         if ((my_rsm == NULL) ||
17883                             (my_rsm->r_rtr_cnt != 1)) {
17884                                 /* retransmission? */
17885                                 goto use_latest;
17886                         }
17887                 } else {
17888                         if (rack->r_ctl.rc_first_appl == NULL) {
17889                                 /*
17890                                  * If rc_first_appl is NULL
17891                                  * then the cnt should be 0.
17892                                  * This is probably an error, maybe
17893                                  * a KASSERT would be approprate.
17894                                  */
17895                                 goto use_latest;
17896                         }
17897                         /*
17898                          * If we have a marker pointer to the last one that is
17899                          * app limited we can use that, but we need to set
17900                          * things up so that when it gets ack'ed we record
17901                          * the ack time (if its not already acked).
17902                          */
17903                         rack->app_limited_needs_set = 1;
17904                         /*
17905                          * We want to get to the rsm that is either
17906                          * next with space i.e. over 1 MSS or the one
17907                          * after that (after the app-limited).
17908                          */
17909                         my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
17910                         if (my_rsm) {
17911                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
17912                                         /* Have to use the next one */
17913                                         my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17914                                 else {
17915                                         /* Use after the first MSS of it is acked */
17916                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
17917                                         goto start_set;
17918                                 }
17919                         }
17920                         if ((my_rsm == NULL) ||
17921                             (my_rsm->r_rtr_cnt != 1)) {
17922                                 /*
17923                                  * Either its a retransmit or
17924                                  * the last is the app-limited one.
17925                                  */
17926                                 goto use_latest;
17927                         }
17928                 }
17929                 tp->gput_seq = my_rsm->r_start;
17930 start_set:
17931                 if (my_rsm->r_flags & RACK_ACKED) {
17932                         /*
17933                          * This one has been acked use the arrival ack time
17934                          */
17935                         struct rack_sendmap *nrsm;
17936
17937                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
17938                         rack->app_limited_needs_set = 0;
17939                         /*
17940                          * Ok in this path we need to use the r_end now
17941                          * since this guy is the starting ack.
17942                          */
17943                         tp->gput_seq = my_rsm->r_end;
17944                         /*
17945                          * We also need to adjust up the sendtime
17946                          * to the send of the next data after my_rsm.
17947                          */
17948                         nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17949                         if (nrsm != NULL)
17950                                 my_rsm = nrsm;
17951                         else {
17952                                 /*
17953                                  * The next as not been sent, thats the
17954                                  * case for using the latest.
17955                                  */
17956                                 goto use_latest;
17957                         }
17958                 }
17959                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
17960                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
17961                 rack->r_ctl.rc_gp_cumack_ts = 0;
17962                 rack_log_pacing_delay_calc(rack,
17963                                            tp->gput_seq,
17964                                            tp->gput_ack,
17965                                            (uint64_t)my_rsm,
17966                                            tp->gput_ts,
17967                                            (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
17968                                            9,
17969                                            __LINE__, my_rsm, 0);
17970                 /* Now lets make sure all are marked as they should be */
17971                 rack_tend_gp_marks(tp, rack);
17972                 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
17973                 return;
17974         }
17975
17976 use_latest:
17977         /*
17978          * We don't know how long we may have been
17979          * idle or if this is the first-send. Lets
17980          * setup the flag so we will trim off
17981          * the first ack'd data so we get a true
17982          * measurement.
17983          */
17984         rack->app_limited_needs_set = 1;
17985         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
17986         rack->r_ctl.rc_gp_cumack_ts = 0;
17987         /* Find this guy so we can pull the send time */
17988         my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
17989         if (my_rsm) {
17990                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
17991                 if (my_rsm->r_flags & RACK_ACKED) {
17992                         /*
17993                          * Unlikely since its probably what was
17994                          * just transmitted (but I am paranoid).
17995                          */
17996                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
17997                         rack->app_limited_needs_set = 0;
17998                 }
17999                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18000                         /* This also is unlikely */
18001                         tp->gput_seq = my_rsm->r_start;
18002                 }
18003         } else {
18004                 /*
18005                  * TSNH unless we have some send-map limit,
18006                  * and even at that it should not be hitting
18007                  * that limit (we should have stopped sending).
18008                  */
18009                 struct timeval tv;
18010
18011                 microuptime(&tv);
18012                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18013         }
18014         rack_tend_gp_marks(tp, rack);
18015         rack_log_pacing_delay_calc(rack,
18016                                    tp->gput_seq,
18017                                    tp->gput_ack,
18018                                    (uint64_t)my_rsm,
18019                                    tp->gput_ts,
18020                                    (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18021                                    9, __LINE__, NULL, 0);
18022         rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18023 }
18024
18025 static inline uint32_t
18026 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
18027     uint32_t avail, int32_t sb_offset)
18028 {
18029         uint32_t len;
18030         uint32_t sendwin;
18031
18032         if (tp->snd_wnd > cwnd_to_use)
18033                 sendwin = cwnd_to_use;
18034         else
18035                 sendwin = tp->snd_wnd;
18036         if (ctf_outstanding(tp) >= tp->snd_wnd) {
18037                 /* We never want to go over our peers rcv-window */
18038                 len = 0;
18039         } else {
18040                 uint32_t flight;
18041
18042                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18043                 if (flight >= sendwin) {
18044                         /*
18045                          * We have in flight what we are allowed by cwnd (if
18046                          * it was rwnd blocking it would have hit above out
18047                          * >= tp->snd_wnd).
18048                          */
18049                         return (0);
18050                 }
18051                 len = sendwin - flight;
18052                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18053                         /* We would send too much (beyond the rwnd) */
18054                         len = tp->snd_wnd - ctf_outstanding(tp);
18055                 }
18056                 if ((len + sb_offset) > avail) {
18057                         /*
18058                          * We don't have that much in the SB, how much is
18059                          * there?
18060                          */
18061                         len = avail - sb_offset;
18062                 }
18063         }
18064         return (len);
18065 }
18066
18067 static void
18068 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18069              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18070              int rsm_is_null, int optlen, int line, uint16_t mode)
18071 {
18072         if (tcp_bblogging_on(rack->rc_tp)) {
18073                 union tcp_log_stackspecific log;
18074                 struct timeval tv;
18075
18076                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18077                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18078                 log.u_bbr.flex1 = error;
18079                 log.u_bbr.flex2 = flags;
18080                 log.u_bbr.flex3 = rsm_is_null;
18081                 log.u_bbr.flex4 = ipoptlen;
18082                 log.u_bbr.flex5 = tp->rcv_numsacks;
18083                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18084                 log.u_bbr.flex7 = optlen;
18085                 log.u_bbr.flex8 = rack->r_fsb_inited;
18086                 log.u_bbr.applimited = rack->r_fast_output;
18087                 log.u_bbr.bw_inuse = rack_get_bw(rack);
18088                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18089                 log.u_bbr.cwnd_gain = mode;
18090                 log.u_bbr.pkts_out = orig_len;
18091                 log.u_bbr.lt_epoch = len;
18092                 log.u_bbr.delivered = line;
18093                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18094                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18095                 tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
18096                                len, &log, false, NULL, __func__, __LINE__, &tv);
18097         }
18098 }
18099
18100
18101 static struct mbuf *
18102 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
18103                    struct rack_fast_send_blk *fsb,
18104                    int32_t seglimit, int32_t segsize, int hw_tls)
18105 {
18106 #ifdef KERN_TLS
18107         struct ktls_session *tls, *ntls;
18108 #ifdef INVARIANTS
18109         struct mbuf *start;
18110 #endif
18111 #endif
18112         struct mbuf *m, *n, **np, *smb;
18113         struct mbuf *top;
18114         int32_t off, soff;
18115         int32_t len = *plen;
18116         int32_t fragsize;
18117         int32_t len_cp = 0;
18118         uint32_t mlen, frags;
18119
18120         soff = off = the_off;
18121         smb = m = the_m;
18122         np = &top;
18123         top = NULL;
18124 #ifdef KERN_TLS
18125         if (hw_tls && (m->m_flags & M_EXTPG))
18126                 tls = m->m_epg_tls;
18127         else
18128                 tls = NULL;
18129 #ifdef INVARIANTS
18130         start = m;
18131 #endif
18132 #endif
18133         while (len > 0) {
18134                 if (m == NULL) {
18135                         *plen = len_cp;
18136                         break;
18137                 }
18138 #ifdef KERN_TLS
18139                 if (hw_tls) {
18140                         if (m->m_flags & M_EXTPG)
18141                                 ntls = m->m_epg_tls;
18142                         else
18143                                 ntls = NULL;
18144
18145                         /*
18146                          * Avoid mixing TLS records with handshake
18147                          * data or TLS records from different
18148                          * sessions.
18149                          */
18150                         if (tls != ntls) {
18151                                 MPASS(m != start);
18152                                 *plen = len_cp;
18153                                 break;
18154                         }
18155                 }
18156 #endif
18157                 mlen = min(len, m->m_len - off);
18158                 if (seglimit) {
18159                         /*
18160                          * For M_EXTPG mbufs, add 3 segments
18161                          * + 1 in case we are crossing page boundaries
18162                          * + 2 in case the TLS hdr/trailer are used
18163                          * It is cheaper to just add the segments
18164                          * than it is to take the cache miss to look
18165                          * at the mbuf ext_pgs state in detail.
18166                          */
18167                         if (m->m_flags & M_EXTPG) {
18168                                 fragsize = min(segsize, PAGE_SIZE);
18169                                 frags = 3;
18170                         } else {
18171                                 fragsize = segsize;
18172                                 frags = 0;
18173                         }
18174
18175                         /* Break if we really can't fit anymore. */
18176                         if ((frags + 1) >= seglimit) {
18177                                 *plen = len_cp;
18178                                 break;
18179                         }
18180
18181                         /*
18182                          * Reduce size if you can't copy the whole
18183                          * mbuf. If we can't copy the whole mbuf, also
18184                          * adjust len so the loop will end after this
18185                          * mbuf.
18186                          */
18187                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
18188                                 mlen = (seglimit - frags - 1) * fragsize;
18189                                 len = mlen;
18190                                 *plen = len_cp + len;
18191                         }
18192                         frags += howmany(mlen, fragsize);
18193                         if (frags == 0)
18194                                 frags++;
18195                         seglimit -= frags;
18196                         KASSERT(seglimit > 0,
18197                             ("%s: seglimit went too low", __func__));
18198                 }
18199                 n = m_get(M_NOWAIT, m->m_type);
18200                 *np = n;
18201                 if (n == NULL)
18202                         goto nospace;
18203                 n->m_len = mlen;
18204                 soff += mlen;
18205                 len_cp += n->m_len;
18206                 if (m->m_flags & (M_EXT|M_EXTPG)) {
18207                         n->m_data = m->m_data + off;
18208                         mb_dupcl(n, m);
18209                 } else {
18210                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
18211                             (u_int)n->m_len);
18212                 }
18213                 len -= n->m_len;
18214                 off = 0;
18215                 m = m->m_next;
18216                 np = &n->m_next;
18217                 if (len || (soff == smb->m_len)) {
18218                         /*
18219                          * We have more so we move forward  or
18220                          * we have consumed the entire mbuf and
18221                          * len has fell to 0.
18222                          */
18223                         soff = 0;
18224                         smb = m;
18225                 }
18226
18227         }
18228         if (fsb != NULL) {
18229                 fsb->m = smb;
18230                 fsb->off = soff;
18231                 if (smb) {
18232                         /*
18233                          * Save off the size of the mbuf. We do
18234                          * this so that we can recognize when it
18235                          * has been trimmed by sbcut() as acks
18236                          * come in.
18237                          */
18238                         fsb->o_m_len = smb->m_len;
18239                         fsb->o_t_len = M_TRAILINGROOM(smb);
18240                 } else {
18241                         /*
18242                          * This is the case where the next mbuf went to NULL. This
18243                          * means with this copy we have sent everything in the sb.
18244                          * In theory we could clear the fast_output flag, but lets
18245                          * not since its possible that we could get more added
18246                          * and acks that call the extend function which would let
18247                          * us send more.
18248                          */
18249                         fsb->o_m_len = 0;
18250                         fsb->o_t_len = 0;
18251                 }
18252         }
18253         return (top);
18254 nospace:
18255         if (top)
18256                 m_freem(top);
18257         return (NULL);
18258
18259 }
18260
18261 /*
18262  * This is a copy of m_copym(), taking the TSO segment size/limit
18263  * constraints into account, and advancing the sndptr as it goes.
18264  */
18265 static struct mbuf *
18266 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
18267                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
18268 {
18269         struct mbuf *m, *n;
18270         int32_t soff;
18271
18272         m = rack->r_ctl.fsb.m;
18273         if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
18274                 /*
18275                  * The trailing space changed, mbufs can grow
18276                  * at the tail but they can't shrink from
18277                  * it, KASSERT that. Adjust the orig_m_len to
18278                  * compensate for this change.
18279                  */
18280                 KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
18281                         ("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
18282                          m,
18283                          rack,
18284                          (intmax_t)M_TRAILINGROOM(m),
18285                          rack->r_ctl.fsb.o_t_len,
18286                          rack->r_ctl.fsb.o_m_len,
18287                          m->m_len));
18288                 rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
18289                 rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
18290         }
18291         if (m->m_len < rack->r_ctl.fsb.o_m_len) {
18292                 /*
18293                  * Mbuf shrank, trimmed off the top by an ack, our
18294                  * offset changes.
18295                  */
18296                 KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
18297                         ("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
18298                          m, m->m_len,
18299                          rack, rack->r_ctl.fsb.o_m_len,
18300                          rack->r_ctl.fsb.off));
18301
18302                 if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
18303                         rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
18304                 else
18305                         rack->r_ctl.fsb.off = 0;
18306                 rack->r_ctl.fsb.o_m_len = m->m_len;
18307 #ifdef INVARIANTS
18308         } else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
18309                 panic("rack:%p m:%p m_len grew outside of t_space compensation",
18310                       rack, m);
18311 #endif
18312         }
18313         soff = rack->r_ctl.fsb.off;
18314         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
18315         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
18316         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
18317                                  __FUNCTION__,
18318                                  rack, *plen, m, m->m_len));
18319         /* Save off the right location before we copy and advance */
18320         *s_soff = soff;
18321         *s_mb = rack->r_ctl.fsb.m;
18322         n = rack_fo_base_copym(m, soff, plen,
18323                                &rack->r_ctl.fsb,
18324                                seglimit, segsize, rack->r_ctl.fsb.hw_tls);
18325         return (n);
18326 }
18327
18328 /* Log the buffer level */
18329 static void
18330 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
18331                      int len, struct timeval *tv,
18332                      uint32_t cts)
18333 {
18334         uint32_t p_rate = 0, p_queue = 0, err = 0;
18335         union tcp_log_stackspecific log;
18336
18337 #ifdef RATELIMIT
18338         err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18339         err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18340 #endif
18341         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18342         log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18343         log.u_bbr.flex1 = p_rate;
18344         log.u_bbr.flex2 = p_queue;
18345         log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18346         log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18347         log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18348         log.u_bbr.flex7 = 99;
18349         log.u_bbr.flex8 = 0;
18350         log.u_bbr.pkts_out = err;
18351         log.u_bbr.delRate = rack->r_ctl.crte->rate;
18352         log.u_bbr.timeStamp = cts;
18353         log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18354         tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18355                        len, &log, false, NULL, __func__, __LINE__, tv);
18356
18357 }
18358
18359 static uint32_t
18360 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
18361                        struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
18362 {
18363         uint64_t lentime = 0;
18364 #ifdef RATELIMIT
18365         uint32_t p_rate = 0, p_queue = 0, err;
18366         union tcp_log_stackspecific log;
18367         uint64_t bw;
18368
18369         err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18370         /* Failed or queue is zero */
18371         if (err || (p_queue == 0)) {
18372                 lentime = 0;
18373                 goto out;
18374         }
18375         err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18376         if (err) {
18377                 lentime = 0;
18378                 goto out;
18379         }
18380         /*
18381          * If we reach here we have some bytes in
18382          * the queue. The number returned is a value
18383          * between 0 and 0xffff where ffff is full
18384          * and 0 is empty. So how best to make this into
18385          * something usable?
18386          *
18387          * The "safer" way is lets take the b/w gotten
18388          * from the query (which should be our b/w rate)
18389          * and pretend that a full send (our rc_pace_max_segs)
18390          * is outstanding. We factor it so its as if a full
18391          * number of our MSS segment is terms of full
18392          * ethernet segments are outstanding.
18393          */
18394         bw = p_rate / 8;
18395         if (bw) {
18396                 lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
18397                 lentime *= ETHERNET_SEGMENT_SIZE;
18398                 lentime *= (uint64_t)HPTS_USEC_IN_SEC;
18399                 lentime /= bw;
18400         } else {
18401                 /* TSNH -- KASSERT? */
18402                 lentime = 0;
18403         }
18404 out:
18405         if (tcp_bblogging_on(tp)) {
18406                 memset(&log, 0, sizeof(log));
18407                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18408                 log.u_bbr.flex1 = p_rate;
18409                 log.u_bbr.flex2 = p_queue;
18410                 log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18411                 log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18412                 log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18413                 log.u_bbr.flex7 = 99;
18414                 log.u_bbr.flex8 = 0;
18415                 log.u_bbr.pkts_out = err;
18416                 log.u_bbr.delRate = rack->r_ctl.crte->rate;
18417                 log.u_bbr.cur_del_rate = lentime;
18418                 log.u_bbr.timeStamp = cts;
18419                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18420                 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18421                                len, &log, false, NULL, __func__, __LINE__,tv);
18422         }
18423 #endif
18424         return ((uint32_t)lentime);
18425 }
18426
18427 static int
18428 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
18429                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
18430 {
18431         /*
18432          * Enter the fast retransmit path. We are given that a sched_pin is
18433          * in place (if accounting is compliled in) and the cycle count taken
18434          * at the entry is in the ts_val. The concept her is that the rsm
18435          * now holds the mbuf offsets and such so we can directly transmit
18436          * without a lot of overhead, the len field is already set for
18437          * us to prohibit us from sending too much (usually its 1MSS).
18438          */
18439         struct ip *ip = NULL;
18440         struct udphdr *udp = NULL;
18441         struct tcphdr *th = NULL;
18442         struct mbuf *m = NULL;
18443         struct inpcb *inp;
18444         uint8_t *cpto;
18445         struct tcp_log_buffer *lgb;
18446 #ifdef TCP_ACCOUNTING
18447         uint64_t crtsc;
18448         int cnt_thru = 1;
18449 #endif
18450         struct tcpopt to;
18451         u_char opt[TCP_MAXOLEN];
18452         uint32_t hdrlen, optlen;
18453         int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
18454         uint16_t flags;
18455         uint32_t if_hw_tsomaxsegcount = 0, startseq;
18456         uint32_t if_hw_tsomaxsegsize;
18457         int32_t ip_sendflag = IP_NO_SND_TAG_RL;
18458
18459 #ifdef INET6
18460         struct ip6_hdr *ip6 = NULL;
18461
18462         if (rack->r_is_v6) {
18463                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18464                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18465         } else
18466 #endif                          /* INET6 */
18467         {
18468                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18469                 hdrlen = sizeof(struct tcpiphdr);
18470         }
18471         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
18472                 goto failed;
18473         }
18474         if (doing_tlp) {
18475                 /* Its a TLP add the flag, it may already be there but be sure */
18476                 rsm->r_flags |= RACK_TLP;
18477         } else {
18478                 /* If it was a TLP it is not not on this retransmit */
18479                 rsm->r_flags &= ~RACK_TLP;
18480         }
18481         startseq = rsm->r_start;
18482         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
18483         inp = rack->rc_inp;
18484         to.to_flags = 0;
18485         flags = tcp_outflags[tp->t_state];
18486         if (flags & (TH_SYN|TH_RST)) {
18487                 goto failed;
18488         }
18489         if (rsm->r_flags & RACK_HAS_FIN) {
18490                 /* We can't send a FIN here */
18491                 goto failed;
18492         }
18493         if (flags & TH_FIN) {
18494                 /* We never send a FIN */
18495                 flags &= ~TH_FIN;
18496         }
18497         if (tp->t_flags & TF_RCVD_TSTMP) {
18498                 to.to_tsval = ms_cts + tp->ts_offset;
18499                 to.to_tsecr = tp->ts_recent;
18500                 to.to_flags = TOF_TS;
18501         }
18502         optlen = tcp_addoptions(&to, opt);
18503         hdrlen += optlen;
18504         udp = rack->r_ctl.fsb.udp;
18505         if (udp)
18506                 hdrlen += sizeof(struct udphdr);
18507         if (rack->r_ctl.rc_pace_max_segs)
18508                 max_val = rack->r_ctl.rc_pace_max_segs;
18509         else if (rack->rc_user_set_max_segs)
18510                 max_val = rack->rc_user_set_max_segs * segsiz;
18511         else
18512                 max_val = len;
18513         if ((tp->t_flags & TF_TSO) &&
18514             V_tcp_do_tso &&
18515             (len > segsiz) &&
18516             (tp->t_port == 0))
18517                 tso = 1;
18518 #ifdef INET6
18519         if (MHLEN < hdrlen + max_linkhdr)
18520                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18521         else
18522 #endif
18523                 m = m_gethdr(M_NOWAIT, MT_DATA);
18524         if (m == NULL)
18525                 goto failed;
18526         m->m_data += max_linkhdr;
18527         m->m_len = hdrlen;
18528         th = rack->r_ctl.fsb.th;
18529         /* Establish the len to send */
18530         if (len > max_val)
18531                 len = max_val;
18532         if ((tso) && (len + optlen > segsiz)) {
18533                 uint32_t if_hw_tsomax;
18534                 int32_t max_len;
18535
18536                 /* extract TSO information */
18537                 if_hw_tsomax = tp->t_tsomax;
18538                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18539                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18540                 /*
18541                  * Check if we should limit by maximum payload
18542                  * length:
18543                  */
18544                 if (if_hw_tsomax != 0) {
18545                         /* compute maximum TSO length */
18546                         max_len = (if_hw_tsomax - hdrlen -
18547                                    max_linkhdr);
18548                         if (max_len <= 0) {
18549                                 goto failed;
18550                         } else if (len > max_len) {
18551                                 len = max_len;
18552                         }
18553                 }
18554                 if (len <= segsiz) {
18555                         /*
18556                          * In case there are too many small fragments don't
18557                          * use TSO:
18558                          */
18559                         tso = 0;
18560                 }
18561         } else {
18562                 tso = 0;
18563         }
18564         if ((tso == 0) && (len > segsiz))
18565                 len = segsiz;
18566         (void)tcp_get_usecs(tv);
18567         if ((len == 0) ||
18568             (len <= MHLEN - hdrlen - max_linkhdr)) {
18569                 goto failed;
18570         }
18571         th->th_seq = htonl(rsm->r_start);
18572         th->th_ack = htonl(tp->rcv_nxt);
18573         /*
18574          * The PUSH bit should only be applied
18575          * if the full retransmission is made. If
18576          * we are sending less than this is the
18577          * left hand edge and should not have
18578          * the PUSH bit.
18579          */
18580         if ((rsm->r_flags & RACK_HAD_PUSH) &&
18581             (len == (rsm->r_end - rsm->r_start)))
18582                 flags |= TH_PUSH;
18583         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
18584         if (th->th_win == 0) {
18585                 tp->t_sndzerowin++;
18586                 tp->t_flags |= TF_RXWIN0SENT;
18587         } else
18588                 tp->t_flags &= ~TF_RXWIN0SENT;
18589         if (rsm->r_flags & RACK_TLP) {
18590                 /*
18591                  * TLP should not count in retran count, but
18592                  * in its own bin
18593                  */
18594                 counter_u64_add(rack_tlp_retran, 1);
18595                 counter_u64_add(rack_tlp_retran_bytes, len);
18596         } else {
18597                 tp->t_sndrexmitpack++;
18598                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18599                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18600         }
18601 #ifdef STATS
18602         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18603                                  len);
18604 #endif
18605         if (rsm->m == NULL)
18606                 goto failed;
18607         if (rsm->m &&
18608             ((rsm->orig_m_len != rsm->m->m_len) ||
18609              (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
18610                 /* Fix up the orig_m_len and possibly the mbuf offset */
18611                 rack_adjust_orig_mlen(rsm);
18612         }
18613         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
18614         if (len <= segsiz) {
18615                 /*
18616                  * Must have ran out of mbufs for the copy
18617                  * shorten it to no longer need tso. Lets
18618                  * not put on sendalot since we are low on
18619                  * mbufs.
18620                  */
18621                 tso = 0;
18622         }
18623         if ((m->m_next == NULL) || (len <= 0)){
18624                 goto failed;
18625         }
18626         if (udp) {
18627                 if (rack->r_is_v6)
18628                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
18629                 else
18630                         ulen = hdrlen + len - sizeof(struct ip);
18631                 udp->uh_ulen = htons(ulen);
18632         }
18633         m->m_pkthdr.rcvif = (struct ifnet *)0;
18634         if (TCPS_HAVERCVDSYN(tp->t_state) &&
18635             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18636                 int ect = tcp_ecn_output_established(tp, &flags, len, true);
18637                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18638                     (tp->t_flags2 & TF2_ECN_SND_ECE))
18639                     tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18640 #ifdef INET6
18641                 if (rack->r_is_v6) {
18642                     ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18643                     ip6->ip6_flow |= htonl(ect << 20);
18644                 }
18645                 else
18646 #endif
18647                 {
18648                     ip->ip_tos &= ~IPTOS_ECN_MASK;
18649                     ip->ip_tos |= ect;
18650                 }
18651         }
18652         if (rack->r_ctl.crte != NULL) {
18653                 /* See if we can send via the hw queue */
18654                 slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
18655                 /* If there is nothing in queue (no pacing time) we can send via the hw queue */
18656                 if (slot == 0)
18657                         ip_sendflag = 0;
18658         }
18659         tcp_set_flags(th, flags);
18660         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
18661 #ifdef INET6
18662         if (rack->r_is_v6) {
18663                 if (tp->t_port) {
18664                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18665                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18666                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18667                         th->th_sum = htons(0);
18668                         UDPSTAT_INC(udps_opackets);
18669                 } else {
18670                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18671                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18672                         th->th_sum = in6_cksum_pseudo(ip6,
18673                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18674                                                       0);
18675                 }
18676         }
18677 #endif
18678 #if defined(INET6) && defined(INET)
18679         else
18680 #endif
18681 #ifdef INET
18682         {
18683                 if (tp->t_port) {
18684                         m->m_pkthdr.csum_flags = CSUM_UDP;
18685                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18686                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18687                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18688                         th->th_sum = htons(0);
18689                         UDPSTAT_INC(udps_opackets);
18690                 } else {
18691                         m->m_pkthdr.csum_flags = CSUM_TCP;
18692                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18693                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
18694                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18695                                                                         IPPROTO_TCP + len + optlen));
18696                 }
18697                 /* IP version must be set here for ipv4/ipv6 checking later */
18698                 KASSERT(ip->ip_v == IPVERSION,
18699                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
18700         }
18701 #endif
18702         if (tso) {
18703                 /*
18704                  * Here we use segsiz since we have no added options besides
18705                  * any standard timestamp options (no DSACKs or SACKS are sent
18706                  * via either fast-path).
18707                  */
18708                 KASSERT(len > segsiz,
18709                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
18710                 m->m_pkthdr.csum_flags |= CSUM_TSO;
18711                 m->m_pkthdr.tso_segsz = segsiz;
18712         }
18713 #ifdef INET6
18714         if (rack->r_is_v6) {
18715                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
18716                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18717                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18718                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18719                 else
18720                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18721         }
18722 #endif
18723 #if defined(INET) && defined(INET6)
18724         else
18725 #endif
18726 #ifdef INET
18727         {
18728                 ip->ip_len = htons(m->m_pkthdr.len);
18729                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
18730                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18731                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18732                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18733                                 ip->ip_off |= htons(IP_DF);
18734                         }
18735                 } else {
18736                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18737                 }
18738         }
18739 #endif
18740         if (doing_tlp == 0) {
18741                 /* Set we retransmitted */
18742                 rack->rc_gp_saw_rec = 1;
18743         } else {
18744                 /* Its a TLP set ca or ss */
18745                 if (tp->snd_cwnd > tp->snd_ssthresh) {
18746                         /* Set we sent in CA */
18747                         rack->rc_gp_saw_ca = 1;
18748                 } else {
18749                         /* Set we sent in SS */
18750                         rack->rc_gp_saw_ss = 1;
18751                 }
18752         }
18753         /* Time to copy in our header */
18754         cpto = mtod(m, uint8_t *);
18755         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18756         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18757         if (optlen) {
18758                 bcopy(opt, th + 1, optlen);
18759                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18760         } else {
18761                 th->th_off = sizeof(struct tcphdr) >> 2;
18762         }
18763         if (tcp_bblogging_on(rack->rc_tp)) {
18764                 union tcp_log_stackspecific log;
18765
18766                 if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18767                         rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18768                         counter_u64_add(rack_collapsed_win_rxt, 1);
18769                         counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18770                 }
18771                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18772                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18773                 if (rack->rack_no_prr)
18774                         log.u_bbr.flex1 = 0;
18775                 else
18776                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18777                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18778                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18779                 log.u_bbr.flex4 = max_val;
18780                 /* Save off the early/late values */
18781                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18782                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18783                 log.u_bbr.bw_inuse = rack_get_bw(rack);
18784                 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
18785                 if (doing_tlp == 0)
18786                         log.u_bbr.flex8 = 1;
18787                 else
18788                         log.u_bbr.flex8 = 2;
18789                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18790                 log.u_bbr.flex7 = 55;
18791                 log.u_bbr.pkts_out = tp->t_maxseg;
18792                 log.u_bbr.timeStamp = cts;
18793                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18794                 if (rsm && (rsm->r_rtr_cnt > 0)) {
18795                         /*
18796                          * When we have a retransmit we want to log the
18797                          * burst at send and flight at send from before.
18798                          */
18799                         log.u_bbr.flex5 = rsm->r_fas;
18800                         log.u_bbr.bbr_substate = rsm->r_bas;
18801                 } else {
18802                         /*
18803                          * This is currently unlikely until we do the
18804                          * packet pair probes but I will add it for completeness.
18805                          */
18806                         log.u_bbr.flex5 = log.u_bbr.inflight;
18807                         log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
18808                 }
18809                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
18810                 log.u_bbr.delivered = 0;
18811                 log.u_bbr.rttProp = (uint64_t)rsm;
18812                 log.u_bbr.delRate = rsm->r_flags;
18813                 log.u_bbr.delRate <<= 31;
18814                 log.u_bbr.delRate |= rack->r_must_retran;
18815                 log.u_bbr.delRate <<= 1;
18816                 log.u_bbr.delRate |= 1;
18817                 log.u_bbr.pkt_epoch = __LINE__;
18818                 lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
18819                                      len, &log, false, NULL, __func__, __LINE__, tv);
18820         } else
18821                 lgb = NULL;
18822         if ((rack->r_ctl.crte != NULL) &&
18823             tcp_bblogging_on(tp)) {
18824                 rack_log_queue_level(tp, rack, len, tv, cts);
18825         }
18826 #ifdef INET6
18827         if (rack->r_is_v6) {
18828                 error = ip6_output(m, NULL,
18829                                    &inp->inp_route6,
18830                                    ip_sendflag, NULL, NULL, inp);
18831         }
18832         else
18833 #endif
18834 #ifdef INET
18835         {
18836                 error = ip_output(m, NULL,
18837                                   &inp->inp_route,
18838                                   ip_sendflag, 0, inp);
18839         }
18840 #endif
18841         m = NULL;
18842         if (lgb) {
18843                 lgb->tlb_errno = error;
18844                 lgb = NULL;
18845         }
18846         if (error) {
18847                 goto failed;
18848         } else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
18849                 rack->rc_hw_nobuf = 0;
18850                 rack->r_ctl.rc_agg_delayed = 0;
18851                 rack->r_early = 0;
18852                 rack->r_late = 0;
18853                 rack->r_ctl.rc_agg_early = 0;
18854         }
18855
18856         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
18857                         rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
18858         if (doing_tlp) {
18859                 rack->rc_tlp_in_progress = 1;
18860                 rack->r_ctl.rc_tlp_cnt_out++;
18861         }
18862         if (error == 0) {
18863                 counter_u64_add(rack_total_bytes, len);
18864                 tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
18865                 if (doing_tlp) {
18866                         rack->rc_last_sent_tlp_past_cumack = 0;
18867                         rack->rc_last_sent_tlp_seq_valid = 1;
18868                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18869                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18870                 }
18871         }
18872         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18873         rack->forced_ack = 0;   /* If we send something zap the FA flag */
18874         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18875                 rack->r_ctl.retran_during_recovery += len;
18876         {
18877                 int idx;
18878
18879                 idx = (len / segsiz) + 3;
18880                 if (idx >= TCP_MSS_ACCT_ATIMER)
18881                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18882                 else
18883                         counter_u64_add(rack_out_size[idx], 1);
18884         }
18885         if (tp->t_rtttime == 0) {
18886                 tp->t_rtttime = ticks;
18887                 tp->t_rtseq = startseq;
18888                 KMOD_TCPSTAT_INC(tcps_segstimed);
18889         }
18890         counter_u64_add(rack_fto_rsm_send, 1);
18891         if (error && (error == ENOBUFS)) {
18892                 if (rack->r_ctl.crte != NULL) {
18893                         tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
18894                         if (tcp_bblogging_on(rack->rc_tp))
18895                                 rack_log_queue_level(tp, rack, len, tv, cts);
18896                 } else
18897                         tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
18898                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18899                 if (rack->rc_enobuf < 0x7f)
18900                         rack->rc_enobuf++;
18901                 if (slot < (10 * HPTS_USEC_IN_MSEC))
18902                         slot = 10 * HPTS_USEC_IN_MSEC;
18903                 if (rack->r_ctl.crte != NULL) {
18904                         counter_u64_add(rack_saw_enobuf_hw, 1);
18905                         tcp_rl_log_enobuf(rack->r_ctl.crte);
18906                 }
18907                 counter_u64_add(rack_saw_enobuf, 1);
18908         } else
18909                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
18910         if ((slot == 0) ||
18911             (rack->rc_always_pace == 0) ||
18912             (rack->r_rr_config == 1)) {
18913                 /*
18914                  * We have no pacing set or we
18915                  * are using old-style rack or
18916                  * we are overridden to use the old 1ms pacing.
18917                  */
18918                 slot = rack->r_ctl.rc_min_to;
18919         }
18920         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
18921 #ifdef TCP_ACCOUNTING
18922         crtsc = get_cyclecount();
18923         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18924                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
18925         }
18926         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18927                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
18928         }
18929         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18930                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
18931         }
18932         sched_unpin();
18933 #endif
18934         return (0);
18935 failed:
18936         if (m)
18937                 m_free(m);
18938         return (-1);
18939 }
18940
18941 static void
18942 rack_sndbuf_autoscale(struct tcp_rack *rack)
18943 {
18944         /*
18945          * Automatic sizing of send socket buffer.  Often the send buffer
18946          * size is not optimally adjusted to the actual network conditions
18947          * at hand (delay bandwidth product).  Setting the buffer size too
18948          * small limits throughput on links with high bandwidth and high
18949          * delay (eg. trans-continental/oceanic links).  Setting the
18950          * buffer size too big consumes too much real kernel memory,
18951          * especially with many connections on busy servers.
18952          *
18953          * The criteria to step up the send buffer one notch are:
18954          *  1. receive window of remote host is larger than send buffer
18955          *     (with a fudge factor of 5/4th);
18956          *  2. send buffer is filled to 7/8th with data (so we actually
18957          *     have data to make use of it);
18958          *  3. send buffer fill has not hit maximal automatic size;
18959          *  4. our send window (slow start and cogestion controlled) is
18960          *     larger than sent but unacknowledged data in send buffer.
18961          *
18962          * Note that the rack version moves things much faster since
18963          * we want to avoid hitting cache lines in the rack_fast_output()
18964          * path so this is called much less often and thus moves
18965          * the SB forward by a percentage.
18966          */
18967         struct socket *so;
18968         struct tcpcb *tp;
18969         uint32_t sendwin, scaleup;
18970
18971         tp = rack->rc_tp;
18972         so = rack->rc_inp->inp_socket;
18973         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
18974         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
18975                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
18976                     sbused(&so->so_snd) >=
18977                     (so->so_snd.sb_hiwat / 8 * 7) &&
18978                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
18979                     sendwin >= (sbused(&so->so_snd) -
18980                     (tp->snd_nxt - tp->snd_una))) {
18981                         if (rack_autosndbuf_inc)
18982                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
18983                         else
18984                                 scaleup = V_tcp_autosndbuf_inc;
18985                         if (scaleup < V_tcp_autosndbuf_inc)
18986                                 scaleup = V_tcp_autosndbuf_inc;
18987                         scaleup += so->so_snd.sb_hiwat;
18988                         if (scaleup > V_tcp_autosndbuf_max)
18989                                 scaleup = V_tcp_autosndbuf_max;
18990                         if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
18991                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
18992                 }
18993         }
18994 }
18995
18996 static int
18997 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
18998                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
18999 {
19000         /*
19001          * Enter to do fast output. We are given that the sched_pin is
19002          * in place (if accounting is compiled in) and the cycle count taken
19003          * at entry is in place in ts_val. The idea here is that
19004          * we know how many more bytes needs to be sent (presumably either
19005          * during pacing or to fill the cwnd and that was greater than
19006          * the max-burst). We have how much to send and all the info we
19007          * need to just send.
19008          */
19009 #ifdef INET
19010         struct ip *ip = NULL;
19011 #endif
19012         struct udphdr *udp = NULL;
19013         struct tcphdr *th = NULL;
19014         struct mbuf *m, *s_mb;
19015         struct inpcb *inp;
19016         uint8_t *cpto;
19017         struct tcp_log_buffer *lgb;
19018 #ifdef TCP_ACCOUNTING
19019         uint64_t crtsc;
19020 #endif
19021         struct tcpopt to;
19022         u_char opt[TCP_MAXOLEN];
19023         uint32_t hdrlen, optlen;
19024 #ifdef TCP_ACCOUNTING
19025         int cnt_thru = 1;
19026 #endif
19027         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19028         uint16_t flags;
19029         uint32_t s_soff;
19030         uint32_t if_hw_tsomaxsegcount = 0, startseq;
19031         uint32_t if_hw_tsomaxsegsize;
19032         uint16_t add_flag = RACK_SENT_FP;
19033 #ifdef INET6
19034         struct ip6_hdr *ip6 = NULL;
19035
19036         if (rack->r_is_v6) {
19037                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19038                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19039         } else
19040 #endif                          /* INET6 */
19041         {
19042 #ifdef INET
19043                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19044                 hdrlen = sizeof(struct tcpiphdr);
19045 #endif
19046         }
19047         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19048                 m = NULL;
19049                 goto failed;
19050         }
19051         startseq = tp->snd_max;
19052         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19053         inp = rack->rc_inp;
19054         len = rack->r_ctl.fsb.left_to_send;
19055         to.to_flags = 0;
19056         flags = rack->r_ctl.fsb.tcp_flags;
19057         if (tp->t_flags & TF_RCVD_TSTMP) {
19058                 to.to_tsval = ms_cts + tp->ts_offset;
19059                 to.to_tsecr = tp->ts_recent;
19060                 to.to_flags = TOF_TS;
19061         }
19062         optlen = tcp_addoptions(&to, opt);
19063         hdrlen += optlen;
19064         udp = rack->r_ctl.fsb.udp;
19065         if (udp)
19066                 hdrlen += sizeof(struct udphdr);
19067         if (rack->r_ctl.rc_pace_max_segs)
19068                 max_val = rack->r_ctl.rc_pace_max_segs;
19069         else if (rack->rc_user_set_max_segs)
19070                 max_val = rack->rc_user_set_max_segs * segsiz;
19071         else
19072                 max_val = len;
19073         if ((tp->t_flags & TF_TSO) &&
19074             V_tcp_do_tso &&
19075             (len > segsiz) &&
19076             (tp->t_port == 0))
19077                 tso = 1;
19078 again:
19079 #ifdef INET6
19080         if (MHLEN < hdrlen + max_linkhdr)
19081                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19082         else
19083 #endif
19084                 m = m_gethdr(M_NOWAIT, MT_DATA);
19085         if (m == NULL)
19086                 goto failed;
19087         m->m_data += max_linkhdr;
19088         m->m_len = hdrlen;
19089         th = rack->r_ctl.fsb.th;
19090         /* Establish the len to send */
19091         if (len > max_val)
19092                 len = max_val;
19093         if ((tso) && (len + optlen > segsiz)) {
19094                 uint32_t if_hw_tsomax;
19095                 int32_t max_len;
19096
19097                 /* extract TSO information */
19098                 if_hw_tsomax = tp->t_tsomax;
19099                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19100                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19101                 /*
19102                  * Check if we should limit by maximum payload
19103                  * length:
19104                  */
19105                 if (if_hw_tsomax != 0) {
19106                         /* compute maximum TSO length */
19107                         max_len = (if_hw_tsomax - hdrlen -
19108                                    max_linkhdr);
19109                         if (max_len <= 0) {
19110                                 goto failed;
19111                         } else if (len > max_len) {
19112                                 len = max_len;
19113                         }
19114                 }
19115                 if (len <= segsiz) {
19116                         /*
19117                          * In case there are too many small fragments don't
19118                          * use TSO:
19119                          */
19120                         tso = 0;
19121                 }
19122         } else {
19123                 tso = 0;
19124         }
19125         if ((tso == 0) && (len > segsiz))
19126                 len = segsiz;
19127         (void)tcp_get_usecs(tv);
19128         if ((len == 0) ||
19129             (len <= MHLEN - hdrlen - max_linkhdr)) {
19130                 goto failed;
19131         }
19132         sb_offset = tp->snd_max - tp->snd_una;
19133         th->th_seq = htonl(tp->snd_max);
19134         th->th_ack = htonl(tp->rcv_nxt);
19135         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19136         if (th->th_win == 0) {
19137                 tp->t_sndzerowin++;
19138                 tp->t_flags |= TF_RXWIN0SENT;
19139         } else
19140                 tp->t_flags &= ~TF_RXWIN0SENT;
19141         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
19142         KMOD_TCPSTAT_INC(tcps_sndpack);
19143         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
19144 #ifdef STATS
19145         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
19146                                  len);
19147 #endif
19148         if (rack->r_ctl.fsb.m == NULL)
19149                 goto failed;
19150
19151         /* s_mb and s_soff are saved for rack_log_output */
19152         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
19153                                     &s_mb, &s_soff);
19154         if (len <= segsiz) {
19155                 /*
19156                  * Must have ran out of mbufs for the copy
19157                  * shorten it to no longer need tso. Lets
19158                  * not put on sendalot since we are low on
19159                  * mbufs.
19160                  */
19161                 tso = 0;
19162         }
19163         if (rack->r_ctl.fsb.rfo_apply_push &&
19164             (len == rack->r_ctl.fsb.left_to_send)) {
19165                 tcp_set_flags(th, flags | TH_PUSH);
19166                 add_flag |= RACK_HAD_PUSH;
19167         }
19168         if ((m->m_next == NULL) || (len <= 0)){
19169                 goto failed;
19170         }
19171         if (udp) {
19172                 if (rack->r_is_v6)
19173                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
19174                 else
19175                         ulen = hdrlen + len - sizeof(struct ip);
19176                 udp->uh_ulen = htons(ulen);
19177         }
19178         m->m_pkthdr.rcvif = (struct ifnet *)0;
19179         if (TCPS_HAVERCVDSYN(tp->t_state) &&
19180             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19181                 int ect = tcp_ecn_output_established(tp, &flags, len, false);
19182                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19183                     (tp->t_flags2 & TF2_ECN_SND_ECE))
19184                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19185 #ifdef INET6
19186                 if (rack->r_is_v6) {
19187                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19188                         ip6->ip6_flow |= htonl(ect << 20);
19189                 }
19190                 else
19191 #endif
19192                 {
19193 #ifdef INET
19194                         ip->ip_tos &= ~IPTOS_ECN_MASK;
19195                         ip->ip_tos |= ect;
19196 #endif
19197                 }
19198         }
19199         tcp_set_flags(th, flags);
19200         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
19201 #ifdef INET6
19202         if (rack->r_is_v6) {
19203                 if (tp->t_port) {
19204                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19205                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19206                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19207                         th->th_sum = htons(0);
19208                         UDPSTAT_INC(udps_opackets);
19209                 } else {
19210                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19211                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19212                         th->th_sum = in6_cksum_pseudo(ip6,
19213                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19214                                                       0);
19215                 }
19216         }
19217 #endif
19218 #if defined(INET6) && defined(INET)
19219         else
19220 #endif
19221 #ifdef INET
19222         {
19223                 if (tp->t_port) {
19224                         m->m_pkthdr.csum_flags = CSUM_UDP;
19225                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19226                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19227                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19228                         th->th_sum = htons(0);
19229                         UDPSTAT_INC(udps_opackets);
19230                 } else {
19231                         m->m_pkthdr.csum_flags = CSUM_TCP;
19232                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19233                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
19234                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19235                                                                         IPPROTO_TCP + len + optlen));
19236                 }
19237                 /* IP version must be set here for ipv4/ipv6 checking later */
19238                 KASSERT(ip->ip_v == IPVERSION,
19239                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
19240         }
19241 #endif
19242         if (tso) {
19243                 /*
19244                  * Here we use segsiz since we have no added options besides
19245                  * any standard timestamp options (no DSACKs or SACKS are sent
19246                  * via either fast-path).
19247                  */
19248                 KASSERT(len > segsiz,
19249                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
19250                 m->m_pkthdr.csum_flags |= CSUM_TSO;
19251                 m->m_pkthdr.tso_segsz = segsiz;
19252         }
19253 #ifdef INET6
19254         if (rack->r_is_v6) {
19255                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19256                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19257                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19258                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19259                 else
19260                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19261         }
19262 #endif
19263 #if defined(INET) && defined(INET6)
19264         else
19265 #endif
19266 #ifdef INET
19267         {
19268                 ip->ip_len = htons(m->m_pkthdr.len);
19269                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19270                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19271                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19272                         if (tp->t_port == 0 || len < V_tcp_minmss) {
19273                                 ip->ip_off |= htons(IP_DF);
19274                         }
19275                 } else {
19276                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19277                 }
19278         }
19279 #endif
19280         if (tp->snd_cwnd > tp->snd_ssthresh) {
19281                 /* Set we sent in CA */
19282                 rack->rc_gp_saw_ca = 1;
19283         } else {
19284                 /* Set we sent in SS */
19285                 rack->rc_gp_saw_ss = 1;
19286         }
19287         /* Time to copy in our header */
19288         cpto = mtod(m, uint8_t *);
19289         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19290         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19291         if (optlen) {
19292                 bcopy(opt, th + 1, optlen);
19293                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19294         } else {
19295                 th->th_off = sizeof(struct tcphdr) >> 2;
19296         }
19297         if ((rack->r_ctl.crte != NULL) &&
19298             tcp_bblogging_on(tp)) {
19299                 rack_log_queue_level(tp, rack, len, tv, cts);
19300         }
19301         if (tcp_bblogging_on(rack->rc_tp)) {
19302                 union tcp_log_stackspecific log;
19303
19304                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19305                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
19306                 if (rack->rack_no_prr)
19307                         log.u_bbr.flex1 = 0;
19308                 else
19309                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19310                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19311                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19312                 log.u_bbr.flex4 = max_val;
19313                 /* Save off the early/late values */
19314                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19315                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19316                 log.u_bbr.bw_inuse = rack_get_bw(rack);
19317                 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19318                 log.u_bbr.flex8 = 0;
19319                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19320                 log.u_bbr.flex7 = 44;
19321                 log.u_bbr.pkts_out = tp->t_maxseg;
19322                 log.u_bbr.timeStamp = cts;
19323                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19324                 log.u_bbr.flex5 = log.u_bbr.inflight;
19325                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19326                 log.u_bbr.delivered = 0;
19327                 log.u_bbr.rttProp = 0;
19328                 log.u_bbr.delRate = rack->r_must_retran;
19329                 log.u_bbr.delRate <<= 1;
19330                 log.u_bbr.pkt_epoch = __LINE__;
19331                 /* For fast output no retrans so just inflight and how many mss we send */
19332                 log.u_bbr.flex5 = log.u_bbr.inflight;
19333                 log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19334                 lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19335                                      len, &log, false, NULL, __func__, __LINE__, tv);
19336         } else
19337                 lgb = NULL;
19338 #ifdef INET6
19339         if (rack->r_is_v6) {
19340                 error = ip6_output(m, NULL,
19341                                    &inp->inp_route6,
19342                                    0, NULL, NULL, inp);
19343         }
19344 #endif
19345 #if defined(INET) && defined(INET6)
19346         else
19347 #endif
19348 #ifdef INET
19349         {
19350                 error = ip_output(m, NULL,
19351                                   &inp->inp_route,
19352                                   0, 0, inp);
19353         }
19354 #endif
19355         if (lgb) {
19356                 lgb->tlb_errno = error;
19357                 lgb = NULL;
19358         }
19359         if (error) {
19360                 *send_err = error;
19361                 m = NULL;
19362                 goto failed;
19363         } else if (rack->rc_hw_nobuf) {
19364                 rack->rc_hw_nobuf = 0;
19365                 rack->r_ctl.rc_agg_delayed = 0;
19366                 rack->r_early = 0;
19367                 rack->r_late = 0;
19368                 rack->r_ctl.rc_agg_early = 0;
19369         }
19370         if ((error == 0) && (rack->lt_bw_up == 0)) {
19371                 /* Unlikely */
19372                 rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(tv);
19373                 rack->r_ctl.lt_seq = tp->snd_una;
19374                 rack->lt_bw_up = 1;
19375         }
19376         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
19377                         NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
19378         m = NULL;
19379         if (tp->snd_una == tp->snd_max) {
19380                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
19381                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19382                 tp->t_acktime = ticks;
19383         }
19384         counter_u64_add(rack_total_bytes, len);
19385         tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
19386
19387         rack->forced_ack = 0;   /* If we send something zap the FA flag */
19388         tot_len += len;
19389         if ((tp->t_flags & TF_GPUTINPROG) == 0)
19390                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
19391         tp->snd_max += len;
19392         tp->snd_nxt = tp->snd_max;
19393         if (rack->rc_new_rnd_needed) {
19394                 /*
19395                  * Update the rnd to start ticking not
19396                  * that from a time perspective all of
19397                  * the preceding idle time is "in the round"
19398                  */
19399                 rack->rc_new_rnd_needed = 0;
19400                 rack->r_ctl.roundends = tp->snd_max;
19401         }
19402         {
19403                 int idx;
19404
19405                 idx = (len / segsiz) + 3;
19406                 if (idx >= TCP_MSS_ACCT_ATIMER)
19407                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19408                 else
19409                         counter_u64_add(rack_out_size[idx], 1);
19410         }
19411         if (len <= rack->r_ctl.fsb.left_to_send)
19412                 rack->r_ctl.fsb.left_to_send -= len;
19413         else
19414                 rack->r_ctl.fsb.left_to_send = 0;
19415         if (rack->r_ctl.fsb.left_to_send < segsiz) {
19416                 rack->r_fast_output = 0;
19417                 rack->r_ctl.fsb.left_to_send = 0;
19418                 /* At the end of fast_output scale up the sb */
19419                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
19420                 rack_sndbuf_autoscale(rack);
19421                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
19422         }
19423         if (tp->t_rtttime == 0) {
19424                 tp->t_rtttime = ticks;
19425                 tp->t_rtseq = startseq;
19426                 KMOD_TCPSTAT_INC(tcps_segstimed);
19427         }
19428         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
19429             (max_val > len) &&
19430             (tso == 0)) {
19431                 max_val -= len;
19432                 len = segsiz;
19433                 th = rack->r_ctl.fsb.th;
19434 #ifdef TCP_ACCOUNTING
19435                 cnt_thru++;
19436 #endif
19437                 goto again;
19438         }
19439         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19440         counter_u64_add(rack_fto_send, 1);
19441         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
19442         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
19443 #ifdef TCP_ACCOUNTING
19444         crtsc = get_cyclecount();
19445         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19446                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19447         }
19448         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19449                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19450         }
19451         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19452                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
19453         }
19454         sched_unpin();
19455 #endif
19456         return (0);
19457 failed:
19458         if (m)
19459                 m_free(m);
19460         rack->r_fast_output = 0;
19461         return (-1);
19462 }
19463
19464 static inline void
19465 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
19466                        struct sockbuf *sb,
19467                        int len, int orig_len, int segsiz, uint32_t pace_max_seg,
19468                        bool hw_tls,
19469                        uint16_t flags)
19470 {
19471         rack->r_fast_output = 1;
19472         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19473         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19474         rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
19475         rack->r_ctl.fsb.tcp_flags = flags;
19476         rack->r_ctl.fsb.left_to_send = orig_len - len;
19477         if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
19478                 /* Less than a full sized pace, lets not  */
19479                 rack->r_fast_output = 0;
19480                 return;
19481         } else {
19482                 /* Round down to the nearest pace_max_seg */
19483                 rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
19484         }
19485         if (hw_tls)
19486                 rack->r_ctl.fsb.hw_tls = 1;
19487         else
19488                 rack->r_ctl.fsb.hw_tls = 0;
19489         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19490                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19491                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19492                  (tp->snd_max - tp->snd_una)));
19493         if (rack->r_ctl.fsb.left_to_send < segsiz)
19494                 rack->r_fast_output = 0;
19495         else {
19496                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19497                         rack->r_ctl.fsb.rfo_apply_push = 1;
19498                 else
19499                         rack->r_ctl.fsb.rfo_apply_push = 0;
19500         }
19501 }
19502
19503 static uint32_t
19504 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
19505 {
19506         uint64_t min_time;
19507         uint32_t maxlen;
19508
19509         min_time = (uint64_t)get_hpts_min_sleep_time();
19510         maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
19511         maxlen = roundup(maxlen, segsiz);
19512         return (maxlen);
19513 }
19514
19515 static struct rack_sendmap *
19516 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
19517 {
19518         struct rack_sendmap *rsm = NULL;
19519         int thresh;
19520
19521 restart:
19522         rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
19523         if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
19524                 /* Nothing, strange turn off validity  */
19525                 rack->r_collapse_point_valid = 0;
19526                 return (NULL);
19527         }
19528         /* Can we send it yet? */
19529         if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
19530                 /*
19531                  * Receiver window has not grown enough for
19532                  * the segment to be put on the wire.
19533                  */
19534                 return (NULL);
19535         }
19536         if (rsm->r_flags & RACK_ACKED) {
19537                 /*
19538                  * It has been sacked, lets move to the
19539                  * next one if possible.
19540                  */
19541                 rack->r_ctl.last_collapse_point = rsm->r_end;
19542                 /* Are we done? */
19543                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19544                             rack->r_ctl.high_collapse_point)) {
19545                         rack->r_collapse_point_valid = 0;
19546                         return (NULL);
19547                 }
19548                 goto restart;
19549         }
19550         /* Now has it been long enough ? */
19551         thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
19552         if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
19553                 rack_log_collapse(rack, rsm->r_start,
19554                                   (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19555                                   thresh, __LINE__, 6, rsm->r_flags, rsm);
19556                 return (rsm);
19557         }
19558         /* Not enough time */
19559         rack_log_collapse(rack, rsm->r_start,
19560                           (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19561                           thresh, __LINE__, 7, rsm->r_flags, rsm);
19562         return (NULL);
19563 }
19564
19565 static inline void
19566 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
19567 {
19568         if ((rack->full_size_rxt == 0) &&
19569             (rack->shape_rxt_to_pacing_min == 0) &&
19570             (*len >= segsiz)) {
19571                 *len = segsiz;
19572         } else if (rack->shape_rxt_to_pacing_min &&
19573                  rack->gp_ready) {
19574                 /* We use pacing min as shaping len req */
19575                 uint32_t maxlen;
19576
19577                 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
19578                 if (*len > maxlen)
19579                         *len = maxlen;
19580         } else {
19581                 /*
19582                  * The else is full_size_rxt is on so send it all
19583                  * note we do need to check this for exceeding
19584                  * our max segment size due to the fact that
19585                  * we do sometimes merge chunks together i.e.
19586                  * we cannot just assume that we will never have
19587                  * a chunk greater than pace_max_seg
19588                  */
19589                 if (*len > pace_max_seg)
19590                         *len = pace_max_seg;
19591         }
19592 }
19593
19594 static int
19595 rack_output(struct tcpcb *tp)
19596 {
19597         struct socket *so;
19598         uint32_t recwin;
19599         uint32_t sb_offset, s_moff = 0;
19600         int32_t len, error = 0;
19601         uint16_t flags;
19602         struct mbuf *m, *s_mb = NULL;
19603         struct mbuf *mb;
19604         uint32_t if_hw_tsomaxsegcount = 0;
19605         uint32_t if_hw_tsomaxsegsize;
19606         int32_t segsiz, minseg;
19607         long tot_len_this_send = 0;
19608 #ifdef INET
19609         struct ip *ip = NULL;
19610 #endif
19611         struct udphdr *udp = NULL;
19612         struct tcp_rack *rack;
19613         struct tcphdr *th;
19614         uint8_t pass = 0;
19615         uint8_t mark = 0;
19616         uint8_t check_done = 0;
19617         uint8_t wanted_cookie = 0;
19618         u_char opt[TCP_MAXOLEN];
19619         unsigned ipoptlen, optlen, hdrlen, ulen=0;
19620         uint32_t rack_seq;
19621
19622 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
19623         unsigned ipsec_optlen = 0;
19624
19625 #endif
19626         int32_t idle, sendalot;
19627         int32_t sub_from_prr = 0;
19628         volatile int32_t sack_rxmit;
19629         struct rack_sendmap *rsm = NULL;
19630         int32_t tso, mtu;
19631         struct tcpopt to;
19632         int32_t slot = 0;
19633         int32_t sup_rack = 0;
19634         uint32_t cts, ms_cts, delayed, early;
19635         uint16_t add_flag = RACK_SENT_SP;
19636         /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
19637         uint8_t hpts_calling,  doing_tlp = 0;
19638         uint32_t cwnd_to_use, pace_max_seg;
19639         int32_t do_a_prefetch = 0;
19640         int32_t prefetch_rsm = 0;
19641         int32_t orig_len = 0;
19642         struct timeval tv;
19643         int32_t prefetch_so_done = 0;
19644         struct tcp_log_buffer *lgb;
19645         struct inpcb *inp = tptoinpcb(tp);
19646         struct sockbuf *sb;
19647         uint64_t ts_val = 0;
19648 #ifdef TCP_ACCOUNTING
19649         uint64_t crtsc;
19650 #endif
19651 #ifdef INET6
19652         struct ip6_hdr *ip6 = NULL;
19653         int32_t isipv6;
19654 #endif
19655         bool hw_tls = false;
19656
19657         NET_EPOCH_ASSERT();
19658         INP_WLOCK_ASSERT(inp);
19659
19660         /* setup and take the cache hits here */
19661         rack = (struct tcp_rack *)tp->t_fb_ptr;
19662 #ifdef TCP_ACCOUNTING
19663         sched_pin();
19664         ts_val = get_cyclecount();
19665 #endif
19666         hpts_calling = inp->inp_hpts_calls;
19667         rack->rc_inp->inp_hpts_calls = 0;
19668 #ifdef TCP_OFFLOAD
19669         if (tp->t_flags & TF_TOE) {
19670 #ifdef TCP_ACCOUNTING
19671                 sched_unpin();
19672 #endif
19673                 return (tcp_offload_output(tp));
19674         }
19675 #endif
19676         if (rack->rack_deferred_inited == 0) {
19677                 /*
19678                  * If we are the connecting socket we will
19679                  * hit rack_init() when no sequence numbers
19680                  * are setup. This makes it so we must defer
19681                  * some initialization. Call that now.
19682                  */
19683                 rack_deferred_init(tp, rack);
19684         }
19685         /*
19686          * For TFO connections in SYN_RECEIVED, only allow the initial
19687          * SYN|ACK and those sent by the retransmit timer.
19688          */
19689         if (IS_FASTOPEN(tp->t_flags) &&
19690             (tp->t_state == TCPS_SYN_RECEIVED) &&
19691             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
19692             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
19693 #ifdef TCP_ACCOUNTING
19694                 sched_unpin();
19695 #endif
19696                 return (0);
19697         }
19698 #ifdef INET6
19699         if (rack->r_state) {
19700                 /* Use the cache line loaded if possible */
19701                 isipv6 = rack->r_is_v6;
19702         } else {
19703                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
19704         }
19705 #endif
19706         early = 0;
19707         cts = tcp_get_usecs(&tv);
19708         ms_cts = tcp_tv_to_mssectick(&tv);
19709         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
19710             tcp_in_hpts(rack->rc_inp)) {
19711                 /*
19712                  * We are on the hpts for some timer but not hptsi output.
19713                  * Remove from the hpts unconditionally.
19714                  */
19715                 rack_timer_cancel(tp, rack, cts, __LINE__);
19716         }
19717         /* Are we pacing and late? */
19718         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19719             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
19720                 /* We are delayed */
19721                 delayed = cts - rack->r_ctl.rc_last_output_to;
19722         } else {
19723                 delayed = 0;
19724         }
19725         /* Do the timers, which may override the pacer */
19726         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
19727                 int retval;
19728
19729                 retval = rack_process_timers(tp, rack, cts, hpts_calling,
19730                                              &doing_tlp);
19731                 if (retval != 0) {
19732                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
19733 #ifdef TCP_ACCOUNTING
19734                         sched_unpin();
19735 #endif
19736                         /*
19737                          * If timers want tcp_drop(), then pass error out,
19738                          * otherwise suppress it.
19739                          */
19740                         return (retval < 0 ? retval : 0);
19741                 }
19742         }
19743         if (rack->rc_in_persist) {
19744                 if (tcp_in_hpts(rack->rc_inp) == 0) {
19745                         /* Timer is not running */
19746                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19747                 }
19748 #ifdef TCP_ACCOUNTING
19749                 sched_unpin();
19750 #endif
19751                 return (0);
19752         }
19753         if ((rack->rc_ack_required == 1) &&
19754             (rack->r_timer_override == 0)){
19755                 /* A timeout occurred and no ack has arrived */
19756                 if (tcp_in_hpts(rack->rc_inp) == 0) {
19757                         /* Timer is not running */
19758                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19759                 }
19760 #ifdef TCP_ACCOUNTING
19761                 sched_unpin();
19762 #endif
19763                 return (0);
19764         }
19765         if ((rack->r_timer_override) ||
19766             (rack->rc_ack_can_sendout_data) ||
19767             (delayed) ||
19768             (tp->t_state < TCPS_ESTABLISHED)) {
19769                 rack->rc_ack_can_sendout_data = 0;
19770                 if (tcp_in_hpts(rack->rc_inp))
19771                         tcp_hpts_remove(rack->rc_inp);
19772         } else if (tcp_in_hpts(rack->rc_inp)) {
19773                 /*
19774                  * On the hpts you can't pass even if ACKNOW is on, we will
19775                  * when the hpts fires.
19776                  */
19777 #ifdef TCP_ACCOUNTING
19778                 crtsc = get_cyclecount();
19779                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19780                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
19781                 }
19782                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19783                         tp->tcp_cnt_counters[SND_BLOCKED]++;
19784                 }
19785                 sched_unpin();
19786 #endif
19787                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
19788                 return (0);
19789         }
19790         /* Finish out both pacing early and late accounting */
19791         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19792             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
19793                 early = rack->r_ctl.rc_last_output_to - cts;
19794         } else
19795                 early = 0;
19796         if (delayed) {
19797                 rack->r_ctl.rc_agg_delayed += delayed;
19798                 rack->r_late = 1;
19799         } else if (early) {
19800                 rack->r_ctl.rc_agg_early += early;
19801                 rack->r_early = 1;
19802         }
19803         /* Now that early/late accounting is done turn off the flag */
19804         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
19805         rack->r_wanted_output = 0;
19806         rack->r_timer_override = 0;
19807         if ((tp->t_state != rack->r_state) &&
19808             TCPS_HAVEESTABLISHED(tp->t_state)) {
19809                 rack_set_state(tp, rack);
19810         }
19811         if ((rack->r_fast_output) &&
19812             (doing_tlp == 0) &&
19813             (tp->rcv_numsacks == 0)) {
19814                 int ret;
19815
19816                 error = 0;
19817                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19818                 if (ret >= 0)
19819                         return(ret);
19820                 else if (error) {
19821                         inp = rack->rc_inp;
19822                         so = inp->inp_socket;
19823                         sb = &so->so_snd;
19824                         goto nomore;
19825                 }
19826         }
19827         inp = rack->rc_inp;
19828         /*
19829          * For TFO connections in SYN_SENT or SYN_RECEIVED,
19830          * only allow the initial SYN or SYN|ACK and those sent
19831          * by the retransmit timer.
19832          */
19833         if (IS_FASTOPEN(tp->t_flags) &&
19834             ((tp->t_state == TCPS_SYN_RECEIVED) ||
19835              (tp->t_state == TCPS_SYN_SENT)) &&
19836             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
19837             (tp->t_rxtshift == 0)) {              /* not a retransmit */
19838                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19839                 so = inp->inp_socket;
19840                 sb = &so->so_snd;
19841                 goto just_return_nolock;
19842         }
19843         /*
19844          * Determine length of data that should be transmitted, and flags
19845          * that will be used. If there is some data or critical controls
19846          * (SYN, RST) to send, then transmit; otherwise, investigate
19847          * further.
19848          */
19849         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
19850         if (tp->t_idle_reduce) {
19851                 if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
19852                         rack_cc_after_idle(rack, tp);
19853         }
19854         tp->t_flags &= ~TF_LASTIDLE;
19855         if (idle) {
19856                 if (tp->t_flags & TF_MORETOCOME) {
19857                         tp->t_flags |= TF_LASTIDLE;
19858                         idle = 0;
19859                 }
19860         }
19861         if ((tp->snd_una == tp->snd_max) &&
19862             rack->r_ctl.rc_went_idle_time &&
19863             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
19864                 idle = cts - rack->r_ctl.rc_went_idle_time;
19865                 if (idle > rack_min_probertt_hold) {
19866                         /* Count as a probe rtt */
19867                         if (rack->in_probe_rtt == 0) {
19868                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
19869                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
19870                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
19871                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
19872                         } else {
19873                                 rack_exit_probertt(rack, cts);
19874                         }
19875                 }
19876                 idle = 0;
19877         }
19878         if (rack_use_fsb &&
19879             (rack->r_ctl.fsb.tcp_ip_hdr) &&
19880             (rack->r_fsb_inited == 0) &&
19881             (rack->r_state != TCPS_CLOSED))
19882                 rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
19883 again:
19884         /*
19885          * If we've recently taken a timeout, snd_max will be greater than
19886          * snd_nxt.  There may be SACK information that allows us to avoid
19887          * resending already delivered data.  Adjust snd_nxt accordingly.
19888          */
19889         sendalot = 0;
19890         cts = tcp_get_usecs(&tv);
19891         ms_cts = tcp_tv_to_mssectick(&tv);
19892         tso = 0;
19893         mtu = 0;
19894         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19895         minseg = segsiz;
19896         if (rack->r_ctl.rc_pace_max_segs == 0)
19897                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
19898         else
19899                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
19900         sb_offset = tp->snd_max - tp->snd_una;
19901         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19902         flags = tcp_outflags[tp->t_state];
19903         while (rack->rc_free_cnt < rack_free_cache) {
19904                 rsm = rack_alloc(rack);
19905                 if (rsm == NULL) {
19906                         if (hpts_calling)
19907                                 /* Retry in a ms */
19908                                 slot = (1 * HPTS_USEC_IN_MSEC);
19909                         so = inp->inp_socket;
19910                         sb = &so->so_snd;
19911                         goto just_return_nolock;
19912                 }
19913                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
19914                 rack->rc_free_cnt++;
19915                 rsm = NULL;
19916         }
19917         sack_rxmit = 0;
19918         len = 0;
19919         rsm = NULL;
19920         if (flags & TH_RST) {
19921                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
19922                 so = inp->inp_socket;
19923                 sb = &so->so_snd;
19924                 goto send;
19925         }
19926         if (rack->r_ctl.rc_resend) {
19927                 /* Retransmit timer */
19928                 rsm = rack->r_ctl.rc_resend;
19929                 rack->r_ctl.rc_resend = NULL;
19930                 len = rsm->r_end - rsm->r_start;
19931                 sack_rxmit = 1;
19932                 sendalot = 0;
19933                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
19934                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
19935                          __func__, __LINE__,
19936                          rsm->r_start, tp->snd_una, tp, rack, rsm));
19937                 sb_offset = rsm->r_start - tp->snd_una;
19938                 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
19939         } else if (rack->r_collapse_point_valid &&
19940                    ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
19941                 /*
19942                  * If an RSM is returned then enough time has passed
19943                  * for us to retransmit it. Move up the collapse point,
19944                  * since this rsm has its chance to retransmit now.
19945                  */
19946                 tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
19947                 rack->r_ctl.last_collapse_point = rsm->r_end;
19948                 /* Are we done? */
19949                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19950                             rack->r_ctl.high_collapse_point))
19951                         rack->r_collapse_point_valid = 0;
19952                 sack_rxmit = 1;
19953                 /* We are not doing a TLP */
19954                 doing_tlp = 0;
19955                 len = rsm->r_end - rsm->r_start;
19956                 sb_offset = rsm->r_start - tp->snd_una;
19957                 sendalot = 0;
19958                 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
19959         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
19960                 /* We have a retransmit that takes precedence */
19961                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
19962                     ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
19963                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
19964                         /* Enter recovery if not induced by a time-out */
19965                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
19966                 }
19967 #ifdef INVARIANTS
19968                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
19969                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
19970                               tp, rack, rsm, rsm->r_start, tp->snd_una);
19971                 }
19972 #endif
19973                 len = rsm->r_end - rsm->r_start;
19974                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
19975                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
19976                          __func__, __LINE__,
19977                          rsm->r_start, tp->snd_una, tp, rack, rsm));
19978                 sb_offset = rsm->r_start - tp->snd_una;
19979                 sendalot = 0;
19980                 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
19981                 if (len > 0) {
19982                         sack_rxmit = 1;
19983                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
19984                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
19985                                          min(len, segsiz));
19986                 }
19987         } else if (rack->r_ctl.rc_tlpsend) {
19988                 /* Tail loss probe */
19989                 long cwin;
19990                 long tlen;
19991
19992                 /*
19993                  * Check if we can do a TLP with a RACK'd packet
19994                  * this can happen if we are not doing the rack
19995                  * cheat and we skipped to a TLP and it
19996                  * went off.
19997                  */
19998                 rsm = rack->r_ctl.rc_tlpsend;
19999                 /* We are doing a TLP make sure the flag is preent */
20000                 rsm->r_flags |= RACK_TLP;
20001                 rack->r_ctl.rc_tlpsend = NULL;
20002                 sack_rxmit = 1;
20003                 tlen = rsm->r_end - rsm->r_start;
20004                 if (tlen > segsiz)
20005                         tlen = segsiz;
20006                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20007                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20008                          __func__, __LINE__,
20009                          rsm->r_start, tp->snd_una, tp, rack, rsm));
20010                 sb_offset = rsm->r_start - tp->snd_una;
20011                 cwin = min(tp->snd_wnd, tlen);
20012                 len = cwin;
20013         }
20014         if (rack->r_must_retran &&
20015             (doing_tlp == 0) &&
20016             (SEQ_GT(tp->snd_max, tp->snd_una)) &&
20017             (rsm == NULL)) {
20018                 /*
20019                  * There are two different ways that we
20020                  * can get into this block:
20021                  * a) This is a non-sack connection, we had a time-out
20022                  *    and thus r_must_retran was set and everything
20023                  *    left outstanding as been marked for retransmit.
20024                  * b) The MTU of the path shrank, so that everything
20025                  *    was marked to be retransmitted with the smaller
20026                  *    mtu and r_must_retran was set.
20027                  *
20028                  * This means that we expect the sendmap (outstanding)
20029                  * to all be marked must. We can use the tmap to
20030                  * look at them.
20031                  *
20032                  */
20033                 int sendwin, flight;
20034
20035                 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
20036                 flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
20037                 if (flight >= sendwin) {
20038                         /*
20039                          * We can't send yet.
20040                          */
20041                         so = inp->inp_socket;
20042                         sb = &so->so_snd;
20043                         goto just_return_nolock;
20044                 }
20045                 /*
20046                  * This is the case a/b mentioned above. All
20047                  * outstanding/not-acked should be marked.
20048                  * We can use the tmap to find them.
20049                  */
20050                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
20051                 if (rsm == NULL) {
20052                         /* TSNH */
20053                         rack->r_must_retran = 0;
20054                         rack->r_ctl.rc_out_at_rto = 0;
20055                         so = inp->inp_socket;
20056                         sb = &so->so_snd;
20057                         goto just_return_nolock;
20058                 }
20059                 if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
20060                         /*
20061                          * The first one does not have the flag, did we collapse
20062                          * further up in our list?
20063                          */
20064                         rack->r_must_retran = 0;
20065                         rack->r_ctl.rc_out_at_rto = 0;
20066                         rsm = NULL;
20067                         sack_rxmit = 0;
20068                 } else {
20069                         sack_rxmit = 1;
20070                         len = rsm->r_end - rsm->r_start;
20071                         sb_offset = rsm->r_start - tp->snd_una;
20072                         sendalot = 0;
20073                         if ((rack->full_size_rxt == 0) &&
20074                             (rack->shape_rxt_to_pacing_min == 0) &&
20075                             (len >= segsiz))
20076                                 len = segsiz;
20077                         else if (rack->shape_rxt_to_pacing_min &&
20078                                  rack->gp_ready) {
20079                                 /* We use pacing min as shaping len req */
20080                                 uint32_t maxlen;
20081
20082                                 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20083                                 if (len > maxlen)
20084                                         len = maxlen;
20085                         }
20086                         /*
20087                          * Delay removing the flag RACK_MUST_RXT so
20088                          * that the fastpath for retransmit will
20089                          * work with this rsm.
20090                          */
20091                 }
20092         }
20093         /*
20094          * Enforce a connection sendmap count limit if set
20095          * as long as we are not retransmiting.
20096          */
20097         if ((rsm == NULL) &&
20098             (rack->do_detection == 0) &&
20099             (V_tcp_map_entries_limit > 0) &&
20100             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
20101                 counter_u64_add(rack_to_alloc_limited, 1);
20102                 if (!rack->alloc_limit_reported) {
20103                         rack->alloc_limit_reported = 1;
20104                         counter_u64_add(rack_alloc_limited_conns, 1);
20105                 }
20106                 so = inp->inp_socket;
20107                 sb = &so->so_snd;
20108                 goto just_return_nolock;
20109         }
20110         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
20111                 /* we are retransmitting the fin */
20112                 len--;
20113                 if (len) {
20114                         /*
20115                          * When retransmitting data do *not* include the
20116                          * FIN. This could happen from a TLP probe.
20117                          */
20118                         flags &= ~TH_FIN;
20119                 }
20120         }
20121         if (rsm && rack->r_fsb_inited &&
20122             rack_use_rsm_rfo &&
20123             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
20124                 int ret;
20125
20126                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
20127                 if (ret == 0)
20128                         return (0);
20129         }
20130         so = inp->inp_socket;
20131         sb = &so->so_snd;
20132         if (do_a_prefetch == 0) {
20133                 kern_prefetch(sb, &do_a_prefetch);
20134                 do_a_prefetch = 1;
20135         }
20136 #ifdef NETFLIX_SHARED_CWND
20137         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
20138             rack->rack_enable_scwnd) {
20139                 /* We are doing cwnd sharing */
20140                 if (rack->gp_ready &&
20141                     (rack->rack_attempted_scwnd == 0) &&
20142                     (rack->r_ctl.rc_scw == NULL) &&
20143                     tp->t_lib) {
20144                         /* The pcbid is in, lets make an attempt */
20145                         counter_u64_add(rack_try_scwnd, 1);
20146                         rack->rack_attempted_scwnd = 1;
20147                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
20148                                                                    &rack->r_ctl.rc_scw_index,
20149                                                                    segsiz);
20150                 }
20151                 if (rack->r_ctl.rc_scw &&
20152                     (rack->rack_scwnd_is_idle == 1) &&
20153                     sbavail(&so->so_snd)) {
20154                         /* we are no longer out of data */
20155                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20156                         rack->rack_scwnd_is_idle = 0;
20157                 }
20158                 if (rack->r_ctl.rc_scw) {
20159                         /* First lets update and get the cwnd */
20160                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
20161                                                                                        rack->r_ctl.rc_scw_index,
20162                                                                                        tp->snd_cwnd, tp->snd_wnd, segsiz);
20163                 }
20164         }
20165 #endif
20166         /*
20167          * Get standard flags, and add SYN or FIN if requested by 'hidden'
20168          * state flags.
20169          */
20170         if (tp->t_flags & TF_NEEDFIN)
20171                 flags |= TH_FIN;
20172         if (tp->t_flags & TF_NEEDSYN)
20173                 flags |= TH_SYN;
20174         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
20175                 void *end_rsm;
20176                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
20177                 if (end_rsm)
20178                         kern_prefetch(end_rsm, &prefetch_rsm);
20179                 prefetch_rsm = 1;
20180         }
20181         SOCKBUF_LOCK(sb);
20182         /*
20183          * If snd_nxt == snd_max and we have transmitted a FIN, the
20184          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
20185          * negative length.  This can also occur when TCP opens up its
20186          * congestion window while receiving additional duplicate acks after
20187          * fast-retransmit because TCP will reset snd_nxt to snd_max after
20188          * the fast-retransmit.
20189          *
20190          * In the normal retransmit-FIN-only case, however, snd_nxt will be
20191          * set to snd_una, the sb_offset will be 0, and the length may wind
20192          * up 0.
20193          *
20194          * If sack_rxmit is true we are retransmitting from the scoreboard
20195          * in which case len is already set.
20196          */
20197         if ((sack_rxmit == 0) &&
20198             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
20199                 uint32_t avail;
20200
20201                 avail = sbavail(sb);
20202                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
20203                         sb_offset = tp->snd_nxt - tp->snd_una;
20204                 else
20205                         sb_offset = 0;
20206                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
20207                         if (rack->r_ctl.rc_tlp_new_data) {
20208                                 /* TLP is forcing out new data */
20209                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
20210                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
20211                                 }
20212                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
20213                                         if (tp->snd_wnd > sb_offset)
20214                                                 len = tp->snd_wnd - sb_offset;
20215                                         else
20216                                                 len = 0;
20217                                 } else {
20218                                         len = rack->r_ctl.rc_tlp_new_data;
20219                                 }
20220                                 rack->r_ctl.rc_tlp_new_data = 0;
20221                         }  else {
20222                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
20223                         }
20224                         if ((rack->r_ctl.crte == NULL) &&
20225                             IN_FASTRECOVERY(tp->t_flags) &&
20226                             (rack->full_size_rxt == 0) &&
20227                             (rack->shape_rxt_to_pacing_min == 0) &&
20228                             (len > segsiz)) {
20229                                 /*
20230                                  * For prr=off, we need to send only 1 MSS
20231                                  * at a time. We do this because another sack could
20232                                  * be arriving that causes us to send retransmits and
20233                                  * we don't want to be on a long pace due to a larger send
20234                                  * that keeps us from sending out the retransmit.
20235                                  */
20236                                 len = segsiz;
20237                         } else if (rack->shape_rxt_to_pacing_min &&
20238                                    rack->gp_ready) {
20239                                 /* We use pacing min as shaping len req */
20240                                 uint32_t maxlen;
20241
20242                                 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20243                                 if (len > maxlen)
20244                                         len = maxlen;
20245                         }/* The else is full_size_rxt is on so send it all */
20246                 } else {
20247                         uint32_t outstanding;
20248                         /*
20249                          * We are inside of a Fast recovery episode, this
20250                          * is caused by a SACK or 3 dup acks. At this point
20251                          * we have sent all the retransmissions and we rely
20252                          * on PRR to dictate what we will send in the form of
20253                          * new data.
20254                          */
20255
20256                         outstanding = tp->snd_max - tp->snd_una;
20257                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
20258                                 if (tp->snd_wnd > outstanding) {
20259                                         len = tp->snd_wnd - outstanding;
20260                                         /* Check to see if we have the data */
20261                                         if ((sb_offset + len) > avail) {
20262                                                 /* It does not all fit */
20263                                                 if (avail > sb_offset)
20264                                                         len = avail - sb_offset;
20265                                                 else
20266                                                         len = 0;
20267                                         }
20268                                 } else {
20269                                         len = 0;
20270                                 }
20271                         } else if (avail > sb_offset) {
20272                                 len = avail - sb_offset;
20273                         } else {
20274                                 len = 0;
20275                         }
20276                         if (len > 0) {
20277                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
20278                                         len = rack->r_ctl.rc_prr_sndcnt;
20279                                 }
20280                                 if (len > 0) {
20281                                         sub_from_prr = 1;
20282                                 }
20283                         }
20284                         if (len > segsiz) {
20285                                 /*
20286                                  * We should never send more than a MSS when
20287                                  * retransmitting or sending new data in prr
20288                                  * mode unless the override flag is on. Most
20289                                  * likely the PRR algorithm is not going to
20290                                  * let us send a lot as well :-)
20291                                  */
20292                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
20293                                         len = segsiz;
20294                                 }
20295                         } else if (len < segsiz) {
20296                                 /*
20297                                  * Do we send any? The idea here is if the
20298                                  * send empty's the socket buffer we want to
20299                                  * do it. However if not then lets just wait
20300                                  * for our prr_sndcnt to get bigger.
20301                                  */
20302                                 long leftinsb;
20303
20304                                 leftinsb = sbavail(sb) - sb_offset;
20305                                 if (leftinsb > len) {
20306                                         /* This send does not empty the sb */
20307                                         len = 0;
20308                                 }
20309                         }
20310                 }
20311         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
20312                 /*
20313                  * If you have not established
20314                  * and are not doing FAST OPEN
20315                  * no data please.
20316                  */
20317                 if ((sack_rxmit == 0) &&
20318                     (!IS_FASTOPEN(tp->t_flags))){
20319                         len = 0;
20320                         sb_offset = 0;
20321                 }
20322         }
20323         if (prefetch_so_done == 0) {
20324                 kern_prefetch(so, &prefetch_so_done);
20325                 prefetch_so_done = 1;
20326         }
20327         /*
20328          * Lop off SYN bit if it has already been sent.  However, if this is
20329          * SYN-SENT state and if segment contains data and if we don't know
20330          * that foreign host supports TAO, suppress sending segment.
20331          */
20332         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
20333             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
20334                 /*
20335                  * When sending additional segments following a TFO SYN|ACK,
20336                  * do not include the SYN bit.
20337                  */
20338                 if (IS_FASTOPEN(tp->t_flags) &&
20339                     (tp->t_state == TCPS_SYN_RECEIVED))
20340                         flags &= ~TH_SYN;
20341         }
20342         /*
20343          * Be careful not to send data and/or FIN on SYN segments. This
20344          * measure is needed to prevent interoperability problems with not
20345          * fully conformant TCP implementations.
20346          */
20347         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
20348                 len = 0;
20349                 flags &= ~TH_FIN;
20350         }
20351         /*
20352          * On TFO sockets, ensure no data is sent in the following cases:
20353          *
20354          *  - When retransmitting SYN|ACK on a passively-created socket
20355          *
20356          *  - When retransmitting SYN on an actively created socket
20357          *
20358          *  - When sending a zero-length cookie (cookie request) on an
20359          *    actively created socket
20360          *
20361          *  - When the socket is in the CLOSED state (RST is being sent)
20362          */
20363         if (IS_FASTOPEN(tp->t_flags) &&
20364             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
20365              ((tp->t_state == TCPS_SYN_SENT) &&
20366               (tp->t_tfo_client_cookie_len == 0)) ||
20367              (flags & TH_RST))) {
20368                 sack_rxmit = 0;
20369                 len = 0;
20370         }
20371         /* Without fast-open there should never be data sent on a SYN */
20372         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
20373                 tp->snd_nxt = tp->iss;
20374                 len = 0;
20375         }
20376         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
20377                 /* We only send 1 MSS if we have a DSACK block */
20378                 add_flag |= RACK_SENT_W_DSACK;
20379                 len = segsiz;
20380         }
20381         orig_len = len;
20382         if (len <= 0) {
20383                 /*
20384                  * If FIN has been sent but not acked, but we haven't been
20385                  * called to retransmit, len will be < 0.  Otherwise, window
20386                  * shrank after we sent into it.  If window shrank to 0,
20387                  * cancel pending retransmit, pull snd_nxt back to (closed)
20388                  * window, and set the persist timer if it isn't already
20389                  * going.  If the window didn't close completely, just wait
20390                  * for an ACK.
20391                  *
20392                  * We also do a general check here to ensure that we will
20393                  * set the persist timer when we have data to send, but a
20394                  * 0-byte window. This makes sure the persist timer is set
20395                  * even if the packet hits one of the "goto send" lines
20396                  * below.
20397                  */
20398                 len = 0;
20399                 if ((tp->snd_wnd == 0) &&
20400                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20401                     (tp->snd_una == tp->snd_max) &&
20402                     (sb_offset < (int)sbavail(sb))) {
20403                         rack_enter_persist(tp, rack, cts, tp->snd_una);
20404                 }
20405         } else if ((rsm == NULL) &&
20406                    (doing_tlp == 0) &&
20407                    (len < pace_max_seg)) {
20408                 /*
20409                  * We are not sending a maximum sized segment for
20410                  * some reason. Should we not send anything (think
20411                  * sws or persists)?
20412                  */
20413                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20414                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20415                     (len < minseg) &&
20416                     (len < (int)(sbavail(sb) - sb_offset))) {
20417                         /*
20418                          * Here the rwnd is less than
20419                          * the minimum pacing size, this is not a retransmit,
20420                          * we are established and
20421                          * the send is not the last in the socket buffer
20422                          * we send nothing, and we may enter persists
20423                          * if nothing is outstanding.
20424                          */
20425                         len = 0;
20426                         if (tp->snd_max == tp->snd_una) {
20427                                 /*
20428                                  * Nothing out we can
20429                                  * go into persists.
20430                                  */
20431                                 rack_enter_persist(tp, rack, cts, tp->snd_una);
20432                         }
20433                 } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
20434                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20435                            (len < (int)(sbavail(sb) - sb_offset)) &&
20436                            (len < minseg)) {
20437                         /*
20438                          * Here we are not retransmitting, and
20439                          * the cwnd is not so small that we could
20440                          * not send at least a min size (rxt timer
20441                          * not having gone off), We have 2 segments or
20442                          * more already in flight, its not the tail end
20443                          * of the socket buffer  and the cwnd is blocking
20444                          * us from sending out a minimum pacing segment size.
20445                          * Lets not send anything.
20446                          */
20447                         len = 0;
20448                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
20449                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20450                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20451                            (len < (int)(sbavail(sb) - sb_offset)) &&
20452                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
20453                         /*
20454                          * Here we have a send window but we have
20455                          * filled it up and we can't send another pacing segment.
20456                          * We also have in flight more than 2 segments
20457                          * and we are not completing the sb i.e. we allow
20458                          * the last bytes of the sb to go out even if
20459                          * its not a full pacing segment.
20460                          */
20461                         len = 0;
20462                 } else if ((rack->r_ctl.crte != NULL) &&
20463                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
20464                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
20465                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
20466                            (len < (int)(sbavail(sb) - sb_offset))) {
20467                         /*
20468                          * Here we are doing hardware pacing, this is not a TLP,
20469                          * we are not sending a pace max segment size, there is rwnd
20470                          * room to send at least N pace_max_seg, the cwnd is greater
20471                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
20472                          * more segments in flight and its not the tail of the socket buffer.
20473                          *
20474                          * We don't want to send instead we need to get more ack's in to
20475                          * allow us to send a full pacing segment. Normally, if we are pacing
20476                          * about the right speed, we should have finished our pacing
20477                          * send as most of the acks have come back if we are at the
20478                          * right rate. This is a bit fuzzy since return path delay
20479                          * can delay the acks, which is why we want to make sure we
20480                          * have cwnd space to have a bit more than a max pace segments in flight.
20481                          *
20482                          * If we have not gotten our acks back we are pacing at too high a
20483                          * rate delaying will not hurt and will bring our GP estimate down by
20484                          * injecting the delay. If we don't do this we will send
20485                          * 2 MSS out in response to the acks being clocked in which
20486                          * defeats the point of hw-pacing (i.e. to help us get
20487                          * larger TSO's out).
20488                          */
20489                         len = 0;
20490                 }
20491
20492         }
20493         /* len will be >= 0 after this point. */
20494         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
20495         rack_sndbuf_autoscale(rack);
20496         /*
20497          * Decide if we can use TCP Segmentation Offloading (if supported by
20498          * hardware).
20499          *
20500          * TSO may only be used if we are in a pure bulk sending state.  The
20501          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
20502          * options prevent using TSO.  With TSO the TCP header is the same
20503          * (except for the sequence number) for all generated packets.  This
20504          * makes it impossible to transmit any options which vary per
20505          * generated segment or packet.
20506          *
20507          * IPv4 handling has a clear separation of ip options and ip header
20508          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
20509          * the right thing below to provide length of just ip options and thus
20510          * checking for ipoptlen is enough to decide if ip options are present.
20511          */
20512         ipoptlen = 0;
20513 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20514         /*
20515          * Pre-calculate here as we save another lookup into the darknesses
20516          * of IPsec that way and can actually decide if TSO is ok.
20517          */
20518 #ifdef INET6
20519         if (isipv6 && IPSEC_ENABLED(ipv6))
20520                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
20521 #ifdef INET
20522         else
20523 #endif
20524 #endif                          /* INET6 */
20525 #ifdef INET
20526                 if (IPSEC_ENABLED(ipv4))
20527                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
20528 #endif                          /* INET */
20529 #endif
20530
20531 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20532         ipoptlen += ipsec_optlen;
20533 #endif
20534         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
20535             (tp->t_port == 0) &&
20536             ((tp->t_flags & TF_SIGNATURE) == 0) &&
20537             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
20538             ipoptlen == 0)
20539                 tso = 1;
20540         {
20541                 uint32_t outstanding __unused;
20542
20543                 outstanding = tp->snd_max - tp->snd_una;
20544                 if (tp->t_flags & TF_SENTFIN) {
20545                         /*
20546                          * If we sent a fin, snd_max is 1 higher than
20547                          * snd_una
20548                          */
20549                         outstanding--;
20550                 }
20551                 if (sack_rxmit) {
20552                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
20553                                 flags &= ~TH_FIN;
20554                 } else {
20555                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
20556                                    sbused(sb)))
20557                                 flags &= ~TH_FIN;
20558                 }
20559         }
20560         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
20561                       (long)TCP_MAXWIN << tp->rcv_scale);
20562
20563         /*
20564          * Sender silly window avoidance.   We transmit under the following
20565          * conditions when len is non-zero:
20566          *
20567          * - We have a full segment (or more with TSO) - This is the last
20568          * buffer in a write()/send() and we are either idle or running
20569          * NODELAY - we've timed out (e.g. persist timer) - we have more
20570          * then 1/2 the maximum send window's worth of data (receiver may be
20571          * limited the window size) - we need to retransmit
20572          */
20573         if (len) {
20574                 if (len >= segsiz) {
20575                         goto send;
20576                 }
20577                 /*
20578                  * NOTE! on localhost connections an 'ack' from the remote
20579                  * end may occur synchronously with the output and cause us
20580                  * to flush a buffer queued with moretocome.  XXX
20581                  *
20582                  */
20583                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
20584                     (idle || (tp->t_flags & TF_NODELAY)) &&
20585                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20586                     (tp->t_flags & TF_NOPUSH) == 0) {
20587                         pass = 2;
20588                         goto send;
20589                 }
20590                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
20591                         pass = 22;
20592                         goto send;
20593                 }
20594                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
20595                         pass = 4;
20596                         goto send;
20597                 }
20598                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
20599                         pass = 5;
20600                         goto send;
20601                 }
20602                 if (sack_rxmit) {
20603                         pass = 6;
20604                         goto send;
20605                 }
20606                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
20607                     (ctf_outstanding(tp) < (segsiz * 2))) {
20608                         /*
20609                          * We have less than two MSS outstanding (delayed ack)
20610                          * and our rwnd will not let us send a full sized
20611                          * MSS. Lets go ahead and let this small segment
20612                          * out because we want to try to have at least two
20613                          * packets inflight to not be caught by delayed ack.
20614                          */
20615                         pass = 12;
20616                         goto send;
20617                 }
20618         }
20619         /*
20620          * Sending of standalone window updates.
20621          *
20622          * Window updates are important when we close our window due to a
20623          * full socket buffer and are opening it again after the application
20624          * reads data from it.  Once the window has opened again and the
20625          * remote end starts to send again the ACK clock takes over and
20626          * provides the most current window information.
20627          *
20628          * We must avoid the silly window syndrome whereas every read from
20629          * the receive buffer, no matter how small, causes a window update
20630          * to be sent.  We also should avoid sending a flurry of window
20631          * updates when the socket buffer had queued a lot of data and the
20632          * application is doing small reads.
20633          *
20634          * Prevent a flurry of pointless window updates by only sending an
20635          * update when we can increase the advertized window by more than
20636          * 1/4th of the socket buffer capacity.  When the buffer is getting
20637          * full or is very small be more aggressive and send an update
20638          * whenever we can increase by two mss sized segments. In all other
20639          * situations the ACK's to new incoming data will carry further
20640          * window increases.
20641          *
20642          * Don't send an independent window update if a delayed ACK is
20643          * pending (it will get piggy-backed on it) or the remote side
20644          * already has done a half-close and won't send more data.  Skip
20645          * this if the connection is in T/TCP half-open state.
20646          */
20647         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
20648             !(tp->t_flags & TF_DELACK) &&
20649             !TCPS_HAVERCVDFIN(tp->t_state)) {
20650                 /*
20651                  * "adv" is the amount we could increase the window, taking
20652                  * into account that we are limited by TCP_MAXWIN <<
20653                  * tp->rcv_scale.
20654                  */
20655                 int32_t adv;
20656                 int oldwin;
20657
20658                 adv = recwin;
20659                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
20660                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
20661                         if (adv > oldwin)
20662                                 adv -= oldwin;
20663                         else {
20664                                 /* We can't increase the window */
20665                                 adv = 0;
20666                         }
20667                 } else
20668                         oldwin = 0;
20669
20670                 /*
20671                  * If the new window size ends up being the same as or less
20672                  * than the old size when it is scaled, then don't force
20673                  * a window update.
20674                  */
20675                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
20676                         goto dontupdate;
20677
20678                 if (adv >= (int32_t)(2 * segsiz) &&
20679                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
20680                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
20681                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
20682                         pass = 7;
20683                         goto send;
20684                 }
20685                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
20686                         pass = 23;
20687                         goto send;
20688                 }
20689         }
20690 dontupdate:
20691
20692         /*
20693          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
20694          * is also a catch-all for the retransmit timer timeout case.
20695          */
20696         if (tp->t_flags & TF_ACKNOW) {
20697                 pass = 8;
20698                 goto send;
20699         }
20700         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
20701                 pass = 9;
20702                 goto send;
20703         }
20704         /*
20705          * If our state indicates that FIN should be sent and we have not
20706          * yet done so, then we need to send.
20707          */
20708         if ((flags & TH_FIN) &&
20709             (tp->snd_nxt == tp->snd_una)) {
20710                 pass = 11;
20711                 goto send;
20712         }
20713         /*
20714          * No reason to send a segment, just return.
20715          */
20716 just_return:
20717         SOCKBUF_UNLOCK(sb);
20718 just_return_nolock:
20719         {
20720                 int app_limited = CTF_JR_SENT_DATA;
20721
20722                 if (tot_len_this_send > 0) {
20723                         /* Make sure snd_nxt is up to max */
20724                         rack->r_ctl.fsb.recwin = recwin;
20725                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
20726                         if ((error == 0) &&
20727                             rack_use_rfo &&
20728                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
20729                             (ipoptlen == 0) &&
20730                             (tp->snd_nxt == tp->snd_max) &&
20731                             (tp->rcv_numsacks == 0) &&
20732                             rack->r_fsb_inited &&
20733                             TCPS_HAVEESTABLISHED(tp->t_state) &&
20734                             ((IN_RECOVERY(tp->t_flags)) == 0) &&
20735                             (rack->r_must_retran == 0) &&
20736                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
20737                             (len > 0) && (orig_len > 0) &&
20738                             (orig_len > len) &&
20739                             ((orig_len - len) >= segsiz) &&
20740                             ((optlen == 0) ||
20741                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
20742                                 /* We can send at least one more MSS using our fsb */
20743                                 rack_setup_fast_output(tp, rack, sb, len, orig_len,
20744                                                        segsiz, pace_max_seg, hw_tls, flags);
20745                         } else
20746                                 rack->r_fast_output = 0;
20747
20748
20749                         rack_log_fsb(rack, tp, so, flags,
20750                                      ipoptlen, orig_len, len, 0,
20751                                      1, optlen, __LINE__, 1);
20752                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
20753                                 tp->snd_nxt = tp->snd_max;
20754                 } else {
20755                         int end_window = 0;
20756                         uint32_t seq = tp->gput_ack;
20757
20758                         rsm = tqhash_max(rack->r_ctl.tqh);
20759                         if (rsm) {
20760                                 /*
20761                                  * Mark the last sent that we just-returned (hinting
20762                                  * that delayed ack may play a role in any rtt measurement).
20763                                  */
20764                                 rsm->r_just_ret = 1;
20765                         }
20766                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
20767                         rack->r_ctl.rc_agg_delayed = 0;
20768                         rack->r_early = 0;
20769                         rack->r_late = 0;
20770                         rack->r_ctl.rc_agg_early = 0;
20771                         if ((ctf_outstanding(tp) +
20772                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
20773                                  minseg)) >= tp->snd_wnd) {
20774                                 /* We are limited by the rwnd */
20775                                 app_limited = CTF_JR_RWND_LIMITED;
20776                                 if (IN_FASTRECOVERY(tp->t_flags))
20777                                         rack->r_ctl.rc_prr_sndcnt = 0;
20778                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
20779                                 /* We are limited by whats available -- app limited */
20780                                 app_limited = CTF_JR_APP_LIMITED;
20781                                 if (IN_FASTRECOVERY(tp->t_flags))
20782                                         rack->r_ctl.rc_prr_sndcnt = 0;
20783                         } else if ((idle == 0) &&
20784                                    ((tp->t_flags & TF_NODELAY) == 0) &&
20785                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20786                                    (len < segsiz)) {
20787                                 /*
20788                                  * No delay is not on and the
20789                                  * user is sending less than 1MSS. This
20790                                  * brings out SWS avoidance so we
20791                                  * don't send. Another app-limited case.
20792                                  */
20793                                 app_limited = CTF_JR_APP_LIMITED;
20794                         } else if (tp->t_flags & TF_NOPUSH) {
20795                                 /*
20796                                  * The user has requested no push of
20797                                  * the last segment and we are
20798                                  * at the last segment. Another app
20799                                  * limited case.
20800                                  */
20801                                 app_limited = CTF_JR_APP_LIMITED;
20802                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
20803                                 /* Its the cwnd */
20804                                 app_limited = CTF_JR_CWND_LIMITED;
20805                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
20806                                    (rack->rack_no_prr == 0) &&
20807                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
20808                                 app_limited = CTF_JR_PRR;
20809                         } else {
20810                                 /* Now why here are we not sending? */
20811 #ifdef NOW
20812 #ifdef INVARIANTS
20813                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
20814 #endif
20815 #endif
20816                                 app_limited = CTF_JR_ASSESSING;
20817                         }
20818                         /*
20819                          * App limited in some fashion, for our pacing GP
20820                          * measurements we don't want any gap (even cwnd).
20821                          * Close  down the measurement window.
20822                          */
20823                         if (rack_cwnd_block_ends_measure &&
20824                             ((app_limited == CTF_JR_CWND_LIMITED) ||
20825                              (app_limited == CTF_JR_PRR))) {
20826                                 /*
20827                                  * The reason we are not sending is
20828                                  * the cwnd (or prr). We have been configured
20829                                  * to end the measurement window in
20830                                  * this case.
20831                                  */
20832                                 end_window = 1;
20833                         } else if (rack_rwnd_block_ends_measure &&
20834                                    (app_limited == CTF_JR_RWND_LIMITED)) {
20835                                 /*
20836                                  * We are rwnd limited and have been
20837                                  * configured to end the measurement
20838                                  * window in this case.
20839                                  */
20840                                 end_window = 1;
20841                         } else if (app_limited == CTF_JR_APP_LIMITED) {
20842                                 /*
20843                                  * A true application limited period, we have
20844                                  * ran out of data.
20845                                  */
20846                                 end_window = 1;
20847                         } else if (app_limited == CTF_JR_ASSESSING) {
20848                                 /*
20849                                  * In the assessing case we hit the end of
20850                                  * the if/else and had no known reason
20851                                  * This will panic us under invariants..
20852                                  *
20853                                  * If we get this out in logs we need to
20854                                  * investagate which reason we missed.
20855                                  */
20856                                 end_window = 1;
20857                         }
20858                         if (end_window) {
20859                                 uint8_t log = 0;
20860
20861                                 /* Adjust the Gput measurement */
20862                                 if ((tp->t_flags & TF_GPUTINPROG) &&
20863                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
20864                                         tp->gput_ack = tp->snd_max;
20865                                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
20866                                                 /*
20867                                                  * There is not enough to measure.
20868                                                  */
20869                                                 tp->t_flags &= ~TF_GPUTINPROG;
20870                                                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
20871                                                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
20872                                                                            tp->gput_seq,
20873                                                                            0, 0, 18, __LINE__, NULL, 0);
20874                                         } else
20875                                                 log = 1;
20876                                 }
20877                                 /* Mark the last packet has app limited */
20878                                 rsm = tqhash_max(rack->r_ctl.tqh);
20879                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
20880                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
20881                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
20882                                         else {
20883                                                 /*
20884                                                  * Go out to the end app limited and mark
20885                                                  * this new one as next and move the end_appl up
20886                                                  * to this guy.
20887                                                  */
20888                                                 if (rack->r_ctl.rc_end_appl)
20889                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
20890                                                 rack->r_ctl.rc_end_appl = rsm;
20891                                         }
20892                                         rsm->r_flags |= RACK_APP_LIMITED;
20893                                         rack->r_ctl.rc_app_limited_cnt++;
20894                                 }
20895                                 if (log)
20896                                         rack_log_pacing_delay_calc(rack,
20897                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
20898                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
20899                         }
20900                 }
20901                 /* Check if we need to go into persists or not */
20902                 if ((tp->snd_max == tp->snd_una) &&
20903                     TCPS_HAVEESTABLISHED(tp->t_state) &&
20904                     sbavail(sb) &&
20905                     (sbavail(sb) > tp->snd_wnd) &&
20906                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
20907                         /* Yes lets make sure to move to persist before timer-start */
20908                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
20909                 }
20910                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
20911                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
20912         }
20913 #ifdef NETFLIX_SHARED_CWND
20914         if ((sbavail(sb) == 0) &&
20915             rack->r_ctl.rc_scw) {
20916                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20917                 rack->rack_scwnd_is_idle = 1;
20918         }
20919 #endif
20920 #ifdef TCP_ACCOUNTING
20921         if (tot_len_this_send > 0) {
20922                 crtsc = get_cyclecount();
20923                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20924                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
20925                 }
20926                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20927                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
20928                 }
20929                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20930                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
20931                 }
20932         } else {
20933                 crtsc = get_cyclecount();
20934                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20935                         tp->tcp_cnt_counters[SND_LIMITED]++;
20936                 }
20937                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20938                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
20939                 }
20940         }
20941         sched_unpin();
20942 #endif
20943         return (0);
20944
20945 send:
20946         if ((rack->r_ctl.crte != NULL) &&
20947             (rsm == NULL) &&
20948             ((rack->rc_hw_nobuf == 1) ||
20949              (rack_hw_check_queue && (check_done == 0)))) {
20950                 /*
20951                  * We only want to do this once with the hw_check_queue,
20952                  * for the enobuf case we would only do it once if
20953                  * we come around to again, the flag will be clear.
20954                  */
20955                 check_done = 1;
20956                 slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
20957                 if (slot) {
20958                         rack->r_ctl.rc_agg_delayed = 0;
20959                         rack->r_ctl.rc_agg_early = 0;
20960                         rack->r_early = 0;
20961                         rack->r_late = 0;
20962                         SOCKBUF_UNLOCK(&so->so_snd);
20963                         goto skip_all_send;
20964                 }
20965         }
20966         if (rsm || sack_rxmit)
20967                 counter_u64_add(rack_nfto_resend, 1);
20968         else
20969                 counter_u64_add(rack_non_fto_send, 1);
20970         if ((flags & TH_FIN) &&
20971             sbavail(sb)) {
20972                 /*
20973                  * We do not transmit a FIN
20974                  * with data outstanding. We
20975                  * need to make it so all data
20976                  * is acked first.
20977                  */
20978                 flags &= ~TH_FIN;
20979         }
20980         /* Enforce stack imposed max seg size if we have one */
20981         if (rack->r_ctl.rc_pace_max_segs &&
20982             (len > rack->r_ctl.rc_pace_max_segs)) {
20983                 mark = 1;
20984                 len = rack->r_ctl.rc_pace_max_segs;
20985         }
20986         SOCKBUF_LOCK_ASSERT(sb);
20987         if (len > 0) {
20988                 if (len >= segsiz)
20989                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
20990                 else
20991                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
20992         }
20993         /*
20994          * Before ESTABLISHED, force sending of initial options unless TCP
20995          * set not to do any options. NOTE: we assume that the IP/TCP header
20996          * plus TCP options always fit in a single mbuf, leaving room for a
20997          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
20998          * + optlen <= MCLBYTES
20999          */
21000         optlen = 0;
21001 #ifdef INET6
21002         if (isipv6)
21003                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
21004         else
21005 #endif
21006                 hdrlen = sizeof(struct tcpiphdr);
21007
21008         /*
21009          * Compute options for segment. We only have to care about SYN and
21010          * established connection segments.  Options for SYN-ACK segments
21011          * are handled in TCP syncache.
21012          */
21013         to.to_flags = 0;
21014         if ((tp->t_flags & TF_NOOPT) == 0) {
21015                 /* Maximum segment size. */
21016                 if (flags & TH_SYN) {
21017                         tp->snd_nxt = tp->iss;
21018                         to.to_mss = tcp_mssopt(&inp->inp_inc);
21019                         if (tp->t_port)
21020                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
21021                         to.to_flags |= TOF_MSS;
21022
21023                         /*
21024                          * On SYN or SYN|ACK transmits on TFO connections,
21025                          * only include the TFO option if it is not a
21026                          * retransmit, as the presence of the TFO option may
21027                          * have caused the original SYN or SYN|ACK to have
21028                          * been dropped by a middlebox.
21029                          */
21030                         if (IS_FASTOPEN(tp->t_flags) &&
21031                             (tp->t_rxtshift == 0)) {
21032                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
21033                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
21034                                         to.to_tfo_cookie =
21035                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
21036                                         to.to_flags |= TOF_FASTOPEN;
21037                                         wanted_cookie = 1;
21038                                 } else if (tp->t_state == TCPS_SYN_SENT) {
21039                                         to.to_tfo_len =
21040                                                 tp->t_tfo_client_cookie_len;
21041                                         to.to_tfo_cookie =
21042                                                 tp->t_tfo_cookie.client;
21043                                         to.to_flags |= TOF_FASTOPEN;
21044                                         wanted_cookie = 1;
21045                                         /*
21046                                          * If we wind up having more data to
21047                                          * send with the SYN than can fit in
21048                                          * one segment, don't send any more
21049                                          * until the SYN|ACK comes back from
21050                                          * the other end.
21051                                          */
21052                                         sendalot = 0;
21053                                 }
21054                         }
21055                 }
21056                 /* Window scaling. */
21057                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
21058                         to.to_wscale = tp->request_r_scale;
21059                         to.to_flags |= TOF_SCALE;
21060                 }
21061                 /* Timestamps. */
21062                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
21063                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
21064                         to.to_tsval = ms_cts + tp->ts_offset;
21065                         to.to_tsecr = tp->ts_recent;
21066                         to.to_flags |= TOF_TS;
21067                 }
21068                 /* Set receive buffer autosizing timestamp. */
21069                 if (tp->rfbuf_ts == 0 &&
21070                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
21071                         tp->rfbuf_ts = tcp_ts_getticks();
21072                 /* Selective ACK's. */
21073                 if (tp->t_flags & TF_SACK_PERMIT) {
21074                         if (flags & TH_SYN)
21075                                 to.to_flags |= TOF_SACKPERM;
21076                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21077                                  tp->rcv_numsacks > 0) {
21078                                 to.to_flags |= TOF_SACK;
21079                                 to.to_nsacks = tp->rcv_numsacks;
21080                                 to.to_sacks = (u_char *)tp->sackblks;
21081                         }
21082                 }
21083 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21084                 /* TCP-MD5 (RFC2385). */
21085                 if (tp->t_flags & TF_SIGNATURE)
21086                         to.to_flags |= TOF_SIGNATURE;
21087 #endif                          /* TCP_SIGNATURE */
21088
21089                 /* Processing the options. */
21090                 hdrlen += optlen = tcp_addoptions(&to, opt);
21091                 /*
21092                  * If we wanted a TFO option to be added, but it was unable
21093                  * to fit, ensure no data is sent.
21094                  */
21095                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
21096                     !(to.to_flags & TOF_FASTOPEN))
21097                         len = 0;
21098         }
21099         if (tp->t_port) {
21100                 if (V_tcp_udp_tunneling_port == 0) {
21101                         /* The port was removed?? */
21102                         SOCKBUF_UNLOCK(&so->so_snd);
21103 #ifdef TCP_ACCOUNTING
21104                         crtsc = get_cyclecount();
21105                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21106                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
21107                         }
21108                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21109                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
21110                         }
21111                         sched_unpin();
21112 #endif
21113                         return (EHOSTUNREACH);
21114                 }
21115                 hdrlen += sizeof(struct udphdr);
21116         }
21117 #ifdef INET6
21118         if (isipv6)
21119                 ipoptlen = ip6_optlen(inp);
21120         else
21121 #endif
21122                 if (inp->inp_options)
21123                         ipoptlen = inp->inp_options->m_len -
21124                                 offsetof(struct ipoption, ipopt_list);
21125                 else
21126                         ipoptlen = 0;
21127 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21128         ipoptlen += ipsec_optlen;
21129 #endif
21130
21131         /*
21132          * Adjust data length if insertion of options will bump the packet
21133          * length beyond the t_maxseg length. Clear the FIN bit because we
21134          * cut off the tail of the segment.
21135          */
21136         if (len + optlen + ipoptlen > tp->t_maxseg) {
21137                 if (tso) {
21138                         uint32_t if_hw_tsomax;
21139                         uint32_t moff;
21140                         int32_t max_len;
21141
21142                         /* extract TSO information */
21143                         if_hw_tsomax = tp->t_tsomax;
21144                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
21145                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
21146                         KASSERT(ipoptlen == 0,
21147                                 ("%s: TSO can't do IP options", __func__));
21148
21149                         /*
21150                          * Check if we should limit by maximum payload
21151                          * length:
21152                          */
21153                         if (if_hw_tsomax != 0) {
21154                                 /* compute maximum TSO length */
21155                                 max_len = (if_hw_tsomax - hdrlen -
21156                                            max_linkhdr);
21157                                 if (max_len <= 0) {
21158                                         len = 0;
21159                                 } else if (len > max_len) {
21160                                         sendalot = 1;
21161                                         len = max_len;
21162                                         mark = 2;
21163                                 }
21164                         }
21165                         /*
21166                          * Prevent the last segment from being fractional
21167                          * unless the send sockbuf can be emptied:
21168                          */
21169                         max_len = (tp->t_maxseg - optlen);
21170                         if ((sb_offset + len) < sbavail(sb)) {
21171                                 moff = len % (u_int)max_len;
21172                                 if (moff != 0) {
21173                                         mark = 3;
21174                                         len -= moff;
21175                                 }
21176                         }
21177                         /*
21178                          * In case there are too many small fragments don't
21179                          * use TSO:
21180                          */
21181                         if (len <= max_len) {
21182                                 mark = 4;
21183                                 tso = 0;
21184                         }
21185                         /*
21186                          * Send the FIN in a separate segment after the bulk
21187                          * sending is done. We don't trust the TSO
21188                          * implementations to clear the FIN flag on all but
21189                          * the last segment.
21190                          */
21191                         if (tp->t_flags & TF_NEEDFIN) {
21192                                 sendalot = 4;
21193                         }
21194                 } else {
21195                         mark = 5;
21196                         if (optlen + ipoptlen >= tp->t_maxseg) {
21197                                 /*
21198                                  * Since we don't have enough space to put
21199                                  * the IP header chain and the TCP header in
21200                                  * one packet as required by RFC 7112, don't
21201                                  * send it. Also ensure that at least one
21202                                  * byte of the payload can be put into the
21203                                  * TCP segment.
21204                                  */
21205                                 SOCKBUF_UNLOCK(&so->so_snd);
21206                                 error = EMSGSIZE;
21207                                 sack_rxmit = 0;
21208                                 goto out;
21209                         }
21210                         len = tp->t_maxseg - optlen - ipoptlen;
21211                         sendalot = 5;
21212                 }
21213         } else {
21214                 tso = 0;
21215                 mark = 6;
21216         }
21217         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
21218                 ("%s: len > IP_MAXPACKET", __func__));
21219 #ifdef DIAGNOSTIC
21220 #ifdef INET6
21221         if (max_linkhdr + hdrlen > MCLBYTES)
21222 #else
21223                 if (max_linkhdr + hdrlen > MHLEN)
21224 #endif
21225                         panic("tcphdr too big");
21226 #endif
21227
21228         /*
21229          * This KASSERT is here to catch edge cases at a well defined place.
21230          * Before, those had triggered (random) panic conditions further
21231          * down.
21232          */
21233         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21234         if ((len == 0) &&
21235             (flags & TH_FIN) &&
21236             (sbused(sb))) {
21237                 /*
21238                  * We have outstanding data, don't send a fin by itself!.
21239                  */
21240                 goto just_return;
21241         }
21242         /*
21243          * Grab a header mbuf, attaching a copy of data to be transmitted,
21244          * and initialize the header from the template for sends on this
21245          * connection.
21246          */
21247         hw_tls = tp->t_nic_ktls_xmit != 0;
21248         if (len) {
21249                 uint32_t max_val;
21250                 uint32_t moff;
21251
21252                 if (rack->r_ctl.rc_pace_max_segs)
21253                         max_val = rack->r_ctl.rc_pace_max_segs;
21254                 else if (rack->rc_user_set_max_segs)
21255                         max_val = rack->rc_user_set_max_segs * segsiz;
21256                 else
21257                         max_val = len;
21258                 /*
21259                  * We allow a limit on sending with hptsi.
21260                  */
21261                 if (len > max_val) {
21262                         mark = 7;
21263                         len = max_val;
21264                 }
21265 #ifdef INET6
21266                 if (MHLEN < hdrlen + max_linkhdr)
21267                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
21268                 else
21269 #endif
21270                         m = m_gethdr(M_NOWAIT, MT_DATA);
21271
21272                 if (m == NULL) {
21273                         SOCKBUF_UNLOCK(sb);
21274                         error = ENOBUFS;
21275                         sack_rxmit = 0;
21276                         goto out;
21277                 }
21278                 m->m_data += max_linkhdr;
21279                 m->m_len = hdrlen;
21280
21281                 /*
21282                  * Start the m_copy functions from the closest mbuf to the
21283                  * sb_offset in the socket buffer chain.
21284                  */
21285                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
21286                 s_mb = mb;
21287                 s_moff = moff;
21288                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
21289                         m_copydata(mb, moff, (int)len,
21290                                    mtod(m, caddr_t)+hdrlen);
21291                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
21292                                 sbsndptr_adv(sb, mb, len);
21293                         m->m_len += len;
21294                 } else {
21295                         struct sockbuf *msb;
21296
21297                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
21298                                 msb = NULL;
21299                         else
21300                                 msb = sb;
21301                         m->m_next = tcp_m_copym(
21302                                 mb, moff, &len,
21303                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
21304                                 ((rsm == NULL) ? hw_tls : 0)
21305 #ifdef NETFLIX_COPY_ARGS
21306                                 , &s_mb, &s_moff
21307 #endif
21308                                 );
21309                         if (len <= (tp->t_maxseg - optlen)) {
21310                                 /*
21311                                  * Must have ran out of mbufs for the copy
21312                                  * shorten it to no longer need tso. Lets
21313                                  * not put on sendalot since we are low on
21314                                  * mbufs.
21315                                  */
21316                                 tso = 0;
21317                         }
21318                         if (m->m_next == NULL) {
21319                                 SOCKBUF_UNLOCK(sb);
21320                                 (void)m_free(m);
21321                                 error = ENOBUFS;
21322                                 sack_rxmit = 0;
21323                                 goto out;
21324                         }
21325                 }
21326                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
21327                         if (rsm && (rsm->r_flags & RACK_TLP)) {
21328                                 /*
21329                                  * TLP should not count in retran count, but
21330                                  * in its own bin
21331                                  */
21332                                 counter_u64_add(rack_tlp_retran, 1);
21333                                 counter_u64_add(rack_tlp_retran_bytes, len);
21334                         } else {
21335                                 tp->t_sndrexmitpack++;
21336                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
21337                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
21338                         }
21339 #ifdef STATS
21340                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
21341                                                  len);
21342 #endif
21343                 } else {
21344                         KMOD_TCPSTAT_INC(tcps_sndpack);
21345                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
21346 #ifdef STATS
21347                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
21348                                                  len);
21349 #endif
21350                 }
21351                 /*
21352                  * If we're sending everything we've got, set PUSH. (This
21353                  * will keep happy those implementations which only give
21354                  * data to the user when a buffer fills or a PUSH comes in.)
21355                  */
21356                 if (sb_offset + len == sbused(sb) &&
21357                     sbused(sb) &&
21358                     !(flags & TH_SYN)) {
21359                         flags |= TH_PUSH;
21360                         add_flag |= RACK_HAD_PUSH;
21361                 }
21362
21363                 SOCKBUF_UNLOCK(sb);
21364         } else {
21365                 SOCKBUF_UNLOCK(sb);
21366                 if (tp->t_flags & TF_ACKNOW)
21367                         KMOD_TCPSTAT_INC(tcps_sndacks);
21368                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
21369                         KMOD_TCPSTAT_INC(tcps_sndctrl);
21370                 else
21371                         KMOD_TCPSTAT_INC(tcps_sndwinup);
21372
21373                 m = m_gethdr(M_NOWAIT, MT_DATA);
21374                 if (m == NULL) {
21375                         error = ENOBUFS;
21376                         sack_rxmit = 0;
21377                         goto out;
21378                 }
21379 #ifdef INET6
21380                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
21381                     MHLEN >= hdrlen) {
21382                         M_ALIGN(m, hdrlen);
21383                 } else
21384 #endif
21385                         m->m_data += max_linkhdr;
21386                 m->m_len = hdrlen;
21387         }
21388         SOCKBUF_UNLOCK_ASSERT(sb);
21389         m->m_pkthdr.rcvif = (struct ifnet *)0;
21390 #ifdef MAC
21391         mac_inpcb_create_mbuf(inp, m);
21392 #endif
21393         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
21394 #ifdef INET6
21395                 if (isipv6)
21396                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
21397                 else
21398 #endif                          /* INET6 */
21399 #ifdef INET
21400                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
21401 #endif
21402                 th = rack->r_ctl.fsb.th;
21403                 udp = rack->r_ctl.fsb.udp;
21404                 if (udp) {
21405 #ifdef INET6
21406                         if (isipv6)
21407                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
21408                         else
21409 #endif                          /* INET6 */
21410                                 ulen = hdrlen + len - sizeof(struct ip);
21411                         udp->uh_ulen = htons(ulen);
21412                 }
21413         } else {
21414 #ifdef INET6
21415                 if (isipv6) {
21416                         ip6 = mtod(m, struct ip6_hdr *);
21417                         if (tp->t_port) {
21418                                 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
21419                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21420                                 udp->uh_dport = tp->t_port;
21421                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
21422                                 udp->uh_ulen = htons(ulen);
21423                                 th = (struct tcphdr *)(udp + 1);
21424                         } else
21425                                 th = (struct tcphdr *)(ip6 + 1);
21426                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
21427                 } else
21428 #endif                          /* INET6 */
21429                 {
21430 #ifdef INET
21431                         ip = mtod(m, struct ip *);
21432                         if (tp->t_port) {
21433                                 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
21434                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21435                                 udp->uh_dport = tp->t_port;
21436                                 ulen = hdrlen + len - sizeof(struct ip);
21437                                 udp->uh_ulen = htons(ulen);
21438                                 th = (struct tcphdr *)(udp + 1);
21439                         } else
21440                                 th = (struct tcphdr *)(ip + 1);
21441                         tcpip_fillheaders(inp, tp->t_port, ip, th);
21442 #endif
21443                 }
21444         }
21445         /*
21446          * Fill in fields, remembering maximum advertised window for use in
21447          * delaying messages about window sizes. If resending a FIN, be sure
21448          * not to use a new sequence number.
21449          */
21450         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
21451             tp->snd_nxt == tp->snd_max)
21452                 tp->snd_nxt--;
21453         /*
21454          * If we are starting a connection, send ECN setup SYN packet. If we
21455          * are on a retransmit, we may resend those bits a number of times
21456          * as per RFC 3168.
21457          */
21458         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
21459                 flags |= tcp_ecn_output_syn_sent(tp);
21460         }
21461         /* Also handle parallel SYN for ECN */
21462         if (TCPS_HAVERCVDSYN(tp->t_state) &&
21463             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
21464                 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
21465                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
21466                     (tp->t_flags2 & TF2_ECN_SND_ECE))
21467                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
21468 #ifdef INET6
21469                 if (isipv6) {
21470                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
21471                         ip6->ip6_flow |= htonl(ect << 20);
21472                 }
21473                 else
21474 #endif
21475                 {
21476 #ifdef INET
21477                         ip->ip_tos &= ~IPTOS_ECN_MASK;
21478                         ip->ip_tos |= ect;
21479 #endif
21480                 }
21481         }
21482         /*
21483          * If we are doing retransmissions, then snd_nxt will not reflect
21484          * the first unsent octet.  For ACK only packets, we do not want the
21485          * sequence number of the retransmitted packet, we want the sequence
21486          * number of the next unsent octet.  So, if there is no data (and no
21487          * SYN or FIN), use snd_max instead of snd_nxt when filling in
21488          * ti_seq.  But if we are in persist state, snd_max might reflect
21489          * one byte beyond the right edge of the window, so use snd_nxt in
21490          * that case, since we know we aren't doing a retransmission.
21491          * (retransmit and persist are mutually exclusive...)
21492          */
21493         if (sack_rxmit == 0) {
21494                 if (len || (flags & (TH_SYN | TH_FIN))) {
21495                         th->th_seq = htonl(tp->snd_nxt);
21496                         rack_seq = tp->snd_nxt;
21497                 } else {
21498                         th->th_seq = htonl(tp->snd_max);
21499                         rack_seq = tp->snd_max;
21500                 }
21501         } else {
21502                 th->th_seq = htonl(rsm->r_start);
21503                 rack_seq = rsm->r_start;
21504         }
21505         th->th_ack = htonl(tp->rcv_nxt);
21506         tcp_set_flags(th, flags);
21507         /*
21508          * Calculate receive window.  Don't shrink window, but avoid silly
21509          * window syndrome.
21510          * If a RST segment is sent, advertise a window of zero.
21511          */
21512         if (flags & TH_RST) {
21513                 recwin = 0;
21514         } else {
21515                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
21516                     recwin < (long)segsiz) {
21517                         recwin = 0;
21518                 }
21519                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
21520                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
21521                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
21522         }
21523
21524         /*
21525          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
21526          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
21527          * handled in syncache.
21528          */
21529         if (flags & TH_SYN)
21530                 th->th_win = htons((u_short)
21531                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
21532         else {
21533                 /* Avoid shrinking window with window scaling. */
21534                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
21535                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
21536         }
21537         /*
21538          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
21539          * window.  This may cause the remote transmitter to stall.  This
21540          * flag tells soreceive() to disable delayed acknowledgements when
21541          * draining the buffer.  This can occur if the receiver is
21542          * attempting to read more data than can be buffered prior to
21543          * transmitting on the connection.
21544          */
21545         if (th->th_win == 0) {
21546                 tp->t_sndzerowin++;
21547                 tp->t_flags |= TF_RXWIN0SENT;
21548         } else
21549                 tp->t_flags &= ~TF_RXWIN0SENT;
21550         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
21551         /* Now are we using fsb?, if so copy the template data to the mbuf */
21552         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21553                 uint8_t *cpto;
21554
21555                 cpto = mtod(m, uint8_t *);
21556                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
21557                 /*
21558                  * We have just copied in:
21559                  * IP/IP6
21560                  * <optional udphdr>
21561                  * tcphdr (no options)
21562                  *
21563                  * We need to grab the correct pointers into the mbuf
21564                  * for both the tcp header, and possibly the udp header (if tunneling).
21565                  * We do this by using the offset in the copy buffer and adding it
21566                  * to the mbuf base pointer (cpto).
21567                  */
21568 #ifdef INET6
21569                 if (isipv6)
21570                         ip6 = mtod(m, struct ip6_hdr *);
21571                 else
21572 #endif                          /* INET6 */
21573 #ifdef INET
21574                         ip = mtod(m, struct ip *);
21575 #endif
21576                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
21577                 /* If we have a udp header lets set it into the mbuf as well */
21578                 if (udp)
21579                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
21580         }
21581 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21582         if (to.to_flags & TOF_SIGNATURE) {
21583                 /*
21584                  * Calculate MD5 signature and put it into the place
21585                  * determined before.
21586                  * NOTE: since TCP options buffer doesn't point into
21587                  * mbuf's data, calculate offset and use it.
21588                  */
21589                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
21590                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
21591                         /*
21592                          * Do not send segment if the calculation of MD5
21593                          * digest has failed.
21594                          */
21595                         goto out;
21596                 }
21597         }
21598 #endif
21599         if (optlen) {
21600                 bcopy(opt, th + 1, optlen);
21601                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
21602         }
21603         /*
21604          * Put TCP length in extended header, and then checksum extended
21605          * header and data.
21606          */
21607         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
21608 #ifdef INET6
21609         if (isipv6) {
21610                 /*
21611                  * ip6_plen is not need to be filled now, and will be filled
21612                  * in ip6_output.
21613                  */
21614                 if (tp->t_port) {
21615                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
21616                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21617                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
21618                         th->th_sum = htons(0);
21619                         UDPSTAT_INC(udps_opackets);
21620                 } else {
21621                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
21622                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21623                         th->th_sum = in6_cksum_pseudo(ip6,
21624                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
21625                                                       0);
21626                 }
21627         }
21628 #endif
21629 #if defined(INET6) && defined(INET)
21630         else
21631 #endif
21632 #ifdef INET
21633         {
21634                 if (tp->t_port) {
21635                         m->m_pkthdr.csum_flags = CSUM_UDP;
21636                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21637                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
21638                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
21639                         th->th_sum = htons(0);
21640                         UDPSTAT_INC(udps_opackets);
21641                 } else {
21642                         m->m_pkthdr.csum_flags = CSUM_TCP;
21643                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21644                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
21645                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
21646                                                                         IPPROTO_TCP + len + optlen));
21647                 }
21648                 /* IP version must be set here for ipv4/ipv6 checking later */
21649                 KASSERT(ip->ip_v == IPVERSION,
21650                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
21651         }
21652 #endif
21653         /*
21654          * Enable TSO and specify the size of the segments. The TCP pseudo
21655          * header checksum is always provided. XXX: Fixme: This is currently
21656          * not the case for IPv6.
21657          */
21658         if (tso) {
21659                 /*
21660                  * Here we must use t_maxseg and the optlen since
21661                  * the optlen may include SACK's (or DSACK).
21662                  */
21663                 KASSERT(len > tp->t_maxseg - optlen,
21664                         ("%s: len <= tso_segsz", __func__));
21665                 m->m_pkthdr.csum_flags |= CSUM_TSO;
21666                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
21667         }
21668         KASSERT(len + hdrlen == m_length(m, NULL),
21669                 ("%s: mbuf chain different than expected: %d + %u != %u",
21670                  __func__, len, hdrlen, m_length(m, NULL)));
21671
21672 #ifdef TCP_HHOOK
21673         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
21674         hhook_run_tcp_est_out(tp, th, &to, len, tso);
21675 #endif
21676         if ((rack->r_ctl.crte != NULL) &&
21677             (rack->rc_hw_nobuf == 0) &&
21678             tcp_bblogging_on(tp)) {
21679                 rack_log_queue_level(tp, rack, len, &tv, cts);
21680         }
21681         /* We're getting ready to send; log now. */
21682         if (tcp_bblogging_on(rack->rc_tp)) {
21683                 union tcp_log_stackspecific log;
21684
21685                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
21686                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
21687                 if (rack->rack_no_prr)
21688                         log.u_bbr.flex1 = 0;
21689                 else
21690                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
21691                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
21692                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
21693                 log.u_bbr.flex4 = orig_len;
21694                 /* Save off the early/late values */
21695                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
21696                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
21697                 log.u_bbr.bw_inuse = rack_get_bw(rack);
21698                 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
21699                 log.u_bbr.flex8 = 0;
21700                 if (rsm) {
21701                         if (rsm->r_flags & RACK_RWND_COLLAPSED) {
21702                                 rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
21703                                 counter_u64_add(rack_collapsed_win_rxt, 1);
21704                                 counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
21705                         }
21706                         if (doing_tlp)
21707                                 log.u_bbr.flex8 = 2;
21708                         else
21709                                 log.u_bbr.flex8 = 1;
21710                 } else {
21711                         if (doing_tlp)
21712                                 log.u_bbr.flex8 = 3;
21713                 }
21714                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
21715                 log.u_bbr.flex7 = mark;
21716                 log.u_bbr.flex7 <<= 8;
21717                 log.u_bbr.flex7 |= pass;
21718                 log.u_bbr.pkts_out = tp->t_maxseg;
21719                 log.u_bbr.timeStamp = cts;
21720                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
21721                 if (rsm && (rsm->r_rtr_cnt > 0)) {
21722                         /*
21723                          * When we have a retransmit we want to log the
21724                          * burst at send and flight at send from before.
21725                          */
21726                         log.u_bbr.flex5 = rsm->r_fas;
21727                         log.u_bbr.bbr_substate = rsm->r_bas;
21728                 } else {
21729                         /*
21730                          * New transmits we log in flex5 the inflight again as
21731                          * well as the number of segments in our send in the
21732                          * substate field.
21733                          */
21734                         log.u_bbr.flex5 = log.u_bbr.inflight;
21735                         log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
21736                 }
21737                 log.u_bbr.lt_epoch = cwnd_to_use;
21738                 log.u_bbr.delivered = sendalot;
21739                 log.u_bbr.rttProp = (uint64_t)rsm;
21740                 log.u_bbr.pkt_epoch = __LINE__;
21741                 if (rsm) {
21742                         log.u_bbr.delRate = rsm->r_flags;
21743                         log.u_bbr.delRate <<= 31;
21744                         log.u_bbr.delRate |= rack->r_must_retran;
21745                         log.u_bbr.delRate <<= 1;
21746                         log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21747                 } else {
21748                         log.u_bbr.delRate = rack->r_must_retran;
21749                         log.u_bbr.delRate <<= 1;
21750                         log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21751                 }
21752                 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
21753                                     len, &log, false, NULL, __func__, __LINE__, &tv);
21754         } else
21755                 lgb = NULL;
21756
21757         /*
21758          * Fill in IP length and desired time to live and send to IP level.
21759          * There should be a better way to handle ttl and tos; we could keep
21760          * them in the template, but need a way to checksum without them.
21761          */
21762         /*
21763          * m->m_pkthdr.len should have been set before cksum calcuration,
21764          * because in6_cksum() need it.
21765          */
21766 #ifdef INET6
21767         if (isipv6) {
21768                 /*
21769                  * we separately set hoplimit for every segment, since the
21770                  * user might want to change the value via setsockopt. Also,
21771                  * desired default hop limit might be changed via Neighbor
21772                  * Discovery.
21773                  */
21774                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
21775
21776                 /*
21777                  * Set the packet size here for the benefit of DTrace
21778                  * probes. ip6_output() will set it properly; it's supposed
21779                  * to include the option header lengths as well.
21780                  */
21781                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
21782
21783                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
21784                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
21785                 else
21786                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
21787
21788                 if (tp->t_state == TCPS_SYN_SENT)
21789                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
21790
21791                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
21792                 /* TODO: IPv6 IP6TOS_ECT bit on */
21793                 error = ip6_output(m,
21794 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21795                                    inp->in6p_outputopts,
21796 #else
21797                                    NULL,
21798 #endif
21799                                    &inp->inp_route6,
21800                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
21801                                    NULL, NULL, inp);
21802
21803                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
21804                         mtu = inp->inp_route6.ro_nh->nh_mtu;
21805         }
21806 #endif                          /* INET6 */
21807 #if defined(INET) && defined(INET6)
21808         else
21809 #endif
21810 #ifdef INET
21811         {
21812                 ip->ip_len = htons(m->m_pkthdr.len);
21813 #ifdef INET6
21814                 if (inp->inp_vflag & INP_IPV6PROTO)
21815                         ip->ip_ttl = in6_selecthlim(inp, NULL);
21816 #endif                          /* INET6 */
21817                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
21818                 /*
21819                  * If we do path MTU discovery, then we set DF on every
21820                  * packet. This might not be the best thing to do according
21821                  * to RFC3390 Section 2. However the tcp hostcache migitates
21822                  * the problem so it affects only the first tcp connection
21823                  * with a host.
21824                  *
21825                  * NB: Don't set DF on small MTU/MSS to have a safe
21826                  * fallback.
21827                  */
21828                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
21829                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
21830                         if (tp->t_port == 0 || len < V_tcp_minmss) {
21831                                 ip->ip_off |= htons(IP_DF);
21832                         }
21833                 } else {
21834                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
21835                 }
21836
21837                 if (tp->t_state == TCPS_SYN_SENT)
21838                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
21839
21840                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
21841
21842                 error = ip_output(m,
21843 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21844                                   inp->inp_options,
21845 #else
21846                                   NULL,
21847 #endif
21848                                   &inp->inp_route,
21849                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
21850                                   inp);
21851                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
21852                         mtu = inp->inp_route.ro_nh->nh_mtu;
21853         }
21854 #endif                          /* INET */
21855
21856 out:
21857         if (lgb) {
21858                 lgb->tlb_errno = error;
21859                 lgb = NULL;
21860         }
21861         /*
21862          * In transmit state, time the transmission and arrange for the
21863          * retransmit.  In persist state, just set snd_max.
21864          */
21865         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
21866                         rack_to_usec_ts(&tv),
21867                         rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
21868         if (error == 0) {
21869                 if (rsm == NULL) {
21870                         if (rack->lt_bw_up == 0) {
21871                                 rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
21872                                 rack->r_ctl.lt_seq = tp->snd_una;
21873                                 rack->lt_bw_up = 1;
21874                         } else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
21875                                 /*
21876                                  * Need to record what we have since we are
21877                                  * approaching seq wrap.
21878                                  */
21879                                 uint64_t tmark;
21880
21881                                 rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
21882                                 rack->r_ctl.lt_seq = tp->snd_una;
21883                                 tmark = tcp_tv_to_lusectick(&tv);
21884                                 rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
21885                                 rack->r_ctl.lt_timemark = tmark;
21886                         }
21887                 }
21888                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
21889                 counter_u64_add(rack_total_bytes, len);
21890                 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
21891                 if (rsm && doing_tlp) {
21892                         rack->rc_last_sent_tlp_past_cumack = 0;
21893                         rack->rc_last_sent_tlp_seq_valid = 1;
21894                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
21895                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
21896                 }
21897                 if (rack->rc_hw_nobuf) {
21898                         rack->rc_hw_nobuf = 0;
21899                         rack->r_ctl.rc_agg_delayed = 0;
21900                         rack->r_early = 0;
21901                         rack->r_late = 0;
21902                         rack->r_ctl.rc_agg_early = 0;
21903                 }
21904                 if (rsm && (doing_tlp == 0)) {
21905                         /* Set we retransmitted */
21906                         rack->rc_gp_saw_rec = 1;
21907                 } else {
21908                         if (cwnd_to_use > tp->snd_ssthresh) {
21909                                 /* Set we sent in CA */
21910                                 rack->rc_gp_saw_ca = 1;
21911                         } else {
21912                                 /* Set we sent in SS */
21913                                 rack->rc_gp_saw_ss = 1;
21914                         }
21915                 }
21916                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21917                     (tp->t_flags & TF_SACK_PERMIT) &&
21918                     tp->rcv_numsacks > 0)
21919                         tcp_clean_dsack_blocks(tp);
21920                 tot_len_this_send += len;
21921                 if (len == 0) {
21922                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
21923                 } else {
21924                         int idx;
21925
21926                         idx = (len / segsiz) + 3;
21927                         if (idx >= TCP_MSS_ACCT_ATIMER)
21928                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
21929                         else
21930                                 counter_u64_add(rack_out_size[idx], 1);
21931                 }
21932         }
21933         if ((rack->rack_no_prr == 0) &&
21934             sub_from_prr &&
21935             (error == 0)) {
21936                 if (rack->r_ctl.rc_prr_sndcnt >= len)
21937                         rack->r_ctl.rc_prr_sndcnt -= len;
21938                 else
21939                         rack->r_ctl.rc_prr_sndcnt = 0;
21940         }
21941         sub_from_prr = 0;
21942         if (doing_tlp) {
21943                 /* Make sure the TLP is added */
21944                 add_flag |= RACK_TLP;
21945         } else if (rsm) {
21946                 /* If its a resend without TLP then it must not have the flag */
21947                 rsm->r_flags &= ~RACK_TLP;
21948         }
21949
21950
21951         if ((error == 0) &&
21952             (len > 0) &&
21953             (tp->snd_una == tp->snd_max))
21954                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
21955         {
21956                 tcp_seq startseq = tp->snd_nxt;
21957
21958                 /* Track our lost count */
21959                 if (rsm && (doing_tlp == 0))
21960                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
21961                 /*
21962                  * Advance snd_nxt over sequence space of this segment.
21963                  */
21964                 if (error)
21965                         /* We don't log or do anything with errors */
21966                         goto nomore;
21967                 if (doing_tlp == 0) {
21968                         if (rsm == NULL) {
21969                                 /*
21970                                  * Not a retransmission of some
21971                                  * sort, new data is going out so
21972                                  * clear our TLP count and flag.
21973                                  */
21974                                 rack->rc_tlp_in_progress = 0;
21975                                 rack->r_ctl.rc_tlp_cnt_out = 0;
21976                         }
21977                 } else {
21978                         /*
21979                          * We have just sent a TLP, mark that it is true
21980                          * and make sure our in progress is set so we
21981                          * continue to check the count.
21982                          */
21983                         rack->rc_tlp_in_progress = 1;
21984                         rack->r_ctl.rc_tlp_cnt_out++;
21985                 }
21986                 if (flags & (TH_SYN | TH_FIN)) {
21987                         if (flags & TH_SYN)
21988                                 tp->snd_nxt++;
21989                         if (flags & TH_FIN) {
21990                                 tp->snd_nxt++;
21991                                 tp->t_flags |= TF_SENTFIN;
21992                         }
21993                 }
21994                 /* In the ENOBUFS case we do *not* update snd_max */
21995                 if (sack_rxmit)
21996                         goto nomore;
21997
21998                 tp->snd_nxt += len;
21999                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
22000                         if (tp->snd_una == tp->snd_max) {
22001                                 /*
22002                                  * Update the time we just added data since
22003                                  * none was outstanding.
22004                                  */
22005                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
22006                                 tp->t_acktime = ticks;
22007                         }
22008                         tp->snd_max = tp->snd_nxt;
22009                         if (rack->rc_new_rnd_needed) {
22010                                 /*
22011                                  * Update the rnd to start ticking not
22012                                  * that from a time perspective all of
22013                                  * the preceding idle time is "in the round"
22014                                  */
22015                                 rack->rc_new_rnd_needed = 0;
22016                                 rack->r_ctl.roundends = tp->snd_max;
22017                         }
22018                         /*
22019                          * Time this transmission if not a retransmission and
22020                          * not currently timing anything.
22021                          * This is only relevant in case of switching back to
22022                          * the base stack.
22023                          */
22024                         if (tp->t_rtttime == 0) {
22025                                 tp->t_rtttime = ticks;
22026                                 tp->t_rtseq = startseq;
22027                                 KMOD_TCPSTAT_INC(tcps_segstimed);
22028                         }
22029                         if (len &&
22030                             ((tp->t_flags & TF_GPUTINPROG) == 0))
22031                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
22032                 }
22033                 /*
22034                  * If we are doing FO we need to update the mbuf position and subtract
22035                  * this happens when the peer sends us duplicate information and
22036                  * we thus want to send a DSACK.
22037                  *
22038                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
22039                  * turned off? If not then we are going to echo multiple DSACK blocks
22040                  * out (with the TSO), which we should not be doing.
22041                  */
22042                 if (rack->r_fast_output && len) {
22043                         if (rack->r_ctl.fsb.left_to_send > len)
22044                                 rack->r_ctl.fsb.left_to_send -= len;
22045                         else
22046                                 rack->r_ctl.fsb.left_to_send = 0;
22047                         if (rack->r_ctl.fsb.left_to_send < segsiz)
22048                                 rack->r_fast_output = 0;
22049                         if (rack->r_fast_output) {
22050                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
22051                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
22052                                 rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
22053                         }
22054                 }
22055         }
22056 nomore:
22057         if (error) {
22058                 rack->r_ctl.rc_agg_delayed = 0;
22059                 rack->r_early = 0;
22060                 rack->r_late = 0;
22061                 rack->r_ctl.rc_agg_early = 0;
22062                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
22063                 /*
22064                  * Failures do not advance the seq counter above. For the
22065                  * case of ENOBUFS we will fall out and retry in 1ms with
22066                  * the hpts. Everything else will just have to retransmit
22067                  * with the timer.
22068                  *
22069                  * In any case, we do not want to loop around for another
22070                  * send without a good reason.
22071                  */
22072                 sendalot = 0;
22073                 switch (error) {
22074                 case EPERM:
22075                         tp->t_softerror = error;
22076 #ifdef TCP_ACCOUNTING
22077                         crtsc = get_cyclecount();
22078                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22079                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22080                         }
22081                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22082                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22083                         }
22084                         sched_unpin();
22085 #endif
22086                         return (error);
22087                 case ENOBUFS:
22088                         /*
22089                          * Pace us right away to retry in a some
22090                          * time
22091                          */
22092                         if (rack->r_ctl.crte != NULL) {
22093                                 tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
22094                                 if (tcp_bblogging_on(rack->rc_tp))
22095                                         rack_log_queue_level(tp, rack, len, &tv, cts);
22096                         } else
22097                                 tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
22098                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
22099                         if (rack->rc_enobuf < 0x7f)
22100                                 rack->rc_enobuf++;
22101                         if (slot < (10 * HPTS_USEC_IN_MSEC))
22102                                 slot = 10 * HPTS_USEC_IN_MSEC;
22103                         if (rack->r_ctl.crte != NULL) {
22104                                 counter_u64_add(rack_saw_enobuf_hw, 1);
22105                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
22106                         }
22107                         counter_u64_add(rack_saw_enobuf, 1);
22108                         goto enobufs;
22109                 case EMSGSIZE:
22110                         /*
22111                          * For some reason the interface we used initially
22112                          * to send segments changed to another or lowered
22113                          * its MTU. If TSO was active we either got an
22114                          * interface without TSO capabilits or TSO was
22115                          * turned off. If we obtained mtu from ip_output()
22116                          * then update it and try again.
22117                          */
22118                         if (tso)
22119                                 tp->t_flags &= ~TF_TSO;
22120                         if (mtu != 0) {
22121                                 int saved_mtu;
22122
22123                                 saved_mtu = tp->t_maxseg;
22124                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
22125                                 if (saved_mtu > tp->t_maxseg) {
22126                                         goto again;
22127                                 }
22128                         }
22129                         slot = 10 * HPTS_USEC_IN_MSEC;
22130                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22131 #ifdef TCP_ACCOUNTING
22132                         crtsc = get_cyclecount();
22133                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22134                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22135                         }
22136                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22137                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22138                         }
22139                         sched_unpin();
22140 #endif
22141                         return (error);
22142                 case ENETUNREACH:
22143                         counter_u64_add(rack_saw_enetunreach, 1);
22144                 case EHOSTDOWN:
22145                 case EHOSTUNREACH:
22146                 case ENETDOWN:
22147                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
22148                                 tp->t_softerror = error;
22149                         }
22150                         /* FALLTHROUGH */
22151                 default:
22152                         slot = 10 * HPTS_USEC_IN_MSEC;
22153                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22154 #ifdef TCP_ACCOUNTING
22155                         crtsc = get_cyclecount();
22156                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22157                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22158                         }
22159                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22160                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22161                         }
22162                         sched_unpin();
22163 #endif
22164                         return (error);
22165                 }
22166         } else {
22167                 rack->rc_enobuf = 0;
22168                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
22169                         rack->r_ctl.retran_during_recovery += len;
22170         }
22171         KMOD_TCPSTAT_INC(tcps_sndtotal);
22172
22173         /*
22174          * Data sent (as far as we can tell). If this advertises a larger
22175          * window than any other segment, then remember the size of the
22176          * advertised window. Any pending ACK has now been sent.
22177          */
22178         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
22179                 tp->rcv_adv = tp->rcv_nxt + recwin;
22180
22181         tp->last_ack_sent = tp->rcv_nxt;
22182         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
22183 enobufs:
22184         if (sendalot) {
22185                 /* Do we need to turn off sendalot? */
22186                 if (rack->r_ctl.rc_pace_max_segs &&
22187                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
22188                         /* We hit our max. */
22189                         sendalot = 0;
22190                 } else if ((rack->rc_user_set_max_segs) &&
22191                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
22192                         /* We hit the user defined max */
22193                         sendalot = 0;
22194                 }
22195         }
22196         if ((error == 0) && (flags & TH_FIN))
22197                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
22198         if (flags & TH_RST) {
22199                 /*
22200                  * We don't send again after sending a RST.
22201                  */
22202                 slot = 0;
22203                 sendalot = 0;
22204                 if (error == 0)
22205                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
22206         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
22207                 /*
22208                  * Get our pacing rate, if an error
22209                  * occurred in sending (ENOBUF) we would
22210                  * hit the else if with slot preset. Other
22211                  * errors return.
22212                  */
22213                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
22214         }
22215         if (rsm &&
22216             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
22217             rack->use_rack_rr) {
22218                 /* Its a retransmit and we use the rack cheat? */
22219                 if ((slot == 0) ||
22220                     (rack->rc_always_pace == 0) ||
22221                     (rack->r_rr_config == 1)) {
22222                         /*
22223                          * We have no pacing set or we
22224                          * are using old-style rack or
22225                          * we are overridden to use the old 1ms pacing.
22226                          */
22227                         slot = rack->r_ctl.rc_min_to;
22228                 }
22229         }
22230         /* We have sent clear the flag */
22231         rack->r_ent_rec_ns = 0;
22232         if (rack->r_must_retran) {
22233                 if (rsm) {
22234                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
22235                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
22236                                 /*
22237                                  * We have retransmitted all.
22238                                  */
22239                                 rack->r_must_retran = 0;
22240                                 rack->r_ctl.rc_out_at_rto = 0;
22241                         }
22242                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22243                         /*
22244                          * Sending new data will also kill
22245                          * the loop.
22246                          */
22247                         rack->r_must_retran = 0;
22248                         rack->r_ctl.rc_out_at_rto = 0;
22249                 }
22250         }
22251         rack->r_ctl.fsb.recwin = recwin;
22252         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
22253             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22254                 /*
22255                  * We hit an RTO and now have past snd_max at the RTO
22256                  * clear all the WAS flags.
22257                  */
22258                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
22259         }
22260         if (slot) {
22261                 /* set the rack tcb into the slot N */
22262                 if ((error == 0) &&
22263                     rack_use_rfo &&
22264                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
22265                     (rsm == NULL) &&
22266                     (tp->snd_nxt == tp->snd_max) &&
22267                     (ipoptlen == 0) &&
22268                     (tp->rcv_numsacks == 0) &&
22269                     rack->r_fsb_inited &&
22270                     TCPS_HAVEESTABLISHED(tp->t_state) &&
22271                     ((IN_RECOVERY(tp->t_flags)) == 0) &&
22272                     (rack->r_must_retran == 0) &&
22273                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
22274                     (len > 0) && (orig_len > 0) &&
22275                     (orig_len > len) &&
22276                     ((orig_len - len) >= segsiz) &&
22277                     ((optlen == 0) ||
22278                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22279                         /* We can send at least one more MSS using our fsb */
22280                         rack_setup_fast_output(tp, rack, sb, len, orig_len,
22281                                                segsiz, pace_max_seg, hw_tls, flags);
22282                 } else
22283                         rack->r_fast_output = 0;
22284                 rack_log_fsb(rack, tp, so, flags,
22285                              ipoptlen, orig_len, len, error,
22286                              (rsm == NULL), optlen, __LINE__, 2);
22287         } else if (sendalot) {
22288                 int ret;
22289
22290                 sack_rxmit = 0;
22291                 if ((error == 0) &&
22292                     rack_use_rfo &&
22293                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
22294                     (rsm == NULL) &&
22295                     (ipoptlen == 0) &&
22296                     (tp->rcv_numsacks == 0) &&
22297                     (tp->snd_nxt == tp->snd_max) &&
22298                     (rack->r_must_retran == 0) &&
22299                     rack->r_fsb_inited &&
22300                     TCPS_HAVEESTABLISHED(tp->t_state) &&
22301                     ((IN_RECOVERY(tp->t_flags)) == 0) &&
22302                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
22303                     (len > 0) && (orig_len > 0) &&
22304                     (orig_len > len) &&
22305                     ((orig_len - len) >= segsiz) &&
22306                     ((optlen == 0) ||
22307                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22308                         /* we can use fast_output for more */
22309                         rack_setup_fast_output(tp, rack, sb, len, orig_len,
22310                                                segsiz, pace_max_seg, hw_tls, flags);
22311                         if (rack->r_fast_output) {
22312                                 error = 0;
22313                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
22314                                 if (ret >= 0)
22315                                         return (ret);
22316                                 else if (error)
22317                                         goto nomore;
22318
22319                         }
22320                 }
22321                 goto again;
22322         }
22323         /* Assure when we leave that snd_nxt will point to top */
22324 skip_all_send:
22325         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22326                 tp->snd_nxt = tp->snd_max;
22327         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
22328 #ifdef TCP_ACCOUNTING
22329         crtsc = get_cyclecount() - ts_val;
22330         if (tot_len_this_send) {
22331                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22332                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
22333                 }
22334                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22335                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
22336                 }
22337                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22338                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
22339                 }
22340         } else {
22341                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22342                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
22343                 }
22344                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22345                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
22346                 }
22347         }
22348         sched_unpin();
22349 #endif
22350         if (error == ENOBUFS)
22351                 error = 0;
22352         return (error);
22353 }
22354
22355 static void
22356 rack_update_seg(struct tcp_rack *rack)
22357 {
22358         uint32_t orig_val;
22359
22360         orig_val = rack->r_ctl.rc_pace_max_segs;
22361         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
22362         if (orig_val != rack->r_ctl.rc_pace_max_segs)
22363                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
22364 }
22365
22366 static void
22367 rack_mtu_change(struct tcpcb *tp)
22368 {
22369         /*
22370          * The MSS may have changed
22371          */
22372         struct tcp_rack *rack;
22373         struct rack_sendmap *rsm;
22374
22375         rack = (struct tcp_rack *)tp->t_fb_ptr;
22376         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
22377                 /*
22378                  * The MTU has changed we need to resend everything
22379                  * since all we have sent is lost. We first fix
22380                  * up the mtu though.
22381                  */
22382                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
22383                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
22384                 rack_remxt_tmr(tp);
22385                 rack->r_fast_output = 0;
22386                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
22387                                                 rack->r_ctl.rc_sacked);
22388                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
22389                 rack->r_must_retran = 1;
22390                 /* Mark all inflight to needing to be rxt'd */
22391                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
22392                         rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
22393                 }
22394         }
22395         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
22396         /* We don't use snd_nxt to retransmit */
22397         tp->snd_nxt = tp->snd_max;
22398 }
22399
22400 static int
22401 rack_set_dgp(struct tcp_rack *rack)
22402 {
22403         /* pace_always=1 */
22404         if (rack->rc_always_pace == 0) {
22405                 if (tcp_can_enable_pacing() == 0)
22406                         return (EBUSY);
22407         }
22408         rack->dgp_on = 1;
22409         rack->rc_always_pace = 1;
22410         rack->use_fixed_rate = 0;
22411         if (rack->gp_ready)
22412                 rack_set_cc_pacing(rack);
22413         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
22414         rack->rack_attempt_hdwr_pace = 0;
22415         /* rxt settings */
22416         rack->full_size_rxt = 1;
22417         rack->shape_rxt_to_pacing_min  = 0;
22418         /* cmpack=1 */
22419         rack->r_use_cmp_ack = 1;
22420         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
22421             rack->r_use_cmp_ack)
22422                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
22423         /* scwnd=1 */
22424         rack->rack_enable_scwnd = 1;
22425         /* dynamic=100 */
22426         rack->rc_gp_dyn_mul = 1;
22427         /* gp_inc_ca */
22428         rack->r_ctl.rack_per_of_gp_ca = 100;
22429         /* rrr_conf=3 */
22430         rack->r_rr_config = 3;
22431         /* npush=2 */
22432         rack->r_ctl.rc_no_push_at_mrtt = 2;
22433         /* fillcw=1 */
22434         if (rack->r_cwnd_was_clamped == 0) {
22435                 rack->rc_pace_to_cwnd = 1;
22436         } else {
22437                 rack->rc_pace_to_cwnd = 0;
22438                 /* Reset all multipliers to 100.0 so just the measured bw */
22439                 rack->r_ctl.rack_per_of_gp_ss = 100;
22440                 rack->r_ctl.rack_per_of_gp_ca = 100;
22441         }
22442         rack->rc_pace_fill_if_rttin_range = 0;
22443         rack->rtt_limit_mul = 0;
22444         /* noprr=1 */
22445         rack->rack_no_prr = 1;
22446         /* lscwnd=1 */
22447         rack->r_limit_scw = 1;
22448         /* gp_inc_rec */
22449         rack->r_ctl.rack_per_of_gp_rec = 90;
22450         rack_client_buffer_level_set(rack);
22451         return (0);
22452 }
22453
22454
22455
22456 static int
22457 rack_set_profile(struct tcp_rack *rack, int prof)
22458 {
22459         int err = EINVAL;
22460         if (prof == 1) {
22461                 /*
22462                  * Profile 1 is "standard" DGP. It ignores
22463                  * client buffer level.
22464                  */
22465                 rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL0;
22466                 err = rack_set_dgp(rack);
22467                 if (err)
22468                         return (err);
22469         } else if (prof == 2) {
22470                 /*
22471                  * Profile 2 is DGP. Less aggressive with
22472                  * respect to client buffer level.
22473                  */
22474                 rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL1;
22475                 err = rack_set_dgp(rack);
22476                 if (err)
22477                         return (err);
22478         } else if (prof == 3) {
22479                 /*
22480                  * Profile 3 is DGP. Even Less aggressive with
22481                  * respect to client buffer level.
22482                  */
22483                 rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL2;
22484                 err = rack_set_dgp(rack);
22485                 if (err)
22486                         return (err);
22487         } else if (prof == 4) {
22488                 /*
22489                  * Profile 4 is DGP with the most responsiveness
22490                  * to client buffer level.
22491                  */
22492                 rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL3;
22493                 err = rack_set_dgp(rack);
22494                 if (err)
22495                         return (err);
22496         } else if (prof == 0) {
22497                 /* This changes things back to the default settings */
22498                 rack->dgp_on = 0;
22499                 rack->rc_hybrid_mode = 0;
22500                 err = 0;
22501                 if (rack_fill_cw_state)
22502                         rack->rc_pace_to_cwnd = 1;
22503                 else
22504                         rack->rc_pace_to_cwnd = 0;
22505                 if (rack->rc_always_pace) {
22506                         tcp_decrement_paced_conn();
22507                         rack_undo_cc_pacing(rack);
22508                         rack->rc_always_pace = 0;
22509                 }
22510                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
22511                         rack->rc_always_pace = 1;
22512                         if ((rack->gp_ready) && (rack->use_fixed_rate == 0))
22513                                 rack_set_cc_pacing(rack);
22514                 } else
22515                         rack->rc_always_pace = 0;
22516                 if (rack_dsack_std_based & 0x1) {
22517                         /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
22518                         rack->rc_rack_tmr_std_based = 1;
22519                 }
22520                 if (rack_dsack_std_based & 0x2) {
22521                         /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
22522                         rack->rc_rack_use_dsack = 1;
22523                 }
22524                 if (rack_use_cmp_acks)
22525                         rack->r_use_cmp_ack = 1;
22526                 else
22527                         rack->r_use_cmp_ack = 0;
22528                 if (rack_disable_prr)
22529                         rack->rack_no_prr = 1;
22530                 else
22531                         rack->rack_no_prr = 0;
22532                 if (rack_gp_no_rec_chg)
22533                         rack->rc_gp_no_rec_chg = 1;
22534                 else
22535                         rack->rc_gp_no_rec_chg = 0;
22536                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
22537                         rack->r_mbuf_queue = 1;
22538                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
22539                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
22540                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
22541                 } else {
22542                         rack->r_mbuf_queue = 0;
22543                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
22544                 }
22545                 if (rack_enable_shared_cwnd)
22546                         rack->rack_enable_scwnd = 1;
22547                 else
22548                         rack->rack_enable_scwnd = 0;
22549                 if (rack_do_dyn_mul) {
22550                         /* When dynamic adjustment is on CA needs to start at 100% */
22551                         rack->rc_gp_dyn_mul = 1;
22552                         if (rack_do_dyn_mul >= 100)
22553                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
22554                 } else {
22555                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
22556                         rack->rc_gp_dyn_mul = 0;
22557                 }
22558                 rack->r_rr_config = 0;
22559                 rack->r_ctl.rc_no_push_at_mrtt = 0;
22560                 rack->rc_pace_to_cwnd = 0;
22561                 rack->rc_pace_fill_if_rttin_range = 0;
22562                 rack->rtt_limit_mul = 0;
22563
22564                 if (rack_enable_hw_pacing)
22565                         rack->rack_hdw_pace_ena = 1;
22566                 else
22567                         rack->rack_hdw_pace_ena = 0;
22568                 if (rack_disable_prr)
22569                         rack->rack_no_prr = 1;
22570                 else
22571                         rack->rack_no_prr = 0;
22572                 if (rack_limits_scwnd)
22573                         rack->r_limit_scw  = 1;
22574                 else
22575                         rack->r_limit_scw  = 0;
22576                 rack_init_retransmit_value(rack, rack_rxt_controls);
22577                 err = 0;
22578         }
22579         return (err);
22580 }
22581
22582 static int
22583 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
22584 {
22585         struct deferred_opt_list *dol;
22586
22587         dol = malloc(sizeof(struct deferred_opt_list),
22588                      M_TCPFSB, M_NOWAIT|M_ZERO);
22589         if (dol == NULL) {
22590                 /*
22591                  * No space yikes -- fail out..
22592                  */
22593                 return (0);
22594         }
22595         dol->optname = sopt_name;
22596         dol->optval = loptval;
22597         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
22598         return (1);
22599 }
22600
22601 static int
22602 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
22603 {
22604 #ifdef TCP_REQUEST_TRK
22605         struct http_sendfile_track *sft;
22606         struct timeval tv;
22607         tcp_seq seq;
22608         int err;
22609
22610         microuptime(&tv);
22611
22612         /*
22613          * If BB logging is not on we need to look at the DTL flag.
22614          * If its on already then those reasons override the DTL input.
22615          * We do this with any request, you can turn DTL on, but it does
22616          * not turn off at least from hybrid pacing requests.
22617          */
22618         if (tcp_bblogging_on(rack->rc_tp) == 0) {
22619                 if (hybrid->hybrid_flags & TCP_HYBRID_PACING_DTL) {
22620                         /* Turn on BB point logging  */
22621                         tcp_set_bblog_state(rack->rc_tp, TCP_LOG_VIA_BBPOINTS,
22622                                             TCP_BBPOINT_REQ_LEVEL_LOGGING);
22623                 }
22624         }
22625         /* Make sure no fixed rate is on */
22626         rack->use_fixed_rate = 0;
22627         rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
22628         rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
22629         rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
22630         /* Now allocate or find our entry that will have these settings */
22631         sft = tcp_http_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusectick(&tv), 0);
22632         if (sft == NULL) {
22633                 rack->rc_tp->tcp_hybrid_error++;
22634                 /* no space, where would it have gone? */
22635                 seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
22636                 rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
22637                 return (ENOSPC);
22638         }
22639         /* The seq will be snd_una + everything in the buffer */
22640         seq = sft->start_seq;
22641         if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
22642                 /* Disabling hybrid pacing */
22643                 if (rack->rc_hybrid_mode) {
22644                         rack_set_profile(rack, 0);
22645                         rack->rc_tp->tcp_hybrid_stop++;
22646                 }
22647                 rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
22648                 return (0);
22649         }
22650         if (rack->dgp_on == 0) {
22651                 /*
22652                  * If we have not yet turned DGP on, do so
22653                  * now setting pure DGP mode, no buffer level
22654                  * response.
22655                  */
22656                 if ((err = rack_set_profile(rack, 1)) != 0){
22657                         /* Failed to turn pacing on */
22658                         rack->rc_tp->tcp_hybrid_error++;
22659                         rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
22660                         return (err);
22661                 }
22662         }
22663         /* Now set in our flags */
22664         sft->hybrid_flags = hybrid->hybrid_flags;
22665         if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
22666                 sft->cspr = hybrid->cspr;
22667         else
22668                 sft->cspr = 0;
22669         if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
22670                 sft->hint_maxseg = hybrid->hint_maxseg;
22671         else
22672                 sft->hint_maxseg = 0;
22673         rack->rc_hybrid_mode = 1;
22674         rack->rc_tp->tcp_hybrid_start++;
22675         rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
22676         return (0);
22677 #else
22678         return (ENOTSUP);
22679 #endif
22680 }
22681
22682 static int
22683 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
22684                     uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
22685
22686 {
22687         struct epoch_tracker et;
22688         struct sockopt sopt;
22689         struct cc_newreno_opts opt;
22690         struct inpcb *inp = tptoinpcb(tp);
22691         uint64_t val;
22692         int error = 0;
22693         uint16_t ca, ss;
22694
22695         switch (sopt_name) {
22696         case TCP_RACK_SET_RXT_OPTIONS:
22697                 if ((optval >= 0) && (optval <= 2)) {
22698                         rack_init_retransmit_value(rack, optval);
22699                 } else {
22700                         /*
22701                          * You must send in 0, 1 or 2 all else is
22702                          * invalid.
22703                          */
22704                         error = EINVAL;
22705                 }
22706                 break;
22707         case TCP_RACK_DSACK_OPT:
22708                 RACK_OPTS_INC(tcp_rack_dsack_opt);
22709                 if (optval & 0x1) {
22710                         rack->rc_rack_tmr_std_based = 1;
22711                 } else {
22712                         rack->rc_rack_tmr_std_based = 0;
22713                 }
22714                 if (optval & 0x2) {
22715                         rack->rc_rack_use_dsack = 1;
22716                 } else {
22717                         rack->rc_rack_use_dsack = 0;
22718                 }
22719                 rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
22720                 break;
22721         case TCP_RACK_PACING_DIVISOR:
22722                 RACK_OPTS_INC(tcp_rack_pacing_divisor);
22723                 if (optval == 0) {
22724                         rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
22725                 } else {
22726                         if (optval < RL_MIN_DIVISOR)
22727                                 rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
22728                         else
22729                                 rack->r_ctl.pace_len_divisor = optval;
22730                 }
22731                 break;
22732         case TCP_RACK_HI_BETA:
22733                 RACK_OPTS_INC(tcp_rack_hi_beta);
22734                 if (optval)
22735                         rack->rack_hibeta = 1;
22736                 else
22737                         rack->rack_hibeta = 0;
22738                 break;
22739         case TCP_RACK_PACING_BETA:
22740                 RACK_OPTS_INC(tcp_rack_beta);
22741                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
22742                         /* This only works for newreno. */
22743                         error = EINVAL;
22744                         break;
22745                 }
22746                 if (rack->rc_pacing_cc_set) {
22747                         /*
22748                          * Set them into the real CC module
22749                          * whats in the rack pcb is the old values
22750                          * to be used on restoral/
22751                          */
22752                         sopt.sopt_dir = SOPT_SET;
22753                         opt.name = CC_NEWRENO_BETA;
22754                         opt.val = optval;
22755                         if (CC_ALGO(tp)->ctl_output != NULL)
22756                                 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
22757                         else {
22758                                 error = ENOENT;
22759                                 break;
22760                         }
22761                 } else {
22762                         /*
22763                          * Not pacing yet so set it into our local
22764                          * rack pcb storage.
22765                          */
22766                         rack->r_ctl.rc_saved_beta.beta = optval;
22767                 }
22768                 break;
22769         case TCP_RACK_TIMER_SLOP:
22770                 RACK_OPTS_INC(tcp_rack_timer_slop);
22771                 rack->r_ctl.timer_slop = optval;
22772                 if (rack->rc_tp->t_srtt) {
22773                         /*
22774                          * If we have an SRTT lets update t_rxtcur
22775                          * to have the new slop.
22776                          */
22777                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
22778                                            rack_rto_min, rack_rto_max,
22779                                            rack->r_ctl.timer_slop);
22780                 }
22781                 break;
22782         case TCP_RACK_PACING_BETA_ECN:
22783                 RACK_OPTS_INC(tcp_rack_beta_ecn);
22784                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
22785                         /* This only works for newreno. */
22786                         error = EINVAL;
22787                         break;
22788                 }
22789                 if (rack->rc_pacing_cc_set) {
22790                         /*
22791                          * Set them into the real CC module
22792                          * whats in the rack pcb is the old values
22793                          * to be used on restoral/
22794                          */
22795                         sopt.sopt_dir = SOPT_SET;
22796                         opt.name = CC_NEWRENO_BETA_ECN;
22797                         opt.val = optval;
22798                         if (CC_ALGO(tp)->ctl_output != NULL)
22799                                 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
22800                         else
22801                                 error = ENOENT;
22802                 } else {
22803                         /*
22804                          * Not pacing yet so set it into our local
22805                          * rack pcb storage.
22806                          */
22807                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
22808                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
22809                 }
22810                 break;
22811         case TCP_DEFER_OPTIONS:
22812                 RACK_OPTS_INC(tcp_defer_opt);
22813                 if (optval) {
22814                         if (rack->gp_ready) {
22815                                 /* Too late */
22816                                 error = EINVAL;
22817                                 break;
22818                         }
22819                         rack->defer_options = 1;
22820                 } else
22821                         rack->defer_options = 0;
22822                 break;
22823         case TCP_RACK_MEASURE_CNT:
22824                 RACK_OPTS_INC(tcp_rack_measure_cnt);
22825                 if (optval && (optval <= 0xff)) {
22826                         rack->r_ctl.req_measurements = optval;
22827                 } else
22828                         error = EINVAL;
22829                 break;
22830         case TCP_REC_ABC_VAL:
22831                 RACK_OPTS_INC(tcp_rec_abc_val);
22832                 if (optval > 0)
22833                         rack->r_use_labc_for_rec = 1;
22834                 else
22835                         rack->r_use_labc_for_rec = 0;
22836                 break;
22837         case TCP_RACK_ABC_VAL:
22838                 RACK_OPTS_INC(tcp_rack_abc_val);
22839                 if ((optval > 0) && (optval < 255))
22840                         rack->rc_labc = optval;
22841                 else
22842                         error = EINVAL;
22843                 break;
22844         case TCP_HDWR_UP_ONLY:
22845                 RACK_OPTS_INC(tcp_pacing_up_only);
22846                 if (optval)
22847                         rack->r_up_only = 1;
22848                 else
22849                         rack->r_up_only = 0;
22850                 break;
22851         case TCP_PACING_RATE_CAP:
22852                 RACK_OPTS_INC(tcp_pacing_rate_cap);
22853                 rack->r_ctl.bw_rate_cap = loptval;
22854                 break;
22855         case TCP_HYBRID_PACING:
22856                 if (hybrid == NULL) {
22857                         error = EINVAL;
22858                         break;
22859                 }
22860                 error = process_hybrid_pacing(rack, hybrid);
22861                 break;
22862         case TCP_RACK_PROFILE:
22863                 RACK_OPTS_INC(tcp_profile);
22864                 error = rack_set_profile(rack, optval);
22865                 break;
22866         case TCP_USE_CMP_ACKS:
22867                 RACK_OPTS_INC(tcp_use_cmp_acks);
22868                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
22869                         /* You can't turn it off once its on! */
22870                         error = EINVAL;
22871                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
22872                         rack->r_use_cmp_ack = 1;
22873                         rack->r_mbuf_queue = 1;
22874                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
22875                 }
22876                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
22877                         inp->inp_flags2 |= INP_MBUF_ACKCMP;
22878                 break;
22879         case TCP_SHARED_CWND_TIME_LIMIT:
22880                 RACK_OPTS_INC(tcp_lscwnd);
22881                 if (optval)
22882                         rack->r_limit_scw = 1;
22883                 else
22884                         rack->r_limit_scw = 0;
22885                 break;
22886         case TCP_RACK_DGP_IN_REC:
22887                 RACK_OPTS_INC(tcp_dgp_in_rec);
22888                 if (optval)
22889                         rack->r_ctl.full_dgp_in_rec = 1;
22890                 else
22891                         rack->r_ctl.full_dgp_in_rec = 0;
22892                 break;
22893         case TCP_RXT_CLAMP:
22894                 RACK_OPTS_INC(tcp_rxt_clamp);
22895                 rack_translate_clamp_value(rack, optval);
22896                 break;
22897         case TCP_RACK_PACE_TO_FILL:
22898                 RACK_OPTS_INC(tcp_fillcw);
22899                 if (optval == 0)
22900                         rack->rc_pace_to_cwnd = 0;
22901                 else {
22902                         rack->rc_pace_to_cwnd = 1;
22903                         if (optval > 1)
22904                                 rack->r_fill_less_agg = 1;
22905                 }
22906                 if ((optval >= rack_gp_rtt_maxmul) &&
22907                     rack_gp_rtt_maxmul &&
22908                     (optval < 0xf)) {
22909                         rack->rc_pace_fill_if_rttin_range = 1;
22910                         rack->rtt_limit_mul = optval;
22911                 } else {
22912                         rack->rc_pace_fill_if_rttin_range = 0;
22913                         rack->rtt_limit_mul = 0;
22914                 }
22915                 break;
22916         case TCP_RACK_NO_PUSH_AT_MAX:
22917                 RACK_OPTS_INC(tcp_npush);
22918                 if (optval == 0)
22919                         rack->r_ctl.rc_no_push_at_mrtt = 0;
22920                 else if (optval < 0xff)
22921                         rack->r_ctl.rc_no_push_at_mrtt = optval;
22922                 else
22923                         error = EINVAL;
22924                 break;
22925         case TCP_SHARED_CWND_ENABLE:
22926                 RACK_OPTS_INC(tcp_rack_scwnd);
22927                 if (optval == 0)
22928                         rack->rack_enable_scwnd = 0;
22929                 else
22930                         rack->rack_enable_scwnd = 1;
22931                 break;
22932         case TCP_RACK_MBUF_QUEUE:
22933                 /* Now do we use the LRO mbuf-queue feature */
22934                 RACK_OPTS_INC(tcp_rack_mbufq);
22935                 if (optval || rack->r_use_cmp_ack)
22936                         rack->r_mbuf_queue = 1;
22937                 else
22938                         rack->r_mbuf_queue = 0;
22939                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
22940                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
22941                 else
22942                         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
22943                 break;
22944         case TCP_RACK_NONRXT_CFG_RATE:
22945                 RACK_OPTS_INC(tcp_rack_cfg_rate);
22946                 if (optval == 0)
22947                         rack->rack_rec_nonrxt_use_cr = 0;
22948                 else
22949                         rack->rack_rec_nonrxt_use_cr = 1;
22950                 break;
22951         case TCP_NO_PRR:
22952                 RACK_OPTS_INC(tcp_rack_noprr);
22953                 if (optval == 0)
22954                         rack->rack_no_prr = 0;
22955                 else if (optval == 1)
22956                         rack->rack_no_prr = 1;
22957                 else if (optval == 2)
22958                         rack->no_prr_addback = 1;
22959                 else
22960                         error = EINVAL;
22961                 break;
22962         case TCP_TIMELY_DYN_ADJ:
22963                 RACK_OPTS_INC(tcp_timely_dyn);
22964                 if (optval == 0)
22965                         rack->rc_gp_dyn_mul = 0;
22966                 else {
22967                         rack->rc_gp_dyn_mul = 1;
22968                         if (optval >= 100) {
22969                                 /*
22970                                  * If the user sets something 100 or more
22971                                  * its the gp_ca value.
22972                                  */
22973                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
22974                         }
22975                 }
22976                 break;
22977         case TCP_RACK_DO_DETECTION:
22978                 RACK_OPTS_INC(tcp_rack_do_detection);
22979                 if (optval == 0)
22980                         rack->do_detection = 0;
22981                 else
22982                         rack->do_detection = 1;
22983                 break;
22984         case TCP_RACK_TLP_USE:
22985                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
22986                         error = EINVAL;
22987                         break;
22988                 }
22989                 RACK_OPTS_INC(tcp_tlp_use);
22990                 rack->rack_tlp_threshold_use = optval;
22991                 break;
22992         case TCP_RACK_TLP_REDUCE:
22993                 /* RACK TLP cwnd reduction (bool) */
22994                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
22995                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
22996                 break;
22997                 /*  Pacing related ones */
22998         case TCP_RACK_PACE_ALWAYS:
22999                 /*
23000                  * zero is old rack method, 1 is new
23001                  * method using a pacing rate.
23002                  */
23003                 RACK_OPTS_INC(tcp_rack_pace_always);
23004                 if (optval > 0) {
23005                         if (rack->rc_always_pace) {
23006                                 error = EALREADY;
23007                                 break;
23008                         } else if (tcp_can_enable_pacing()) {
23009                                 rack->rc_always_pace = 1;
23010                                 if ((rack->gp_ready) && (rack->use_fixed_rate == 0))
23011                                         rack_set_cc_pacing(rack);
23012                         }
23013                         else {
23014                                 error = ENOSPC;
23015                                 break;
23016                         }
23017                 } else {
23018                         if (rack->rc_always_pace) {
23019                                 tcp_decrement_paced_conn();
23020                                 rack->rc_always_pace = 0;
23021                                 rack_undo_cc_pacing(rack);
23022                         }
23023                 }
23024                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23025                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
23026                 else
23027                         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
23028                 /* A rate may be set irate or other, if so set seg size */
23029                 rack_update_seg(rack);
23030                 break;
23031         case TCP_BBR_RACK_INIT_RATE:
23032                 RACK_OPTS_INC(tcp_initial_rate);
23033                 val = optval;
23034                 /* Change from kbits per second to bytes per second */
23035                 val *= 1000;
23036                 val /= 8;
23037                 rack->r_ctl.init_rate = val;
23038                 if (rack->rc_init_win != rack_default_init_window) {
23039                         uint32_t win, snt;
23040
23041                         /*
23042                          * Options don't always get applied
23043                          * in the order you think. So in order
23044                          * to assure we update a cwnd we need
23045                          * to check and see if we are still
23046                          * where we should raise the cwnd.
23047                          */
23048                         win = rc_init_window(rack);
23049                         if (SEQ_GT(tp->snd_max, tp->iss))
23050                                 snt = tp->snd_max - tp->iss;
23051                         else
23052                                 snt = 0;
23053                         if ((snt < win) &&
23054                             (tp->snd_cwnd < win))
23055                                 tp->snd_cwnd = win;
23056                 }
23057                 if (rack->rc_always_pace)
23058                         rack_update_seg(rack);
23059                 break;
23060         case TCP_BBR_IWINTSO:
23061                 RACK_OPTS_INC(tcp_initial_win);
23062                 if (optval && (optval <= 0xff)) {
23063                         uint32_t win, snt;
23064
23065                         rack->rc_init_win = optval;
23066                         win = rc_init_window(rack);
23067                         if (SEQ_GT(tp->snd_max, tp->iss))
23068                                 snt = tp->snd_max - tp->iss;
23069                         else
23070                                 snt = 0;
23071                         if ((snt < win) &&
23072                             (tp->t_srtt |
23073                              rack->r_ctl.init_rate)) {
23074                                 /*
23075                                  * We are not past the initial window
23076                                  * and we have some bases for pacing,
23077                                  * so we need to possibly adjust up
23078                                  * the cwnd. Note even if we don't set
23079                                  * the cwnd, its still ok to raise the rc_init_win
23080                                  * which can be used coming out of idle when we
23081                                  * would have a rate.
23082                                  */
23083                                 if (tp->snd_cwnd < win)
23084                                         tp->snd_cwnd = win;
23085                         }
23086                         if (rack->rc_always_pace)
23087                                 rack_update_seg(rack);
23088                 } else
23089                         error = EINVAL;
23090                 break;
23091         case TCP_RACK_FORCE_MSEG:
23092                 RACK_OPTS_INC(tcp_rack_force_max_seg);
23093                 if (optval)
23094                         rack->rc_force_max_seg = 1;
23095                 else
23096                         rack->rc_force_max_seg = 0;
23097                 break;
23098         case TCP_RACK_PACE_MIN_SEG:
23099                 RACK_OPTS_INC(tcp_rack_min_seg);
23100                 rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
23101                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
23102                 break;
23103         case TCP_RACK_PACE_MAX_SEG:
23104                 /* Max segments size in a pace in bytes */
23105                 RACK_OPTS_INC(tcp_rack_max_seg);
23106                 rack->rc_user_set_max_segs = optval;
23107                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
23108                 break;
23109         case TCP_RACK_PACE_RATE_REC:
23110                 /* Set the fixed pacing rate in Bytes per second ca */
23111                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
23112                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23113                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23114                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23115                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23116                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23117                 rack->use_fixed_rate = 1;
23118                 if (rack->rc_always_pace && rack->gp_ready && rack->rack_hibeta)
23119                         rack_set_cc_pacing(rack);
23120                 rack_log_pacing_delay_calc(rack,
23121                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
23122                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
23123                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23124                                            __LINE__, NULL,0);
23125                 break;
23126
23127         case TCP_RACK_PACE_RATE_SS:
23128                 /* Set the fixed pacing rate in Bytes per second ca */
23129                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
23130                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23131                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23132                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23133                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23134                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23135                 rack->use_fixed_rate = 1;
23136                 if (rack->rc_always_pace && rack->gp_ready && rack->rack_hibeta)
23137                         rack_set_cc_pacing(rack);
23138                 rack_log_pacing_delay_calc(rack,
23139                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
23140                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
23141                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23142                                            __LINE__, NULL, 0);
23143                 break;
23144
23145         case TCP_RACK_PACE_RATE_CA:
23146                 /* Set the fixed pacing rate in Bytes per second ca */
23147                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
23148                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23149                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23150                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23151                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23152                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23153                 rack->use_fixed_rate = 1;
23154                 if (rack->rc_always_pace && rack->gp_ready && rack->rack_hibeta)
23155                         rack_set_cc_pacing(rack);
23156                 rack_log_pacing_delay_calc(rack,
23157                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
23158                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
23159                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23160                                            __LINE__, NULL, 0);
23161                 break;
23162         case TCP_RACK_GP_INCREASE_REC:
23163                 RACK_OPTS_INC(tcp_gp_inc_rec);
23164                 rack->r_ctl.rack_per_of_gp_rec = optval;
23165                 rack_log_pacing_delay_calc(rack,
23166                                            rack->r_ctl.rack_per_of_gp_ss,
23167                                            rack->r_ctl.rack_per_of_gp_ca,
23168                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23169                                            __LINE__, NULL, 0);
23170                 break;
23171         case TCP_RACK_GP_INCREASE_CA:
23172                 RACK_OPTS_INC(tcp_gp_inc_ca);
23173                 ca = optval;
23174                 if (ca < 100) {
23175                         /*
23176                          * We don't allow any reduction
23177                          * over the GP b/w.
23178                          */
23179                         error = EINVAL;
23180                         break;
23181                 }
23182                 rack->r_ctl.rack_per_of_gp_ca = ca;
23183                 rack_log_pacing_delay_calc(rack,
23184                                            rack->r_ctl.rack_per_of_gp_ss,
23185                                            rack->r_ctl.rack_per_of_gp_ca,
23186                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23187                                            __LINE__, NULL, 0);
23188                 break;
23189         case TCP_RACK_GP_INCREASE_SS:
23190                 RACK_OPTS_INC(tcp_gp_inc_ss);
23191                 ss = optval;
23192                 if (ss < 100) {
23193                         /*
23194                          * We don't allow any reduction
23195                          * over the GP b/w.
23196                          */
23197                         error = EINVAL;
23198                         break;
23199                 }
23200                 rack->r_ctl.rack_per_of_gp_ss = ss;
23201                 rack_log_pacing_delay_calc(rack,
23202                                            rack->r_ctl.rack_per_of_gp_ss,
23203                                            rack->r_ctl.rack_per_of_gp_ca,
23204                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23205                                            __LINE__, NULL, 0);
23206                 break;
23207         case TCP_RACK_RR_CONF:
23208                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
23209                 if (optval && optval <= 3)
23210                         rack->r_rr_config = optval;
23211                 else
23212                         rack->r_rr_config = 0;
23213                 break;
23214         case TCP_PACING_DND:                    /*  URL:dnd */
23215                 if (optval > 0)
23216                         rack->rc_pace_dnd = 1;
23217                 else
23218                         rack->rc_pace_dnd = 0;
23219                 break;
23220         case TCP_HDWR_RATE_CAP:
23221                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
23222                 if (optval) {
23223                         if (rack->r_rack_hw_rate_caps == 0)
23224                                 rack->r_rack_hw_rate_caps = 1;
23225                         else
23226                                 error = EALREADY;
23227                 } else {
23228                         rack->r_rack_hw_rate_caps = 0;
23229                 }
23230                 break;
23231         case TCP_RACK_SPLIT_LIMIT:
23232                 RACK_OPTS_INC(tcp_split_limit);
23233                 rack->r_ctl.rc_split_limit = optval;
23234                 break;
23235         case TCP_BBR_HDWR_PACE:
23236                 RACK_OPTS_INC(tcp_hdwr_pacing);
23237                 if (optval){
23238                         if (rack->rack_hdrw_pacing == 0) {
23239                                 rack->rack_hdw_pace_ena = 1;
23240                                 rack->rack_attempt_hdwr_pace = 0;
23241                         } else
23242                                 error = EALREADY;
23243                 } else {
23244                         rack->rack_hdw_pace_ena = 0;
23245 #ifdef RATELIMIT
23246                         if (rack->r_ctl.crte != NULL) {
23247                                 rack->rack_hdrw_pacing = 0;
23248                                 rack->rack_attempt_hdwr_pace = 0;
23249                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
23250                                 rack->r_ctl.crte = NULL;
23251                         }
23252 #endif
23253                 }
23254                 break;
23255                 /*  End Pacing related ones */
23256         case TCP_RACK_PRR_SENDALOT:
23257                 /* Allow PRR to send more than one seg */
23258                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
23259                 rack->r_ctl.rc_prr_sendalot = optval;
23260                 break;
23261         case TCP_RACK_MIN_TO:
23262                 /* Minimum time between rack t-o's in ms */
23263                 RACK_OPTS_INC(tcp_rack_min_to);
23264                 rack->r_ctl.rc_min_to = optval;
23265                 break;
23266         case TCP_RACK_EARLY_SEG:
23267                 /* If early recovery max segments */
23268                 RACK_OPTS_INC(tcp_rack_early_seg);
23269                 rack->r_ctl.rc_early_recovery_segs = optval;
23270                 break;
23271         case TCP_RACK_ENABLE_HYSTART:
23272         {
23273                 if (optval) {
23274                         tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23275                         if (rack_do_hystart > RACK_HYSTART_ON)
23276                                 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23277                         if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23278                                 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23279                 } else {
23280                         tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23281                 }
23282         }
23283         break;
23284         case TCP_RACK_REORD_THRESH:
23285                 /* RACK reorder threshold (shift amount) */
23286                 RACK_OPTS_INC(tcp_rack_reord_thresh);
23287                 if ((optval > 0) && (optval < 31))
23288                         rack->r_ctl.rc_reorder_shift = optval;
23289                 else
23290                         error = EINVAL;
23291                 break;
23292         case TCP_RACK_REORD_FADE:
23293                 /* Does reordering fade after ms time */
23294                 RACK_OPTS_INC(tcp_rack_reord_fade);
23295                 rack->r_ctl.rc_reorder_fade = optval;
23296                 break;
23297         case TCP_RACK_TLP_THRESH:
23298                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
23299                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
23300                 if (optval)
23301                         rack->r_ctl.rc_tlp_threshold = optval;
23302                 else
23303                         error = EINVAL;
23304                 break;
23305         case TCP_BBR_USE_RACK_RR:
23306                 RACK_OPTS_INC(tcp_rack_rr);
23307                 if (optval)
23308                         rack->use_rack_rr = 1;
23309                 else
23310                         rack->use_rack_rr = 0;
23311                 break;
23312         case TCP_RACK_PKT_DELAY:
23313                 /* RACK added ms i.e. rack-rtt + reord + N */
23314                 RACK_OPTS_INC(tcp_rack_pkt_delay);
23315                 rack->r_ctl.rc_pkt_delay = optval;
23316                 break;
23317         case TCP_DELACK:
23318                 RACK_OPTS_INC(tcp_rack_delayed_ack);
23319                 if (optval == 0)
23320                         tp->t_delayed_ack = 0;
23321                 else
23322                         tp->t_delayed_ack = 1;
23323                 if (tp->t_flags & TF_DELACK) {
23324                         tp->t_flags &= ~TF_DELACK;
23325                         tp->t_flags |= TF_ACKNOW;
23326                         NET_EPOCH_ENTER(et);
23327                         rack_output(tp);
23328                         NET_EPOCH_EXIT(et);
23329                 }
23330                 break;
23331
23332         case TCP_BBR_RACK_RTT_USE:
23333                 RACK_OPTS_INC(tcp_rack_rtt_use);
23334                 if ((optval != USE_RTT_HIGH) &&
23335                     (optval != USE_RTT_LOW) &&
23336                     (optval != USE_RTT_AVG))
23337                         error = EINVAL;
23338                 else
23339                         rack->r_ctl.rc_rate_sample_method = optval;
23340                 break;
23341         case TCP_DATA_AFTER_CLOSE:
23342                 RACK_OPTS_INC(tcp_data_after_close);
23343                 if (optval)
23344                         rack->rc_allow_data_af_clo = 1;
23345                 else
23346                         rack->rc_allow_data_af_clo = 0;
23347                 break;
23348         default:
23349                 break;
23350         }
23351         tcp_log_socket_option(tp, sopt_name, optval, error);
23352         return (error);
23353 }
23354
23355
23356 static void
23357 rack_apply_deferred_options(struct tcp_rack *rack)
23358 {
23359         struct deferred_opt_list *dol, *sdol;
23360         uint32_t s_optval;
23361
23362         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
23363                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
23364                 /* Disadvantage of deferal is you loose the error return */
23365                 s_optval = (uint32_t)dol->optval;
23366                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
23367                 free(dol, M_TCPDO);
23368         }
23369 }
23370
23371 static void
23372 rack_hw_tls_change(struct tcpcb *tp, int chg)
23373 {
23374         /* Update HW tls state */
23375         struct tcp_rack *rack;
23376
23377         rack = (struct tcp_rack *)tp->t_fb_ptr;
23378         if (chg)
23379                 rack->r_ctl.fsb.hw_tls = 1;
23380         else
23381                 rack->r_ctl.fsb.hw_tls = 0;
23382 }
23383
23384 static int
23385 rack_pru_options(struct tcpcb *tp, int flags)
23386 {
23387         if (flags & PRUS_OOB)
23388                 return (EOPNOTSUPP);
23389         return (0);
23390 }
23391
23392 static bool
23393 rack_wake_check(struct tcpcb *tp)
23394 {
23395         struct tcp_rack *rack;
23396         struct timeval tv;
23397         uint32_t cts;
23398
23399         rack = (struct tcp_rack *)tp->t_fb_ptr;
23400         if (rack->r_ctl.rc_hpts_flags) {
23401                 cts = tcp_get_usecs(&tv);
23402                 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
23403                         /*
23404                          * Pacing timer is up, check if we are ready.
23405                          */
23406                         if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
23407                                 return (true);
23408                 } else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
23409                         /*
23410                          * A timer is up, check if we are ready.
23411                          */
23412                         if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
23413                                 return (true);
23414                 }
23415         }
23416         return (false);
23417 }
23418
23419 static struct tcp_function_block __tcp_rack = {
23420         .tfb_tcp_block_name = __XSTRING(STACKNAME),
23421         .tfb_tcp_output = rack_output,
23422         .tfb_do_queued_segments = ctf_do_queued_segments,
23423         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
23424         .tfb_tcp_do_segment = rack_do_segment,
23425         .tfb_tcp_ctloutput = rack_ctloutput,
23426         .tfb_tcp_fb_init = rack_init,
23427         .tfb_tcp_fb_fini = rack_fini,
23428         .tfb_tcp_timer_stop_all = rack_stopall,
23429         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
23430         .tfb_tcp_handoff_ok = rack_handoff_ok,
23431         .tfb_tcp_mtu_chg = rack_mtu_change,
23432         .tfb_pru_options = rack_pru_options,
23433         .tfb_hwtls_change = rack_hw_tls_change,
23434         .tfb_chg_query = rack_chg_query,
23435         .tfb_switch_failed = rack_switch_failed,
23436         .tfb_early_wake_check = rack_wake_check,
23437         .tfb_compute_pipe = rack_compute_pipe,
23438         .tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
23439 };
23440
23441 /*
23442  * rack_ctloutput() must drop the inpcb lock before performing copyin on
23443  * socket option arguments.  When it re-acquires the lock after the copy, it
23444  * has to revalidate that the connection is still valid for the socket
23445  * option.
23446  */
23447 static int
23448 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
23449 {
23450         struct inpcb *inp = tptoinpcb(tp);
23451 #ifdef INET6
23452         struct ip6_hdr *ip6;
23453         int32_t mask, tclass;
23454 #endif
23455 #ifdef INET
23456         struct ip *ip;
23457 #endif
23458         struct tcp_rack *rack;
23459         struct tcp_hybrid_req hybrid;
23460         uint64_t loptval;
23461         int32_t error = 0, optval;
23462
23463         rack = (struct tcp_rack *)tp->t_fb_ptr;
23464         if (rack == NULL) {
23465                 INP_WUNLOCK(inp);
23466                 return (EINVAL);
23467         }
23468 #ifdef INET6
23469         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
23470 #endif
23471 #ifdef INET
23472         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
23473 #endif
23474
23475         switch (sopt->sopt_level) {
23476 #ifdef INET6
23477         case IPPROTO_IPV6:
23478                 MPASS(inp->inp_vflag & INP_IPV6PROTO);
23479                 switch (sopt->sopt_name) {
23480                 case IPV6_USE_MIN_MTU:
23481                         tcp6_use_min_mtu(tp);
23482                         break;
23483                 case IPV6_TCLASS:
23484                         /*
23485                          * The DSCP codepoint has changed, update the fsb
23486                          * by overwriting any previous traffic class.
23487                          */
23488                         if (inp->in6p_outputopts) {
23489                                 mask = 0xfc;
23490                                 tclass = inp->in6p_outputopts->ip6po_tclass;
23491                                 ip6->ip6_flow &= htonl((~mask) << 20);
23492                                 ip6->ip6_flow |= htonl((tclass & mask) << 20);
23493                         }
23494                         break;
23495                 }
23496                 INP_WUNLOCK(inp);
23497                 return (0);
23498 #endif
23499 #ifdef INET
23500         case IPPROTO_IP:
23501                 switch (sopt->sopt_name) {
23502                 case IP_TOS:
23503                         /*
23504                          * The DSCP codepoint has changed, update the fsb.
23505                          */
23506                         ip->ip_tos = rack->rc_inp->inp_ip_tos;
23507                         break;
23508                 case IP_TTL:
23509                         /*
23510                          * The TTL has changed, update the fsb.
23511                          */
23512                         ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
23513                         break;
23514                 }
23515                 INP_WUNLOCK(inp);
23516                 return (0);
23517 #endif
23518 #ifdef SO_PEERPRIO
23519         case SOL_SOCKET:
23520                 switch (sopt->sopt_name) {
23521                 case SO_PEERPRIO:                       /*  SC-URL:bs */
23522                         /* Already read in and sanity checked in sosetopt(). */
23523                         if (inp->inp_socket) {
23524                                 rack->client_bufferlvl = inp->inp_socket->so_peerprio;
23525                                 rack_client_buffer_level_set(rack);
23526                         }
23527                         break;
23528                 }
23529                 INP_WUNLOCK(inp);
23530                 return (0);
23531 #endif
23532         case IPPROTO_TCP:
23533                 switch (sopt->sopt_name) {
23534                 case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
23535                 /*  Pacing related ones */
23536                 case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
23537                 case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
23538                 case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
23539                 case TCP_RACK_PACE_MIN_SEG:             /*  URL:pace_min_seg */
23540                 case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
23541                 case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
23542                 case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
23543                 case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
23544                 case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
23545                 case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
23546                 case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
23547                 case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
23548                 case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
23549                 case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
23550                 case TCP_HDWR_RATE_CAP:                 /*  URL:hdwrcap boolean */
23551                 case TCP_PACING_RATE_CAP:               /*  URL:cap  -- used by side-channel */
23552                 case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
23553                 case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
23554                 case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
23555                 case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
23556                 case TCP_RACK_DGP_IN_REC:               /*  URL:dgpinrec */
23557                         /* End pacing related */
23558                 case TCP_RXT_CLAMP:                     /*  URL:rxtclamp */
23559                 case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
23560                 case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
23561                 case TCP_RACK_MIN_TO:                   /*  URL:min_to */
23562                 case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
23563                 case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
23564                 case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
23565                 case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
23566                 case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
23567                 case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
23568                 case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
23569                 case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
23570                 case TCP_RACK_DO_DETECTION:             /*  URL:detect */
23571                 case TCP_NO_PRR:                        /*  URL:noprr */
23572                 case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
23573                 case TCP_DATA_AFTER_CLOSE:              /*  no URL */
23574                 case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
23575                 case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
23576                 case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
23577                 case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
23578                 case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
23579                 case TCP_RACK_PROFILE:                  /*  URL:profile */
23580                 case TCP_HYBRID_PACING:                 /*  URL:hybrid */
23581                 case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
23582                 case TCP_RACK_ABC_VAL:                  /*  URL:labc */
23583                 case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
23584                 case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
23585                 case TCP_DEFER_OPTIONS:                 /*  URL:defer */
23586                 case TCP_RACK_DSACK_OPT:                /*  URL:dsack */
23587                 case TCP_RACK_TIMER_SLOP:               /*  URL:timer_slop */
23588                 case TCP_RACK_ENABLE_HYSTART:           /*  URL:hystart */
23589                 case TCP_RACK_SET_RXT_OPTIONS:          /*  URL:rxtsz */
23590                 case TCP_RACK_HI_BETA:                  /*  URL:hibeta */
23591                 case TCP_RACK_SPLIT_LIMIT:              /*  URL:split */
23592                 case TCP_RACK_PACING_DIVISOR:           /*  URL:divisor */
23593                 case TCP_PACING_DND:                    /*  URL:dnd */
23594                         goto process_opt;
23595                         break;
23596                 default:
23597                         /* Filter off all unknown options to the base stack */
23598                         return (tcp_default_ctloutput(tp, sopt));
23599                         break;
23600                 }
23601
23602         default:
23603                 INP_WUNLOCK(inp);
23604                 return (0);
23605         }
23606 process_opt:
23607         INP_WUNLOCK(inp);
23608         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
23609                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
23610                 /*
23611                  * We truncate it down to 32 bits for the socket-option trace this
23612                  * means rates > 34Gbps won't show right, but thats probably ok.
23613                  */
23614                 optval = (uint32_t)loptval;
23615         } else if (sopt->sopt_name == TCP_HYBRID_PACING) {
23616                 error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
23617         } else {
23618                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
23619                 /* Save it in 64 bit form too */
23620                 loptval = optval;
23621         }
23622         if (error)
23623                 return (error);
23624         INP_WLOCK(inp);
23625         if (tp->t_fb != &__tcp_rack) {
23626                 INP_WUNLOCK(inp);
23627                 return (ENOPROTOOPT);
23628         }
23629         if (rack->defer_options && (rack->gp_ready == 0) &&
23630             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
23631             (sopt->sopt_name != TCP_HYBRID_PACING) &&
23632             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
23633             (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
23634             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
23635             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
23636                 /* Options are beind deferred */
23637                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
23638                         INP_WUNLOCK(inp);
23639                         return (0);
23640                 } else {
23641                         /* No memory to defer, fail */
23642                         INP_WUNLOCK(inp);
23643                         return (ENOMEM);
23644                 }
23645         }
23646         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
23647         INP_WUNLOCK(inp);
23648         return (error);
23649 }
23650
23651 static void
23652 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
23653 {
23654
23655         INP_WLOCK_ASSERT(tptoinpcb(tp));
23656         bzero(ti, sizeof(*ti));
23657
23658         ti->tcpi_state = tp->t_state;
23659         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
23660                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
23661         if (tp->t_flags & TF_SACK_PERMIT)
23662                 ti->tcpi_options |= TCPI_OPT_SACK;
23663         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
23664                 ti->tcpi_options |= TCPI_OPT_WSCALE;
23665                 ti->tcpi_snd_wscale = tp->snd_scale;
23666                 ti->tcpi_rcv_wscale = tp->rcv_scale;
23667         }
23668         if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
23669                 ti->tcpi_options |= TCPI_OPT_ECN;
23670         if (tp->t_flags & TF_FASTOPEN)
23671                 ti->tcpi_options |= TCPI_OPT_TFO;
23672         /* still kept in ticks is t_rcvtime */
23673         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
23674         /* Since we hold everything in precise useconds this is easy */
23675         ti->tcpi_rtt = tp->t_srtt;
23676         ti->tcpi_rttvar = tp->t_rttvar;
23677         ti->tcpi_rto = tp->t_rxtcur;
23678         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
23679         ti->tcpi_snd_cwnd = tp->snd_cwnd;
23680         /*
23681          * FreeBSD-specific extension fields for tcp_info.
23682          */
23683         ti->tcpi_rcv_space = tp->rcv_wnd;
23684         ti->tcpi_rcv_nxt = tp->rcv_nxt;
23685         ti->tcpi_snd_wnd = tp->snd_wnd;
23686         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
23687         ti->tcpi_snd_nxt = tp->snd_nxt;
23688         ti->tcpi_snd_mss = tp->t_maxseg;
23689         ti->tcpi_rcv_mss = tp->t_maxseg;
23690         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
23691         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
23692         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
23693         ti->tcpi_total_tlp = tp->t_sndtlppack;
23694         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
23695 #ifdef NETFLIX_STATS
23696         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
23697 #endif
23698 #ifdef TCP_OFFLOAD
23699         if (tp->t_flags & TF_TOE) {
23700                 ti->tcpi_options |= TCPI_OPT_TOE;
23701                 tcp_offload_tcp_info(tp, ti);
23702         }
23703 #endif
23704 }
23705
23706 static int
23707 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
23708 {
23709         struct inpcb *inp = tptoinpcb(tp);
23710         struct tcp_rack *rack;
23711         int32_t error, optval;
23712         uint64_t val, loptval;
23713         struct  tcp_info ti;
23714         /*
23715          * Because all our options are either boolean or an int, we can just
23716          * pull everything into optval and then unlock and copy. If we ever
23717          * add a option that is not a int, then this will have quite an
23718          * impact to this routine.
23719          */
23720         error = 0;
23721         rack = (struct tcp_rack *)tp->t_fb_ptr;
23722         if (rack == NULL) {
23723                 INP_WUNLOCK(inp);
23724                 return (EINVAL);
23725         }
23726         switch (sopt->sopt_name) {
23727         case TCP_INFO:
23728                 /* First get the info filled */
23729                 rack_fill_info(tp, &ti);
23730                 /* Fix up the rtt related fields if needed */
23731                 INP_WUNLOCK(inp);
23732                 error = sooptcopyout(sopt, &ti, sizeof ti);
23733                 return (error);
23734         /*
23735          * Beta is the congestion control value for NewReno that influences how
23736          * much of a backoff happens when loss is detected. It is normally set
23737          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
23738          * when you exit recovery.
23739          */
23740         case TCP_RACK_PACING_BETA:
23741                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
23742                         error = EINVAL;
23743                 else if (rack->rc_pacing_cc_set == 0)
23744                         optval = rack->r_ctl.rc_saved_beta.beta;
23745                 else {
23746                         /*
23747                          * Reach out into the CC data and report back what
23748                          * I have previously set. Yeah it looks hackish but
23749                          * we don't want to report the saved values.
23750                          */
23751                         if (tp->t_ccv.cc_data)
23752                                 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
23753                         else
23754                                 error = EINVAL;
23755                 }
23756                 break;
23757                 /*
23758                  * Beta_ecn is the congestion control value for NewReno that influences how
23759                  * much of a backoff happens when a ECN mark is detected. It is normally set
23760                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
23761                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
23762                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
23763                  */
23764
23765         case TCP_RACK_PACING_BETA_ECN:
23766                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
23767                         error = EINVAL;
23768                 else if (rack->rc_pacing_cc_set == 0)
23769                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
23770                 else {
23771                         /*
23772                          * Reach out into the CC data and report back what
23773                          * I have previously set. Yeah it looks hackish but
23774                          * we don't want to report the saved values.
23775                          */
23776                         if (tp->t_ccv.cc_data)
23777                                 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
23778                         else
23779                                 error = EINVAL;
23780                 }
23781                 break;
23782         case TCP_RACK_DSACK_OPT:
23783                 optval = 0;
23784                 if (rack->rc_rack_tmr_std_based) {
23785                         optval |= 1;
23786                 }
23787                 if (rack->rc_rack_use_dsack) {
23788                         optval |= 2;
23789                 }
23790                 break;
23791         case TCP_RACK_ENABLE_HYSTART:
23792         {
23793                 if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
23794                         optval = RACK_HYSTART_ON;
23795                         if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
23796                                 optval = RACK_HYSTART_ON_W_SC;
23797                         if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
23798                                 optval = RACK_HYSTART_ON_W_SC_C;
23799                 } else {
23800                         optval = RACK_HYSTART_OFF;
23801                 }
23802         }
23803         break;
23804         case TCP_RACK_DGP_IN_REC:
23805                 optval = rack->r_ctl.full_dgp_in_rec;
23806                 break;
23807         case TCP_RACK_HI_BETA:
23808                 optval = rack->rack_hibeta;
23809                 break;
23810         case TCP_RXT_CLAMP:
23811                 optval = rack->r_ctl.saved_rxt_clamp_val;
23812                 break;
23813         case TCP_DEFER_OPTIONS:
23814                 optval = rack->defer_options;
23815                 break;
23816         case TCP_RACK_MEASURE_CNT:
23817                 optval = rack->r_ctl.req_measurements;
23818                 break;
23819         case TCP_REC_ABC_VAL:
23820                 optval = rack->r_use_labc_for_rec;
23821                 break;
23822         case TCP_RACK_ABC_VAL:
23823                 optval = rack->rc_labc;
23824                 break;
23825         case TCP_HDWR_UP_ONLY:
23826                 optval= rack->r_up_only;
23827                 break;
23828         case TCP_PACING_RATE_CAP:
23829                 loptval = rack->r_ctl.bw_rate_cap;
23830                 break;
23831         case TCP_RACK_PROFILE:
23832                 /* You cannot retrieve a profile, its write only */
23833                 error = EINVAL;
23834                 break;
23835         case TCP_HYBRID_PACING:
23836                 /* You cannot retrieve hybrid pacing information, its write only */
23837                 error = EINVAL;
23838                 break;
23839         case TCP_USE_CMP_ACKS:
23840                 optval = rack->r_use_cmp_ack;
23841                 break;
23842         case TCP_RACK_PACE_TO_FILL:
23843                 optval = rack->rc_pace_to_cwnd;
23844                 if (optval && rack->r_fill_less_agg)
23845                         optval++;
23846                 break;
23847         case TCP_RACK_NO_PUSH_AT_MAX:
23848                 optval = rack->r_ctl.rc_no_push_at_mrtt;
23849                 break;
23850         case TCP_SHARED_CWND_ENABLE:
23851                 optval = rack->rack_enable_scwnd;
23852                 break;
23853         case TCP_RACK_NONRXT_CFG_RATE:
23854                 optval = rack->rack_rec_nonrxt_use_cr;
23855                 break;
23856         case TCP_NO_PRR:
23857                 if (rack->rack_no_prr  == 1)
23858                         optval = 1;
23859                 else if (rack->no_prr_addback == 1)
23860                         optval = 2;
23861                 else
23862                         optval = 0;
23863                 break;
23864         case TCP_RACK_DO_DETECTION:
23865                 optval = rack->do_detection;
23866                 break;
23867         case TCP_RACK_MBUF_QUEUE:
23868                 /* Now do we use the LRO mbuf-queue feature */
23869                 optval = rack->r_mbuf_queue;
23870                 break;
23871         case TCP_TIMELY_DYN_ADJ:
23872                 optval = rack->rc_gp_dyn_mul;
23873                 break;
23874         case TCP_BBR_IWINTSO:
23875                 optval = rack->rc_init_win;
23876                 break;
23877         case TCP_RACK_TLP_REDUCE:
23878                 /* RACK TLP cwnd reduction (bool) */
23879                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
23880                 break;
23881         case TCP_BBR_RACK_INIT_RATE:
23882                 val = rack->r_ctl.init_rate;
23883                 /* convert to kbits per sec */
23884                 val *= 8;
23885                 val /= 1000;
23886                 optval = (uint32_t)val;
23887                 break;
23888         case TCP_RACK_FORCE_MSEG:
23889                 optval = rack->rc_force_max_seg;
23890                 break;
23891         case TCP_RACK_PACE_MIN_SEG:
23892                 optval = rack->r_ctl.rc_user_set_min_segs;
23893                 break;
23894         case TCP_RACK_PACE_MAX_SEG:
23895                 /* Max segments in a pace */
23896                 optval = rack->rc_user_set_max_segs;
23897                 break;
23898         case TCP_RACK_PACE_ALWAYS:
23899                 /* Use the always pace method */
23900                 optval = rack->rc_always_pace;
23901                 break;
23902         case TCP_RACK_PRR_SENDALOT:
23903                 /* Allow PRR to send more than one seg */
23904                 optval = rack->r_ctl.rc_prr_sendalot;
23905                 break;
23906         case TCP_RACK_MIN_TO:
23907                 /* Minimum time between rack t-o's in ms */
23908                 optval = rack->r_ctl.rc_min_to;
23909                 break;
23910         case TCP_RACK_SPLIT_LIMIT:
23911                 optval = rack->r_ctl.rc_split_limit;
23912                 break;
23913         case TCP_RACK_EARLY_SEG:
23914                 /* If early recovery max segments */
23915                 optval = rack->r_ctl.rc_early_recovery_segs;
23916                 break;
23917         case TCP_RACK_REORD_THRESH:
23918                 /* RACK reorder threshold (shift amount) */
23919                 optval = rack->r_ctl.rc_reorder_shift;
23920                 break;
23921         case TCP_RACK_REORD_FADE:
23922                 /* Does reordering fade after ms time */
23923                 optval = rack->r_ctl.rc_reorder_fade;
23924                 break;
23925         case TCP_BBR_USE_RACK_RR:
23926                 /* Do we use the rack cheat for rxt */
23927                 optval = rack->use_rack_rr;
23928                 break;
23929         case TCP_RACK_RR_CONF:
23930                 optval = rack->r_rr_config;
23931                 break;
23932         case TCP_HDWR_RATE_CAP:
23933                 optval = rack->r_rack_hw_rate_caps;
23934                 break;
23935         case TCP_BBR_HDWR_PACE:
23936                 optval = rack->rack_hdw_pace_ena;
23937                 break;
23938         case TCP_RACK_TLP_THRESH:
23939                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
23940                 optval = rack->r_ctl.rc_tlp_threshold;
23941                 break;
23942         case TCP_RACK_PKT_DELAY:
23943                 /* RACK added ms i.e. rack-rtt + reord + N */
23944                 optval = rack->r_ctl.rc_pkt_delay;
23945                 break;
23946         case TCP_RACK_TLP_USE:
23947                 optval = rack->rack_tlp_threshold_use;
23948                 break;
23949         case TCP_PACING_DND:
23950                 optval = rack->rc_pace_dnd;
23951                 break;
23952         case TCP_RACK_PACE_RATE_CA:
23953                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
23954                 break;
23955         case TCP_RACK_PACE_RATE_SS:
23956                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
23957                 break;
23958         case TCP_RACK_PACE_RATE_REC:
23959                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
23960                 break;
23961         case TCP_RACK_GP_INCREASE_SS:
23962                 optval = rack->r_ctl.rack_per_of_gp_ca;
23963                 break;
23964         case TCP_RACK_GP_INCREASE_CA:
23965                 optval = rack->r_ctl.rack_per_of_gp_ss;
23966                 break;
23967         case TCP_RACK_PACING_DIVISOR:
23968                 optval = rack->r_ctl.pace_len_divisor;
23969                 break;
23970         case TCP_BBR_RACK_RTT_USE:
23971                 optval = rack->r_ctl.rc_rate_sample_method;
23972                 break;
23973         case TCP_DELACK:
23974                 optval = tp->t_delayed_ack;
23975                 break;
23976         case TCP_DATA_AFTER_CLOSE:
23977                 optval = rack->rc_allow_data_af_clo;
23978                 break;
23979         case TCP_SHARED_CWND_TIME_LIMIT:
23980                 optval = rack->r_limit_scw;
23981                 break;
23982         case TCP_RACK_TIMER_SLOP:
23983                 optval = rack->r_ctl.timer_slop;
23984                 break;
23985         default:
23986                 return (tcp_default_ctloutput(tp, sopt));
23987                 break;
23988         }
23989         INP_WUNLOCK(inp);
23990         if (error == 0) {
23991                 if (TCP_PACING_RATE_CAP)
23992                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
23993                 else
23994                         error = sooptcopyout(sopt, &optval, sizeof optval);
23995         }
23996         return (error);
23997 }
23998
23999 static int
24000 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
24001 {
24002         if (sopt->sopt_dir == SOPT_SET) {
24003                 return (rack_set_sockopt(tp, sopt));
24004         } else if (sopt->sopt_dir == SOPT_GET) {
24005                 return (rack_get_sockopt(tp, sopt));
24006         } else {
24007                 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
24008         }
24009 }
24010
24011 static const char *rack_stack_names[] = {
24012         __XSTRING(STACKNAME),
24013 #ifdef STACKALIAS
24014         __XSTRING(STACKALIAS),
24015 #endif
24016 };
24017
24018 static int
24019 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
24020 {
24021         memset(mem, 0, size);
24022         return (0);
24023 }
24024
24025 static void
24026 rack_dtor(void *mem, int32_t size, void *arg)
24027 {
24028
24029 }
24030
24031 static bool rack_mod_inited = false;
24032
24033 static int
24034 tcp_addrack(module_t mod, int32_t type, void *data)
24035 {
24036         int32_t err = 0;
24037         int num_stacks;
24038
24039         switch (type) {
24040         case MOD_LOAD:
24041                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
24042                     sizeof(struct rack_sendmap),
24043                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
24044
24045                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
24046                     sizeof(struct tcp_rack),
24047                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
24048
24049                 sysctl_ctx_init(&rack_sysctl_ctx);
24050                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
24051                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
24052                     OID_AUTO,
24053 #ifdef STACKALIAS
24054                     __XSTRING(STACKALIAS),
24055 #else
24056                     __XSTRING(STACKNAME),
24057 #endif
24058                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
24059                     "");
24060                 if (rack_sysctl_root == NULL) {
24061                         printf("Failed to add sysctl node\n");
24062                         err = EFAULT;
24063                         goto free_uma;
24064                 }
24065                 rack_init_sysctls();
24066                 num_stacks = nitems(rack_stack_names);
24067                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
24068                     rack_stack_names, &num_stacks);
24069                 if (err) {
24070                         printf("Failed to register %s stack name for "
24071                             "%s module\n", rack_stack_names[num_stacks],
24072                             __XSTRING(MODNAME));
24073                         sysctl_ctx_free(&rack_sysctl_ctx);
24074 free_uma:
24075                         uma_zdestroy(rack_zone);
24076                         uma_zdestroy(rack_pcb_zone);
24077                         rack_counter_destroy();
24078                         printf("Failed to register rack module -- err:%d\n", err);
24079                         return (err);
24080                 }
24081                 tcp_lro_reg_mbufq();
24082                 rack_mod_inited = true;
24083                 break;
24084         case MOD_QUIESCE:
24085                 err = deregister_tcp_functions(&__tcp_rack, true, false);
24086                 break;
24087         case MOD_UNLOAD:
24088                 err = deregister_tcp_functions(&__tcp_rack, false, true);
24089                 if (err == EBUSY)
24090                         break;
24091                 if (rack_mod_inited) {
24092                         uma_zdestroy(rack_zone);
24093                         uma_zdestroy(rack_pcb_zone);
24094                         sysctl_ctx_free(&rack_sysctl_ctx);
24095                         rack_counter_destroy();
24096                         rack_mod_inited = false;
24097                 }
24098                 tcp_lro_dereg_mbufq();
24099                 err = 0;
24100                 break;
24101         default:
24102                 return (EOPNOTSUPP);
24103         }
24104         return (err);
24105 }
24106
24107 static moduledata_t tcp_rack = {
24108         .name = __XSTRING(MODNAME),
24109         .evhand = tcp_addrack,
24110         .priv = 0
24111 };
24112
24113 MODULE_VERSION(MODNAME, 1);
24114 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
24115 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
24116
24117 #endif /* #if !defined(INET) && !defined(INET6) */