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_req_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 tcp_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_tcpreq_info[0]);
2053                         log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_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 tcp_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_TRK_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_tp);
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_tp->t_hpts_slot;
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_tp);
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_tp);
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_tp);
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_tp);
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_tp);
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_tp);
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_tp);
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_tp);
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_tp);
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_tp) &&
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_tp);
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_tp)) {
6510                 tcp_hpts_remove(rack->rc_tp);
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(tp)) {
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          * TF2_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          * TF2_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         tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_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                 tp->t_flags2 |= TF2_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                                 tp->t_flags2 |= TF2_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                                         tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6877                                 }
6878                         }
6879                 }
6880                 /* For sack attackers we want to ignore sack */
6881                 if (rack->sack_attack_disable == 1) {
6882                         tp->t_flags2 |= (TF2_DONT_SACK_QUEUE |
6883                             TF2_MBUF_QUEUE_READY);
6884                 } else if (rack->rc_ack_can_sendout_data) {
6885                         /*
6886                          * Ahh but wait, this is that special case
6887                          * where the pacing timer can be disturbed
6888                          * backout the changes (used for non-paced
6889                          * burst limiting).
6890                          */
6891                         tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
6892                             TF2_MBUF_QUEUE_READY);
6893                 }
6894                 if ((rack->use_rack_rr) &&
6895                     (rack->r_rr_config < 2) &&
6896                     ((hpts_timeout) && (hpts_timeout < slot))) {
6897                         /*
6898                          * Arrange for the hpts to kick back in after the
6899                          * t-o if the t-o does not cause a send.
6900                          */
6901                         (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6902                                                    __LINE__, &diag);
6903                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6904                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6905                 } else {
6906                         (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
6907                                                    __LINE__, &diag);
6908                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6909                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
6910                 }
6911         } else if (hpts_timeout) {
6912                 /*
6913                  * With respect to t_flags2(?) here, lets let any new acks wake
6914                  * us up here. Since we are not pacing (no pacing timer), output
6915                  * can happen so we should let it. If its a Rack timer, then any inbound
6916                  * packet probably won't change the sending (we will be blocked)
6917                  * but it may change the prr stats so letting it in (the set defaults
6918                  * at the start of this block) are good enough.
6919                  */
6920                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6921                 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6922                                            __LINE__, &diag);
6923                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6924                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6925         } else {
6926                 /* No timer starting */
6927 #ifdef INVARIANTS
6928                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
6929                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
6930                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
6931                 }
6932 #endif
6933         }
6934         rack->rc_tmr_stopped = 0;
6935         if (slot)
6936                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6937 }
6938
6939 /*
6940  * RACK Timer, here we simply do logging and house keeping.
6941  * the normal rack_output() function will call the
6942  * appropriate thing to check if we need to do a RACK retransmit.
6943  * We return 1, saying don't proceed with rack_output only
6944  * when all timers have been stopped (destroyed PCB?).
6945  */
6946 static int
6947 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6948 {
6949         /*
6950          * This timer simply provides an internal trigger to send out data.
6951          * The check_recovery_mode call will see if there are needed
6952          * retransmissions, if so we will enter fast-recovery. The output
6953          * call may or may not do the same thing depending on sysctl
6954          * settings.
6955          */
6956         struct rack_sendmap *rsm;
6957
6958         counter_u64_add(rack_to_tot, 1);
6959         if (rack->r_state && (rack->r_state != tp->t_state))
6960                 rack_set_state(tp, rack);
6961         rack->rc_on_min_to = 0;
6962         rsm = rack_check_recovery_mode(tp, cts);
6963         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
6964         if (rsm) {
6965                 rack->r_ctl.rc_resend = rsm;
6966                 rack->r_timer_override = 1;
6967                 if (rack->use_rack_rr) {
6968                         /*
6969                          * Don't accumulate extra pacing delay
6970                          * we are allowing the rack timer to
6971                          * over-ride pacing i.e. rrr takes precedence
6972                          * if the pacing interval is longer than the rrr
6973                          * time (in other words we get the min pacing
6974                          * time versus rrr pacing time).
6975                          */
6976                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6977                 }
6978         }
6979         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
6980         if (rsm == NULL) {
6981                 /* restart a timer and return 1 */
6982                 rack_start_hpts_timer(rack, tp, cts,
6983                                       0, 0, 0);
6984                 return (1);
6985         }
6986         return (0);
6987 }
6988
6989
6990
6991 static void
6992 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
6993 {
6994
6995         if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
6996                 /*
6997                  * The trailing space changed, mbufs can grow
6998                  * at the tail but they can't shrink from
6999                  * it, KASSERT that. Adjust the orig_m_len to
7000                  * compensate for this change.
7001                  */
7002                 KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
7003                         ("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
7004                          rsm->m,
7005                          rsm,
7006                          (intmax_t)M_TRAILINGROOM(rsm->m),
7007                          rsm->orig_t_space,
7008                          rsm->orig_m_len,
7009                          rsm->m->m_len));
7010                 rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
7011                 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7012         }
7013         if (rsm->m->m_len < rsm->orig_m_len) {
7014                 /*
7015                  * Mbuf shrank, trimmed off the top by an ack, our
7016                  * offset changes.
7017                  */
7018                 KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
7019                         ("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
7020                          rsm->m, rsm->m->m_len,
7021                          rsm, rsm->orig_m_len,
7022                          rsm->soff));
7023                 if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
7024                         rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
7025                 else
7026                         rsm->soff = 0;
7027                 rsm->orig_m_len = rsm->m->m_len;
7028 #ifdef INVARIANTS
7029         } else if (rsm->m->m_len > rsm->orig_m_len) {
7030                 panic("rsm:%p m:%p m_len grew outside of t_space compensation",
7031                       rsm, rsm->m);
7032 #endif
7033         }
7034 }
7035
7036 static void
7037 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
7038 {
7039         struct mbuf *m;
7040         uint32_t soff;
7041
7042         if (src_rsm->m &&
7043             ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
7044              (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
7045                 /* Fix up the orig_m_len and possibly the mbuf offset */
7046                 rack_adjust_orig_mlen(src_rsm);
7047         }
7048         m = src_rsm->m;
7049         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
7050         while (soff >= m->m_len) {
7051                 /* Move out past this mbuf */
7052                 soff -= m->m_len;
7053                 m = m->m_next;
7054                 KASSERT((m != NULL),
7055                         ("rsm:%p nrsm:%p hit at soff:%u null m",
7056                          src_rsm, rsm, soff));
7057                 if (m == NULL) {
7058                         /* This should *not* happen which is why there is a kassert */
7059                         src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7060                                                (src_rsm->r_start - rack->rc_tp->snd_una),
7061                                                &src_rsm->soff);
7062                         src_rsm->orig_m_len = src_rsm->m->m_len;
7063                         src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7064                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7065                                            (rsm->r_start - rack->rc_tp->snd_una),
7066                                            &rsm->soff);
7067                         rsm->orig_m_len = rsm->m->m_len;
7068                         rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7069                         return;
7070                 }
7071         }
7072         rsm->m = m;
7073         rsm->soff = soff;
7074         rsm->orig_m_len = m->m_len;
7075         rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7076 }
7077
7078 static __inline void
7079 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7080                struct rack_sendmap *rsm, uint32_t start)
7081 {
7082         int idx;
7083
7084         nrsm->r_start = start;
7085         nrsm->r_end = rsm->r_end;
7086         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7087         nrsm->r_flags = rsm->r_flags;
7088         nrsm->r_dupack = rsm->r_dupack;
7089         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7090         nrsm->r_rtr_bytes = 0;
7091         nrsm->r_fas = rsm->r_fas;
7092         nrsm->r_bas = rsm->r_bas;
7093         rsm->r_end = nrsm->r_start;
7094         nrsm->r_just_ret = rsm->r_just_ret;
7095         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7096                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7097         }
7098         /* Now if we have SYN flag we keep it on the left edge */
7099         if (nrsm->r_flags & RACK_HAS_SYN)
7100                 nrsm->r_flags &= ~RACK_HAS_SYN;
7101         /* Now if we have a FIN flag we keep it on the right edge */
7102         if (rsm->r_flags & RACK_HAS_FIN)
7103                 rsm->r_flags &= ~RACK_HAS_FIN;
7104         /* Push bit must go to the right edge as well */
7105         if (rsm->r_flags & RACK_HAD_PUSH)
7106                 rsm->r_flags &= ~RACK_HAD_PUSH;
7107         /* Clone over the state of the hw_tls flag */
7108         nrsm->r_hw_tls = rsm->r_hw_tls;
7109         /*
7110          * Now we need to find nrsm's new location in the mbuf chain
7111          * we basically calculate a new offset, which is soff +
7112          * how much is left in original rsm. Then we walk out the mbuf
7113          * chain to find the righ position, it may be the same mbuf
7114          * or maybe not.
7115          */
7116         KASSERT(((rsm->m != NULL) ||
7117                  (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7118                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7119         if (rsm->m)
7120                 rack_setup_offset_for_rsm(rack, rsm, nrsm);
7121 }
7122
7123 static struct rack_sendmap *
7124 rack_merge_rsm(struct tcp_rack *rack,
7125                struct rack_sendmap *l_rsm,
7126                struct rack_sendmap *r_rsm)
7127 {
7128         /*
7129          * We are merging two ack'd RSM's,
7130          * the l_rsm is on the left (lower seq
7131          * values) and the r_rsm is on the right
7132          * (higher seq value). The simplest way
7133          * to merge these is to move the right
7134          * one into the left. I don't think there
7135          * is any reason we need to try to find
7136          * the oldest (or last oldest retransmitted).
7137          */
7138         rack_log_map_chg(rack->rc_tp, rack, NULL,
7139                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7140         l_rsm->r_end = r_rsm->r_end;
7141         if (l_rsm->r_dupack < r_rsm->r_dupack)
7142                 l_rsm->r_dupack = r_rsm->r_dupack;
7143         if (r_rsm->r_rtr_bytes)
7144                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7145         if (r_rsm->r_in_tmap) {
7146                 /* This really should not happen */
7147                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7148                 r_rsm->r_in_tmap = 0;
7149         }
7150
7151         /* Now the flags */
7152         if (r_rsm->r_flags & RACK_HAS_FIN)
7153                 l_rsm->r_flags |= RACK_HAS_FIN;
7154         if (r_rsm->r_flags & RACK_TLP)
7155                 l_rsm->r_flags |= RACK_TLP;
7156         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7157                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7158         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
7159             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7160                 /*
7161                  * If both are app-limited then let the
7162                  * free lower the count. If right is app
7163                  * limited and left is not, transfer.
7164                  */
7165                 l_rsm->r_flags |= RACK_APP_LIMITED;
7166                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
7167                 if (r_rsm == rack->r_ctl.rc_first_appl)
7168                         rack->r_ctl.rc_first_appl = l_rsm;
7169         }
7170         tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7171         /*
7172          * We keep the largest value, which is the newest
7173          * send. We do this in case a segment that is
7174          * joined together and not part of a GP estimate
7175          * later gets expanded into the GP estimate.
7176          *
7177          * We prohibit the merging of unlike kinds i.e.
7178          * all pieces that are in the GP estimate can be
7179          * merged and all pieces that are not in a GP estimate
7180          * can be merged, but not disimilar pieces. Combine
7181          * this with taking the highest here and we should
7182          * be ok unless of course the client reneges. Then
7183          * all bets are off.
7184          */
7185         if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7186            r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7187                 l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7188         }
7189         /*
7190          * When merging two RSM's we also need to consider the ack time and keep
7191          * newest. If the ack gets merged into a measurement then that is the
7192          * one we will want to be using.
7193          */
7194         if(l_rsm->r_ack_arrival  < r_rsm->r_ack_arrival)
7195                 l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7196
7197         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7198                 /* Transfer the split limit to the map we free */
7199                 r_rsm->r_limit_type = l_rsm->r_limit_type;
7200                 l_rsm->r_limit_type = 0;
7201         }
7202         rack_free(rack, r_rsm);
7203         l_rsm->r_flags |= RACK_MERGED;
7204         return (l_rsm);
7205 }
7206
7207 /*
7208  * TLP Timer, here we simply setup what segment we want to
7209  * have the TLP expire on, the normal rack_output() will then
7210  * send it out.
7211  *
7212  * We return 1, saying don't proceed with rack_output only
7213  * when all timers have been stopped (destroyed PCB?).
7214  */
7215 static int
7216 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7217 {
7218         /*
7219          * Tail Loss Probe.
7220          */
7221         struct rack_sendmap *rsm = NULL;
7222         int insret __diagused;
7223         struct socket *so = tptosocket(tp);
7224         uint32_t amm;
7225         uint32_t out, avail;
7226         int collapsed_win = 0;
7227
7228         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7229                 /* Its not time yet */
7230                 return (0);
7231         }
7232         if (ctf_progress_timeout_check(tp, true)) {
7233                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7234                 return (-ETIMEDOUT);    /* tcp_drop() */
7235         }
7236         /*
7237          * A TLP timer has expired. We have been idle for 2 rtts. So we now
7238          * need to figure out how to force a full MSS segment out.
7239          */
7240         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7241         rack->r_ctl.retran_during_recovery = 0;
7242         rack->r_ctl.dsack_byte_cnt = 0;
7243         counter_u64_add(rack_tlp_tot, 1);
7244         if (rack->r_state && (rack->r_state != tp->t_state))
7245                 rack_set_state(tp, rack);
7246         avail = sbavail(&so->so_snd);
7247         out = tp->snd_max - tp->snd_una;
7248         if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7249                 /* special case, we need a retransmission */
7250                 collapsed_win = 1;
7251                 goto need_retran;
7252         }
7253         if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7254                 rack->r_ctl.dsack_persist--;
7255                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7256                         rack->r_ctl.num_dsack = 0;
7257                 }
7258                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7259         }
7260         if ((tp->t_flags & TF_GPUTINPROG) &&
7261             (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7262                 /*
7263                  * If this is the second in a row
7264                  * TLP and we are doing a measurement
7265                  * its time to abandon the measurement.
7266                  * Something is likely broken on
7267                  * the clients network and measuring a
7268                  * broken network does us no good.
7269                  */
7270                 tp->t_flags &= ~TF_GPUTINPROG;
7271                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7272                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
7273                                            tp->gput_seq,
7274                                            0, 0, 18, __LINE__, NULL, 0);
7275         }
7276         /*
7277          * Check our send oldest always settings, and if
7278          * there is an oldest to send jump to the need_retran.
7279          */
7280         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7281                 goto need_retran;
7282
7283         if (avail > out) {
7284                 /* New data is available */
7285                 amm = avail - out;
7286                 if (amm > ctf_fixed_maxseg(tp)) {
7287                         amm = ctf_fixed_maxseg(tp);
7288                         if ((amm + out) > tp->snd_wnd) {
7289                                 /* We are rwnd limited */
7290                                 goto need_retran;
7291                         }
7292                 } else if (amm < ctf_fixed_maxseg(tp)) {
7293                         /* not enough to fill a MTU */
7294                         goto need_retran;
7295                 }
7296                 if (IN_FASTRECOVERY(tp->t_flags)) {
7297                         /* Unlikely */
7298                         if (rack->rack_no_prr == 0) {
7299                                 if (out + amm <= tp->snd_wnd) {
7300                                         rack->r_ctl.rc_prr_sndcnt = amm;
7301                                         rack->r_ctl.rc_tlp_new_data = amm;
7302                                         rack_log_to_prr(rack, 4, 0, __LINE__);
7303                                 }
7304                         } else
7305                                 goto need_retran;
7306                 } else {
7307                         /* Set the send-new override */
7308                         if (out + amm <= tp->snd_wnd)
7309                                 rack->r_ctl.rc_tlp_new_data = amm;
7310                         else
7311                                 goto need_retran;
7312                 }
7313                 rack->r_ctl.rc_tlpsend = NULL;
7314                 counter_u64_add(rack_tlp_newdata, 1);
7315                 goto send;
7316         }
7317 need_retran:
7318         /*
7319          * Ok we need to arrange the last un-acked segment to be re-sent, or
7320          * optionally the first un-acked segment.
7321          */
7322         if (collapsed_win == 0) {
7323                 if (rack_always_send_oldest)
7324                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7325                 else {
7326                         rsm = tqhash_max(rack->r_ctl.tqh);
7327                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7328                                 rsm = rack_find_high_nonack(rack, rsm);
7329                         }
7330                 }
7331                 if (rsm == NULL) {
7332 #ifdef TCP_BLACKBOX
7333                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7334 #endif
7335                         goto out;
7336                 }
7337         } else {
7338                 /*
7339                  * We had a collapsed window, lets find
7340                  * the point before the collapse.
7341                  */
7342                 if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7343                         rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7344                 else {
7345                         rsm = tqhash_min(rack->r_ctl.tqh);
7346                 }
7347                 if (rsm == NULL) {
7348                         /* Huh */
7349                         goto out;
7350                 }
7351         }
7352         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7353                 /*
7354                  * We need to split this the last segment in two.
7355                  */
7356                 struct rack_sendmap *nrsm;
7357
7358                 nrsm = rack_alloc_full_limit(rack);
7359                 if (nrsm == NULL) {
7360                         /*
7361                          * No memory to split, we will just exit and punt
7362                          * off to the RXT timer.
7363                          */
7364                         goto out;
7365                 }
7366                 rack_clone_rsm(rack, nrsm, rsm,
7367                                (rsm->r_end - ctf_fixed_maxseg(tp)));
7368                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7369 #ifndef INVARIANTS
7370                 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7371 #else
7372                 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7373                         panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
7374                               nrsm, insret, rack, rsm);
7375                 }
7376 #endif
7377                 if (rsm->r_in_tmap) {
7378                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7379                         nrsm->r_in_tmap = 1;
7380                 }
7381                 rsm = nrsm;
7382         }
7383         rack->r_ctl.rc_tlpsend = rsm;
7384 send:
7385         /* Make sure output path knows we are doing a TLP */
7386         *doing_tlp = 1;
7387         rack->r_timer_override = 1;
7388         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7389         return (0);
7390 out:
7391         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7392         return (0);
7393 }
7394
7395 /*
7396  * Delayed ack Timer, here we simply need to setup the
7397  * ACK_NOW flag and remove the DELACK flag. From there
7398  * the output routine will send the ack out.
7399  *
7400  * We only return 1, saying don't proceed, if all timers
7401  * are stopped (destroyed PCB?).
7402  */
7403 static int
7404 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7405 {
7406
7407         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7408         tp->t_flags &= ~TF_DELACK;
7409         tp->t_flags |= TF_ACKNOW;
7410         KMOD_TCPSTAT_INC(tcps_delack);
7411         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7412         return (0);
7413 }
7414
7415 /*
7416  * Persists timer, here we simply send the
7417  * same thing as a keepalive will.
7418  * the one byte send.
7419  *
7420  * We only return 1, saying don't proceed, if all timers
7421  * are stopped (destroyed PCB?).
7422  */
7423 static int
7424 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7425 {
7426         struct tcptemp *t_template;
7427         int32_t retval = 1;
7428
7429         if (rack->rc_in_persist == 0)
7430                 return (0);
7431         if (ctf_progress_timeout_check(tp, false)) {
7432                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7433                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7434                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7435                 return (-ETIMEDOUT);    /* tcp_drop() */
7436         }
7437         /*
7438          * Persistence timer into zero window. Force a byte to be output, if
7439          * possible.
7440          */
7441         KMOD_TCPSTAT_INC(tcps_persisttimeo);
7442         /*
7443          * Hack: if the peer is dead/unreachable, we do not time out if the
7444          * window is closed.  After a full backoff, drop the connection if
7445          * the idle time (no responses to probes) reaches the maximum
7446          * backoff that we would use if retransmitting.
7447          */
7448         if (tp->t_rxtshift >= V_tcp_retries &&
7449             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
7450              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
7451                 KMOD_TCPSTAT_INC(tcps_persistdrop);
7452                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7453                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7454                 retval = -ETIMEDOUT;    /* tcp_drop() */
7455                 goto out;
7456         }
7457         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
7458             tp->snd_una == tp->snd_max)
7459                 rack_exit_persist(tp, rack, cts);
7460         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
7461         /*
7462          * If the user has closed the socket then drop a persisting
7463          * connection after a much reduced timeout.
7464          */
7465         if (tp->t_state > TCPS_CLOSE_WAIT &&
7466             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
7467                 KMOD_TCPSTAT_INC(tcps_persistdrop);
7468                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7469                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7470                 retval = -ETIMEDOUT;    /* tcp_drop() */
7471                 goto out;
7472         }
7473         t_template = tcpip_maketemplate(rack->rc_inp);
7474         if (t_template) {
7475                 /* only set it if we were answered */
7476                 if (rack->forced_ack == 0) {
7477                         rack->forced_ack = 1;
7478                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7479                 } else {
7480                         rack->probe_not_answered = 1;
7481                         counter_u64_add(rack_persists_loss, 1);
7482                         rack->r_ctl.persist_lost_ends++;
7483                 }
7484                 counter_u64_add(rack_persists_sends, 1);
7485                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
7486                 tcp_respond(tp, t_template->tt_ipgen,
7487                             &t_template->tt_t, (struct mbuf *)NULL,
7488                             tp->rcv_nxt, tp->snd_una - 1, 0);
7489                 /* This sends an ack */
7490                 if (tp->t_flags & TF_DELACK)
7491                         tp->t_flags &= ~TF_DELACK;
7492                 free(t_template, M_TEMP);
7493         }
7494         if (tp->t_rxtshift < V_tcp_retries)
7495                 tp->t_rxtshift++;
7496 out:
7497         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
7498         rack_start_hpts_timer(rack, tp, cts,
7499                               0, 0, 0);
7500         return (retval);
7501 }
7502
7503 /*
7504  * If a keepalive goes off, we had no other timers
7505  * happening. We always return 1 here since this
7506  * routine either drops the connection or sends
7507  * out a segment with respond.
7508  */
7509 static int
7510 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7511 {
7512         struct tcptemp *t_template;
7513         struct inpcb *inp = tptoinpcb(tp);
7514
7515         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
7516         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
7517         /*
7518          * Keep-alive timer went off; send something or drop connection if
7519          * idle for too long.
7520          */
7521         KMOD_TCPSTAT_INC(tcps_keeptimeo);
7522         if (tp->t_state < TCPS_ESTABLISHED)
7523                 goto dropit;
7524         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7525             tp->t_state <= TCPS_CLOSING) {
7526                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
7527                         goto dropit;
7528                 /*
7529                  * Send a packet designed to force a response if the peer is
7530                  * up and reachable: either an ACK if the connection is
7531                  * still alive, or an RST if the peer has closed the
7532                  * connection due to timeout or reboot. Using sequence
7533                  * number tp->snd_una-1 causes the transmitted zero-length
7534                  * segment to lie outside the receive window; by the
7535                  * protocol spec, this requires the correspondent TCP to
7536                  * respond.
7537                  */
7538                 KMOD_TCPSTAT_INC(tcps_keepprobe);
7539                 t_template = tcpip_maketemplate(inp);
7540                 if (t_template) {
7541                         if (rack->forced_ack == 0) {
7542                                 rack->forced_ack = 1;
7543                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7544                         } else {
7545                                 rack->probe_not_answered = 1;
7546                         }
7547                         tcp_respond(tp, t_template->tt_ipgen,
7548                             &t_template->tt_t, (struct mbuf *)NULL,
7549                             tp->rcv_nxt, tp->snd_una - 1, 0);
7550                         free(t_template, M_TEMP);
7551                 }
7552         }
7553         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
7554         return (1);
7555 dropit:
7556         KMOD_TCPSTAT_INC(tcps_keepdrops);
7557         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7558         return (-ETIMEDOUT);    /* tcp_drop() */
7559 }
7560
7561 /*
7562  * Retransmit helper function, clear up all the ack
7563  * flags and take care of important book keeping.
7564  */
7565 static void
7566 rack_remxt_tmr(struct tcpcb *tp)
7567 {
7568         /*
7569          * The retransmit timer went off, all sack'd blocks must be
7570          * un-acked.
7571          */
7572         struct rack_sendmap *rsm, *trsm = NULL;
7573         struct tcp_rack *rack;
7574
7575         rack = (struct tcp_rack *)tp->t_fb_ptr;
7576         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
7577         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
7578         if (rack->r_state && (rack->r_state != tp->t_state))
7579                 rack_set_state(tp, rack);
7580         /*
7581          * Ideally we would like to be able to
7582          * mark SACK-PASS on anything not acked here.
7583          *
7584          * However, if we do that we would burst out
7585          * all that data 1ms apart. This would be unwise,
7586          * so for now we will just let the normal rxt timer
7587          * and tlp timer take care of it.
7588          *
7589          * Also we really need to stick them back in sequence
7590          * order. This way we send in the proper order and any
7591          * sacks that come floating in will "re-ack" the data.
7592          * To do this we zap the tmap with an INIT and then
7593          * walk through and place every rsm in the RB tree
7594          * back in its seq ordered place.
7595          */
7596         TAILQ_INIT(&rack->r_ctl.rc_tmap);
7597
7598         TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
7599                 rsm->r_dupack = 0;
7600                 if (rack_verbose_logging)
7601                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7602                 /* We must re-add it back to the tlist */
7603                 if (trsm == NULL) {
7604                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7605                 } else {
7606                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
7607                 }
7608                 rsm->r_in_tmap = 1;
7609                 trsm = rsm;
7610                 if (rsm->r_flags & RACK_ACKED)
7611                         rsm->r_flags |= RACK_WAS_ACKED;
7612                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
7613                 rsm->r_flags |= RACK_MUST_RXT;
7614         }
7615         /* Clear the count (we just un-acked them) */
7616         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
7617         rack->r_ctl.rc_sacked = 0;
7618         rack->r_ctl.rc_sacklast = NULL;
7619         rack->r_ctl.rc_agg_delayed = 0;
7620         rack->r_early = 0;
7621         rack->r_ctl.rc_agg_early = 0;
7622         rack->r_late = 0;
7623         /* Clear the tlp rtx mark */
7624         rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7625         if (rack->r_ctl.rc_resend != NULL)
7626                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7627         rack->r_ctl.rc_prr_sndcnt = 0;
7628         rack_log_to_prr(rack, 6, 0, __LINE__);
7629         rack->r_timer_override = 1;
7630         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
7631 #ifdef TCP_SAD_DETECTION
7632             || (rack->sack_attack_disable != 0)
7633 #endif
7634                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
7635                 /*
7636                  * For non-sack customers new data
7637                  * needs to go out as retransmits until
7638                  * we retransmit up to snd_max.
7639                  */
7640                 rack->r_must_retran = 1;
7641                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
7642                                                 rack->r_ctl.rc_sacked);
7643         }
7644         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
7645 }
7646
7647 static void
7648 rack_convert_rtts(struct tcpcb *tp)
7649 {
7650         tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
7651         tp->t_rxtcur = RACK_REXMTVAL(tp);
7652         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7653                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
7654         }
7655         if (tp->t_rxtcur > rack_rto_max) {
7656                 tp->t_rxtcur = rack_rto_max;
7657         }
7658 }
7659
7660 static void
7661 rack_cc_conn_init(struct tcpcb *tp)
7662 {
7663         struct tcp_rack *rack;
7664         uint32_t srtt;
7665
7666         rack = (struct tcp_rack *)tp->t_fb_ptr;
7667         srtt = tp->t_srtt;
7668         cc_conn_init(tp);
7669         /*
7670          * Now convert to rack's internal format,
7671          * if required.
7672          */
7673         if ((srtt == 0) && (tp->t_srtt != 0))
7674                 rack_convert_rtts(tp);
7675         /*
7676          * We want a chance to stay in slowstart as
7677          * we create a connection. TCP spec says that
7678          * initially ssthresh is infinite. For our
7679          * purposes that is the snd_wnd.
7680          */
7681         if (tp->snd_ssthresh < tp->snd_wnd) {
7682                 tp->snd_ssthresh = tp->snd_wnd;
7683         }
7684         /*
7685          * We also want to assure a IW worth of
7686          * data can get inflight.
7687          */
7688         if (rc_init_window(rack) < tp->snd_cwnd)
7689                 tp->snd_cwnd = rc_init_window(rack);
7690 }
7691
7692 /*
7693  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
7694  * we will setup to retransmit the lowest seq number outstanding.
7695  */
7696 static int
7697 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7698 {
7699         struct inpcb *inp = tptoinpcb(tp);
7700         int32_t rexmt;
7701         int32_t retval = 0;
7702         bool isipv6;
7703
7704         if ((tp->t_flags & TF_GPUTINPROG) &&
7705             (tp->t_rxtshift)) {
7706                 /*
7707                  * We have had a second timeout
7708                  * measurements on successive rxt's are not profitable.
7709                  * It is unlikely to be of any use (the network is
7710                  * broken or the client went away).
7711                  */
7712                 tp->t_flags &= ~TF_GPUTINPROG;
7713                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7714                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
7715                                            tp->gput_seq,
7716                                            0, 0, 18, __LINE__, NULL, 0);
7717         }
7718         if (ctf_progress_timeout_check(tp, false)) {
7719                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7720                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7721                 return (-ETIMEDOUT);    /* tcp_drop() */
7722         }
7723         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
7724         rack->r_ctl.retran_during_recovery = 0;
7725         rack->rc_ack_required = 1;
7726         rack->r_ctl.dsack_byte_cnt = 0;
7727         if (IN_FASTRECOVERY(tp->t_flags))
7728                 tp->t_flags |= TF_WASFRECOVERY;
7729         else
7730                 tp->t_flags &= ~TF_WASFRECOVERY;
7731         if (IN_CONGRECOVERY(tp->t_flags))
7732                 tp->t_flags |= TF_WASCRECOVERY;
7733         else
7734                 tp->t_flags &= ~TF_WASCRECOVERY;
7735         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7736             (tp->snd_una == tp->snd_max)) {
7737                 /* Nothing outstanding .. nothing to do */
7738                 return (0);
7739         }
7740         if (rack->r_ctl.dsack_persist) {
7741                 rack->r_ctl.dsack_persist--;
7742                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7743                         rack->r_ctl.num_dsack = 0;
7744                 }
7745                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7746         }
7747         /*
7748          * Rack can only run one timer  at a time, so we cannot
7749          * run a KEEPINIT (gating SYN sending) and a retransmit
7750          * timer for the SYN. So if we are in a front state and
7751          * have a KEEPINIT timer we need to check the first transmit
7752          * against now to see if we have exceeded the KEEPINIT time
7753          * (if one is set).
7754          */
7755         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
7756             (TP_KEEPINIT(tp) != 0)) {
7757                 struct rack_sendmap *rsm;
7758
7759                 rsm = tqhash_min(rack->r_ctl.tqh);
7760                 if (rsm) {
7761                         /* Ok we have something outstanding to test keepinit with */
7762                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
7763                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
7764                                 /* We have exceeded the KEEPINIT time */
7765                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7766                                 goto drop_it;
7767                         }
7768                 }
7769         }
7770         /*
7771          * Retransmission timer went off.  Message has not been acked within
7772          * retransmit interval.  Back off to a longer retransmit interval
7773          * and retransmit one segment.
7774          */
7775         rack_remxt_tmr(tp);
7776         if ((rack->r_ctl.rc_resend == NULL) ||
7777             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
7778                 /*
7779                  * If the rwnd collapsed on
7780                  * the one we are retransmitting
7781                  * it does not count against the
7782                  * rxt count.
7783                  */
7784                 tp->t_rxtshift++;
7785         }
7786         if (tp->t_rxtshift > V_tcp_retries) {
7787                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7788 drop_it:
7789                 tp->t_rxtshift = V_tcp_retries;
7790                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
7791                 /* XXXGL: previously t_softerror was casted to uint16_t */
7792                 MPASS(tp->t_softerror >= 0);
7793                 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
7794                 goto out;       /* tcp_drop() */
7795         }
7796         if (tp->t_state == TCPS_SYN_SENT) {
7797                 /*
7798                  * If the SYN was retransmitted, indicate CWND to be limited
7799                  * to 1 segment in cc_conn_init().
7800                  */
7801                 tp->snd_cwnd = 1;
7802         } else if (tp->t_rxtshift == 1) {
7803                 /*
7804                  * first retransmit; record ssthresh and cwnd so they can be
7805                  * recovered if this turns out to be a "bad" retransmit. A
7806                  * retransmit is considered "bad" if an ACK for this segment
7807                  * is received within RTT/2 interval; the assumption here is
7808                  * that the ACK was already in flight.  See "On Estimating
7809                  * End-to-End Network Path Properties" by Allman and Paxson
7810                  * for more details.
7811                  */
7812                 tp->snd_cwnd_prev = tp->snd_cwnd;
7813                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
7814                 tp->snd_recover_prev = tp->snd_recover;
7815                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
7816                 tp->t_flags |= TF_PREVVALID;
7817         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
7818                 tp->t_flags &= ~TF_PREVVALID;
7819         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
7820         if ((tp->t_state == TCPS_SYN_SENT) ||
7821             (tp->t_state == TCPS_SYN_RECEIVED))
7822                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
7823         else
7824                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
7825
7826         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
7827            max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
7828         /*
7829          * We enter the path for PLMTUD if connection is established or, if
7830          * connection is FIN_WAIT_1 status, reason for the last is that if
7831          * amount of data we send is very small, we could send it in couple
7832          * of packets and process straight to FIN. In that case we won't
7833          * catch ESTABLISHED state.
7834          */
7835 #ifdef INET6
7836         isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
7837 #else
7838         isipv6 = false;
7839 #endif
7840         if (((V_tcp_pmtud_blackhole_detect == 1) ||
7841             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
7842             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
7843             ((tp->t_state == TCPS_ESTABLISHED) ||
7844             (tp->t_state == TCPS_FIN_WAIT_1))) {
7845                 /*
7846                  * Idea here is that at each stage of mtu probe (usually,
7847                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
7848                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
7849                  * should take care of that.
7850                  */
7851                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
7852                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
7853                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
7854                     tp->t_rxtshift % 2 == 0)) {
7855                         /*
7856                          * Enter Path MTU Black-hole Detection mechanism: -
7857                          * Disable Path MTU Discovery (IP "DF" bit). -
7858                          * Reduce MTU to lower value than what we negotiated
7859                          * with peer.
7860                          */
7861                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7862                                 /* Record that we may have found a black hole. */
7863                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7864                                 /* Keep track of previous MSS. */
7865                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7866                         }
7867
7868                         /*
7869                          * Reduce the MSS to blackhole value or to the
7870                          * default in an attempt to retransmit.
7871                          */
7872 #ifdef INET6
7873                         if (isipv6 &&
7874                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7875                                 /* Use the sysctl tuneable blackhole MSS. */
7876                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7877                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7878                         } else if (isipv6) {
7879                                 /* Use the default MSS. */
7880                                 tp->t_maxseg = V_tcp_v6mssdflt;
7881                                 /*
7882                                  * Disable Path MTU Discovery when we switch
7883                                  * to minmss.
7884                                  */
7885                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7886                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7887                         }
7888 #endif
7889 #if defined(INET6) && defined(INET)
7890                         else
7891 #endif
7892 #ifdef INET
7893                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7894                                 /* Use the sysctl tuneable blackhole MSS. */
7895                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
7896                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7897                         } else {
7898                                 /* Use the default MSS. */
7899                                 tp->t_maxseg = V_tcp_mssdflt;
7900                                 /*
7901                                  * Disable Path MTU Discovery when we switch
7902                                  * to minmss.
7903                                  */
7904                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7905                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7906                         }
7907 #endif
7908                 } else {
7909                         /*
7910                          * If further retransmissions are still unsuccessful
7911                          * with a lowered MTU, maybe this isn't a blackhole
7912                          * and we restore the previous MSS and blackhole
7913                          * detection flags. The limit '6' is determined by
7914                          * giving each probe stage (1448, 1188, 524) 2
7915                          * chances to recover.
7916                          */
7917                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
7918                             (tp->t_rxtshift >= 6)) {
7919                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
7920                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
7921                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
7922                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
7923                         }
7924                 }
7925         }
7926         /*
7927          * Disable RFC1323 and SACK if we haven't got any response to
7928          * our third SYN to work-around some broken terminal servers
7929          * (most of which have hopefully been retired) that have bad VJ
7930          * header compression code which trashes TCP segments containing
7931          * unknown-to-them TCP options.
7932          */
7933         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
7934             (tp->t_rxtshift == 3))
7935                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
7936         /*
7937          * If we backed off this far, our srtt estimate is probably bogus.
7938          * Clobber it so we'll take the next rtt measurement as our srtt;
7939          * move the current srtt into rttvar to keep the current retransmit
7940          * times until then.
7941          */
7942         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
7943 #ifdef INET6
7944                 if ((inp->inp_vflag & INP_IPV6) != 0)
7945                         in6_losing(inp);
7946                 else
7947 #endif
7948                         in_losing(inp);
7949                 tp->t_rttvar += tp->t_srtt;
7950                 tp->t_srtt = 0;
7951         }
7952         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
7953         tp->snd_recover = tp->snd_max;
7954         tp->t_flags |= TF_ACKNOW;
7955         tp->t_rtttime = 0;
7956         rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
7957 out:
7958         return (retval);
7959 }
7960
7961 static int
7962 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
7963 {
7964         int32_t ret = 0;
7965         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
7966
7967         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7968             (tp->t_flags & TF_GPUTINPROG)) {
7969                 /*
7970                  * We have a goodput in progress
7971                  * and we have entered a late state.
7972                  * Do we have enough data in the sb
7973                  * to handle the GPUT request?
7974                  */
7975                 uint32_t bytes;
7976
7977                 bytes = tp->gput_ack - tp->gput_seq;
7978                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
7979                         bytes += tp->gput_seq - tp->snd_una;
7980                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
7981                         /*
7982                          * There are not enough bytes in the socket
7983                          * buffer that have been sent to cover this
7984                          * measurement. Cancel it.
7985                          */
7986                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7987                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
7988                                                    tp->gput_seq,
7989                                                    0, 0, 18, __LINE__, NULL, 0);
7990                         tp->t_flags &= ~TF_GPUTINPROG;
7991                 }
7992         }
7993         if (timers == 0) {
7994                 return (0);
7995         }
7996         if (tp->t_state == TCPS_LISTEN) {
7997                 /* no timers on listen sockets */
7998                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
7999                         return (0);
8000                 return (1);
8001         }
8002         if ((timers & PACE_TMR_RACK) &&
8003             rack->rc_on_min_to) {
8004                 /*
8005                  * For the rack timer when we
8006                  * are on a min-timeout (which means rrr_conf = 3)
8007                  * we don't want to check the timer. It may
8008                  * be going off for a pace and thats ok we
8009                  * want to send the retransmit (if its ready).
8010                  *
8011                  * If its on a normal rack timer (non-min) then
8012                  * we will check if its expired.
8013                  */
8014                 goto skip_time_check;
8015         }
8016         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
8017                 uint32_t left;
8018
8019                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
8020                         ret = -1;
8021                         rack_log_to_processing(rack, cts, ret, 0);
8022                         return (0);
8023                 }
8024                 if (hpts_calling == 0) {
8025                         /*
8026                          * A user send or queued mbuf (sack) has called us? We
8027                          * return 0 and let the pacing guards
8028                          * deal with it if they should or
8029                          * should not cause a send.
8030                          */
8031                         ret = -2;
8032                         rack_log_to_processing(rack, cts, ret, 0);
8033                         return (0);
8034                 }
8035                 /*
8036                  * Ok our timer went off early and we are not paced false
8037                  * alarm, go back to sleep. We make sure we don't have
8038                  * no-sack wakeup on since we no longer have a PKT_OUTPUT
8039                  * flag in place.
8040                  */
8041                 rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8042                 ret = -3;
8043                 left = rack->r_ctl.rc_timer_exp - cts;
8044                 tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8045                 rack_log_to_processing(rack, cts, ret, left);
8046                 return (1);
8047         }
8048 skip_time_check:
8049         rack->rc_tmr_stopped = 0;
8050         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8051         if (timers & PACE_TMR_DELACK) {
8052                 ret = rack_timeout_delack(tp, rack, cts);
8053         } else if (timers & PACE_TMR_RACK) {
8054                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8055                 rack->r_fast_output = 0;
8056                 ret = rack_timeout_rack(tp, rack, cts);
8057         } else if (timers & PACE_TMR_TLP) {
8058                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8059                 ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8060         } else if (timers & PACE_TMR_RXT) {
8061                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8062                 rack->r_fast_output = 0;
8063                 ret = rack_timeout_rxt(tp, rack, cts);
8064         } else if (timers & PACE_TMR_PERSIT) {
8065                 ret = rack_timeout_persist(tp, rack, cts);
8066         } else if (timers & PACE_TMR_KEEP) {
8067                 ret = rack_timeout_keepalive(tp, rack, cts);
8068         }
8069         rack_log_to_processing(rack, cts, ret, timers);
8070         return (ret);
8071 }
8072
8073 static void
8074 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8075 {
8076         struct timeval tv;
8077         uint32_t us_cts, flags_on_entry;
8078         uint8_t hpts_removed = 0;
8079
8080         flags_on_entry = rack->r_ctl.rc_hpts_flags;
8081         us_cts = tcp_get_usecs(&tv);
8082         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8083             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8084              ((tp->snd_max - tp->snd_una) == 0))) {
8085                 tcp_hpts_remove(rack->rc_tp);
8086                 hpts_removed = 1;
8087                 /* If we were not delayed cancel out the flag. */
8088                 if ((tp->snd_max - tp->snd_una) == 0)
8089                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8090                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8091         }
8092         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8093                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8094                 if (tcp_in_hpts(rack->rc_tp) &&
8095                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8096                         /*
8097                          * Canceling timer's when we have no output being
8098                          * paced. We also must remove ourselves from the
8099                          * hpts.
8100                          */
8101                         tcp_hpts_remove(rack->rc_tp);
8102                         hpts_removed = 1;
8103                 }
8104                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8105         }
8106         if (hpts_removed == 0)
8107                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8108 }
8109
8110 static int
8111 rack_stopall(struct tcpcb *tp)
8112 {
8113         struct tcp_rack *rack;
8114         rack = (struct tcp_rack *)tp->t_fb_ptr;
8115         rack->t_timers_stopped = 1;
8116         return (0);
8117 }
8118
8119 static void
8120 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8121 {
8122         /*
8123          * Assure no timers are running.
8124          */
8125         if (tcp_timer_active(tp, TT_PERSIST)) {
8126                 /* We enter in persists, set the flag appropriately */
8127                 rack->rc_in_persist = 1;
8128         }
8129         if (tcp_in_hpts(rack->rc_tp)) {
8130                 tcp_hpts_remove(rack->rc_tp);
8131         }
8132 }
8133
8134 static void
8135 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8136     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag, int segsiz)
8137 {
8138         int32_t idx;
8139
8140         rsm->r_rtr_cnt++;
8141         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8142         rsm->r_dupack = 0;
8143         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8144                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8145                 rsm->r_flags |= RACK_OVERMAX;
8146         }
8147         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8148                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8149                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8150         }
8151         idx = rsm->r_rtr_cnt - 1;
8152         rsm->r_tim_lastsent[idx] = ts;
8153         /*
8154          * Here we don't add in the len of send, since its already
8155          * in snduna <->snd_max.
8156          */
8157         rsm->r_fas = ctf_flight_size(rack->rc_tp,
8158                                      rack->r_ctl.rc_sacked);
8159         if (rsm->r_flags & RACK_ACKED) {
8160                 /* Problably MTU discovery messing with us */
8161                 rsm->r_flags &= ~RACK_ACKED;
8162                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8163         }
8164         if (rsm->r_in_tmap) {
8165                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8166                 rsm->r_in_tmap = 0;
8167         }
8168         /* Lets make sure it really is in or not the GP window */
8169         rack_mark_in_gp_win(tp, rsm);
8170         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8171         rsm->r_in_tmap = 1;
8172         rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8173         /* Take off the must retransmit flag, if its on */
8174         if (rsm->r_flags & RACK_MUST_RXT) {
8175                 if (rack->r_must_retran)
8176                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8177                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8178                         /*
8179                          * We have retransmitted all we need. Clear
8180                          * any must retransmit flags.
8181                          */
8182                         rack->r_must_retran = 0;
8183                         rack->r_ctl.rc_out_at_rto = 0;
8184                 }
8185                 rsm->r_flags &= ~RACK_MUST_RXT;
8186         }
8187         /* Remove any collapsed flag */
8188         rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8189         if (rsm->r_flags & RACK_SACK_PASSED) {
8190                 /* We have retransmitted due to the SACK pass */
8191                 rsm->r_flags &= ~RACK_SACK_PASSED;
8192                 rsm->r_flags |= RACK_WAS_SACKPASS;
8193         }
8194 }
8195
8196 static uint32_t
8197 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8198     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag, int segsiz)
8199 {
8200         /*
8201          * We (re-)transmitted starting at rsm->r_start for some length
8202          * (possibly less than r_end.
8203          */
8204         struct rack_sendmap *nrsm;
8205         int insret __diagused;
8206         uint32_t c_end;
8207         int32_t len;
8208
8209         len = *lenp;
8210         c_end = rsm->r_start + len;
8211         if (SEQ_GEQ(c_end, rsm->r_end)) {
8212                 /*
8213                  * We retransmitted the whole piece or more than the whole
8214                  * slopping into the next rsm.
8215                  */
8216                 rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8217                 if (c_end == rsm->r_end) {
8218                         *lenp = 0;
8219                         return (0);
8220                 } else {
8221                         int32_t act_len;
8222
8223                         /* Hangs over the end return whats left */
8224                         act_len = rsm->r_end - rsm->r_start;
8225                         *lenp = (len - act_len);
8226                         return (rsm->r_end);
8227                 }
8228                 /* We don't get out of this block. */
8229         }
8230         /*
8231          * Here we retransmitted less than the whole thing which means we
8232          * have to split this into what was transmitted and what was not.
8233          */
8234         nrsm = rack_alloc_full_limit(rack);
8235         if (nrsm == NULL) {
8236                 /*
8237                  * We can't get memory, so lets not proceed.
8238                  */
8239                 *lenp = 0;
8240                 return (0);
8241         }
8242         /*
8243          * So here we are going to take the original rsm and make it what we
8244          * retransmitted. nrsm will be the tail portion we did not
8245          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8246          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8247          * 1, 6 and the new piece will be 6, 11.
8248          */
8249         rack_clone_rsm(rack, nrsm, rsm, c_end);
8250         nrsm->r_dupack = 0;
8251         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8252 #ifndef INVARIANTS
8253         (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8254 #else
8255         if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8256                 panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
8257                       nrsm, insret, rack, rsm);
8258         }
8259 #endif
8260         if (rsm->r_in_tmap) {
8261                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8262                 nrsm->r_in_tmap = 1;
8263         }
8264         rsm->r_flags &= (~RACK_HAS_FIN);
8265         rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8266         /* Log a split of rsm into rsm and nrsm */
8267         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8268         *lenp = 0;
8269         return (0);
8270 }
8271
8272 static void
8273 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8274                 uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8275                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb,
8276                 uint32_t s_moff, int hw_tls, int segsiz)
8277 {
8278         struct tcp_rack *rack;
8279         struct rack_sendmap *rsm, *nrsm;
8280         int insret __diagused;
8281
8282         register uint32_t snd_max, snd_una;
8283
8284         /*
8285          * Add to the RACK log of packets in flight or retransmitted. If
8286          * there is a TS option we will use the TS echoed, if not we will
8287          * grab a TS.
8288          *
8289          * Retransmissions will increment the count and move the ts to its
8290          * proper place. Note that if options do not include TS's then we
8291          * won't be able to effectively use the ACK for an RTT on a retran.
8292          *
8293          * Notes about r_start and r_end. Lets consider a send starting at
8294          * sequence 1 for 10 bytes. In such an example the r_start would be
8295          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8296          * This means that r_end is actually the first sequence for the next
8297          * slot (11).
8298          *
8299          */
8300         /*
8301          * If err is set what do we do XXXrrs? should we not add the thing?
8302          * -- i.e. return if err != 0 or should we pretend we sent it? --
8303          * i.e. proceed with add ** do this for now.
8304          */
8305         INP_WLOCK_ASSERT(tptoinpcb(tp));
8306         if (err)
8307                 /*
8308                  * We don't log errors -- we could but snd_max does not
8309                  * advance in this case either.
8310                  */
8311                 return;
8312
8313         if (th_flags & TH_RST) {
8314                 /*
8315                  * We don't log resets and we return immediately from
8316                  * sending
8317                  */
8318                 return;
8319         }
8320         rack = (struct tcp_rack *)tp->t_fb_ptr;
8321         snd_una = tp->snd_una;
8322         snd_max = tp->snd_max;
8323         if (th_flags & (TH_SYN | TH_FIN)) {
8324                 /*
8325                  * The call to rack_log_output is made before bumping
8326                  * snd_max. This means we can record one extra byte on a SYN
8327                  * or FIN if seq_out is adding more on and a FIN is present
8328                  * (and we are not resending).
8329                  */
8330                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
8331                         len++;
8332                 if (th_flags & TH_FIN)
8333                         len++;
8334                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
8335                         /*
8336                          * The add/update as not been done for the FIN/SYN
8337                          * yet.
8338                          */
8339                         snd_max = tp->snd_nxt;
8340                 }
8341         }
8342         if (SEQ_LEQ((seq_out + len), snd_una)) {
8343                 /* Are sending an old segment to induce an ack (keep-alive)? */
8344                 return;
8345         }
8346         if (SEQ_LT(seq_out, snd_una)) {
8347                 /* huh? should we panic? */
8348                 uint32_t end;
8349
8350                 end = seq_out + len;
8351                 seq_out = snd_una;
8352                 if (SEQ_GEQ(end, seq_out))
8353                         len = end - seq_out;
8354                 else
8355                         len = 0;
8356         }
8357         if (len == 0) {
8358                 /* We don't log zero window probes */
8359                 return;
8360         }
8361         if (IN_FASTRECOVERY(tp->t_flags)) {
8362                 rack->r_ctl.rc_prr_out += len;
8363         }
8364         /* First question is it a retransmission or new? */
8365         if (seq_out == snd_max) {
8366                 /* Its new */
8367                 rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
8368 again:
8369                 rsm = rack_alloc(rack);
8370                 if (rsm == NULL) {
8371                         /*
8372                          * Hmm out of memory and the tcb got destroyed while
8373                          * we tried to wait.
8374                          */
8375                         return;
8376                 }
8377                 if (th_flags & TH_FIN) {
8378                         rsm->r_flags = RACK_HAS_FIN|add_flag;
8379                 } else {
8380                         rsm->r_flags = add_flag;
8381                 }
8382                 if (hw_tls)
8383                         rsm->r_hw_tls = 1;
8384                 rsm->r_tim_lastsent[0] = cts;
8385                 rsm->r_rtr_cnt = 1;
8386                 rsm->r_rtr_bytes = 0;
8387                 if (th_flags & TH_SYN) {
8388                         /* The data space is one beyond snd_una */
8389                         rsm->r_flags |= RACK_HAS_SYN;
8390                 }
8391                 rsm->r_start = seq_out;
8392                 rsm->r_end = rsm->r_start + len;
8393                 rack_mark_in_gp_win(tp, rsm);
8394                 rsm->r_dupack = 0;
8395                 /*
8396                  * save off the mbuf location that
8397                  * sndmbuf_noadv returned (which is
8398                  * where we started copying from)..
8399                  */
8400                 rsm->m = s_mb;
8401                 rsm->soff = s_moff;
8402                 /*
8403                  * Here we do add in the len of send, since its not yet
8404                  * reflected in in snduna <->snd_max
8405                  */
8406                 rsm->r_fas = (ctf_flight_size(rack->rc_tp,
8407                                               rack->r_ctl.rc_sacked) +
8408                               (rsm->r_end - rsm->r_start));
8409                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
8410                 if (rsm->m) {
8411                         if (rsm->m->m_len <= rsm->soff) {
8412                                 /*
8413                                  * XXXrrs Question, will this happen?
8414                                  *
8415                                  * If sbsndptr is set at the correct place
8416                                  * then s_moff should always be somewhere
8417                                  * within rsm->m. But if the sbsndptr was
8418                                  * off then that won't be true. If it occurs
8419                                  * we need to walkout to the correct location.
8420                                  */
8421                                 struct mbuf *lm;
8422
8423                                 lm = rsm->m;
8424                                 while (lm->m_len <= rsm->soff) {
8425                                         rsm->soff -= lm->m_len;
8426                                         lm = lm->m_next;
8427                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
8428                                                              __func__, rack, s_moff, s_mb, rsm->soff));
8429                                 }
8430                                 rsm->m = lm;
8431                         }
8432                         rsm->orig_m_len = rsm->m->m_len;
8433                         rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
8434                 } else {
8435                         rsm->orig_m_len = 0;
8436                         rsm->orig_t_space = 0;
8437                 }
8438                 rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
8439                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8440                 /* Log a new rsm */
8441                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
8442 #ifndef INVARIANTS
8443                 (void)tqhash_insert(rack->r_ctl.tqh, rsm);
8444 #else
8445                 if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
8446                         panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
8447                               nrsm, insret, rack, rsm);
8448                 }
8449 #endif
8450                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8451                 rsm->r_in_tmap = 1;
8452                 /*
8453                  * Special case detection, is there just a single
8454                  * packet outstanding when we are not in recovery?
8455                  *
8456                  * If this is true mark it so.
8457                  */
8458                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
8459                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
8460                         struct rack_sendmap *prsm;
8461
8462                         prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
8463                         if (prsm)
8464                                 prsm->r_one_out_nr = 1;
8465                 }
8466                 return;
8467         }
8468         /*
8469          * If we reach here its a retransmission and we need to find it.
8470          */
8471 more:
8472         if (hintrsm && (hintrsm->r_start == seq_out)) {
8473                 rsm = hintrsm;
8474                 hintrsm = NULL;
8475         } else {
8476                 /* No hints sorry */
8477                 rsm = NULL;
8478         }
8479         if ((rsm) && (rsm->r_start == seq_out)) {
8480                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8481                 if (len == 0) {
8482                         return;
8483                 } else {
8484                         goto more;
8485                 }
8486         }
8487         /* Ok it was not the last pointer go through it the hard way. */
8488 refind:
8489         rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
8490         if (rsm) {
8491                 if (rsm->r_start == seq_out) {
8492                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8493                         if (len == 0) {
8494                                 return;
8495                         } else {
8496                                 goto refind;
8497                         }
8498                 }
8499                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
8500                         /* Transmitted within this piece */
8501                         /*
8502                          * Ok we must split off the front and then let the
8503                          * update do the rest
8504                          */
8505                         nrsm = rack_alloc_full_limit(rack);
8506                         if (nrsm == NULL) {
8507                                 rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
8508                                 return;
8509                         }
8510                         /*
8511                          * copy rsm to nrsm and then trim the front of rsm
8512                          * to not include this part.
8513                          */
8514                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
8515                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8516 #ifndef INVARIANTS
8517                         (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8518 #else
8519                         if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8520                                 panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
8521                                       nrsm, insret, rack, rsm);
8522                         }
8523 #endif
8524                         if (rsm->r_in_tmap) {
8525                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8526                                 nrsm->r_in_tmap = 1;
8527                         }
8528                         rsm->r_flags &= (~RACK_HAS_FIN);
8529                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
8530                         if (len == 0) {
8531                                 return;
8532                         } else if (len > 0)
8533                                 goto refind;
8534                 }
8535         }
8536         /*
8537          * Hmm not found in map did they retransmit both old and on into the
8538          * new?
8539          */
8540         if (seq_out == tp->snd_max) {
8541                 goto again;
8542         } else if (SEQ_LT(seq_out, tp->snd_max)) {
8543 #ifdef INVARIANTS
8544                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
8545                        seq_out, len, tp->snd_una, tp->snd_max);
8546                 printf("Starting Dump of all rack entries\n");
8547                 TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
8548                         printf("rsm:%p start:%u end:%u\n",
8549                                rsm, rsm->r_start, rsm->r_end);
8550                 }
8551                 printf("Dump complete\n");
8552                 panic("seq_out not found rack:%p tp:%p",
8553                       rack, tp);
8554 #endif
8555         } else {
8556 #ifdef INVARIANTS
8557                 /*
8558                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
8559                  * flag)
8560                  */
8561                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
8562                       seq_out, len, tp->snd_max, tp);
8563 #endif
8564         }
8565 }
8566
8567 /*
8568  * Record one of the RTT updates from an ack into
8569  * our sample structure.
8570  */
8571
8572 static void
8573 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
8574                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
8575 {
8576         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8577             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
8578                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
8579         }
8580         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8581             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
8582                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
8583         }
8584         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
8585             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
8586                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
8587             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
8588                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
8589         }
8590         if ((confidence == 1) &&
8591             ((rsm == NULL) ||
8592              (rsm->r_just_ret) ||
8593              (rsm->r_one_out_nr &&
8594               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
8595                 /*
8596                  * If the rsm had a just return
8597                  * hit it then we can't trust the
8598                  * rtt measurement for buffer deterimination
8599                  * Note that a confidence of 2, indicates
8600                  * SACK'd which overrides the r_just_ret or
8601                  * the r_one_out_nr. If it was a CUM-ACK and
8602                  * we had only two outstanding, but get an
8603                  * ack for only 1. Then that also lowers our
8604                  * confidence.
8605                  */
8606                 confidence = 0;
8607         }
8608         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8609             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
8610                 if (rack->r_ctl.rack_rs.confidence == 0) {
8611                         /*
8612                          * We take anything with no current confidence
8613                          * saved.
8614                          */
8615                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8616                         rack->r_ctl.rack_rs.confidence = confidence;
8617                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8618                 } else if (confidence != 0) {
8619                         /*
8620                          * Once we have a confident number,
8621                          * we can update it with a smaller
8622                          * value since this confident number
8623                          * may include the DSACK time until
8624                          * the next segment (the second one) arrived.
8625                          */
8626                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8627                         rack->r_ctl.rack_rs.confidence = confidence;
8628                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8629                 }
8630         }
8631         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
8632         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
8633         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
8634         rack->r_ctl.rack_rs.rs_rtt_cnt++;
8635 }
8636
8637 /*
8638  * Collect new round-trip time estimate
8639  * and update averages and current timeout.
8640  */
8641 static void
8642 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
8643 {
8644         int32_t delta;
8645         int32_t rtt;
8646
8647         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
8648                 /* No valid sample */
8649                 return;
8650         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
8651                 /* We are to use the lowest RTT seen in a single ack */
8652                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
8653         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
8654                 /* We are to use the highest RTT seen in a single ack */
8655                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
8656         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
8657                 /* We are to use the average RTT seen in a single ack */
8658                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
8659                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
8660         } else {
8661 #ifdef INVARIANTS
8662                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
8663 #endif
8664                 return;
8665         }
8666         if (rtt == 0)
8667                 rtt = 1;
8668         if (rack->rc_gp_rtt_set == 0) {
8669                 /*
8670                  * With no RTT we have to accept
8671                  * even one we are not confident of.
8672                  */
8673                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
8674                 rack->rc_gp_rtt_set = 1;
8675         } else if (rack->r_ctl.rack_rs.confidence) {
8676                 /* update the running gp srtt */
8677                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
8678                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
8679         }
8680         if (rack->r_ctl.rack_rs.confidence) {
8681                 /*
8682                  * record the low and high for highly buffered path computation,
8683                  * we only do this if we are confident (not a retransmission).
8684                  */
8685                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
8686                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8687                 }
8688                 if (rack->rc_highly_buffered == 0) {
8689                         /*
8690                          * Currently once we declare a path has
8691                          * highly buffered there is no going
8692                          * back, which may be a problem...
8693                          */
8694                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
8695                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
8696                                                      rack->r_ctl.rc_highest_us_rtt,
8697                                                      rack->r_ctl.rc_lowest_us_rtt,
8698                                                      RACK_RTTS_SEEHBP);
8699                                 rack->rc_highly_buffered = 1;
8700                         }
8701                 }
8702         }
8703         if ((rack->r_ctl.rack_rs.confidence) ||
8704             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
8705                 /*
8706                  * If we are highly confident of it <or> it was
8707                  * never retransmitted we accept it as the last us_rtt.
8708                  */
8709                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8710                 /* The lowest rtt can be set if its was not retransmited */
8711                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
8712                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8713                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
8714                                 rack->r_ctl.rc_lowest_us_rtt = 1;
8715                 }
8716         }
8717         rack = (struct tcp_rack *)tp->t_fb_ptr;
8718         if (tp->t_srtt != 0) {
8719                 /*
8720                  * We keep a simple srtt in microseconds, like our rtt
8721                  * measurement. We don't need to do any tricks with shifting
8722                  * etc. Instead we just add in 1/8th of the new measurement
8723                  * and subtract out 1/8 of the old srtt. We do the same with
8724                  * the variance after finding the absolute value of the
8725                  * difference between this sample and the current srtt.
8726                  */
8727                 delta = tp->t_srtt - rtt;
8728                 /* Take off 1/8th of the current sRTT */
8729                 tp->t_srtt -= (tp->t_srtt >> 3);
8730                 /* Add in 1/8th of the new RTT just measured */
8731                 tp->t_srtt += (rtt >> 3);
8732                 if (tp->t_srtt <= 0)
8733                         tp->t_srtt = 1;
8734                 /* Now lets make the absolute value of the variance */
8735                 if (delta < 0)
8736                         delta = -delta;
8737                 /* Subtract out 1/8th */
8738                 tp->t_rttvar -= (tp->t_rttvar >> 3);
8739                 /* Add in 1/8th of the new variance we just saw */
8740                 tp->t_rttvar += (delta >> 3);
8741                 if (tp->t_rttvar <= 0)
8742                         tp->t_rttvar = 1;
8743         } else {
8744                 /*
8745                  * No rtt measurement yet - use the unsmoothed rtt. Set the
8746                  * variance to half the rtt (so our first retransmit happens
8747                  * at 3*rtt).
8748                  */
8749                 tp->t_srtt = rtt;
8750                 tp->t_rttvar = rtt >> 1;
8751         }
8752         rack->rc_srtt_measure_made = 1;
8753         KMOD_TCPSTAT_INC(tcps_rttupdated);
8754         if (tp->t_rttupdated < UCHAR_MAX)
8755                 tp->t_rttupdated++;
8756 #ifdef STATS
8757         if (rack_stats_gets_ms_rtt == 0) {
8758                 /* Send in the microsecond rtt used for rxt timeout purposes */
8759                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
8760         } else if (rack_stats_gets_ms_rtt == 1) {
8761                 /* Send in the millisecond rtt used for rxt timeout purposes */
8762                 int32_t ms_rtt;
8763
8764                 /* Round up */
8765                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8766                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8767         } else if (rack_stats_gets_ms_rtt == 2) {
8768                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
8769                 int32_t ms_rtt;
8770
8771                 /* Round up */
8772                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8773                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8774         }  else {
8775                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
8776                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8777         }
8778         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8779 #endif
8780         /*
8781          * the retransmit should happen at rtt + 4 * rttvar. Because of the
8782          * way we do the smoothing, srtt and rttvar will each average +1/2
8783          * tick of bias.  When we compute the retransmit timer, we want 1/2
8784          * tick of rounding and 1 extra tick because of +-1/2 tick
8785          * uncertainty in the firing of the timer.  The bias will give us
8786          * exactly the 1.5 tick we need.  But, because the bias is
8787          * statistical, we have to test that we don't drop below the minimum
8788          * feasible timer (which is 2 ticks).
8789          */
8790         tp->t_rxtshift = 0;
8791         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8792                       max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
8793         rack_log_rtt_sample(rack, rtt);
8794         tp->t_softerror = 0;
8795 }
8796
8797
8798 static void
8799 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
8800 {
8801         /*
8802          * Apply to filter the inbound us-rtt at us_cts.
8803          */
8804         uint32_t old_rtt;
8805
8806         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
8807         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
8808                                us_rtt, us_cts);
8809         if (old_rtt > us_rtt) {
8810                 /* We just hit a new lower rtt time */
8811                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
8812                                      __LINE__, RACK_RTTS_NEWRTT);
8813                 /*
8814                  * Only count it if its lower than what we saw within our
8815                  * calculated range.
8816                  */
8817                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
8818                         if (rack_probertt_lower_within &&
8819                             rack->rc_gp_dyn_mul &&
8820                             (rack->use_fixed_rate == 0) &&
8821                             (rack->rc_always_pace)) {
8822                                 /*
8823                                  * We are seeing a new lower rtt very close
8824                                  * to the time that we would have entered probe-rtt.
8825                                  * This is probably due to the fact that a peer flow
8826                                  * has entered probe-rtt. Lets go in now too.
8827                                  */
8828                                 uint32_t val;
8829
8830                                 val = rack_probertt_lower_within * rack_time_between_probertt;
8831                                 val /= 100;
8832                                 if ((rack->in_probe_rtt == 0)  &&
8833                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
8834                                         rack_enter_probertt(rack, us_cts);
8835                                 }
8836                         }
8837                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
8838                 }
8839         }
8840 }
8841
8842 static int
8843 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
8844     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
8845 {
8846         uint32_t us_rtt;
8847         int32_t i, all;
8848         uint32_t t, len_acked;
8849
8850         if ((rsm->r_flags & RACK_ACKED) ||
8851             (rsm->r_flags & RACK_WAS_ACKED))
8852                 /* Already done */
8853                 return (0);
8854         if (rsm->r_no_rtt_allowed) {
8855                 /* Not allowed */
8856                 return (0);
8857         }
8858         if (ack_type == CUM_ACKED) {
8859                 if (SEQ_GT(th_ack, rsm->r_end)) {
8860                         len_acked = rsm->r_end - rsm->r_start;
8861                         all = 1;
8862                 } else {
8863                         len_acked = th_ack - rsm->r_start;
8864                         all = 0;
8865                 }
8866         } else {
8867                 len_acked = rsm->r_end - rsm->r_start;
8868                 all = 0;
8869         }
8870         if (rsm->r_rtr_cnt == 1) {
8871
8872                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8873                 if ((int)t <= 0)
8874                         t = 1;
8875                 if (!tp->t_rttlow || tp->t_rttlow > t)
8876                         tp->t_rttlow = t;
8877                 if (!rack->r_ctl.rc_rack_min_rtt ||
8878                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8879                         rack->r_ctl.rc_rack_min_rtt = t;
8880                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
8881                                 rack->r_ctl.rc_rack_min_rtt = 1;
8882                         }
8883                 }
8884                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8885                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8886                 else
8887                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8888                 if (us_rtt == 0)
8889                         us_rtt = 1;
8890                 if (CC_ALGO(tp)->rttsample != NULL) {
8891                         /* Kick the RTT to the CC */
8892                         CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8893                 }
8894                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
8895                 if (ack_type == SACKED) {
8896                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
8897                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
8898                 } else {
8899                         /*
8900                          * We need to setup what our confidence
8901                          * is in this ack.
8902                          *
8903                          * If the rsm was app limited and it is
8904                          * less than a mss in length (the end
8905                          * of the send) then we have a gap. If we
8906                          * were app limited but say we were sending
8907                          * multiple MSS's then we are more confident
8908                          * int it.
8909                          *
8910                          * When we are not app-limited then we see if
8911                          * the rsm is being included in the current
8912                          * measurement, we tell this by the app_limited_needs_set
8913                          * flag.
8914                          *
8915                          * Note that being cwnd blocked is not applimited
8916                          * as well as the pacing delay between packets which
8917                          * are sending only 1 or 2 MSS's also will show up
8918                          * in the RTT. We probably need to examine this algorithm
8919                          * a bit more and enhance it to account for the delay
8920                          * between rsm's. We could do that by saving off the
8921                          * pacing delay of each rsm (in an rsm) and then
8922                          * factoring that in somehow though for now I am
8923                          * not sure how :)
8924                          */
8925                         int calc_conf = 0;
8926
8927                         if (rsm->r_flags & RACK_APP_LIMITED) {
8928                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
8929                                         calc_conf = 0;
8930                                 else
8931                                         calc_conf = 1;
8932                         } else if (rack->app_limited_needs_set == 0) {
8933                                 calc_conf = 1;
8934                         } else {
8935                                 calc_conf = 0;
8936                         }
8937                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
8938                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
8939                                             calc_conf, rsm, rsm->r_rtr_cnt);
8940                 }
8941                 if ((rsm->r_flags & RACK_TLP) &&
8942                     (!IN_FASTRECOVERY(tp->t_flags))) {
8943                         /* Segment was a TLP and our retrans matched */
8944                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
8945                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
8946                         }
8947                 }
8948                 if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
8949                     (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
8950                             (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
8951                         /* New more recent rack_tmit_time */
8952                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8953                         if (rack->r_ctl.rc_rack_tmit_time == 0)
8954                                 rack->r_ctl.rc_rack_tmit_time = 1;
8955                         rack->rc_rack_rtt = t;
8956                 }
8957                 return (1);
8958         }
8959         /*
8960          * We clear the soft/rxtshift since we got an ack.
8961          * There is no assurance we will call the commit() function
8962          * so we need to clear these to avoid incorrect handling.
8963          */
8964         tp->t_rxtshift = 0;
8965         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8966                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
8967         tp->t_softerror = 0;
8968         if (to && (to->to_flags & TOF_TS) &&
8969             (ack_type == CUM_ACKED) &&
8970             (to->to_tsecr) &&
8971             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
8972                 /*
8973                  * Now which timestamp does it match? In this block the ACK
8974                  * must be coming from a previous transmission.
8975                  */
8976                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
8977                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
8978                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8979                                 if ((int)t <= 0)
8980                                         t = 1;
8981                                 if (CC_ALGO(tp)->rttsample != NULL) {
8982                                         /*
8983                                          * Kick the RTT to the CC, here
8984                                          * we lie a bit in that we know the
8985                                          * retransmission is correct even though
8986                                          * we retransmitted. This is because
8987                                          * we match the timestamps.
8988                                          */
8989                                         if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
8990                                                 us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
8991                                         else
8992                                                 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
8993                                         CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8994                                 }
8995                                 if ((i + 1) < rsm->r_rtr_cnt) {
8996                                         /*
8997                                          * The peer ack'd from our previous
8998                                          * transmission. We have a spurious
8999                                          * retransmission and thus we dont
9000                                          * want to update our rack_rtt.
9001                                          *
9002                                          * Hmm should there be a CC revert here?
9003                                          *
9004                                          */
9005                                         return (0);
9006                                 }
9007                                 if (!tp->t_rttlow || tp->t_rttlow > t)
9008                                         tp->t_rttlow = t;
9009                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9010                                         rack->r_ctl.rc_rack_min_rtt = t;
9011                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
9012                                                 rack->r_ctl.rc_rack_min_rtt = 1;
9013                                         }
9014                                 }
9015                                 if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9016                                     (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9017                                             (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9018                                         /* New more recent rack_tmit_time */
9019                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9020                                         if (rack->r_ctl.rc_rack_tmit_time == 0)
9021                                                 rack->r_ctl.rc_rack_tmit_time = 1;
9022                                         rack->rc_rack_rtt = t;
9023                                 }
9024                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9025                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9026                                                     rsm->r_rtr_cnt);
9027                                 return (1);
9028                         }
9029                 }
9030                 /* If we are logging log out the sendmap */
9031                 if (tcp_bblogging_on(rack->rc_tp)) {
9032                         for (i = 0; i < rsm->r_rtr_cnt; i++) {
9033                                 rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9034                         }
9035                 }
9036                 goto ts_not_found;
9037         } else {
9038                 /*
9039                  * Ok its a SACK block that we retransmitted. or a windows
9040                  * machine without timestamps. We can tell nothing from the
9041                  * time-stamp since its not there or the time the peer last
9042                  * recieved a segment that moved forward its cum-ack point.
9043                  */
9044 ts_not_found:
9045                 i = rsm->r_rtr_cnt - 1;
9046                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9047                 if ((int)t <= 0)
9048                         t = 1;
9049                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9050                         /*
9051                          * We retransmitted and the ack came back in less
9052                          * than the smallest rtt we have observed. We most
9053                          * likely did an improper retransmit as outlined in
9054                          * 6.2 Step 2 point 2 in the rack-draft so we
9055                          * don't want to update our rack_rtt. We in
9056                          * theory (in future) might want to think about reverting our
9057                          * cwnd state but we won't for now.
9058                          */
9059                         return (0);
9060                 } else if (rack->r_ctl.rc_rack_min_rtt) {
9061                         /*
9062                          * We retransmitted it and the retransmit did the
9063                          * job.
9064                          */
9065                         if (!rack->r_ctl.rc_rack_min_rtt ||
9066                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9067                                 rack->r_ctl.rc_rack_min_rtt = t;
9068                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
9069                                         rack->r_ctl.rc_rack_min_rtt = 1;
9070                                 }
9071                         }
9072                         if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9073                             (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9074                                     (uint32_t)rsm->r_tim_lastsent[i]))) {
9075                                 /* New more recent rack_tmit_time */
9076                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9077                                 if (rack->r_ctl.rc_rack_tmit_time == 0)
9078                                         rack->r_ctl.rc_rack_tmit_time = 1;
9079                                 rack->rc_rack_rtt = t;
9080                         }
9081                         return (1);
9082                 }
9083         }
9084         return (0);
9085 }
9086
9087 /*
9088  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9089  */
9090 static void
9091 rack_log_sack_passed(struct tcpcb *tp,
9092     struct tcp_rack *rack, struct rack_sendmap *rsm)
9093 {
9094         struct rack_sendmap *nrsm;
9095
9096         nrsm = rsm;
9097         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9098             rack_head, r_tnext) {
9099                 if (nrsm == rsm) {
9100                         /* Skip original segment he is acked */
9101                         continue;
9102                 }
9103                 if (nrsm->r_flags & RACK_ACKED) {
9104                         /*
9105                          * Skip ack'd segments, though we
9106                          * should not see these, since tmap
9107                          * should not have ack'd segments.
9108                          */
9109                         continue;
9110                 }
9111                 if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9112                         /*
9113                          * If the peer dropped the rwnd on
9114                          * these then we don't worry about them.
9115                          */
9116                         continue;
9117                 }
9118                 if (nrsm->r_flags & RACK_SACK_PASSED) {
9119                         /*
9120                          * We found one that is already marked
9121                          * passed, we have been here before and
9122                          * so all others below this are marked.
9123                          */
9124                         break;
9125                 }
9126                 nrsm->r_flags |= RACK_SACK_PASSED;
9127                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9128         }
9129 }
9130
9131 static void
9132 rack_need_set_test(struct tcpcb *tp,
9133                    struct tcp_rack *rack,
9134                    struct rack_sendmap *rsm,
9135                    tcp_seq th_ack,
9136                    int line,
9137                    int use_which)
9138 {
9139         struct rack_sendmap *s_rsm;
9140
9141         if ((tp->t_flags & TF_GPUTINPROG) &&
9142             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9143                 /*
9144                  * We were app limited, and this ack
9145                  * butts up or goes beyond the point where we want
9146                  * to start our next measurement. We need
9147                  * to record the new gput_ts as here and
9148                  * possibly update the start sequence.
9149                  */
9150                 uint32_t seq, ts;
9151
9152                 if (rsm->r_rtr_cnt > 1) {
9153                         /*
9154                          * This is a retransmit, can we
9155                          * really make any assessment at this
9156                          * point?  We are not really sure of
9157                          * the timestamp, is it this or the
9158                          * previous transmission?
9159                          *
9160                          * Lets wait for something better that
9161                          * is not retransmitted.
9162                          */
9163                         return;
9164                 }
9165                 seq = tp->gput_seq;
9166                 ts = tp->gput_ts;
9167                 rack->app_limited_needs_set = 0;
9168                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9169                 /* Do we start at a new end? */
9170                 if ((use_which == RACK_USE_BEG) &&
9171                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9172                         /*
9173                          * When we get an ACK that just eats
9174                          * up some of the rsm, we set RACK_USE_BEG
9175                          * since whats at r_start (i.e. th_ack)
9176                          * is left unacked and thats where the
9177                          * measurement now starts.
9178                          */
9179                         tp->gput_seq = rsm->r_start;
9180                 }
9181                 if ((use_which == RACK_USE_END) &&
9182                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9183                         /*
9184                          * We use the end when the cumack
9185                          * is moving forward and completely
9186                          * deleting the rsm passed so basically
9187                          * r_end holds th_ack.
9188                          *
9189                          * For SACK's we also want to use the end
9190                          * since this piece just got sacked and
9191                          * we want to target anything after that
9192                          * in our measurement.
9193                          */
9194                         tp->gput_seq = rsm->r_end;
9195                 }
9196                 if (use_which == RACK_USE_END_OR_THACK) {
9197                         /*
9198                          * special case for ack moving forward,
9199                          * not a sack, we need to move all the
9200                          * way up to where this ack cum-ack moves
9201                          * to.
9202                          */
9203                         if (SEQ_GT(th_ack, rsm->r_end))
9204                                 tp->gput_seq = th_ack;
9205                         else
9206                                 tp->gput_seq = rsm->r_end;
9207                 }
9208                 if (SEQ_LT(tp->gput_seq, tp->snd_max))
9209                         s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9210                 else
9211                         s_rsm = NULL;
9212                 /*
9213                  * Pick up the correct send time if we can the rsm passed in
9214                  * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9215                  * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9216                  * find a different seq i.e. the next send up.
9217                  *
9218                  * If that has not been sent, s_rsm will be NULL and we must
9219                  * arrange it so this function will get called again by setting
9220                  * app_limited_needs_set.
9221                  */
9222                 if (s_rsm)
9223                         rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9224                 else {
9225                         /* If we hit here we have to have *not* sent tp->gput_seq */
9226                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9227                         /* Set it up so we will go through here again */
9228                         rack->app_limited_needs_set = 1;
9229                 }
9230                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9231                         /*
9232                          * We moved beyond this guy's range, re-calculate
9233                          * the new end point.
9234                          */
9235                         if (rack->rc_gp_filled == 0) {
9236                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9237                         } else {
9238                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9239                         }
9240                 }
9241                 /*
9242                  * We are moving the goal post, we may be able to clear the
9243                  * measure_saw_probe_rtt flag.
9244                  */
9245                 if ((rack->in_probe_rtt == 0) &&
9246                     (rack->measure_saw_probe_rtt) &&
9247                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9248                         rack->measure_saw_probe_rtt = 0;
9249                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9250                                            seq, tp->gput_seq,
9251                                            (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9252                                             (uint64_t)rack->r_ctl.rc_gp_output_ts),
9253                                            5, line, NULL, 0);
9254                 if (rack->rc_gp_filled &&
9255                     ((tp->gput_ack - tp->gput_seq) <
9256                      max(rc_init_window(rack), (MIN_GP_WIN *
9257                                                 ctf_fixed_maxseg(tp))))) {
9258                         uint32_t ideal_amount;
9259
9260                         ideal_amount = rack_get_measure_window(tp, rack);
9261                         if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9262                                 /*
9263                                  * There is no sense of continuing this measurement
9264                                  * because its too small to gain us anything we
9265                                  * trust. Skip it and that way we can start a new
9266                                  * measurement quicker.
9267                                  */
9268                                 tp->t_flags &= ~TF_GPUTINPROG;
9269                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9270                                                            0, 0,
9271                                                            (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9272                                                             (uint64_t)rack->r_ctl.rc_gp_output_ts),
9273                                                            6, __LINE__, NULL, 0);
9274                         } else {
9275                                 /*
9276                                  * Reset the window further out.
9277                                  */
9278                                 tp->gput_ack = tp->gput_seq + ideal_amount;
9279                         }
9280                 }
9281                 rack_tend_gp_marks(tp, rack);
9282                 rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9283         }
9284 }
9285
9286 static inline int
9287 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9288 {
9289         if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
9290                 /* Behind our TLP definition or right at */
9291                 return (0);
9292         }
9293         if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
9294                 /* The start is beyond or right at our end of TLP definition */
9295                 return (0);
9296         }
9297         /* It has to be a sub-part of the original TLP recorded */
9298         return (1);
9299 }
9300
9301
9302
9303 static uint32_t
9304 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
9305                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
9306                    int *no_extra,
9307                    int *moved_two, uint32_t segsiz)
9308 {
9309         uint32_t start, end, changed = 0;
9310         struct rack_sendmap stack_map;
9311         struct rack_sendmap *rsm, *nrsm, *prev, *next;
9312         int insret __diagused;
9313         int32_t used_ref = 1;
9314         int moved = 0;
9315 #ifdef TCP_SAD_DETECTION
9316         int allow_segsiz;
9317         int first_time_through = 1;
9318 #endif
9319         int noextra = 0;
9320         int can_use_hookery = 0;
9321
9322         start = sack->start;
9323         end = sack->end;
9324         rsm = *prsm;
9325
9326 #ifdef TCP_SAD_DETECTION
9327         /*
9328          * There are a strange number of proxys and meddle boxes in the world
9329          * that seem to cut up segments on different boundaries. This gets us
9330          * smaller sacks that are still ok in terms of it being an attacker.
9331          * We use the base segsiz to calculate an allowable smallness but
9332          * also enforce a min on the segsiz in case it is an attacker playing
9333          * games with MSS. So basically if the sack arrives and it is
9334          * larger than a worse case 960 bytes, we don't classify the guy
9335          * as supicious.
9336          */
9337         allow_segsiz = max(segsiz, 1200) * sad_seg_size_per;
9338         allow_segsiz /= 1000;
9339 #endif
9340 do_rest_ofb:
9341         if ((rsm == NULL) ||
9342             (SEQ_LT(end, rsm->r_start)) ||
9343             (SEQ_GEQ(start, rsm->r_end)) ||
9344             (SEQ_LT(start, rsm->r_start))) {
9345                 /*
9346                  * We are not in the right spot,
9347                  * find the correct spot in the tree.
9348                  */
9349                 used_ref = 0;
9350                 rsm = tqhash_find(rack->r_ctl.tqh, start);
9351                 moved++;
9352         }
9353         if (rsm == NULL) {
9354                 /* TSNH */
9355                 goto out;
9356         }
9357 #ifdef TCP_SAD_DETECTION
9358         /* Now we must check for suspicous activity */
9359         if ((first_time_through == 1) &&
9360             ((end - start) < min((rsm->r_end - rsm->r_start), allow_segsiz)) &&
9361             ((rsm->r_flags & RACK_PMTU_CHG) == 0) &&
9362             ((rsm->r_flags & RACK_TLP) == 0)) {
9363                 /*
9364                  * Its less than a full MSS or the segment being acked
9365                  * this should only happen if the rsm in question had the
9366                  * r_just_ret flag set <and> the end matches the end of
9367                  * the rsm block.
9368                  *
9369                  * Note we do not look at segments that have had TLP's on
9370                  * them since we can get un-reported rwnd collapses that
9371                  * basically we TLP on and then we get back a sack block
9372                  * that goes from the start to only a small way.
9373                  *
9374                  */
9375                 int loss, ok;
9376
9377                 ok = 0;
9378                 if (SEQ_GEQ(end, rsm->r_end)) {
9379                         if (rsm->r_just_ret == 1) {
9380                                 /* This was at the end of a send which is ok */
9381                                 ok = 1;
9382                         } else {
9383                                 /* A bit harder was it the end of our segment */
9384                                 int segs, len;
9385
9386                                 len = (rsm->r_end - rsm->r_start);
9387                                 segs = len / segsiz;
9388                                 segs *= segsiz;
9389                                 if ((segs + (rsm->r_end - start)) == len) {
9390                                         /*
9391                                          * So this last bit was the
9392                                          * end of our send if we cut it
9393                                          * up into segsiz pieces so its ok.
9394                                          */
9395                                         ok = 1;
9396                                 }
9397                         }
9398                 }
9399                 if (ok == 0) {
9400                         /*
9401                          * This guy is doing something suspicious
9402                          * lets start detection.
9403                          */
9404                         if (rack->rc_suspicious == 0) {
9405                                 tcp_trace_point(rack->rc_tp, TCP_TP_SAD_SUSPECT);
9406                                 counter_u64_add(rack_sack_attacks_suspect, 1);
9407                                 rack->rc_suspicious = 1;
9408                                 rack_log_sad(rack, 4);
9409                                 if (tcp_bblogging_on(rack->rc_tp)) {
9410                                         union tcp_log_stackspecific log;
9411                                         struct timeval tv;
9412
9413                                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
9414                                         log.u_bbr.flex1 = end;
9415                                         log.u_bbr.flex2 = start;
9416                                         log.u_bbr.flex3 = rsm->r_end;
9417                                         log.u_bbr.flex4 = rsm->r_start;
9418                                         log.u_bbr.flex5 = segsiz;
9419                                         log.u_bbr.flex6 = rsm->r_fas;
9420                                         log.u_bbr.flex7 = rsm->r_bas;
9421                                         log.u_bbr.flex8 = 5;
9422                                         log.u_bbr.pkts_out = rsm->r_flags;
9423                                         log.u_bbr.bbr_state = rack->rc_suspicious;
9424                                         log.u_bbr.bbr_substate = rsm->r_just_ret;
9425                                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
9426                                         log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9427                                         TCP_LOG_EVENTP(rack->rc_tp, NULL,
9428                                                        &rack->rc_inp->inp_socket->so_rcv,
9429                                                        &rack->rc_inp->inp_socket->so_snd,
9430                                                        TCP_SAD_DETECTION, 0,
9431                                                        0, &log, false, &tv);
9432                                 }
9433                         }
9434                         /* You loose some ack count every time you sack
9435                          * a small bit that is not butting to the end of
9436                          * what we have sent. This is because we never
9437                          * send small bits unless its the end of the sb.
9438                          * Anyone sending a sack that is not at the end
9439                          * is thus very very suspicious.
9440                          */
9441                         loss = (segsiz/2) / (end - start);
9442                         if (loss < rack->r_ctl.ack_count)
9443                                 rack->r_ctl.ack_count -= loss;
9444                         else
9445                                 rack->r_ctl.ack_count = 0;
9446                 }
9447         }
9448         first_time_through = 0;
9449 #endif
9450         /* Ok we have an ACK for some piece of this rsm */
9451         if (rsm->r_start != start) {
9452                 if ((rsm->r_flags & RACK_ACKED) == 0) {
9453                         /*
9454                          * Before any splitting or hookery is
9455                          * done is it a TLP of interest i.e. rxt?
9456                          */
9457                         if ((rsm->r_flags & RACK_TLP) &&
9458                             (rsm->r_rtr_cnt > 1)) {
9459                                 /*
9460                                  * We are splitting a rxt TLP, check
9461                                  * if we need to save off the start/end
9462                                  */
9463                                 if (rack->rc_last_tlp_acked_set &&
9464                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9465                                         /*
9466                                          * We already turned this on since we are inside
9467                                          * the previous one was a partially sack now we
9468                                          * are getting another one (maybe all of it).
9469                                          *
9470                                          */
9471                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9472                                         /*
9473                                          * Lets make sure we have all of it though.
9474                                          */
9475                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9476                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9477                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9478                                                                      rack->r_ctl.last_tlp_acked_end);
9479                                         }
9480                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9481                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9482                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9483                                                                      rack->r_ctl.last_tlp_acked_end);
9484                                         }
9485                                 } else {
9486                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9487                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9488                                         rack->rc_last_tlp_past_cumack = 0;
9489                                         rack->rc_last_tlp_acked_set = 1;
9490                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9491                                 }
9492                         }
9493                         /**
9494                          * Need to split this in two pieces the before and after,
9495                          * the before remains in the map, the after must be
9496                          * added. In other words we have:
9497                          * rsm        |--------------|
9498                          * sackblk        |------->
9499                          * rsm will become
9500                          *     rsm    |---|
9501                          * and nrsm will be  the sacked piece
9502                          *     nrsm       |----------|
9503                          *
9504                          * But before we start down that path lets
9505                          * see if the sack spans over on top of
9506                          * the next guy and it is already sacked.
9507                          *
9508                          */
9509                         /*
9510                          * Hookery can only be used if the two entries
9511                          * are in the same bucket and neither one of
9512                          * them staddle the bucket line.
9513                          */
9514                         next = tqhash_next(rack->r_ctl.tqh, rsm);
9515                         if (next &&
9516                             (rsm->bindex == next->bindex) &&
9517                             ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9518                             ((next->r_flags & RACK_STRADDLE) == 0) &&
9519                             (rsm->r_flags & RACK_IN_GP_WIN) &&
9520                             (next->r_flags & RACK_IN_GP_WIN))
9521                                 can_use_hookery = 1;
9522                         else if (next &&
9523                                  (rsm->bindex == next->bindex) &&
9524                                  ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9525                                  ((next->r_flags & RACK_STRADDLE) == 0) &&
9526                                  ((rsm->r_flags & RACK_IN_GP_WIN) == 0) &&
9527                                  ((next->r_flags & RACK_IN_GP_WIN) == 0))
9528                                 can_use_hookery = 1;
9529                         else
9530                                 can_use_hookery = 0;
9531                         if (next && can_use_hookery &&
9532                             (next->r_flags & RACK_ACKED) &&
9533                             SEQ_GEQ(end, next->r_start)) {
9534                                 /**
9535                                  * So the next one is already acked, and
9536                                  * we can thus by hookery use our stack_map
9537                                  * to reflect the piece being sacked and
9538                                  * then adjust the two tree entries moving
9539                                  * the start and ends around. So we start like:
9540                                  *  rsm     |------------|             (not-acked)
9541                                  *  next                 |-----------| (acked)
9542                                  *  sackblk        |-------->
9543                                  *  We want to end like so:
9544                                  *  rsm     |------|                   (not-acked)
9545                                  *  next           |-----------------| (acked)
9546                                  *  nrsm           |-----|
9547                                  * Where nrsm is a temporary stack piece we
9548                                  * use to update all the gizmos.
9549                                  */
9550                                 /* Copy up our fudge block */
9551                                 noextra++;
9552                                 nrsm = &stack_map;
9553                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9554                                 /* Now adjust our tree blocks */
9555                                 rsm->r_end = start;
9556                                 next->r_start = start;
9557                                 rsm->r_flags |= RACK_SHUFFLED;
9558                                 next->r_flags |= RACK_SHUFFLED;
9559                                 /* Now we must adjust back where next->m is */
9560                                 rack_setup_offset_for_rsm(rack, rsm, next);
9561                                 /*
9562                                  * Which timestamp do we keep? It is rather
9563                                  * important in GP measurements to have the
9564                                  * accurate end of the send window.
9565                                  *
9566                                  * We keep the largest value, which is the newest
9567                                  * send. We do this in case a segment that is
9568                                  * joined together and not part of a GP estimate
9569                                  * later gets expanded into the GP estimate.
9570                                  *
9571                                  * We prohibit the merging of unlike kinds i.e.
9572                                  * all pieces that are in the GP estimate can be
9573                                  * merged and all pieces that are not in a GP estimate
9574                                  * can be merged, but not disimilar pieces. Combine
9575                                  * this with taking the highest here and we should
9576                                  * be ok unless of course the client reneges. Then
9577                                  * all bets are off.
9578                                  */
9579                                 if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
9580                                     nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
9581                                         next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
9582                                 /*
9583                                  * And we must keep the newest ack arrival time.
9584                                  */
9585                                 if (next->r_ack_arrival <
9586                                     rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9587                                         next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9588
9589
9590                                 /* We don't need to adjust rsm, it did not change */
9591                                 /* Clear out the dup ack count of the remainder */
9592                                 rsm->r_dupack = 0;
9593                                 rsm->r_just_ret = 0;
9594                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9595                                 /* Now lets make sure our fudge block is right */
9596                                 nrsm->r_start = start;
9597                                 /* Now lets update all the stats and such */
9598                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9599                                 if (rack->app_limited_needs_set)
9600                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9601                                 changed += (nrsm->r_end - nrsm->r_start);
9602                                 /* You get a count for acking a whole segment or more */
9603                                 if ((nrsm->r_end - nrsm->r_start) >= segsiz)
9604                                         rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
9605                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9606                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
9607                                         rack->r_ctl.rc_reorder_ts = cts;
9608                                         if (rack->r_ctl.rc_reorder_ts == 0)
9609                                                 rack->r_ctl.rc_reorder_ts = 1;
9610                                 }
9611                                 /*
9612                                  * Now we want to go up from rsm (the
9613                                  * one left un-acked) to the next one
9614                                  * in the tmap. We do this so when
9615                                  * we walk backwards we include marking
9616                                  * sack-passed on rsm (The one passed in
9617                                  * is skipped since it is generally called
9618                                  * on something sacked before removing it
9619                                  * from the tmap).
9620                                  */
9621                                 if (rsm->r_in_tmap) {
9622                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
9623                                         /*
9624                                          * Now that we have the next
9625                                          * one walk backwards from there.
9626                                          */
9627                                         if (nrsm && nrsm->r_in_tmap)
9628                                                 rack_log_sack_passed(tp, rack, nrsm);
9629                                 }
9630                                 /* Now are we done? */
9631                                 if (SEQ_LT(end, next->r_end) ||
9632                                     (end == next->r_end)) {
9633                                         /* Done with block */
9634                                         goto out;
9635                                 }
9636                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
9637                                 counter_u64_add(rack_sack_used_next_merge, 1);
9638                                 /* Postion for the next block */
9639                                 start = next->r_end;
9640                                 rsm = tqhash_next(rack->r_ctl.tqh, next);
9641                                 if (rsm == NULL)
9642                                         goto out;
9643                         } else {
9644                                 /**
9645                                  * We can't use any hookery here, so we
9646                                  * need to split the map. We enter like
9647                                  * so:
9648                                  *  rsm      |--------|
9649                                  *  sackblk       |----->
9650                                  * We will add the new block nrsm and
9651                                  * that will be the new portion, and then
9652                                  * fall through after reseting rsm. So we
9653                                  * split and look like this:
9654                                  *  rsm      |----|
9655                                  *  sackblk       |----->
9656                                  *  nrsm          |---|
9657                                  * We then fall through reseting
9658                                  * rsm to nrsm, so the next block
9659                                  * picks it up.
9660                                  */
9661                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9662                                 if (nrsm == NULL) {
9663                                         /*
9664                                          * failed XXXrrs what can we do but loose the sack
9665                                          * info?
9666                                          */
9667                                         goto out;
9668                                 }
9669                                 counter_u64_add(rack_sack_splits, 1);
9670                                 rack_clone_rsm(rack, nrsm, rsm, start);
9671                                 moved++;
9672                                 rsm->r_just_ret = 0;
9673 #ifndef INVARIANTS
9674                                 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9675 #else
9676                                 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9677                                         panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
9678                                               nrsm, insret, rack, rsm);
9679                                 }
9680 #endif
9681                                 if (rsm->r_in_tmap) {
9682                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9683                                         nrsm->r_in_tmap = 1;
9684                                 }
9685                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
9686                                 rsm->r_flags &= (~RACK_HAS_FIN);
9687                                 /* Position us to point to the new nrsm that starts the sack blk */
9688                                 rsm = nrsm;
9689                         }
9690                 } else {
9691                         /* Already sacked this piece */
9692                         counter_u64_add(rack_sack_skipped_acked, 1);
9693                         moved++;
9694                         if (end == rsm->r_end) {
9695                                 /* Done with block */
9696                                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9697                                 goto out;
9698                         } else if (SEQ_LT(end, rsm->r_end)) {
9699                                 /* A partial sack to a already sacked block */
9700                                 moved++;
9701                                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9702                                 goto out;
9703                         } else {
9704                                 /*
9705                                  * The end goes beyond this guy
9706                                  * reposition the start to the
9707                                  * next block.
9708                                  */
9709                                 start = rsm->r_end;
9710                                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9711                                 if (rsm == NULL)
9712                                         goto out;
9713                         }
9714                 }
9715         }
9716         if (SEQ_GEQ(end, rsm->r_end)) {
9717                 /**
9718                  * The end of this block is either beyond this guy or right
9719                  * at this guy. I.e.:
9720                  *  rsm ---                 |-----|
9721                  *  end                     |-----|
9722                  *  <or>
9723                  *  end                     |---------|
9724                  */
9725                 if ((rsm->r_flags & RACK_ACKED) == 0) {
9726                         /*
9727                          * Is it a TLP of interest?
9728                          */
9729                         if ((rsm->r_flags & RACK_TLP) &&
9730                             (rsm->r_rtr_cnt > 1)) {
9731                                 /*
9732                                  * We are splitting a rxt TLP, check
9733                                  * if we need to save off the start/end
9734                                  */
9735                                 if (rack->rc_last_tlp_acked_set &&
9736                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9737                                         /*
9738                                          * We already turned this on since we are inside
9739                                          * the previous one was a partially sack now we
9740                                          * are getting another one (maybe all of it).
9741                                          */
9742                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9743                                         /*
9744                                          * Lets make sure we have all of it though.
9745                                          */
9746                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9747                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9748                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9749                                                                      rack->r_ctl.last_tlp_acked_end);
9750                                         }
9751                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9752                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9753                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9754                                                                      rack->r_ctl.last_tlp_acked_end);
9755                                         }
9756                                 } else {
9757                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9758                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9759                                         rack->rc_last_tlp_past_cumack = 0;
9760                                         rack->rc_last_tlp_acked_set = 1;
9761                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9762                                 }
9763                         }
9764                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9765                         changed += (rsm->r_end - rsm->r_start);
9766                         /* You get a count for acking a whole segment or more */
9767                         if ((rsm->r_end - rsm->r_start) >= segsiz)
9768                                 rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
9769                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9770                         if (rsm->r_in_tmap) /* should be true */
9771                                 rack_log_sack_passed(tp, rack, rsm);
9772                         /* Is Reordering occuring? */
9773                         if (rsm->r_flags & RACK_SACK_PASSED) {
9774                                 rsm->r_flags &= ~RACK_SACK_PASSED;
9775                                 rack->r_ctl.rc_reorder_ts = cts;
9776                                 if (rack->r_ctl.rc_reorder_ts == 0)
9777                                         rack->r_ctl.rc_reorder_ts = 1;
9778                         }
9779                         if (rack->app_limited_needs_set)
9780                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9781                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9782                         rsm->r_flags |= RACK_ACKED;
9783                         if (rsm->r_in_tmap) {
9784                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9785                                 rsm->r_in_tmap = 0;
9786                         }
9787                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
9788                 } else {
9789                         counter_u64_add(rack_sack_skipped_acked, 1);
9790                         moved++;
9791                 }
9792                 if (end == rsm->r_end) {
9793                         /* This block only - done, setup for next */
9794                         goto out;
9795                 }
9796                 /*
9797                  * There is more not coverend by this rsm move on
9798                  * to the next block in the RB tree.
9799                  */
9800                 nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
9801                 start = rsm->r_end;
9802                 rsm = nrsm;
9803                 if (rsm == NULL)
9804                         goto out;
9805                 goto do_rest_ofb;
9806         }
9807         /**
9808          * The end of this sack block is smaller than
9809          * our rsm i.e.:
9810          *  rsm ---                 |-----|
9811          *  end                     |--|
9812          */
9813         if ((rsm->r_flags & RACK_ACKED) == 0) {
9814                 /*
9815                  * Is it a TLP of interest?
9816                  */
9817                 if ((rsm->r_flags & RACK_TLP) &&
9818                     (rsm->r_rtr_cnt > 1)) {
9819                         /*
9820                          * We are splitting a rxt TLP, check
9821                          * if we need to save off the start/end
9822                          */
9823                         if (rack->rc_last_tlp_acked_set &&
9824                             (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9825                                 /*
9826                                  * We already turned this on since we are inside
9827                                  * the previous one was a partially sack now we
9828                                  * are getting another one (maybe all of it).
9829                                  */
9830                                 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9831                                 /*
9832                                  * Lets make sure we have all of it though.
9833                                  */
9834                                 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9835                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9836                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9837                                                              rack->r_ctl.last_tlp_acked_end);
9838                                 }
9839                                 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9840                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9841                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9842                                                              rack->r_ctl.last_tlp_acked_end);
9843                                 }
9844                         } else {
9845                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9846                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9847                                 rack->rc_last_tlp_past_cumack = 0;
9848                                 rack->rc_last_tlp_acked_set = 1;
9849                                 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9850                         }
9851                 }
9852                 /*
9853                  * Hookery can only be used if the two entries
9854                  * are in the same bucket and neither one of
9855                  * them staddle the bucket line.
9856                  */
9857                 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
9858                 if (prev &&
9859                     (rsm->bindex == prev->bindex) &&
9860                     ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9861                     ((prev->r_flags & RACK_STRADDLE) == 0) &&
9862                     (rsm->r_flags & RACK_IN_GP_WIN) &&
9863                     (prev->r_flags & RACK_IN_GP_WIN))
9864                         can_use_hookery = 1;
9865                 else if (prev &&
9866                          (rsm->bindex == prev->bindex) &&
9867                          ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9868                          ((prev->r_flags & RACK_STRADDLE) == 0) &&
9869                          ((rsm->r_flags & RACK_IN_GP_WIN) == 0) &&
9870                          ((prev->r_flags & RACK_IN_GP_WIN) == 0))
9871                         can_use_hookery = 1;
9872                 else
9873                         can_use_hookery = 0;
9874
9875                 if (prev && can_use_hookery &&
9876                     (prev->r_flags & RACK_ACKED)) {
9877                         /**
9878                          * Goal, we want the right remainder of rsm to shrink
9879                          * in place and span from (rsm->r_start = end) to rsm->r_end.
9880                          * We want to expand prev to go all the way
9881                          * to prev->r_end <- end.
9882                          * so in the tree we have before:
9883                          *   prev     |--------|         (acked)
9884                          *   rsm               |-------| (non-acked)
9885                          *   sackblk           |-|
9886                          * We churn it so we end up with
9887                          *   prev     |----------|       (acked)
9888                          *   rsm                 |-----| (non-acked)
9889                          *   nrsm              |-| (temporary)
9890                          *
9891                          * Note if either prev/rsm is a TLP we don't
9892                          * do this.
9893                          */
9894                         noextra++;
9895                         nrsm = &stack_map;
9896                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9897                         prev->r_end = end;
9898                         rsm->r_start = end;
9899                         rsm->r_flags |= RACK_SHUFFLED;
9900                         prev->r_flags |= RACK_SHUFFLED;
9901                         /* Now adjust nrsm (stack copy) to be
9902                          * the one that is the small
9903                          * piece that was "sacked".
9904                          */
9905                         nrsm->r_end = end;
9906                         rsm->r_dupack = 0;
9907                         /*
9908                          * Which timestamp do we keep? It is rather
9909                          * important in GP measurements to have the
9910                          * accurate end of the send window.
9911                          *
9912                          * We keep the largest value, which is the newest
9913                          * send. We do this in case a segment that is
9914                          * joined together and not part of a GP estimate
9915                          * later gets expanded into the GP estimate.
9916                          *
9917                          * We prohibit the merging of unlike kinds i.e.
9918                          * all pieces that are in the GP estimate can be
9919                          * merged and all pieces that are not in a GP estimate
9920                          * can be merged, but not disimilar pieces. Combine
9921                          * this with taking the highest here and we should
9922                          * be ok unless of course the client reneges. Then
9923                          * all bets are off.
9924                          */
9925                         if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
9926                            nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
9927                                 prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9928                         }
9929                         /*
9930                          * And we must keep the newest ack arrival time.
9931                          */
9932
9933                         if(prev->r_ack_arrival <
9934                            rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9935                                 prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9936
9937                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9938                         /*
9939                          * Now that the rsm has had its start moved forward
9940                          * lets go ahead and get its new place in the world.
9941                          */
9942                         rack_setup_offset_for_rsm(rack, prev, rsm);
9943                         /*
9944                          * Now nrsm is our new little piece
9945                          * that is acked (which was merged
9946                          * to prev). Update the rtt and changed
9947                          * based on that. Also check for reordering.
9948                          */
9949                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9950                         if (rack->app_limited_needs_set)
9951                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9952                         changed += (nrsm->r_end - nrsm->r_start);
9953                         /* You get a count for acking a whole segment or more */
9954                         if ((nrsm->r_end - nrsm->r_start) >= segsiz)
9955                                 rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
9956
9957                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9958                         if (nrsm->r_flags & RACK_SACK_PASSED) {
9959                                 rack->r_ctl.rc_reorder_ts = cts;
9960                                 if (rack->r_ctl.rc_reorder_ts == 0)
9961                                         rack->r_ctl.rc_reorder_ts = 1;
9962                         }
9963                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
9964                         rsm = prev;
9965                         counter_u64_add(rack_sack_used_prev_merge, 1);
9966                 } else {
9967                         /**
9968                          * This is the case where our previous
9969                          * block is not acked either, so we must
9970                          * split the block in two.
9971                          */
9972                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9973                         if (nrsm == NULL) {
9974                                 /* failed rrs what can we do but loose the sack info? */
9975                                 goto out;
9976                         }
9977                         if ((rsm->r_flags & RACK_TLP) &&
9978                             (rsm->r_rtr_cnt > 1)) {
9979                                 /*
9980                                  * We are splitting a rxt TLP, check
9981                                  * if we need to save off the start/end
9982                                  */
9983                                 if (rack->rc_last_tlp_acked_set &&
9984                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9985                                         /*
9986                                          * We already turned this on since this block is inside
9987                                          * the previous one was a partially sack now we
9988                                          * are getting another one (maybe all of it).
9989                                          */
9990                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9991                                         /*
9992                                          * Lets make sure we have all of it though.
9993                                          */
9994                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9995                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9996                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9997                                                                      rack->r_ctl.last_tlp_acked_end);
9998                                         }
9999                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10000                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10001                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10002                                                                      rack->r_ctl.last_tlp_acked_end);
10003                                         }
10004                                 } else {
10005                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10006                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10007                                         rack->rc_last_tlp_acked_set = 1;
10008                                         rack->rc_last_tlp_past_cumack = 0;
10009                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10010                                 }
10011                         }
10012                         /**
10013                          * In this case nrsm becomes
10014                          * nrsm->r_start = end;
10015                          * nrsm->r_end = rsm->r_end;
10016                          * which is un-acked.
10017                          * <and>
10018                          * rsm->r_end = nrsm->r_start;
10019                          * i.e. the remaining un-acked
10020                          * piece is left on the left
10021                          * hand side.
10022                          *
10023                          * So we start like this
10024                          * rsm      |----------| (not acked)
10025                          * sackblk  |---|
10026                          * build it so we have
10027                          * rsm      |---|         (acked)
10028                          * nrsm         |------|  (not acked)
10029                          */
10030                         counter_u64_add(rack_sack_splits, 1);
10031                         rack_clone_rsm(rack, nrsm, rsm, end);
10032                         moved++;
10033                         rsm->r_flags &= (~RACK_HAS_FIN);
10034                         rsm->r_just_ret = 0;
10035 #ifndef INVARIANTS
10036                         (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
10037 #else
10038                         if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
10039                                 panic("Insert in rb tree of %p fails ret:% rack:%p rsm:%p",
10040                                       nrsm, insret, rack, rsm);
10041                         }
10042 #endif
10043                         if (rsm->r_in_tmap) {
10044                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10045                                 nrsm->r_in_tmap = 1;
10046                         }
10047                         nrsm->r_dupack = 0;
10048                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
10049                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
10050                         changed += (rsm->r_end - rsm->r_start);
10051                         /* You get a count for acking a whole segment or more */
10052                         if ((rsm->r_end - rsm->r_start) >= segsiz)
10053                                 rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
10054
10055                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
10056                         if (rsm->r_in_tmap) /* should be true */
10057                                 rack_log_sack_passed(tp, rack, rsm);
10058                         /* Is Reordering occuring? */
10059                         if (rsm->r_flags & RACK_SACK_PASSED) {
10060                                 rsm->r_flags &= ~RACK_SACK_PASSED;
10061                                 rack->r_ctl.rc_reorder_ts = cts;
10062                                 if (rack->r_ctl.rc_reorder_ts == 0)
10063                                         rack->r_ctl.rc_reorder_ts = 1;
10064                         }
10065                         if (rack->app_limited_needs_set)
10066                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10067                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10068                         rsm->r_flags |= RACK_ACKED;
10069                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
10070                         if (rsm->r_in_tmap) {
10071                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10072                                 rsm->r_in_tmap = 0;
10073                         }
10074                 }
10075         } else if (start != end){
10076                 /*
10077                  * The block was already acked.
10078                  */
10079                 counter_u64_add(rack_sack_skipped_acked, 1);
10080                 moved++;
10081         }
10082 out:
10083         if (rsm &&
10084             ((rsm->r_flags & RACK_TLP) == 0) &&
10085             (rsm->r_flags & RACK_ACKED)) {
10086                 /*
10087                  * Now can we merge where we worked
10088                  * with either the previous or
10089                  * next block?
10090                  */
10091                 next = tqhash_next(rack->r_ctl.tqh, rsm);
10092                 while (next) {
10093                         if (next->r_flags & RACK_TLP)
10094                                 break;
10095                         /* Only allow merges between ones in or out of GP window */
10096                         if ((next->r_flags & RACK_IN_GP_WIN) &&
10097                             ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10098                                 break;
10099                         }
10100                         if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10101                             ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10102                                 break;
10103                         }
10104                         if (rsm->bindex != next->bindex)
10105                                 break;
10106                         if (rsm->r_flags & RACK_STRADDLE)
10107                                 break;
10108                         if (next->r_flags & RACK_STRADDLE)
10109                                 break;
10110                         if (next->r_flags & RACK_ACKED) {
10111                                 /* yep this and next can be merged */
10112                                 rsm = rack_merge_rsm(rack, rsm, next);
10113                                 noextra++;
10114                                 next = tqhash_next(rack->r_ctl.tqh, rsm);
10115                         } else
10116                                 break;
10117                 }
10118                 /* Now what about the previous? */
10119                 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10120                 while (prev) {
10121                         if (prev->r_flags & RACK_TLP)
10122                                 break;
10123                         /* Only allow merges between ones in or out of GP window */
10124                         if ((prev->r_flags & RACK_IN_GP_WIN) &&
10125                             ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10126                                 break;
10127                         }
10128                         if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10129                             ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10130                                 break;
10131                         }
10132                         if (rsm->bindex != prev->bindex)
10133                                 break;
10134                         if (rsm->r_flags & RACK_STRADDLE)
10135                                 break;
10136                         if (prev->r_flags & RACK_STRADDLE)
10137                                 break;
10138                         if (prev->r_flags & RACK_ACKED) {
10139                                 /* yep the previous and this can be merged */
10140                                 rsm = rack_merge_rsm(rack, prev, rsm);
10141                                 noextra++;
10142                                 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10143                         } else
10144                                 break;
10145                 }
10146         }
10147         if (used_ref == 0) {
10148                 counter_u64_add(rack_sack_proc_all, 1);
10149         } else {
10150                 counter_u64_add(rack_sack_proc_short, 1);
10151         }
10152         /* Save off the next one for quick reference. */
10153         nrsm = tqhash_find(rack->r_ctl.tqh, end);
10154         *prsm = rack->r_ctl.rc_sacklast = nrsm;
10155         /* Pass back the moved. */
10156         *moved_two = moved;
10157         *no_extra = noextra;
10158         return (changed);
10159 }
10160
10161 static void inline
10162 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10163 {
10164         struct rack_sendmap *tmap;
10165
10166         tmap = NULL;
10167         while (rsm && (rsm->r_flags & RACK_ACKED)) {
10168                 /* Its no longer sacked, mark it so */
10169                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10170 #ifdef INVARIANTS
10171                 if (rsm->r_in_tmap) {
10172                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
10173                               rack, rsm, rsm->r_flags);
10174                 }
10175 #endif
10176                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10177                 /* Rebuild it into our tmap */
10178                 if (tmap == NULL) {
10179                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10180                         tmap = rsm;
10181                 } else {
10182                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10183                         tmap = rsm;
10184                 }
10185                 tmap->r_in_tmap = 1;
10186                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10187         }
10188         /*
10189          * Now lets possibly clear the sack filter so we start
10190          * recognizing sacks that cover this area.
10191          */
10192         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10193
10194 }
10195
10196 static void
10197 rack_do_decay(struct tcp_rack *rack)
10198 {
10199         struct timeval res;
10200
10201 #define timersub(tvp, uvp, vvp)                                         \
10202         do {                                                            \
10203                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
10204                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
10205                 if ((vvp)->tv_usec < 0) {                               \
10206                         (vvp)->tv_sec--;                                \
10207                         (vvp)->tv_usec += 1000000;                      \
10208                 }                                                       \
10209         } while (0)
10210
10211         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
10212 #undef timersub
10213
10214         rack->r_ctl.input_pkt++;
10215         if ((rack->rc_in_persist) ||
10216             (res.tv_sec >= 1) ||
10217             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
10218                 /*
10219                  * Check for decay of non-SAD,
10220                  * we want all SAD detection metrics to
10221                  * decay 1/4 per second (or more) passed.
10222                  * Current default is 800 so it decays
10223                  * 80% every second.
10224                  */
10225 #ifdef TCP_SAD_DETECTION
10226                 uint32_t pkt_delta;
10227
10228                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
10229 #endif
10230                 /* Update our saved tracking values */
10231                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
10232                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
10233                 /* Now do we escape without decay? */
10234 #ifdef TCP_SAD_DETECTION
10235                 if (rack->rc_in_persist ||
10236                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
10237                     (pkt_delta < tcp_sad_low_pps)){
10238                         /*
10239                          * We don't decay idle connections
10240                          * or ones that have a low input pps.
10241                          */
10242                         return;
10243                 }
10244                 /* Decay the counters */
10245                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
10246                                                         tcp_sad_decay_val);
10247                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
10248                                                          tcp_sad_decay_val);
10249                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
10250                                                                tcp_sad_decay_val);
10251                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
10252                                                                 tcp_sad_decay_val);
10253 #endif
10254         }
10255 }
10256
10257 static void inline
10258 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10259 {
10260         /*
10261          * We look at advancing the end send time for our GP
10262          * measurement tracking only as the cumulative acknowledgment
10263          * moves forward. You might wonder about this, why not
10264          * at every transmission or retransmission within the
10265          * GP window update the rc_gp_cumack_ts? Well its rather
10266          * nuanced but basically the GP window *may* expand (as
10267          * it does below) or worse and harder to track it may shrink.
10268          *
10269          * This last makes it impossible to track at the time of
10270          * the send, since you may set forward your rc_gp_cumack_ts
10271          * when you send, because that send *is* in your currently
10272          * "guessed" window, but then it shrinks. Now which was
10273          * the send time of the last bytes in the window, by the
10274          * time you ask that question that part of the sendmap
10275          * is freed. So you don't know and you will have too
10276          * long of send window. Instead by updating the time
10277          * marker only when the cumack advances this assures us
10278          * that we will have only the sends in the window of our
10279          * GP measurement.
10280          *
10281          * Another complication from this is the
10282          * merging of sendmap entries. During SACK processing this
10283          * can happen to conserve the sendmap size. That breaks
10284          * everything down in tracking the send window of the GP
10285          * estimate. So to prevent that and keep it working with
10286          * a tiny bit more limited merging, we only allow like
10287          * types to be merged. I.e. if two sends are in the GP window
10288          * then its ok to merge them together. If two sends are not
10289          * in the GP window its ok to merge them together too. Though
10290          * one send in and one send out cannot be merged. We combine
10291          * this with never allowing the shrinking of the GP window when
10292          * we are in recovery so that we can properly calculate the
10293          * sending times.
10294          *
10295          * This all of course seems complicated, because it is.. :)
10296          *
10297          * The cum-ack is being advanced upon the sendmap.
10298          * If we are not doing a GP estimate don't
10299          * proceed.
10300          */
10301         uint64_t ts;
10302
10303         if ((tp->t_flags & TF_GPUTINPROG) == 0)
10304                 return;
10305         /*
10306          * If this sendmap entry is going
10307          * beyond the measurement window we had picked,
10308          * expand the measurement window by that much.
10309          */
10310         if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10311                 tp->gput_ack = rsm->r_end;
10312         }
10313         /*
10314          * If we have not setup a ack, then we
10315          * have no idea if the newly acked pieces
10316          * will be "in our seq measurement range". If
10317          * it is when we clear the app_limited_needs_set
10318          * flag the timestamp will be updated.
10319          */
10320         if (rack->app_limited_needs_set)
10321                 return;
10322         /*
10323          * Finally, we grab out the latest timestamp
10324          * that this packet was sent and then see
10325          * if:
10326          *  a) The packet touches are newly defined GP range.
10327          *  b) The time is greater than (newer) than the
10328          *     one we currently have. If so we update
10329          *     our sending end time window.
10330          *
10331          * Note we *do not* do this at send time. The reason
10332          * is that if you do you *may* pick up a newer timestamp
10333          * for a range you are not going to measure. We project
10334          * out how far and then sometimes modify that to be
10335          * smaller. If that occurs then you will have a send
10336          * that does not belong to the range included.
10337          */
10338         if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10339             rack->r_ctl.rc_gp_cumack_ts)
10340                 return;
10341         if (rack_in_gp_window(tp, rsm)) {
10342                 rack->r_ctl.rc_gp_cumack_ts = ts;
10343                 rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10344                                __LINE__, from, rsm);
10345         }
10346 }
10347
10348 static void
10349 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)
10350 {
10351         struct rack_sendmap *rsm;
10352         /*
10353          * The ACK point is advancing to th_ack, we must drop off
10354          * the packets in the rack log and calculate any eligble
10355          * RTT's.
10356          */
10357
10358         rack->r_wanted_output = 1;
10359         if (SEQ_GT(th_ack, tp->snd_una))
10360             rack->r_ctl.last_cumack_advance = acktime;
10361
10362         /* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
10363         if ((rack->rc_last_tlp_acked_set == 1)&&
10364             (rack->rc_last_tlp_past_cumack == 1) &&
10365             (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10366                 /*
10367                  * We have reached the point where our last rack
10368                  * tlp retransmit sequence is ahead of the cum-ack.
10369                  * This can only happen when the cum-ack moves all
10370                  * the way around (its been a full 2^^31+1 bytes
10371                  * or more since we sent a retransmitted TLP). Lets
10372                  * turn off the valid flag since its not really valid.
10373                  *
10374                  * Note since sack's also turn on this event we have
10375                  * a complication, we have to wait to age it out until
10376                  * the cum-ack is by the TLP before checking which is
10377                  * what the next else clause does.
10378                  */
10379                 rack_log_dsack_event(rack, 9, __LINE__,
10380                                      rack->r_ctl.last_tlp_acked_start,
10381                                      rack->r_ctl.last_tlp_acked_end);
10382                 rack->rc_last_tlp_acked_set = 0;
10383                 rack->rc_last_tlp_past_cumack = 0;
10384         } else if ((rack->rc_last_tlp_acked_set == 1) &&
10385                    (rack->rc_last_tlp_past_cumack == 0) &&
10386                    (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10387                 /*
10388                  * It is safe to start aging TLP's out.
10389                  */
10390                 rack->rc_last_tlp_past_cumack = 1;
10391         }
10392         /* We do the same for the tlp send seq as well */
10393         if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10394             (rack->rc_last_sent_tlp_past_cumack == 1) &&
10395             (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
10396                 rack_log_dsack_event(rack, 9, __LINE__,
10397                                      rack->r_ctl.last_sent_tlp_seq,
10398                                      (rack->r_ctl.last_sent_tlp_seq +
10399                                       rack->r_ctl.last_sent_tlp_len));
10400                 rack->rc_last_sent_tlp_seq_valid = 0;
10401                 rack->rc_last_sent_tlp_past_cumack = 0;
10402         } else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10403                    (rack->rc_last_sent_tlp_past_cumack == 0) &&
10404                    (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10405                 /*
10406                  * It is safe to start aging TLP's send.
10407                  */
10408                 rack->rc_last_sent_tlp_past_cumack = 1;
10409         }
10410 more:
10411         rsm = tqhash_min(rack->r_ctl.tqh);
10412         if (rsm == NULL) {
10413                 if ((th_ack - 1) == tp->iss) {
10414                         /*
10415                          * For the SYN incoming case we will not
10416                          * have called tcp_output for the sending of
10417                          * the SYN, so there will be no map. All
10418                          * other cases should probably be a panic.
10419                          */
10420                         return;
10421                 }
10422                 if (tp->t_flags & TF_SENTFIN) {
10423                         /* if we sent a FIN we often will not have map */
10424                         return;
10425                 }
10426 #ifdef INVARIANTS
10427                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
10428                       tp,
10429                       tp->t_state, th_ack, rack,
10430                       tp->snd_una, tp->snd_max, tp->snd_nxt);
10431 #endif
10432                 return;
10433         }
10434         if (SEQ_LT(th_ack, rsm->r_start)) {
10435                 /* Huh map is missing this */
10436 #ifdef INVARIANTS
10437                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
10438                        rsm->r_start,
10439                        th_ack, tp->t_state, rack->r_state);
10440 #endif
10441                 return;
10442         }
10443         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
10444
10445         /* Now was it a retransmitted TLP? */
10446         if ((rsm->r_flags & RACK_TLP) &&
10447             (rsm->r_rtr_cnt > 1)) {
10448                 /*
10449                  * Yes, this rsm was a TLP and retransmitted, remember that
10450                  * since if a DSACK comes back on this we don't want
10451                  * to think of it as a reordered segment. This may
10452                  * get updated again with possibly even other TLPs
10453                  * in flight, but thats ok. Only when we don't send
10454                  * a retransmitted TLP for 1/2 the sequences space
10455                  * will it get turned off (above).
10456                  */
10457                 if (rack->rc_last_tlp_acked_set &&
10458                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10459                         /*
10460                          * We already turned this on since the end matches,
10461                          * the previous one was a partially ack now we
10462                          * are getting another one (maybe all of it).
10463                          */
10464                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10465                         /*
10466                          * Lets make sure we have all of it though.
10467                          */
10468                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10469                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10470                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10471                                                      rack->r_ctl.last_tlp_acked_end);
10472                         }
10473                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10474                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10475                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10476                                                      rack->r_ctl.last_tlp_acked_end);
10477                         }
10478                 } else {
10479                         rack->rc_last_tlp_past_cumack = 1;
10480                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10481                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10482                         rack->rc_last_tlp_acked_set = 1;
10483                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10484                 }
10485         }
10486         /* Now do we consume the whole thing? */
10487         if (SEQ_GEQ(th_ack, rsm->r_end)) {
10488                 /* Its all consumed. */
10489                 uint32_t left;
10490                 uint8_t newly_acked;
10491
10492                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
10493                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
10494                 rsm->r_rtr_bytes = 0;
10495                 /*
10496                  * Record the time of highest cumack sent if its in our measurement
10497                  * window and possibly bump out the end.
10498                  */
10499                 rack_rsm_sender_update(rack, tp, rsm, 4);
10500                 tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
10501                 if (rsm->r_in_tmap) {
10502                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10503                         rsm->r_in_tmap = 0;
10504                 }
10505                 newly_acked = 1;
10506                 if (rsm->r_flags & RACK_ACKED) {
10507                         /*
10508                          * It was acked on the scoreboard -- remove
10509                          * it from total
10510                          */
10511                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10512                         newly_acked = 0;
10513                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
10514                         /*
10515                          * There are segments ACKED on the
10516                          * scoreboard further up. We are seeing
10517                          * reordering.
10518                          */
10519                         rsm->r_flags &= ~RACK_SACK_PASSED;
10520                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10521                         rsm->r_flags |= RACK_ACKED;
10522                         rack->r_ctl.rc_reorder_ts = cts;
10523                         if (rack->r_ctl.rc_reorder_ts == 0)
10524                                 rack->r_ctl.rc_reorder_ts = 1;
10525                         if (rack->r_ent_rec_ns) {
10526                                 /*
10527                                  * We have sent no more, and we saw an sack
10528                                  * then ack arrive.
10529                                  */
10530                                 rack->r_might_revert = 1;
10531                         }
10532                 }
10533                 if ((rsm->r_flags & RACK_TO_REXT) &&
10534                     (tp->t_flags & TF_RCVD_TSTMP) &&
10535                     (to->to_flags & TOF_TS) &&
10536                     (to->to_tsecr != 0) &&
10537                     (tp->t_flags & TF_PREVVALID)) {
10538                         /*
10539                          * We can use the timestamp to see
10540                          * if this retransmission was from the
10541                          * first transmit. If so we made a mistake.
10542                          */
10543                         tp->t_flags &= ~TF_PREVVALID;
10544                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
10545                                 /* The first transmit is what this ack is for */
10546                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
10547                         }
10548                 }
10549                 left = th_ack - rsm->r_end;
10550                 if (rack->app_limited_needs_set && newly_acked)
10551                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
10552                 /* Free back to zone */
10553                 rack_free(rack, rsm);
10554                 if (left) {
10555                         goto more;
10556                 }
10557                 /* Check for reneging */
10558                 rsm = tqhash_min(rack->r_ctl.tqh);
10559                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
10560                         /*
10561                          * The peer has moved snd_una up to
10562                          * the edge of this send, i.e. one
10563                          * that it had previously acked. The only
10564                          * way that can be true if the peer threw
10565                          * away data (space issues) that it had
10566                          * previously sacked (else it would have
10567                          * given us snd_una up to (rsm->r_end).
10568                          * We need to undo the acked markings here.
10569                          *
10570                          * Note we have to look to make sure th_ack is
10571                          * our rsm->r_start in case we get an old ack
10572                          * where th_ack is behind snd_una.
10573                          */
10574                         rack_peer_reneges(rack, rsm, th_ack);
10575                 }
10576                 return;
10577         }
10578         if (rsm->r_flags & RACK_ACKED) {
10579                 /*
10580                  * It was acked on the scoreboard -- remove it from
10581                  * total for the part being cum-acked.
10582                  */
10583                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
10584         }
10585         /*
10586          * Clear the dup ack count for
10587          * the piece that remains.
10588          */
10589         rsm->r_dupack = 0;
10590         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10591         if (rsm->r_rtr_bytes) {
10592                 /*
10593                  * It was retransmitted adjust the
10594                  * sack holes for what was acked.
10595                  */
10596                 int ack_am;
10597
10598                 ack_am = (th_ack - rsm->r_start);
10599                 if (ack_am >= rsm->r_rtr_bytes) {
10600                         rack->r_ctl.rc_holes_rxt -= ack_am;
10601                         rsm->r_rtr_bytes -= ack_am;
10602                 }
10603         }
10604         /*
10605          * Update where the piece starts and record
10606          * the time of send of highest cumack sent if
10607          * its in our GP range.
10608          */
10609         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
10610         /* Now we need to move our offset forward too */
10611         if (rsm->m &&
10612             ((rsm->orig_m_len != rsm->m->m_len) ||
10613              (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
10614                 /* Fix up the orig_m_len and possibly the mbuf offset */
10615                 rack_adjust_orig_mlen(rsm);
10616         }
10617         rsm->soff += (th_ack - rsm->r_start);
10618         rack_rsm_sender_update(rack, tp, rsm, 5);
10619         /* The trim will move th_ack into r_start for us */
10620         tqhash_trim(rack->r_ctl.tqh, th_ack);
10621         /* Now do we need to move the mbuf fwd too? */
10622         if (rsm->m) {
10623                 while (rsm->soff >= rsm->m->m_len) {
10624                         rsm->soff -= rsm->m->m_len;
10625                         rsm->m = rsm->m->m_next;
10626                         KASSERT((rsm->m != NULL),
10627                                 (" nrsm:%p hit at soff:%u null m",
10628                                  rsm, rsm->soff));
10629                 }
10630                 rsm->orig_m_len = rsm->m->m_len;
10631                 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
10632         }
10633         if (rack->app_limited_needs_set &&
10634             SEQ_GEQ(th_ack, tp->gput_seq))
10635                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
10636 }
10637
10638 static void
10639 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
10640 {
10641         struct rack_sendmap *rsm;
10642         int sack_pass_fnd = 0;
10643
10644         if (rack->r_might_revert) {
10645                 /*
10646                  * Ok we have reordering, have not sent anything, we
10647                  * might want to revert the congestion state if nothing
10648                  * further has SACK_PASSED on it. Lets check.
10649                  *
10650                  * We also get here when we have DSACKs come in for
10651                  * all the data that we FR'd. Note that a rxt or tlp
10652                  * timer clears this from happening.
10653                  */
10654
10655                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
10656                         if (rsm->r_flags & RACK_SACK_PASSED) {
10657                                 sack_pass_fnd = 1;
10658                                 break;
10659                         }
10660                 }
10661                 if (sack_pass_fnd == 0) {
10662                         /*
10663                          * We went into recovery
10664                          * incorrectly due to reordering!
10665                          */
10666                         int orig_cwnd;
10667
10668                         rack->r_ent_rec_ns = 0;
10669                         orig_cwnd = tp->snd_cwnd;
10670                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
10671                         tp->snd_recover = tp->snd_una;
10672                         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
10673                         EXIT_RECOVERY(tp->t_flags);
10674                 }
10675                 rack->r_might_revert = 0;
10676         }
10677 }
10678
10679 #ifdef TCP_SAD_DETECTION
10680
10681 static void
10682 rack_merge_out_sacks(struct tcp_rack *rack)
10683 {
10684         struct rack_sendmap *cur, *next, *rsm, *trsm = NULL;
10685
10686         cur = tqhash_min(rack->r_ctl.tqh);
10687         while(cur) {
10688                 next = tqhash_next(rack->r_ctl.tqh, cur);
10689                 /*
10690                  * The idea is to go through all and merge back
10691                  * together the pieces sent together,
10692                  */
10693                 if ((next != NULL) &&
10694                     (cur->r_tim_lastsent[0] == next->r_tim_lastsent[0])) {
10695                         rack_merge_rsm(rack, cur, next);
10696                 } else {
10697                         cur = next;
10698                 }
10699         }
10700         /*
10701          * now treat it like a rxt event, everything is outstanding
10702          * and sent nothing acvked and dupacks are all zero. If this
10703          * is not an attacker it will have to dupack its way through
10704          * it all.
10705          */
10706         TAILQ_INIT(&rack->r_ctl.rc_tmap);
10707         TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
10708                 rsm->r_dupack = 0;
10709                 /* We must re-add it back to the tlist */
10710                 if (trsm == NULL) {
10711                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10712                 } else {
10713                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
10714                 }
10715                 rsm->r_in_tmap = 1;
10716                 trsm = rsm;
10717                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
10718         }
10719         sack_filter_clear(&rack->r_ctl.rack_sf, rack->rc_tp->snd_una);
10720 }
10721
10722 static void
10723 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
10724 {
10725         int do_detection = 0;
10726
10727         if (rack->sack_attack_disable || rack->rc_suspicious) {
10728                 /*
10729                  * If we have been disabled we must detect
10730                  * to possibly reverse it. Or if the guy has
10731                  * sent in suspicious sacks we want to do detection too.
10732                  */
10733                 do_detection = 1;
10734
10735         } else if  ((rack->do_detection || tcp_force_detection) &&
10736                     (tcp_sack_to_ack_thresh > 0) &&
10737                     (tcp_sack_to_move_thresh > 0) &&
10738                     (rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum)) {
10739                 /*
10740                  * We only detect here if:
10741                  * 1) System wide forcing is on <or> do_detection is on
10742                  *   <and>
10743                  * 2) We have thresholds for move and ack (set one to 0 and we are off)
10744                  *   <and>
10745                  * 3) We have maps allocated larger than our min (500).
10746                  */
10747                 do_detection = 1;
10748         }
10749         if (do_detection > 0) {
10750                 /*
10751                  * We have thresholds set to find
10752                  * possible attackers and disable sack.
10753                  * Check them.
10754                  */
10755                 uint64_t ackratio, moveratio, movetotal;
10756
10757                 /* Log detecting */
10758                 rack_log_sad(rack, 1);
10759                 /* Do we establish a ack ratio */
10760                 if ((rack->r_ctl.sack_count > tcp_map_minimum)  ||
10761                     (rack->rc_suspicious == 1) ||
10762                     (rack->sack_attack_disable > 0)) {
10763                         ackratio = (uint64_t)(rack->r_ctl.sack_count);
10764                         ackratio *= (uint64_t)(1000);
10765                         if (rack->r_ctl.ack_count)
10766                                 ackratio /= (uint64_t)(rack->r_ctl.ack_count);
10767                         else {
10768                                 /* We can hit this due to ack totals degregation (via small sacks) */
10769                                 ackratio = 1000;
10770                         }
10771                 } else {
10772                         /*
10773                          * No ack ratio needed if we have not
10774                          * seen more sacks then the number of map entries.
10775                          * The exception to that is if we have disabled sack then
10776                          * we need to find a ratio.
10777                          */
10778                         ackratio = 0;
10779                 }
10780
10781                 if ((rack->sack_attack_disable == 0) &&
10782                     (ackratio > rack_highest_sack_thresh_seen))
10783                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
10784                 /* Do we establish a move ratio? */
10785                 if ((rack->r_ctl.sack_moved_extra > tcp_map_minimum) ||
10786                     (rack->rc_suspicious == 1) ||
10787                     (rack->sack_attack_disable > 0)) {
10788                         /*
10789                          * We need to have more sack moves than maps
10790                          * allocated to have a move ratio considered.
10791                          */
10792                         movetotal = rack->r_ctl.sack_moved_extra;
10793                         movetotal += rack->r_ctl.sack_noextra_move;
10794                         moveratio = rack->r_ctl.sack_moved_extra;
10795                         moveratio *= (uint64_t)1000;
10796                         if (movetotal)
10797                                 moveratio /= movetotal;
10798                         else {
10799                                 /* No moves, thats pretty good */
10800                                 moveratio = 0;
10801                         }
10802                 } else {
10803                         /*
10804                          * Not enough moves have occured to consider
10805                          * if we are out of whack in that ratio.
10806                          * The exception to that is if we have disabled sack then
10807                          * we need to find a ratio.
10808                          */
10809                         moveratio = 0;
10810                 }
10811                 if ((rack->sack_attack_disable == 0) &&
10812                     (moveratio > rack_highest_move_thresh_seen))
10813                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
10814                 /* Now the tests */
10815                 if (rack->sack_attack_disable == 0) {
10816                         /* Not disabled, do we need to disable? */
10817                         if ((ackratio > tcp_sack_to_ack_thresh) &&
10818                             (moveratio > tcp_sack_to_move_thresh)) {
10819                                 /* Disable sack processing */
10820                                 tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
10821                                 rack->sack_attack_disable = 1;
10822                                 /* set it so we have the built in delay */
10823                                 rack->r_ctl.ack_during_sd = 1;
10824                                 if (rack_merge_out_sacks_on_attack)
10825                                         rack_merge_out_sacks(rack);
10826                                 counter_u64_add(rack_sack_attacks_detected, 1);
10827                                 tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
10828                                 /* Clamp the cwnd at flight size */
10829                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
10830                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10831                                 rack_log_sad(rack, 2);
10832                         }
10833                 } else {
10834                         /* We are sack-disabled check for false positives */
10835                         if ((ackratio <= tcp_restoral_thresh) ||
10836                             ((rack_merge_out_sacks_on_attack == 0) &&
10837                              (rack->rc_suspicious == 0) &&
10838                              (rack->r_ctl.rc_num_maps_alloced <= (tcp_map_minimum/2)))) {
10839                                 rack->sack_attack_disable = 0;
10840                                 rack_log_sad(rack, 3);
10841                                 /* Restart counting */
10842                                 rack->r_ctl.sack_count = 0;
10843                                 rack->r_ctl.sack_moved_extra = 0;
10844                                 rack->r_ctl.sack_noextra_move = 1;
10845                                 rack->rc_suspicious = 0;
10846                                 rack->r_ctl.ack_count = max(1,
10847                                                             (bytes_this_ack / segsiz));
10848
10849                                 counter_u64_add(rack_sack_attacks_reversed, 1);
10850                                 /* Restore the cwnd */
10851                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
10852                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
10853                         }
10854                 }
10855         }
10856 }
10857 #endif
10858
10859 static int
10860 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
10861 {
10862
10863         uint32_t am, l_end;
10864         int was_tlp = 0;
10865
10866         if (SEQ_GT(end, start))
10867                 am = end - start;
10868         else
10869                 am = 0;
10870         if ((rack->rc_last_tlp_acked_set ) &&
10871             (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
10872             (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
10873                 /*
10874                  * The DSACK is because of a TLP which we don't
10875                  * do anything with the reordering window over since
10876                  * it was not reordering that caused the DSACK but
10877                  * our previous retransmit TLP.
10878                  */
10879                 rack_log_dsack_event(rack, 7, __LINE__, start, end);
10880                 was_tlp = 1;
10881                 goto skip_dsack_round;
10882         }
10883         if (rack->rc_last_sent_tlp_seq_valid) {
10884                 l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
10885                 if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
10886                     (SEQ_LEQ(end, l_end))) {
10887                         /*
10888                          * This dsack is from the last sent TLP, ignore it
10889                          * for reordering purposes.
10890                          */
10891                         rack_log_dsack_event(rack, 7, __LINE__, start, end);
10892                         was_tlp = 1;
10893                         goto skip_dsack_round;
10894                 }
10895         }
10896         if (rack->rc_dsack_round_seen == 0) {
10897                 rack->rc_dsack_round_seen = 1;
10898                 rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
10899                 rack->r_ctl.num_dsack++;
10900                 rack->r_ctl.dsack_persist = 16; /* 16 is from the standard */
10901                 rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
10902         }
10903 skip_dsack_round:
10904         /*
10905          * We keep track of how many DSACK blocks we get
10906          * after a recovery incident.
10907          */
10908         rack->r_ctl.dsack_byte_cnt += am;
10909         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
10910             rack->r_ctl.retran_during_recovery &&
10911             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
10912                 /*
10913                  * False recovery most likely culprit is reordering. If
10914                  * nothing else is missing we need to revert.
10915                  */
10916                 rack->r_might_revert = 1;
10917                 rack_handle_might_revert(rack->rc_tp, rack);
10918                 rack->r_might_revert = 0;
10919                 rack->r_ctl.retran_during_recovery = 0;
10920                 rack->r_ctl.dsack_byte_cnt = 0;
10921         }
10922         return (was_tlp);
10923 }
10924
10925 static uint32_t
10926 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
10927 {
10928         return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
10929 }
10930
10931 static int32_t
10932 rack_compute_pipe(struct tcpcb *tp)
10933 {
10934         return ((int32_t)do_rack_compute_pipe(tp,
10935                                               (struct tcp_rack *)tp->t_fb_ptr,
10936                                               tp->snd_una));
10937 }
10938
10939 static void
10940 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
10941 {
10942         /* Deal with changed and PRR here (in recovery only) */
10943         uint32_t pipe, snd_una;
10944
10945         rack->r_ctl.rc_prr_delivered += changed;
10946
10947         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
10948                 /*
10949                  * It is all outstanding, we are application limited
10950                  * and thus we don't need more room to send anything.
10951                  * Note we use tp->snd_una here and not th_ack because
10952                  * the data as yet not been cut from the sb.
10953                  */
10954                 rack->r_ctl.rc_prr_sndcnt = 0;
10955                 return;
10956         }
10957         /* Compute prr_sndcnt */
10958         if (SEQ_GT(tp->snd_una, th_ack)) {
10959                 snd_una = tp->snd_una;
10960         } else {
10961                 snd_una = th_ack;
10962         }
10963         pipe = do_rack_compute_pipe(tp, rack, snd_una);
10964         if (pipe > tp->snd_ssthresh) {
10965                 long sndcnt;
10966
10967                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
10968                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
10969                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
10970                 else {
10971                         rack->r_ctl.rc_prr_sndcnt = 0;
10972                         rack_log_to_prr(rack, 9, 0, __LINE__);
10973                         sndcnt = 0;
10974                 }
10975                 sndcnt++;
10976                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
10977                         sndcnt -= rack->r_ctl.rc_prr_out;
10978                 else
10979                         sndcnt = 0;
10980                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
10981                 rack_log_to_prr(rack, 10, 0, __LINE__);
10982         } else {
10983                 uint32_t limit;
10984
10985                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
10986                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
10987                 else
10988                         limit = 0;
10989                 if (changed > limit)
10990                         limit = changed;
10991                 limit += ctf_fixed_maxseg(tp);
10992                 if (tp->snd_ssthresh > pipe) {
10993                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
10994                         rack_log_to_prr(rack, 11, 0, __LINE__);
10995                 } else {
10996                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
10997                         rack_log_to_prr(rack, 12, 0, __LINE__);
10998                 }
10999         }
11000 }
11001
11002 static void
11003 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
11004              int *dsack_seen, int *sacks_seen)
11005 {
11006         uint32_t changed;
11007         struct tcp_rack *rack;
11008         struct rack_sendmap *rsm;
11009         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
11010         register uint32_t th_ack;
11011         int32_t i, j, k, num_sack_blks = 0;
11012         uint32_t cts, acked, ack_point;
11013         int loop_start = 0, moved_two = 0, no_extra = 0;
11014         uint32_t tsused;
11015         uint32_t segsiz, o_cnt;
11016
11017
11018         INP_WLOCK_ASSERT(tptoinpcb(tp));
11019         if (tcp_get_flags(th) & TH_RST) {
11020                 /* We don't log resets */
11021                 return;
11022         }
11023         rack = (struct tcp_rack *)tp->t_fb_ptr;
11024         cts = tcp_get_usecs(NULL);
11025         rsm = tqhash_min(rack->r_ctl.tqh);
11026         changed = 0;
11027         th_ack = th->th_ack;
11028         if (rack->sack_attack_disable == 0)
11029                 rack_do_decay(rack);
11030         segsiz = ctf_fixed_maxseg(rack->rc_tp);
11031         if (BYTES_THIS_ACK(tp, th) >=  segsiz) {
11032                 /*
11033                  * You only get credit for
11034                  * MSS and greater (and you get extra
11035                  * credit for larger cum-ack moves).
11036                  */
11037                 int ac;
11038
11039                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
11040                 rack->r_ctl.ack_count += ac;
11041                 counter_u64_add(rack_ack_total, ac);
11042         }
11043         if (rack->r_ctl.ack_count > 0xfff00000) {
11044                 /*
11045                  * reduce the number to keep us under
11046                  * a uint32_t.
11047                  */
11048                 rack->r_ctl.ack_count /= 2;
11049                 rack->r_ctl.sack_count /= 2;
11050         }
11051         if (SEQ_GT(th_ack, tp->snd_una)) {
11052                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
11053                 tp->t_acktime = ticks;
11054         }
11055         if (rsm && SEQ_GT(th_ack, rsm->r_start))
11056                 changed = th_ack - rsm->r_start;
11057         if (changed) {
11058                 rack_process_to_cumack(tp, rack, th_ack, cts, to,
11059                                        tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
11060         }
11061         if ((to->to_flags & TOF_SACK) == 0) {
11062                 /* We are done nothing left and no sack. */
11063                 rack_handle_might_revert(tp, rack);
11064                 /*
11065                  * For cases where we struck a dup-ack
11066                  * with no SACK, add to the changes so
11067                  * PRR will work right.
11068                  */
11069                 if (dup_ack_struck && (changed == 0)) {
11070                         changed += ctf_fixed_maxseg(rack->rc_tp);
11071                 }
11072                 goto out;
11073         }
11074         /* Sack block processing */
11075         if (SEQ_GT(th_ack, tp->snd_una))
11076                 ack_point = th_ack;
11077         else
11078                 ack_point = tp->snd_una;
11079         for (i = 0; i < to->to_nsacks; i++) {
11080                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
11081                       &sack, sizeof(sack));
11082                 sack.start = ntohl(sack.start);
11083                 sack.end = ntohl(sack.end);
11084                 if (SEQ_GT(sack.end, sack.start) &&
11085                     SEQ_GT(sack.start, ack_point) &&
11086                     SEQ_LT(sack.start, tp->snd_max) &&
11087                     SEQ_GT(sack.end, ack_point) &&
11088                     SEQ_LEQ(sack.end, tp->snd_max)) {
11089                         sack_blocks[num_sack_blks] = sack;
11090                         num_sack_blks++;
11091                 } else if (SEQ_LEQ(sack.start, th_ack) &&
11092                            SEQ_LEQ(sack.end, th_ack)) {
11093                         int was_tlp;
11094
11095                         if (dsack_seen != NULL)
11096                                 *dsack_seen = 1;
11097                         was_tlp = rack_note_dsack(rack, sack.start, sack.end);
11098                         /*
11099                          * Its a D-SACK block.
11100                          */
11101                         tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
11102                 }
11103         }
11104         if (rack->rc_dsack_round_seen) {
11105                 /* Is the dsack roound over? */
11106                 if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
11107                         /* Yes it is */
11108                         rack->rc_dsack_round_seen = 0;
11109                         rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
11110                 }
11111         }
11112         /*
11113          * Sort the SACK blocks so we can update the rack scoreboard with
11114          * just one pass.
11115          */
11116         o_cnt = num_sack_blks;
11117         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
11118                                          num_sack_blks, th->th_ack);
11119         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
11120         if (sacks_seen != NULL)
11121                 *sacks_seen = num_sack_blks;
11122         if (num_sack_blks == 0) {
11123                 /* Nothing to sack, but we need to update counts */
11124                 if ((o_cnt == 1) &&
11125                     (*dsack_seen != 1))
11126                         rack->r_ctl.sack_count++;
11127                 else if (o_cnt > 1)
11128                         rack->r_ctl.sack_count++;
11129                 goto out_with_totals;
11130         }
11131         if (rack->sack_attack_disable) {
11132                 /*
11133                  * An attacker disablement is in place, for
11134                  * every sack block that is not at least a full MSS
11135                  * count up sack_count.
11136                  */
11137                 for (i = 0; i < num_sack_blks; i++) {
11138                         if ((sack_blocks[i].end - sack_blocks[i].start) < segsiz) {
11139                                 rack->r_ctl.sack_count++;
11140                         }
11141                         if (rack->r_ctl.sack_count > 0xfff00000) {
11142                                 /*
11143                                  * reduce the number to keep us under
11144                                  * a uint32_t.
11145                                  */
11146                                 rack->r_ctl.ack_count /= 2;
11147                                 rack->r_ctl.sack_count /= 2;
11148                         }
11149                 }
11150                 goto out;
11151         }
11152         /* Its a sack of some sort */
11153         rack->r_ctl.sack_count += num_sack_blks;
11154         if (rack->r_ctl.sack_count > 0xfff00000) {
11155                 /*
11156                  * reduce the number to keep us under
11157                  * a uint32_t.
11158                  */
11159                 rack->r_ctl.ack_count /= 2;
11160                 rack->r_ctl.sack_count /= 2;
11161         }
11162         if (num_sack_blks < 2) {
11163                 /* Only one, we don't need to sort */
11164                 goto do_sack_work;
11165         }
11166         /* Sort the sacks */
11167         for (i = 0; i < num_sack_blks; i++) {
11168                 for (j = i + 1; j < num_sack_blks; j++) {
11169                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
11170                                 sack = sack_blocks[i];
11171                                 sack_blocks[i] = sack_blocks[j];
11172                                 sack_blocks[j] = sack;
11173                         }
11174                 }
11175         }
11176         /*
11177          * Now are any of the sack block ends the same (yes some
11178          * implementations send these)?
11179          */
11180 again:
11181         if (num_sack_blks == 0)
11182                 goto out_with_totals;
11183         if (num_sack_blks > 1) {
11184                 for (i = 0; i < num_sack_blks; i++) {
11185                         for (j = i + 1; j < num_sack_blks; j++) {
11186                                 if (sack_blocks[i].end == sack_blocks[j].end) {
11187                                         /*
11188                                          * Ok these two have the same end we
11189                                          * want the smallest end and then
11190                                          * throw away the larger and start
11191                                          * again.
11192                                          */
11193                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
11194                                                 /*
11195                                                  * The second block covers
11196                                                  * more area use that
11197                                                  */
11198                                                 sack_blocks[i].start = sack_blocks[j].start;
11199                                         }
11200                                         /*
11201                                          * Now collapse out the dup-sack and
11202                                          * lower the count
11203                                          */
11204                                         for (k = (j + 1); k < num_sack_blks; k++) {
11205                                                 sack_blocks[j].start = sack_blocks[k].start;
11206                                                 sack_blocks[j].end = sack_blocks[k].end;
11207                                                 j++;
11208                                         }
11209                                         num_sack_blks--;
11210                                         goto again;
11211                                 }
11212                         }
11213                 }
11214         }
11215 do_sack_work:
11216         /*
11217          * First lets look to see if
11218          * we have retransmitted and
11219          * can use the transmit next?
11220          */
11221         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11222         if (rsm &&
11223             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
11224             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
11225                 /*
11226                  * We probably did the FR and the next
11227                  * SACK in continues as we would expect.
11228                  */
11229                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &no_extra, &moved_two, segsiz);
11230                 if (acked) {
11231                         rack->r_wanted_output = 1;
11232                         changed += acked;
11233                 }
11234                 if (num_sack_blks == 1) {
11235                         /*
11236                          * This is what we would expect from
11237                          * a normal implementation to happen
11238                          * after we have retransmitted the FR,
11239                          * i.e the sack-filter pushes down
11240                          * to 1 block and the next to be retransmitted
11241                          * is the sequence in the sack block (has more
11242                          * are acked). Count this as ACK'd data to boost
11243                          * up the chances of recovering any false positives.
11244                          */
11245                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
11246                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
11247                         counter_u64_add(rack_express_sack, 1);
11248                         if (rack->r_ctl.ack_count > 0xfff00000) {
11249                                 /*
11250                                  * reduce the number to keep us under
11251                                  * a uint32_t.
11252                                  */
11253                                 rack->r_ctl.ack_count /= 2;
11254                                 rack->r_ctl.sack_count /= 2;
11255                         }
11256                         if (moved_two) {
11257                                 /*
11258                                  * If we did not get a SACK for at least a MSS and
11259                                  * had to move at all, or if we moved more than our
11260                                  * threshold, it counts against the "extra" move.
11261                                  */
11262                                 rack->r_ctl.sack_moved_extra += moved_two;
11263                                 rack->r_ctl.sack_noextra_move += no_extra;
11264                                 counter_u64_add(rack_move_some, 1);
11265                         } else {
11266                                 /*
11267                                  * else we did not have to move
11268                                  * any more than we would expect.
11269                                  */
11270                                 rack->r_ctl.sack_noextra_move += no_extra;
11271                                 rack->r_ctl.sack_noextra_move++;
11272                                 counter_u64_add(rack_move_none, 1);
11273                         }
11274                         if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
11275                             (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
11276                                 rack->r_ctl.sack_moved_extra /= 2;
11277                                 rack->r_ctl.sack_noextra_move /= 2;
11278                         }
11279                         goto out_with_totals;
11280                 } else {
11281                         /*
11282                          * Start the loop through the
11283                          * rest of blocks, past the first block.
11284                          */
11285                         loop_start = 1;
11286                 }
11287         }
11288         counter_u64_add(rack_sack_total, 1);
11289         rsm = rack->r_ctl.rc_sacklast;
11290         for (i = loop_start; i < num_sack_blks; i++) {
11291                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &no_extra, &moved_two, segsiz);
11292                 if (acked) {
11293                         rack->r_wanted_output = 1;
11294                         changed += acked;
11295                 }
11296                 if (moved_two) {
11297                         /*
11298                          * If we did not get a SACK for at least a MSS and
11299                          * had to move at all, or if we moved more than our
11300                          * threshold, it counts against the "extra" move.
11301                          */
11302                         rack->r_ctl.sack_moved_extra += moved_two;
11303                         rack->r_ctl.sack_noextra_move += no_extra;
11304                         counter_u64_add(rack_move_some, 1);
11305                 } else {
11306                         /*
11307                          * else we did not have to move
11308                          * any more than we would expect.
11309                          */
11310                         rack->r_ctl.sack_noextra_move += no_extra;
11311                         rack->r_ctl.sack_noextra_move++;
11312                         counter_u64_add(rack_move_none, 1);
11313                 }
11314                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
11315                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
11316                         rack->r_ctl.sack_moved_extra /= 2;
11317                         rack->r_ctl.sack_noextra_move /= 2;
11318                 }
11319                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
11320                         /*
11321                          * If the SACK was not a full MSS then
11322                          * we add to sack_count the number of
11323                          * MSS's (or possibly more than
11324                          * a MSS if its a TSO send) we had to skip by.
11325                          */
11326                         rack->r_ctl.sack_count += moved_two;
11327                         if (rack->r_ctl.sack_count > 0xfff00000) {
11328                                 rack->r_ctl.ack_count /= 2;
11329                                 rack->r_ctl.sack_count /= 2;
11330                         }
11331                         counter_u64_add(rack_sack_total, moved_two);
11332                 }
11333                 /*
11334                  * Now we need to setup for the next
11335                  * round. First we make sure we won't
11336                  * exceed the size of our uint32_t on
11337                  * the various counts, and then clear out
11338                  * moved_two.
11339                  */
11340                 moved_two = 0;
11341                 no_extra = 0;
11342         }
11343 out_with_totals:
11344         if (num_sack_blks > 1) {
11345                 /*
11346                  * You get an extra stroke if
11347                  * you have more than one sack-blk, this
11348                  * could be where we are skipping forward
11349                  * and the sack-filter is still working, or
11350                  * it could be an attacker constantly
11351                  * moving us.
11352                  */
11353                 rack->r_ctl.sack_moved_extra++;
11354                 counter_u64_add(rack_move_some, 1);
11355         }
11356 out:
11357 #ifdef TCP_SAD_DETECTION
11358         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
11359 #endif
11360         if (changed) {
11361                 /* Something changed cancel the rack timer */
11362                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11363         }
11364         tsused = tcp_get_usecs(NULL);
11365         rsm = tcp_rack_output(tp, rack, tsused);
11366         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11367             rsm &&
11368             ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11369                 /* Enter recovery */
11370                 entered_recovery = 1;
11371                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
11372                 /*
11373                  * When we enter recovery we need to assure we send
11374                  * one packet.
11375                  */
11376                 if (rack->rack_no_prr == 0) {
11377                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11378                         rack_log_to_prr(rack, 8, 0, __LINE__);
11379                 }
11380                 rack->r_timer_override = 1;
11381                 rack->r_early = 0;
11382                 rack->r_ctl.rc_agg_early = 0;
11383         } else if (IN_FASTRECOVERY(tp->t_flags) &&
11384                    rsm &&
11385                    (rack->r_rr_config == 3)) {
11386                 /*
11387                  * Assure we can output and we get no
11388                  * remembered pace time except the retransmit.
11389                  */
11390                 rack->r_timer_override = 1;
11391                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11392                 rack->r_ctl.rc_resend = rsm;
11393         }
11394         if (IN_FASTRECOVERY(tp->t_flags) &&
11395             (rack->rack_no_prr == 0) &&
11396             (entered_recovery == 0)) {
11397                 rack_update_prr(tp, rack, changed, th_ack);
11398                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11399                      ((tcp_in_hpts(rack->rc_tp) == 0) &&
11400                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11401                         /*
11402                          * If you are pacing output you don't want
11403                          * to override.
11404                          */
11405                         rack->r_early = 0;
11406                         rack->r_ctl.rc_agg_early = 0;
11407                         rack->r_timer_override = 1;
11408                 }
11409         }
11410 }
11411
11412 static void
11413 rack_strike_dupack(struct tcp_rack *rack)
11414 {
11415         struct rack_sendmap *rsm;
11416
11417         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11418         while (rsm) {
11419                 /*
11420                  * We need to skip anything already set
11421                  * to be retransmitted.
11422                  */
11423                 if ((rsm->r_dupack >= DUP_ACK_THRESHOLD)  ||
11424                     (rsm->r_flags & RACK_MUST_RXT)) {
11425                         rsm = TAILQ_NEXT(rsm, r_tnext);
11426                         continue;
11427                 }
11428                 break;
11429         }
11430         if (rsm && (rsm->r_dupack < 0xff)) {
11431                 rsm->r_dupack++;
11432                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11433                         struct timeval tv;
11434                         uint32_t cts;
11435                         /*
11436                          * Here we see if we need to retransmit. For
11437                          * a SACK type connection if enough time has passed
11438                          * we will get a return of the rsm. For a non-sack
11439                          * connection we will get the rsm returned if the
11440                          * dupack value is 3 or more.
11441                          */
11442                         cts = tcp_get_usecs(&tv);
11443                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11444                         if (rack->r_ctl.rc_resend != NULL) {
11445                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11446                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11447                                                          rack->rc_tp->snd_una, __LINE__);
11448                                 }
11449                                 rack->r_wanted_output = 1;
11450                                 rack->r_timer_override = 1;
11451                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11452                         }
11453                 } else {
11454                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11455                 }
11456         }
11457 }
11458
11459 static void
11460 rack_check_bottom_drag(struct tcpcb *tp,
11461                        struct tcp_rack *rack,
11462                        struct socket *so)
11463 {
11464         uint32_t segsiz, minseg;
11465
11466         segsiz = ctf_fixed_maxseg(tp);
11467         minseg = segsiz;
11468         if (tp->snd_max == tp->snd_una) {
11469                 /*
11470                  * We are doing dynamic pacing and we are way
11471                  * under. Basically everything got acked while
11472                  * we were still waiting on the pacer to expire.
11473                  *
11474                  * This means we need to boost the b/w in
11475                  * addition to any earlier boosting of
11476                  * the multiplier.
11477                  */
11478                 uint64_t lt_bw;
11479
11480                 lt_bw = rack_get_lt_bw(rack);
11481                 rack->rc_dragged_bottom = 1;
11482                 rack_validate_multipliers_at_or_above100(rack);
11483                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11484                     (lt_bw > 0)) {
11485                         /*
11486                          * Lets use the long-term b/w we have
11487                          * been getting as a base.
11488                          */
11489                         if (rack->rc_gp_filled == 0) {
11490                                 if (lt_bw > ONE_POINT_TWO_MEG) {
11491                                         /*
11492                                          * If we have no measurement
11493                                          * don't let us set in more than
11494                                          * 1.2Mbps. If we are still too
11495                                          * low after pacing with this we
11496                                          * will hopefully have a max b/w
11497                                          * available to sanity check things.
11498                                          */
11499                                         lt_bw = ONE_POINT_TWO_MEG;
11500                                 }
11501                                 rack->r_ctl.rc_rtt_diff = 0;
11502                                 rack->r_ctl.gp_bw = lt_bw;
11503                                 rack->rc_gp_filled = 1;
11504                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11505                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
11506                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11507                         } else if (lt_bw > rack->r_ctl.gp_bw) {
11508                                 rack->r_ctl.rc_rtt_diff = 0;
11509                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11510                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
11511                                 rack->r_ctl.gp_bw = lt_bw;
11512                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11513                         } else
11514                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
11515                         if ((rack->gp_ready == 0) &&
11516                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11517                                 /* We have enough measurements now */
11518                                 rack->gp_ready = 1;
11519                                 if ((rack->rc_always_pace && (rack->use_fixed_rate == 0)) ||
11520                                     rack->rack_hibeta)
11521                                         rack_set_cc_pacing(rack);
11522                                 if (rack->defer_options)
11523                                         rack_apply_deferred_options(rack);
11524                         }
11525                 } else {
11526                         /*
11527                          * zero rtt possibly?, settle for just an old increase.
11528                          */
11529                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
11530                 }
11531         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11532                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11533                                                minseg)) &&
11534                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11535                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11536                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11537                     (segsiz * rack_req_segs))) {
11538                 /*
11539                  * We are doing dynamic GP pacing and
11540                  * we have everything except 1MSS or less
11541                  * bytes left out. We are still pacing away.
11542                  * And there is data that could be sent, This
11543                  * means we are inserting delayed ack time in
11544                  * our measurements because we are pacing too slow.
11545                  */
11546                 rack_validate_multipliers_at_or_above100(rack);
11547                 rack->rc_dragged_bottom = 1;
11548                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
11549         }
11550 }
11551
11552 #ifdef TCP_REQUEST_TRK
11553 static void
11554 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11555                 struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
11556 {
11557         int do_log;
11558
11559         do_log = tcp_bblogging_on(rack->rc_tp);
11560         if (do_log == 0) {
11561                 if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11562                         return;
11563                 /* We only allow the three below with point logging on */
11564                 if ((mod != HYBRID_LOG_RULES_APP) &&
11565                     (mod != HYBRID_LOG_RULES_SET) &&
11566                     (mod != HYBRID_LOG_REQ_COMP))
11567                         return;
11568
11569         }
11570         if (do_log) {
11571                 union tcp_log_stackspecific log;
11572                 struct timeval tv;
11573
11574                 /* Convert our ms to a microsecond */
11575                 memset(&log, 0, sizeof(log));
11576                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11577                 log.u_bbr.flex1 = seq;
11578                 log.u_bbr.cwnd_gain = line;
11579                 if (cur != NULL) {
11580                         uint64_t off;
11581
11582                         log.u_bbr.flex2 = cur->start_seq;
11583                         log.u_bbr.flex3 = cur->end_seq;
11584                         log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11585                         log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11586                         log.u_bbr.flex6 = cur->flags;
11587                         log.u_bbr.pkts_out = cur->hybrid_flags;
11588                         log.u_bbr.rttProp = cur->timestamp;
11589                         log.u_bbr.cur_del_rate = cur->cspr;
11590                         log.u_bbr.bw_inuse = cur->start;
11591                         log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11592                         log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11593                         log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11594                         log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11595                         log.u_bbr.bbr_state = 1;
11596                         off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
11597                         log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
11598                 } else {
11599                         log.u_bbr.flex2 = err;
11600                 }
11601                 /*
11602                  * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11603                  */
11604                 log.u_bbr.flex7 = rack->rc_catch_up;
11605                 log.u_bbr.flex7 <<= 1;
11606                 log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11607                 log.u_bbr.flex7 <<= 1;
11608                 log.u_bbr.flex7 |= rack->dgp_on;
11609                 log.u_bbr.flex8 = mod;
11610                 log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11611                 log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11612                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11613                 log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
11614                 log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
11615                 log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
11616                 tcp_log_event(rack->rc_tp, NULL,
11617                     &rack->rc_inp->inp_socket->so_rcv,
11618                     &rack->rc_inp->inp_socket->so_snd,
11619                     TCP_HYBRID_PACING_LOG, 0,
11620                     0, &log, false, NULL, __func__, __LINE__, &tv);
11621         }
11622 }
11623 #endif
11624
11625 #ifdef TCP_REQUEST_TRK
11626 static void
11627 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len)
11628 {
11629         struct tcp_sendfile_track *rc_cur;
11630         struct tcpcb *tp;
11631         int err = 0;
11632
11633         rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
11634         if (rc_cur == NULL) {
11635                 /* If not in the beginning what about the end piece */
11636                 rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11637                 rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
11638         } else {
11639                 err = 12345;
11640         }
11641         /* If we find no parameters we are in straight DGP mode */
11642         if(rc_cur == NULL) {
11643                 /* None found for this seq, just DGP for now */
11644                 rack->r_ctl.client_suggested_maxseg = 0;
11645                 rack->rc_catch_up = 0;
11646                 rack->r_ctl.bw_rate_cap = 0;
11647                 rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11648                 if (rack->r_ctl.rc_last_sft) {
11649                         rack->r_ctl.rc_last_sft = NULL;
11650                 }
11651                 return;
11652         }
11653         /*
11654          * Ok if we have a new entry *or* have never
11655          * set up an entry we need to proceed. If
11656          * we have already set it up this entry we
11657          * just continue along with what we already
11658          * setup.
11659          */
11660         tp = rack->rc_tp;
11661         if ((rack->r_ctl.rc_last_sft != NULL) &&
11662             (rack->r_ctl.rc_last_sft == rc_cur)) {
11663                 /* Its already in place */
11664                 rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
11665                 return;
11666         }
11667         if (rack->rc_hybrid_mode == 0) {
11668                 rack->r_ctl.rc_last_sft = rc_cur;
11669                 rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11670                 return;
11671         }
11672         if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
11673                 /* Compensate for all the header overhead's */
11674                 rack->r_ctl.bw_rate_cap = rack_compensate_for_linerate(rack, rc_cur->cspr);
11675         } else
11676                 rack->r_ctl.bw_rate_cap = 0;
11677         if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
11678                 rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
11679         else
11680                 rack->r_ctl.client_suggested_maxseg = 0;
11681         if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
11682             (rc_cur->cspr > 0)) {
11683                 uint64_t len;
11684
11685                 rack->rc_catch_up = 1;
11686                 /*
11687                  * Calculate the deadline time, first set the
11688                  * time to when the request arrived.
11689                  */
11690                 rc_cur->deadline = rc_cur->localtime;
11691                 /*
11692                  * Next calculate the length and compensate for
11693                  * TLS if need be.
11694                  */
11695                 len = rc_cur->end - rc_cur->start;
11696                 if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
11697                         /*
11698                          * This session is doing TLS. Take a swag guess
11699                          * at the overhead.
11700                          */
11701                         len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
11702                 }
11703                 /*
11704                  * Now considering the size, and the cspr, what is the time that
11705                  * would be required at the cspr rate. Here we use the raw
11706                  * cspr value since the client only looks at the raw data. We
11707                  * do use len which includes TLS overhead, but not the TCP/IP etc.
11708                  * That will get made up for in the CU pacing rate set.
11709                  */
11710                 len *= HPTS_USEC_IN_SEC;
11711                 len /= rc_cur->cspr;
11712                 rc_cur->deadline += len;
11713         } else {
11714                 rack->rc_catch_up = 0;
11715                 rc_cur->deadline = 0;
11716         }
11717         if (rack->r_ctl.client_suggested_maxseg != 0) {
11718                 /*
11719                  * We need to reset the max pace segs if we have a
11720                  * client_suggested_maxseg.
11721                  */
11722                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
11723         }
11724         rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11725         /* Remember it for next time and for CU mode */
11726         rack->r_ctl.rc_last_sft = rc_cur;
11727 }
11728 #endif
11729
11730 static void
11731 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11732 {
11733 #ifdef TCP_REQUEST_TRK
11734         struct tcp_sendfile_track *ent;
11735
11736         ent = rack->r_ctl.rc_last_sft;
11737         if ((ent == NULL) ||
11738             (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
11739             (SEQ_GEQ(seq, ent->end_seq))) {
11740                 /* Time to update the track. */
11741                 rack_set_dgp_hybrid_mode(rack, seq, len);
11742                 ent = rack->r_ctl.rc_last_sft;
11743         }
11744         /* Out of all */
11745         if (ent == NULL) {
11746                 return;
11747         }
11748         if (SEQ_LT(ent->end_seq, (seq + len))) {
11749                 /*
11750                  * This is the case where our end_seq guess
11751                  * was wrong. This is usually due to TLS having
11752                  * more bytes then our guess. It could also be the
11753                  * case that the client sent in two requests closely
11754                  * and the SB is full of both so we are sending part
11755                  * of each (end|beg). In such a case lets move this
11756                  * guys end to match the end of this send. That
11757                  * way it will complete when all of it is acked.
11758                  */
11759                 ent->end_seq = (seq + len);
11760                 rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent);
11761         }
11762         /* Now validate we have set the send time of this one */
11763         if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11764                 ent->flags |= TCP_TRK_TRACK_FLG_FSND;
11765                 ent->first_send = cts;
11766                 ent->sent_at_fs = rack->rc_tp->t_sndbytes;
11767                 ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11768         }
11769 #endif
11770 }
11771
11772 static void
11773 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
11774 {
11775         /*
11776          * The fast output path is enabled and we
11777          * have moved the cumack forward. Lets see if
11778          * we can expand forward the fast path length by
11779          * that amount. What we would ideally like to
11780          * do is increase the number of bytes in the
11781          * fast path block (left_to_send) by the
11782          * acked amount. However we have to gate that
11783          * by two factors:
11784          * 1) The amount outstanding and the rwnd of the peer
11785          *    (i.e. we don't want to exceed the rwnd of the peer).
11786          *    <and>
11787          * 2) The amount of data left in the socket buffer (i.e.
11788          *    we can't send beyond what is in the buffer).
11789          *
11790          * Note that this does not take into account any increase
11791          * in the cwnd. We will only extend the fast path by
11792          * what was acked.
11793          */
11794         uint32_t new_total, gating_val;
11795
11796         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
11797         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
11798                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
11799         if (new_total <= gating_val) {
11800                 /* We can increase left_to_send by the acked amount */
11801                 counter_u64_add(rack_extended_rfo, 1);
11802                 rack->r_ctl.fsb.left_to_send = new_total;
11803                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
11804                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
11805                          rack, rack->r_ctl.fsb.left_to_send,
11806                          sbavail(&rack->rc_inp->inp_socket->so_snd),
11807                          (tp->snd_max - tp->snd_una)));
11808
11809         }
11810 }
11811
11812 static void
11813 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
11814 {
11815         /*
11816          * Here any sendmap entry that points to the
11817          * beginning mbuf must be adjusted to the correct
11818          * offset. This must be called with:
11819          * 1) The socket buffer locked
11820          * 2) snd_una adjusted to its new position.
11821          *
11822          * Note that (2) implies rack_ack_received has also
11823          * been called and all the sbcut's have been done.
11824          *
11825          * We grab the first mbuf in the socket buffer and
11826          * then go through the front of the sendmap, recalculating
11827          * the stored offset for any sendmap entry that has
11828          * that mbuf. We must use the sb functions to do this
11829          * since its possible an add was done has well as
11830          * the subtraction we may have just completed. This should
11831          * not be a penalty though, since we just referenced the sb
11832          * to go in and trim off the mbufs that we freed (of course
11833          * there will be a penalty for the sendmap references though).
11834          *
11835          * Note also with INVARIANT on, we validate with a KASSERT
11836          * that the first sendmap entry has a soff of 0.
11837          *
11838          */
11839         struct mbuf *m;
11840         struct rack_sendmap *rsm;
11841         tcp_seq snd_una;
11842 #ifdef INVARIANTS
11843         int first_processed = 0;
11844 #endif
11845
11846         snd_una = rack->rc_tp->snd_una;
11847         SOCKBUF_LOCK_ASSERT(sb);
11848         m = sb->sb_mb;
11849         rsm = tqhash_min(rack->r_ctl.tqh);
11850         if ((rsm == NULL) || (m == NULL)) {
11851                 /* Nothing outstanding */
11852                 return;
11853         }
11854         /* The very first RSM's mbuf must point to the head mbuf in the sb */
11855         KASSERT((rsm->m == m),
11856                 ("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
11857                  rack, sb, rsm));
11858         while (rsm->m && (rsm->m == m)) {
11859                 /* one to adjust */
11860 #ifdef INVARIANTS
11861                 struct mbuf *tm;
11862                 uint32_t soff;
11863
11864                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
11865                 if ((rsm->orig_m_len != m->m_len) ||
11866                     (rsm->orig_t_space != M_TRAILINGROOM(m))){
11867                         rack_adjust_orig_mlen(rsm);
11868                 }
11869                 if (first_processed == 0) {
11870                         KASSERT((rsm->soff == 0),
11871                                 ("Rack:%p rsm:%p -- rsm at head but soff not zero",
11872                                  rack, rsm));
11873                         first_processed = 1;
11874                 }
11875                 if ((rsm->soff != soff) || (rsm->m != tm)) {
11876                         /*
11877                          * This is not a fatal error, we anticipate it
11878                          * might happen (the else code), so we count it here
11879                          * so that under invariant we can see that it really
11880                          * does happen.
11881                          */
11882                         counter_u64_add(rack_adjust_map_bw, 1);
11883                 }
11884                 rsm->m = tm;
11885                 rsm->soff = soff;
11886                 if (tm) {
11887                         rsm->orig_m_len = rsm->m->m_len;
11888                         rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11889                 } else {
11890                         rsm->orig_m_len = 0;
11891                         rsm->orig_t_space = 0;
11892                 }
11893 #else
11894                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
11895                 if (rsm->m) {
11896                         rsm->orig_m_len = rsm->m->m_len;
11897                         rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11898                 } else {
11899                         rsm->orig_m_len = 0;
11900                         rsm->orig_t_space = 0;
11901                 }
11902 #endif
11903                 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
11904                 if (rsm == NULL)
11905                         break;
11906         }
11907 }
11908
11909 #ifdef TCP_REQUEST_TRK
11910 static inline void
11911 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
11912 {
11913         struct tcp_sendfile_track *ent;
11914         int i;
11915
11916         if ((rack->rc_hybrid_mode == 0) &&
11917             (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
11918                 /*
11919                  * Just do normal completions hybrid pacing is not on
11920                  * and CLDL is off as well.
11921                  */
11922                 tcp_req_check_for_comp(rack->rc_tp, th_ack);
11923                 return;
11924         }
11925         /*
11926          * Originally I was just going to find the th_ack associated
11927          * with an entry. But then I realized a large strech ack could
11928          * in theory ack two or more requests at once. So instead we
11929          * need to find all entries that are completed by th_ack not
11930          * just a single entry and do our logging.
11931          */
11932         ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11933         while (ent != NULL) {
11934                 /*
11935                  * We may be doing hybrid pacing or CLDL and need more details possibly
11936                  * so we do it manually instead of calling
11937                  * tcp_req_check_for_comp()
11938                  */
11939                 uint64_t laa, tim, data, cbw, ftim;
11940
11941                 /* Ok this ack frees it */
11942                 rack_log_hybrid(rack, th_ack,
11943                                 ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
11944                 /* calculate the time based on the ack arrival */
11945                 data = ent->end - ent->start;
11946                 laa = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
11947                 if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
11948                         if (ent->first_send > ent->localtime)
11949                                 ftim = ent->first_send;
11950                         else
11951                                 ftim = ent->localtime;
11952                 } else {
11953                         /* TSNH */
11954                         ftim = ent->localtime;
11955                 }
11956                 if (laa > ent->localtime)
11957                         tim = laa - ftim;
11958                 else
11959                         tim = 0;
11960                 cbw = data * HPTS_USEC_IN_SEC;
11961                 if (tim > 0)
11962                         cbw /= tim;
11963                 else
11964                         cbw = 0;
11965                 rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent);
11966                 /*
11967                  * Check to see if we are freeing what we are pointing to send wise
11968                  * if so be sure to NULL the pointer so we know we are no longer
11969                  * set to anything.
11970                  */
11971                 if (ent == rack->r_ctl.rc_last_sft)
11972                         rack->r_ctl.rc_last_sft = NULL;
11973                 /* Generate the log that the tcp_netflix call would have */
11974                 tcp_req_log_req_info(rack->rc_tp, ent,
11975                                       i, TCP_TRK_REQ_LOG_FREED, 0, 0);
11976                 /* Free it and see if there is another one */
11977                 tcp_req_free_a_slot(rack->rc_tp, ent);
11978                 ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11979         }
11980 }
11981 #endif
11982
11983
11984 /*
11985  * Return value of 1, we do not need to call rack_process_data().
11986  * return value of 0, rack_process_data can be called.
11987  * For ret_val if its 0 the TCP is locked, if its non-zero
11988  * its unlocked and probably unsafe to touch the TCB.
11989  */
11990 static int
11991 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11992     struct tcpcb *tp, struct tcpopt *to,
11993     uint32_t tiwin, int32_t tlen,
11994     int32_t * ofia, int32_t thflags, int32_t *ret_val)
11995 {
11996         int32_t ourfinisacked = 0;
11997         int32_t nsegs, acked_amount;
11998         int32_t acked;
11999         struct mbuf *mfree;
12000         struct tcp_rack *rack;
12001         int32_t under_pacing = 0;
12002         int32_t recovery = 0;
12003
12004         INP_WLOCK_ASSERT(tptoinpcb(tp));
12005
12006         rack = (struct tcp_rack *)tp->t_fb_ptr;
12007         if (SEQ_GT(th->th_ack, tp->snd_max)) {
12008                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
12009                                       &rack->r_ctl.challenge_ack_ts,
12010                                       &rack->r_ctl.challenge_ack_cnt);
12011                 rack->r_wanted_output = 1;
12012                 return (1);
12013         }
12014         if (rack->gp_ready &&
12015             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12016                 under_pacing = 1;
12017         }
12018         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
12019                 int in_rec, dup_ack_struck = 0;
12020                 int dsack_seen = 0, sacks_seen = 0;
12021
12022                 in_rec = IN_FASTRECOVERY(tp->t_flags);
12023                 if (rack->rc_in_persist) {
12024                         tp->t_rxtshift = 0;
12025                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12026                                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12027                 }
12028
12029                 if ((th->th_ack == tp->snd_una) &&
12030                     (tiwin == tp->snd_wnd) &&
12031                     ((to->to_flags & TOF_SACK) == 0)) {
12032                         rack_strike_dupack(rack);
12033                         dup_ack_struck = 1;
12034                 }
12035                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
12036                              dup_ack_struck, &dsack_seen, &sacks_seen);
12037                 if ((rack->sack_attack_disable > 0) &&
12038                     (th->th_ack == tp->snd_una) &&
12039                     (tiwin == tp->snd_wnd) &&
12040                     (dsack_seen == 0) &&
12041                     (sacks_seen > 0)) {
12042                         /*
12043                          * If sacks have been disabled we may
12044                          * want to strike a dup-ack "ignoring" the
12045                          * sack as long as the sack was not a "dsack". Note
12046                          * that if no sack is sent (TOF_SACK is off) then the
12047                          * normal dsack code above rack_log_ack() would have
12048                          * already struck. So this is just to catch the case
12049                          * were we are ignoring sacks from this guy due to
12050                          * it being a suspected attacker.
12051                          */
12052                         rack_strike_dupack(rack);
12053                 }
12054
12055         }
12056         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12057                 /*
12058                  * Old ack, behind (or duplicate to) the last one rcv'd
12059                  * Note: We mark reordering is occuring if its
12060                  * less than and we have not closed our window.
12061                  */
12062                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
12063                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12064                         if (rack->r_ctl.rc_reorder_ts == 0)
12065                                 rack->r_ctl.rc_reorder_ts = 1;
12066                 }
12067                 return (0);
12068         }
12069         /*
12070          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
12071          * something we sent.
12072          */
12073         if (tp->t_flags & TF_NEEDSYN) {
12074                 /*
12075                  * T/TCP: Connection was half-synchronized, and our SYN has
12076                  * been ACK'd (so connection is now fully synchronized).  Go
12077                  * to non-starred state, increment snd_una for ACK of SYN,
12078                  * and check if we can do window scaling.
12079                  */
12080                 tp->t_flags &= ~TF_NEEDSYN;
12081                 tp->snd_una++;
12082                 /* Do window scaling? */
12083                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12084                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12085                         tp->rcv_scale = tp->request_r_scale;
12086                         /* Send window already scaled. */
12087                 }
12088         }
12089         nsegs = max(1, m->m_pkthdr.lro_nsegs);
12090
12091         acked = BYTES_THIS_ACK(tp, th);
12092         if (acked) {
12093                 /*
12094                  * Any time we move the cum-ack forward clear
12095                  * keep-alive tied probe-not-answered. The
12096                  * persists clears its own on entry.
12097                  */
12098                 rack->probe_not_answered = 0;
12099         }
12100         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12101         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12102         /*
12103          * If we just performed our first retransmit, and the ACK arrives
12104          * within our recovery window, then it was a mistake to do the
12105          * retransmit in the first place.  Recover our original cwnd and
12106          * ssthresh, and proceed to transmit where we left off.
12107          */
12108         if ((tp->t_flags & TF_PREVVALID) &&
12109             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12110                 tp->t_flags &= ~TF_PREVVALID;
12111                 if (tp->t_rxtshift == 1 &&
12112                     (int)(ticks - tp->t_badrxtwin) < 0)
12113                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12114         }
12115         if (acked) {
12116                 /* assure we are not backed off */
12117                 tp->t_rxtshift = 0;
12118                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12119                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12120                 rack->rc_tlp_in_progress = 0;
12121                 rack->r_ctl.rc_tlp_cnt_out = 0;
12122                 /*
12123                  * If it is the RXT timer we want to
12124                  * stop it, so we can restart a TLP.
12125                  */
12126                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12127                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12128 #ifdef TCP_REQUEST_TRK
12129                 rack_req_check_for_comp(rack, th->th_ack);
12130 #endif
12131         }
12132         /*
12133          * If we have a timestamp reply, update smoothed round trip time. If
12134          * no timestamp is present but transmit timer is running and timed
12135          * sequence number was acked, update smoothed round trip time. Since
12136          * we now have an rtt measurement, cancel the timer backoff (cf.,
12137          * Phil Karn's retransmit alg.). Recompute the initial retransmit
12138          * timer.
12139          *
12140          * Some boxes send broken timestamp replies during the SYN+ACK
12141          * phase, ignore timestamps of 0 or we could calculate a huge RTT
12142          * and blow up the retransmit timer.
12143          */
12144         /*
12145          * If all outstanding data is acked, stop retransmit timer and
12146          * remember to restart (more output or persist). If there is more
12147          * data to be acked, restart retransmit timer, using current
12148          * (possibly backed-off) value.
12149          */
12150         if (acked == 0) {
12151                 if (ofia)
12152                         *ofia = ourfinisacked;
12153                 return (0);
12154         }
12155         if (IN_RECOVERY(tp->t_flags)) {
12156                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
12157                     (SEQ_LT(th->th_ack, tp->snd_max))) {
12158                         tcp_rack_partialack(tp);
12159                 } else {
12160                         rack_post_recovery(tp, th->th_ack);
12161                         recovery = 1;
12162                 }
12163         }
12164         /*
12165          * Let the congestion control algorithm update congestion control
12166          * related information. This typically means increasing the
12167          * congestion window.
12168          */
12169         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
12170         SOCKBUF_LOCK(&so->so_snd);
12171         acked_amount = min(acked, (int)sbavail(&so->so_snd));
12172         tp->snd_wnd -= acked_amount;
12173         mfree = sbcut_locked(&so->so_snd, acked_amount);
12174         if ((sbused(&so->so_snd) == 0) &&
12175             (acked > acked_amount) &&
12176             (tp->t_state >= TCPS_FIN_WAIT_1) &&
12177             (tp->t_flags & TF_SENTFIN)) {
12178                 /*
12179                  * We must be sure our fin
12180                  * was sent and acked (we can be
12181                  * in FIN_WAIT_1 without having
12182                  * sent the fin).
12183                  */
12184                 ourfinisacked = 1;
12185         }
12186         tp->snd_una = th->th_ack;
12187         /* wakeups? */
12188         if (acked_amount && sbavail(&so->so_snd))
12189                 rack_adjust_sendmap_head(rack, &so->so_snd);
12190         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12191         /* NB: sowwakeup_locked() does an implicit unlock. */
12192         sowwakeup_locked(so);
12193         /* now check the rxt clamps */
12194         if ((recovery == 1) &&
12195             (rack->excess_rxt_on) &&
12196             (rack->r_cwnd_was_clamped == 0))  {
12197                 do_rack_excess_rxt(tp, rack);
12198         } else if (rack->r_cwnd_was_clamped)
12199                 do_rack_check_for_unclamp(tp, rack);
12200         m_freem(mfree);
12201         if (SEQ_GT(tp->snd_una, tp->snd_recover))
12202                 tp->snd_recover = tp->snd_una;
12203
12204         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
12205                 tp->snd_nxt = tp->snd_una;
12206         }
12207         if (under_pacing &&
12208             (rack->use_fixed_rate == 0) &&
12209             (rack->in_probe_rtt == 0) &&
12210             rack->rc_gp_dyn_mul &&
12211             rack->rc_always_pace) {
12212                 /* Check if we are dragging bottom */
12213                 rack_check_bottom_drag(tp, rack, so);
12214         }
12215         if (tp->snd_una == tp->snd_max) {
12216                 /* Nothing left outstanding */
12217                 tp->t_flags &= ~TF_PREVVALID;
12218                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12219                 rack->r_ctl.retran_during_recovery = 0;
12220                 rack->r_ctl.dsack_byte_cnt = 0;
12221                 if (rack->r_ctl.rc_went_idle_time == 0)
12222                         rack->r_ctl.rc_went_idle_time = 1;
12223                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12224                 if (sbavail(&tptosocket(tp)->so_snd) == 0)
12225                         tp->t_acktime = 0;
12226                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12227                 rack->rc_suspicious = 0;
12228                 /* Set need output so persist might get set */
12229                 rack->r_wanted_output = 1;
12230                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12231                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12232                     (sbavail(&so->so_snd) == 0) &&
12233                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12234                         /*
12235                          * The socket was gone and the
12236                          * peer sent data (now or in the past), time to
12237                          * reset him.
12238                          */
12239                         *ret_val = 1;
12240                         /* tcp_close will kill the inp pre-log the Reset */
12241                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12242                         tp = tcp_close(tp);
12243                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
12244                         return (1);
12245                 }
12246         }
12247         if (ofia)
12248                 *ofia = ourfinisacked;
12249         return (0);
12250 }
12251
12252
12253 static void
12254 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12255                   int dir, uint32_t flags, struct rack_sendmap *rsm)
12256 {
12257         if (tcp_bblogging_on(rack->rc_tp)) {
12258                 union tcp_log_stackspecific log;
12259                 struct timeval tv;
12260
12261                 memset(&log, 0, sizeof(log));
12262                 log.u_bbr.flex1 = cnt;
12263                 log.u_bbr.flex2 = split;
12264                 log.u_bbr.flex3 = out;
12265                 log.u_bbr.flex4 = line;
12266                 log.u_bbr.flex5 = rack->r_must_retran;
12267                 log.u_bbr.flex6 = flags;
12268                 log.u_bbr.flex7 = rack->rc_has_collapsed;
12269                 log.u_bbr.flex8 = dir;  /*
12270                                          * 1 is collapsed, 0 is uncollapsed,
12271                                          * 2 is log of a rsm being marked, 3 is a split.
12272                                          */
12273                 if (rsm == NULL)
12274                         log.u_bbr.rttProp = 0;
12275                 else
12276                         log.u_bbr.rttProp = (uint64_t)rsm;
12277                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12278                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12279                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
12280                     &rack->rc_inp->inp_socket->so_rcv,
12281                     &rack->rc_inp->inp_socket->so_snd,
12282                     TCP_RACK_LOG_COLLAPSE, 0,
12283                     0, &log, false, &tv);
12284         }
12285 }
12286
12287 static void
12288 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12289 {
12290         /*
12291          * Here all we do is mark the collapsed point and set the flag.
12292          * This may happen again and again, but there is no
12293          * sense splitting our map until we know where the
12294          * peer finally lands in the collapse.
12295          */
12296         tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12297         if ((rack->rc_has_collapsed == 0) ||
12298             (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12299                 counter_u64_add(rack_collapsed_win_seen, 1);
12300         rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12301         rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12302         rack->rc_has_collapsed = 1;
12303         rack->r_collapse_point_valid = 1;
12304         rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12305 }
12306
12307 static void
12308 rack_un_collapse_window(struct tcp_rack *rack, int line)
12309 {
12310         struct rack_sendmap *nrsm, *rsm;
12311         int cnt = 0, split = 0;
12312         int insret __diagused;
12313
12314
12315         tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12316         rack->rc_has_collapsed = 0;
12317         rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12318         if (rsm == NULL) {
12319                 /* Nothing to do maybe the peer ack'ed it all */
12320                 rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12321                 return;
12322         }
12323         /* Now do we need to split this one? */
12324         if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12325                 rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12326                                   rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12327                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12328                 if (nrsm == NULL) {
12329                         /* We can't get a rsm, mark all? */
12330                         nrsm = rsm;
12331                         goto no_split;
12332                 }
12333                 /* Clone it */
12334                 split = 1;
12335                 rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12336 #ifndef INVARIANTS
12337                 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12338 #else
12339                 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12340                         panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
12341                               nrsm, insret, rack, rsm);
12342                 }
12343 #endif
12344                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12345                                  rack->r_ctl.last_collapse_point, __LINE__);
12346                 if (rsm->r_in_tmap) {
12347                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12348                         nrsm->r_in_tmap = 1;
12349                 }
12350                 /*
12351                  * Set in the new RSM as the
12352                  * collapsed starting point
12353                  */
12354                 rsm = nrsm;
12355         }
12356
12357 no_split:
12358         TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm)  {
12359                 cnt++;
12360                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
12361                 rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12362                 cnt++;
12363         }
12364         if (cnt) {
12365                 counter_u64_add(rack_collapsed_win, 1);
12366         }
12367         rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12368 }
12369
12370 static void
12371 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12372                         int32_t tlen, int32_t tfo_syn)
12373 {
12374         if (DELAY_ACK(tp, tlen) || tfo_syn) {
12375                 rack_timer_cancel(tp, rack,
12376                                   rack->r_ctl.rc_rcvtime, __LINE__);
12377                 tp->t_flags |= TF_DELACK;
12378         } else {
12379                 rack->r_wanted_output = 1;
12380                 tp->t_flags |= TF_ACKNOW;
12381         }
12382 }
12383
12384 static void
12385 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12386 {
12387         /*
12388          * If fast output is in progress, lets validate that
12389          * the new window did not shrink on us and make it
12390          * so fast output should end.
12391          */
12392         if (rack->r_fast_output) {
12393                 uint32_t out;
12394
12395                 /*
12396                  * Calculate what we will send if left as is
12397                  * and compare that to our send window.
12398                  */
12399                 out = ctf_outstanding(tp);
12400                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12401                         /* ok we have an issue */
12402                         if (out >= tp->snd_wnd) {
12403                                 /* Turn off fast output the window is met or collapsed */
12404                                 rack->r_fast_output = 0;
12405                         } else {
12406                                 /* we have some room left */
12407                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12408                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12409                                         /* If not at least 1 full segment never mind */
12410                                         rack->r_fast_output = 0;
12411                                 }
12412                         }
12413                 }
12414         }
12415 }
12416
12417
12418 /*
12419  * Return value of 1, the TCB is unlocked and most
12420  * likely gone, return value of 0, the TCP is still
12421  * locked.
12422  */
12423 static int
12424 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12425     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12426     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12427 {
12428         /*
12429          * Update window information. Don't look at window if no ACK: TAC's
12430          * send garbage on first SYN.
12431          */
12432         int32_t nsegs;
12433         int32_t tfo_syn;
12434         struct tcp_rack *rack;
12435
12436         INP_WLOCK_ASSERT(tptoinpcb(tp));
12437
12438         rack = (struct tcp_rack *)tp->t_fb_ptr;
12439         nsegs = max(1, m->m_pkthdr.lro_nsegs);
12440         if ((thflags & TH_ACK) &&
12441             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12442             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12443             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12444                 /* keep track of pure window updates */
12445                 if (tlen == 0 &&
12446                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12447                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12448                 tp->snd_wnd = tiwin;
12449                 rack_validate_fo_sendwin_up(tp, rack);
12450                 tp->snd_wl1 = th->th_seq;
12451                 tp->snd_wl2 = th->th_ack;
12452                 if (tp->snd_wnd > tp->max_sndwnd)
12453                         tp->max_sndwnd = tp->snd_wnd;
12454                 rack->r_wanted_output = 1;
12455         } else if (thflags & TH_ACK) {
12456                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12457                         tp->snd_wnd = tiwin;
12458                         rack_validate_fo_sendwin_up(tp, rack);
12459                         tp->snd_wl1 = th->th_seq;
12460                         tp->snd_wl2 = th->th_ack;
12461                 }
12462         }
12463         if (tp->snd_wnd < ctf_outstanding(tp))
12464                 /* The peer collapsed the window */
12465                 rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12466         else if (rack->rc_has_collapsed)
12467                 rack_un_collapse_window(rack, __LINE__);
12468         if ((rack->r_collapse_point_valid) &&
12469             (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12470                 rack->r_collapse_point_valid = 0;
12471         /* Was persist timer active and now we have window space? */
12472         if ((rack->rc_in_persist != 0) &&
12473             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12474                                 rack->r_ctl.rc_pace_min_segs))) {
12475                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12476                 tp->snd_nxt = tp->snd_max;
12477                 /* Make sure we output to start the timer */
12478                 rack->r_wanted_output = 1;
12479         }
12480         /* Do we enter persists? */
12481         if ((rack->rc_in_persist == 0) &&
12482             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12483             TCPS_HAVEESTABLISHED(tp->t_state) &&
12484             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12485             sbavail(&tptosocket(tp)->so_snd) &&
12486             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12487                 /*
12488                  * Here the rwnd is less than
12489                  * the pacing size, we are established,
12490                  * nothing is outstanding, and there is
12491                  * data to send. Enter persists.
12492                  */
12493                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12494         }
12495         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12496                 m_freem(m);
12497                 return (0);
12498         }
12499         /*
12500          * don't process the URG bit, ignore them drag
12501          * along the up.
12502          */
12503         tp->rcv_up = tp->rcv_nxt;
12504
12505         /*
12506          * Process the segment text, merging it into the TCP sequencing
12507          * queue, and arranging for acknowledgment of receipt if necessary.
12508          * This process logically involves adjusting tp->rcv_wnd as data is
12509          * presented to the user (this happens in tcp_usrreq.c, case
12510          * PRU_RCVD).  If a FIN has already been received on this connection
12511          * then we just ignore the text.
12512          */
12513         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12514                    IS_FASTOPEN(tp->t_flags));
12515         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12516             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12517                 tcp_seq save_start = th->th_seq;
12518                 tcp_seq save_rnxt  = tp->rcv_nxt;
12519                 int     save_tlen  = tlen;
12520
12521                 m_adj(m, drop_hdrlen);  /* delayed header drop */
12522                 /*
12523                  * Insert segment which includes th into TCP reassembly
12524                  * queue with control block tp.  Set thflags to whether
12525                  * reassembly now includes a segment with FIN.  This handles
12526                  * the common case inline (segment is the next to be
12527                  * received on an established connection, and the queue is
12528                  * empty), avoiding linkage into and removal from the queue
12529                  * and repetition of various conversions. Set DELACK for
12530                  * segments received in order, but ack immediately when
12531                  * segments are out of order (so fast retransmit can work).
12532                  */
12533                 if (th->th_seq == tp->rcv_nxt &&
12534                     SEGQ_EMPTY(tp) &&
12535                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
12536                     tfo_syn)) {
12537 #ifdef NETFLIX_SB_LIMITS
12538                         u_int mcnt, appended;
12539
12540                         if (so->so_rcv.sb_shlim) {
12541                                 mcnt = m_memcnt(m);
12542                                 appended = 0;
12543                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12544                                     CFO_NOSLEEP, NULL) == false) {
12545                                         counter_u64_add(tcp_sb_shlim_fails, 1);
12546                                         m_freem(m);
12547                                         return (0);
12548                                 }
12549                         }
12550 #endif
12551                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
12552                         tp->rcv_nxt += tlen;
12553                         if (tlen &&
12554                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12555                             (tp->t_fbyte_in == 0)) {
12556                                 tp->t_fbyte_in = ticks;
12557                                 if (tp->t_fbyte_in == 0)
12558                                         tp->t_fbyte_in = 1;
12559                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
12560                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12561                         }
12562                         thflags = tcp_get_flags(th) & TH_FIN;
12563                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12564                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12565                         SOCKBUF_LOCK(&so->so_rcv);
12566                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12567                                 m_freem(m);
12568                         } else
12569 #ifdef NETFLIX_SB_LIMITS
12570                                 appended =
12571 #endif
12572                                         sbappendstream_locked(&so->so_rcv, m, 0);
12573
12574                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12575                         /* NB: sorwakeup_locked() does an implicit unlock. */
12576                         sorwakeup_locked(so);
12577 #ifdef NETFLIX_SB_LIMITS
12578                         if (so->so_rcv.sb_shlim && appended != mcnt)
12579                                 counter_fo_release(so->so_rcv.sb_shlim,
12580                                     mcnt - appended);
12581 #endif
12582                 } else {
12583                         /*
12584                          * XXX: Due to the header drop above "th" is
12585                          * theoretically invalid by now.  Fortunately
12586                          * m_adj() doesn't actually frees any mbufs when
12587                          * trimming from the head.
12588                          */
12589                         tcp_seq temp = save_start;
12590
12591                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
12592                         tp->t_flags |= TF_ACKNOW;
12593                         if (tp->t_flags & TF_WAKESOR) {
12594                                 tp->t_flags &= ~TF_WAKESOR;
12595                                 /* NB: sorwakeup_locked() does an implicit unlock. */
12596                                 sorwakeup_locked(so);
12597                         }
12598                 }
12599                 if ((tp->t_flags & TF_SACK_PERMIT) &&
12600                     (save_tlen > 0) &&
12601                     TCPS_HAVEESTABLISHED(tp->t_state)) {
12602                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
12603                                 /*
12604                                  * DSACK actually handled in the fastpath
12605                                  * above.
12606                                  */
12607                                 tcp_update_sack_list(tp, save_start,
12608                                     save_start + save_tlen);
12609                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
12610                                 if ((tp->rcv_numsacks >= 1) &&
12611                                     (tp->sackblks[0].end == save_start)) {
12612                                         /*
12613                                          * Partial overlap, recorded at todrop
12614                                          * above.
12615                                          */
12616                                         tcp_update_sack_list(tp,
12617                                             tp->sackblks[0].start,
12618                                             tp->sackblks[0].end);
12619                                 } else {
12620                                         tcp_update_dsack_list(tp, save_start,
12621                                             save_start + save_tlen);
12622                                 }
12623                         } else if (tlen >= save_tlen) {
12624                                 /* Update of sackblks. */
12625                                 tcp_update_dsack_list(tp, save_start,
12626                                     save_start + save_tlen);
12627                         } else if (tlen > 0) {
12628                                 tcp_update_dsack_list(tp, save_start,
12629                                     save_start + tlen);
12630                         }
12631                 }
12632         } else {
12633                 m_freem(m);
12634                 thflags &= ~TH_FIN;
12635         }
12636
12637         /*
12638          * If FIN is received ACK the FIN and let the user know that the
12639          * connection is closing.
12640          */
12641         if (thflags & TH_FIN) {
12642                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12643                         /* The socket upcall is handled by socantrcvmore. */
12644                         socantrcvmore(so);
12645                         /*
12646                          * If connection is half-synchronized (ie NEEDSYN
12647                          * flag on) then delay ACK, so it may be piggybacked
12648                          * when SYN is sent. Otherwise, since we received a
12649                          * FIN then no more input can be expected, send ACK
12650                          * now.
12651                          */
12652                         if (tp->t_flags & TF_NEEDSYN) {
12653                                 rack_timer_cancel(tp, rack,
12654                                     rack->r_ctl.rc_rcvtime, __LINE__);
12655                                 tp->t_flags |= TF_DELACK;
12656                         } else {
12657                                 tp->t_flags |= TF_ACKNOW;
12658                         }
12659                         tp->rcv_nxt++;
12660                 }
12661                 switch (tp->t_state) {
12662                         /*
12663                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
12664                          * CLOSE_WAIT state.
12665                          */
12666                 case TCPS_SYN_RECEIVED:
12667                         tp->t_starttime = ticks;
12668                         /* FALLTHROUGH */
12669                 case TCPS_ESTABLISHED:
12670                         rack_timer_cancel(tp, rack,
12671                             rack->r_ctl.rc_rcvtime, __LINE__);
12672                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
12673                         break;
12674
12675                         /*
12676                          * If still in FIN_WAIT_1 STATE FIN has not been
12677                          * acked so enter the CLOSING state.
12678                          */
12679                 case TCPS_FIN_WAIT_1:
12680                         rack_timer_cancel(tp, rack,
12681                             rack->r_ctl.rc_rcvtime, __LINE__);
12682                         tcp_state_change(tp, TCPS_CLOSING);
12683                         break;
12684
12685                         /*
12686                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
12687                          * starting the time-wait timer, turning off the
12688                          * other standard timers.
12689                          */
12690                 case TCPS_FIN_WAIT_2:
12691                         rack_timer_cancel(tp, rack,
12692                             rack->r_ctl.rc_rcvtime, __LINE__);
12693                         tcp_twstart(tp);
12694                         return (1);
12695                 }
12696         }
12697         /*
12698          * Return any desired output.
12699          */
12700         if ((tp->t_flags & TF_ACKNOW) ||
12701             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
12702                 rack->r_wanted_output = 1;
12703         }
12704         return (0);
12705 }
12706
12707 /*
12708  * Here nothing is really faster, its just that we
12709  * have broken out the fast-data path also just like
12710  * the fast-ack.
12711  */
12712 static int
12713 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
12714     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12715     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
12716 {
12717         int32_t nsegs;
12718         int32_t newsize = 0;    /* automatic sockbuf scaling */
12719         struct tcp_rack *rack;
12720 #ifdef NETFLIX_SB_LIMITS
12721         u_int mcnt, appended;
12722 #endif
12723
12724         /*
12725          * If last ACK falls within this segment's sequence numbers, record
12726          * the timestamp. NOTE that the test is modified according to the
12727          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12728          */
12729         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
12730                 return (0);
12731         }
12732         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
12733                 return (0);
12734         }
12735         if (tiwin && tiwin != tp->snd_wnd) {
12736                 return (0);
12737         }
12738         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
12739                 return (0);
12740         }
12741         if (__predict_false((to->to_flags & TOF_TS) &&
12742             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
12743                 return (0);
12744         }
12745         if (__predict_false((th->th_ack != tp->snd_una))) {
12746                 return (0);
12747         }
12748         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
12749                 return (0);
12750         }
12751         if ((to->to_flags & TOF_TS) != 0 &&
12752             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12753                 tp->ts_recent_age = tcp_ts_getticks();
12754                 tp->ts_recent = to->to_tsval;
12755         }
12756         rack = (struct tcp_rack *)tp->t_fb_ptr;
12757         /*
12758          * This is a pure, in-sequence data packet with nothing on the
12759          * reassembly queue and we have enough buffer space to take it.
12760          */
12761         nsegs = max(1, m->m_pkthdr.lro_nsegs);
12762
12763 #ifdef NETFLIX_SB_LIMITS
12764         if (so->so_rcv.sb_shlim) {
12765                 mcnt = m_memcnt(m);
12766                 appended = 0;
12767                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12768                     CFO_NOSLEEP, NULL) == false) {
12769                         counter_u64_add(tcp_sb_shlim_fails, 1);
12770                         m_freem(m);
12771                         return (1);
12772                 }
12773         }
12774 #endif
12775         /* Clean receiver SACK report if present */
12776         if (tp->rcv_numsacks)
12777                 tcp_clean_sackreport(tp);
12778         KMOD_TCPSTAT_INC(tcps_preddat);
12779         tp->rcv_nxt += tlen;
12780         if (tlen &&
12781             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12782             (tp->t_fbyte_in == 0)) {
12783                 tp->t_fbyte_in = ticks;
12784                 if (tp->t_fbyte_in == 0)
12785                         tp->t_fbyte_in = 1;
12786                 if (tp->t_fbyte_out && tp->t_fbyte_in)
12787                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12788         }
12789         /*
12790          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
12791          */
12792         tp->snd_wl1 = th->th_seq;
12793         /*
12794          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
12795          */
12796         tp->rcv_up = tp->rcv_nxt;
12797         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12798         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12799         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12800
12801         /* Add data to socket buffer. */
12802         SOCKBUF_LOCK(&so->so_rcv);
12803         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12804                 m_freem(m);
12805         } else {
12806                 /*
12807                  * Set new socket buffer size. Give up when limit is
12808                  * reached.
12809                  */
12810                 if (newsize)
12811                         if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12812                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12813                 m_adj(m, drop_hdrlen);  /* delayed header drop */
12814 #ifdef NETFLIX_SB_LIMITS
12815                 appended =
12816 #endif
12817                         sbappendstream_locked(&so->so_rcv, m, 0);
12818                 ctf_calc_rwin(so, tp);
12819         }
12820         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12821         /* NB: sorwakeup_locked() does an implicit unlock. */
12822         sorwakeup_locked(so);
12823 #ifdef NETFLIX_SB_LIMITS
12824         if (so->so_rcv.sb_shlim && mcnt != appended)
12825                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
12826 #endif
12827         rack_handle_delayed_ack(tp, rack, tlen, 0);
12828         if (tp->snd_una == tp->snd_max)
12829                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12830         return (1);
12831 }
12832
12833 /*
12834  * This subfunction is used to try to highly optimize the
12835  * fast path. We again allow window updates that are
12836  * in sequence to remain in the fast-path. We also add
12837  * in the __predict's to attempt to help the compiler.
12838  * Note that if we return a 0, then we can *not* process
12839  * it and the caller should push the packet into the
12840  * slow-path.
12841  */
12842 static int
12843 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12844     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12845     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
12846 {
12847         int32_t acked;
12848         int32_t nsegs;
12849         int32_t under_pacing = 0;
12850         struct tcp_rack *rack;
12851
12852         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12853                 /* Old ack, behind (or duplicate to) the last one rcv'd */
12854                 return (0);
12855         }
12856         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
12857                 /* Above what we have sent? */
12858                 return (0);
12859         }
12860         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
12861                 /* We are retransmitting */
12862                 return (0);
12863         }
12864         if (__predict_false(tiwin == 0)) {
12865                 /* zero window */
12866                 return (0);
12867         }
12868         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
12869                 /* We need a SYN or a FIN, unlikely.. */
12870                 return (0);
12871         }
12872         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
12873                 /* Timestamp is behind .. old ack with seq wrap? */
12874                 return (0);
12875         }
12876         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
12877                 /* Still recovering */
12878                 return (0);
12879         }
12880         rack = (struct tcp_rack *)tp->t_fb_ptr;
12881         if (rack->r_ctl.rc_sacked) {
12882                 /* We have sack holes on our scoreboard */
12883                 return (0);
12884         }
12885         /* Ok if we reach here, we can process a fast-ack */
12886         if (rack->gp_ready &&
12887             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12888                 under_pacing = 1;
12889         }
12890         nsegs = max(1, m->m_pkthdr.lro_nsegs);
12891         rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
12892         /* Did the window get updated? */
12893         if (tiwin != tp->snd_wnd) {
12894                 tp->snd_wnd = tiwin;
12895                 rack_validate_fo_sendwin_up(tp, rack);
12896                 tp->snd_wl1 = th->th_seq;
12897                 if (tp->snd_wnd > tp->max_sndwnd)
12898                         tp->max_sndwnd = tp->snd_wnd;
12899         }
12900         /* Do we exit persists? */
12901         if ((rack->rc_in_persist != 0) &&
12902             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12903                                rack->r_ctl.rc_pace_min_segs))) {
12904                 rack_exit_persist(tp, rack, cts);
12905         }
12906         /* Do we enter persists? */
12907         if ((rack->rc_in_persist == 0) &&
12908             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12909             TCPS_HAVEESTABLISHED(tp->t_state) &&
12910             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12911             sbavail(&tptosocket(tp)->so_snd) &&
12912             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12913                 /*
12914                  * Here the rwnd is less than
12915                  * the pacing size, we are established,
12916                  * nothing is outstanding, and there is
12917                  * data to send. Enter persists.
12918                  */
12919                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
12920         }
12921         /*
12922          * If last ACK falls within this segment's sequence numbers, record
12923          * the timestamp. NOTE that the test is modified according to the
12924          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12925          */
12926         if ((to->to_flags & TOF_TS) != 0 &&
12927             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12928                 tp->ts_recent_age = tcp_ts_getticks();
12929                 tp->ts_recent = to->to_tsval;
12930         }
12931         /*
12932          * This is a pure ack for outstanding data.
12933          */
12934         KMOD_TCPSTAT_INC(tcps_predack);
12935
12936         /*
12937          * "bad retransmit" recovery.
12938          */
12939         if ((tp->t_flags & TF_PREVVALID) &&
12940             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12941                 tp->t_flags &= ~TF_PREVVALID;
12942                 if (tp->t_rxtshift == 1 &&
12943                     (int)(ticks - tp->t_badrxtwin) < 0)
12944                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12945         }
12946         /*
12947          * Recalculate the transmit timer / rtt.
12948          *
12949          * Some boxes send broken timestamp replies during the SYN+ACK
12950          * phase, ignore timestamps of 0 or we could calculate a huge RTT
12951          * and blow up the retransmit timer.
12952          */
12953         acked = BYTES_THIS_ACK(tp, th);
12954
12955 #ifdef TCP_HHOOK
12956         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
12957         hhook_run_tcp_est_in(tp, th, to);
12958 #endif
12959         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12960         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12961         if (acked) {
12962                 struct mbuf *mfree;
12963
12964                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
12965                 SOCKBUF_LOCK(&so->so_snd);
12966                 mfree = sbcut_locked(&so->so_snd, acked);
12967                 tp->snd_una = th->th_ack;
12968                 /* Note we want to hold the sb lock through the sendmap adjust */
12969                 rack_adjust_sendmap_head(rack, &so->so_snd);
12970                 /* Wake up the socket if we have room to write more */
12971                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12972                 sowwakeup_locked(so);
12973                 m_freem(mfree);
12974                 tp->t_rxtshift = 0;
12975                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12976                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12977                 rack->rc_tlp_in_progress = 0;
12978                 rack->r_ctl.rc_tlp_cnt_out = 0;
12979                 /*
12980                  * If it is the RXT timer we want to
12981                  * stop it, so we can restart a TLP.
12982                  */
12983                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12984                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12985
12986 #ifdef TCP_REQUEST_TRK
12987                 rack_req_check_for_comp(rack, th->th_ack);
12988 #endif
12989         }
12990         /*
12991          * Let the congestion control algorithm update congestion control
12992          * related information. This typically means increasing the
12993          * congestion window.
12994          */
12995         if (tp->snd_wnd < ctf_outstanding(tp)) {
12996                 /* The peer collapsed the window */
12997                 rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12998         } else if (rack->rc_has_collapsed)
12999                 rack_un_collapse_window(rack, __LINE__);
13000         if ((rack->r_collapse_point_valid) &&
13001             (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
13002                 rack->r_collapse_point_valid = 0;
13003         /*
13004          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
13005          */
13006         tp->snd_wl2 = th->th_ack;
13007         tp->t_dupacks = 0;
13008         m_freem(m);
13009         /* ND6_HINT(tp);         *//* Some progress has been made. */
13010
13011         /*
13012          * If all outstanding data are acked, stop retransmit timer,
13013          * otherwise restart timer using current (possibly backed-off)
13014          * value. If process is waiting for space, wakeup/selwakeup/signal.
13015          * If data are ready to send, let tcp_output decide between more
13016          * output or persist.
13017          */
13018         if (under_pacing &&
13019             (rack->use_fixed_rate == 0) &&
13020             (rack->in_probe_rtt == 0) &&
13021             rack->rc_gp_dyn_mul &&
13022             rack->rc_always_pace) {
13023                 /* Check if we are dragging bottom */
13024                 rack_check_bottom_drag(tp, rack, so);
13025         }
13026         if (tp->snd_una == tp->snd_max) {
13027                 tp->t_flags &= ~TF_PREVVALID;
13028                 rack->r_ctl.retran_during_recovery = 0;
13029                 rack->rc_suspicious = 0;
13030                 rack->r_ctl.dsack_byte_cnt = 0;
13031                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13032                 if (rack->r_ctl.rc_went_idle_time == 0)
13033                         rack->r_ctl.rc_went_idle_time = 1;
13034                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13035                 if (sbavail(&tptosocket(tp)->so_snd) == 0)
13036                         tp->t_acktime = 0;
13037                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13038         }
13039         if (acked && rack->r_fast_output)
13040                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
13041         if (sbavail(&so->so_snd)) {
13042                 rack->r_wanted_output = 1;
13043         }
13044         return (1);
13045 }
13046
13047 /*
13048  * Return value of 1, the TCB is unlocked and most
13049  * likely gone, return value of 0, the TCP is still
13050  * locked.
13051  */
13052 static int
13053 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
13054     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13055     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13056 {
13057         int32_t ret_val = 0;
13058         int32_t todrop;
13059         int32_t ourfinisacked = 0;
13060         struct tcp_rack *rack;
13061
13062         INP_WLOCK_ASSERT(tptoinpcb(tp));
13063
13064         ctf_calc_rwin(so, tp);
13065         /*
13066          * If the state is SYN_SENT: if seg contains an ACK, but not for our
13067          * SYN, drop the input. if seg contains a RST, then drop the
13068          * connection. if seg does not contain SYN, then drop it. Otherwise
13069          * this is an acceptable SYN segment initialize tp->rcv_nxt and
13070          * tp->irs if seg contains ack then advance tp->snd_una if seg
13071          * contains an ECE and ECN support is enabled, the stream is ECN
13072          * capable. if SYN has been acked change to ESTABLISHED else
13073          * SYN_RCVD state arrange for segment to be acked (eventually)
13074          * continue processing rest of data/controls.
13075          */
13076         if ((thflags & TH_ACK) &&
13077             (SEQ_LEQ(th->th_ack, tp->iss) ||
13078             SEQ_GT(th->th_ack, tp->snd_max))) {
13079                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13080                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13081                 return (1);
13082         }
13083         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
13084                 TCP_PROBE5(connect__refused, NULL, tp,
13085                     mtod(m, const char *), tp, th);
13086                 tp = tcp_drop(tp, ECONNREFUSED);
13087                 ctf_do_drop(m, tp);
13088                 return (1);
13089         }
13090         if (thflags & TH_RST) {
13091                 ctf_do_drop(m, tp);
13092                 return (1);
13093         }
13094         if (!(thflags & TH_SYN)) {
13095                 ctf_do_drop(m, tp);
13096                 return (1);
13097         }
13098         tp->irs = th->th_seq;
13099         tcp_rcvseqinit(tp);
13100         rack = (struct tcp_rack *)tp->t_fb_ptr;
13101         if (thflags & TH_ACK) {
13102                 int tfo_partial = 0;
13103
13104                 KMOD_TCPSTAT_INC(tcps_connects);
13105                 soisconnected(so);
13106 #ifdef MAC
13107                 mac_socketpeer_set_from_mbuf(m, so);
13108 #endif
13109                 /* Do window scaling on this connection? */
13110                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13111                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13112                         tp->rcv_scale = tp->request_r_scale;
13113                 }
13114                 tp->rcv_adv += min(tp->rcv_wnd,
13115                     TCP_MAXWIN << tp->rcv_scale);
13116                 /*
13117                  * If not all the data that was sent in the TFO SYN
13118                  * has been acked, resend the remainder right away.
13119                  */
13120                 if (IS_FASTOPEN(tp->t_flags) &&
13121                     (tp->snd_una != tp->snd_max)) {
13122                         tp->snd_nxt = th->th_ack;
13123                         tfo_partial = 1;
13124                 }
13125                 /*
13126                  * If there's data, delay ACK; if there's also a FIN ACKNOW
13127                  * will be turned on later.
13128                  */
13129                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
13130                         rack_timer_cancel(tp, rack,
13131                                           rack->r_ctl.rc_rcvtime, __LINE__);
13132                         tp->t_flags |= TF_DELACK;
13133                 } else {
13134                         rack->r_wanted_output = 1;
13135                         tp->t_flags |= TF_ACKNOW;
13136                 }
13137
13138                 tcp_ecn_input_syn_sent(tp, thflags, iptos);
13139
13140                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
13141                         /*
13142                          * We advance snd_una for the
13143                          * fast open case. If th_ack is
13144                          * acknowledging data beyond
13145                          * snd_una we can't just call
13146                          * ack-processing since the
13147                          * data stream in our send-map
13148                          * will start at snd_una + 1 (one
13149                          * beyond the SYN). If its just
13150                          * equal we don't need to do that
13151                          * and there is no send_map.
13152                          */
13153                         tp->snd_una++;
13154                 }
13155                 /*
13156                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
13157                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
13158                  */
13159                 tp->t_starttime = ticks;
13160                 if (tp->t_flags & TF_NEEDFIN) {
13161                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
13162                         tp->t_flags &= ~TF_NEEDFIN;
13163                         thflags &= ~TH_SYN;
13164                 } else {
13165                         tcp_state_change(tp, TCPS_ESTABLISHED);
13166                         TCP_PROBE5(connect__established, NULL, tp,
13167                             mtod(m, const char *), tp, th);
13168                         rack_cc_conn_init(tp);
13169                 }
13170         } else {
13171                 /*
13172                  * Received initial SYN in SYN-SENT[*] state => simultaneous
13173                  * open.  If segment contains CC option and there is a
13174                  * cached CC, apply TAO test. If it succeeds, connection is *
13175                  * half-synchronized. Otherwise, do 3-way handshake:
13176                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
13177                  * there was no CC option, clear cached CC value.
13178                  */
13179                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
13180                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
13181         }
13182         /*
13183          * Advance th->th_seq to correspond to first data byte. If data,
13184          * trim to stay within window, dropping FIN if necessary.
13185          */
13186         th->th_seq++;
13187         if (tlen > tp->rcv_wnd) {
13188                 todrop = tlen - tp->rcv_wnd;
13189                 m_adj(m, -todrop);
13190                 tlen = tp->rcv_wnd;
13191                 thflags &= ~TH_FIN;
13192                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13193                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13194         }
13195         tp->snd_wl1 = th->th_seq - 1;
13196         tp->rcv_up = th->th_seq;
13197         /*
13198          * Client side of transaction: already sent SYN and data. If the
13199          * remote host used T/TCP to validate the SYN, our data will be
13200          * ACK'd; if so, enter normal data segment processing in the middle
13201          * of step 5, ack processing. Otherwise, goto step 6.
13202          */
13203         if (thflags & TH_ACK) {
13204                 /* For syn-sent we need to possibly update the rtt */
13205                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13206                         uint32_t t, mcts;
13207
13208                         mcts = tcp_ts_getticks();
13209                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13210                         if (!tp->t_rttlow || tp->t_rttlow > t)
13211                                 tp->t_rttlow = t;
13212                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13213                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13214                         tcp_rack_xmit_timer_commit(rack, tp);
13215                 }
13216                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
13217                         return (ret_val);
13218                 /* We may have changed to FIN_WAIT_1 above */
13219                 if (tp->t_state == TCPS_FIN_WAIT_1) {
13220                         /*
13221                          * In FIN_WAIT_1 STATE in addition to the processing
13222                          * for the ESTABLISHED state if our FIN is now
13223                          * acknowledged then enter FIN_WAIT_2.
13224                          */
13225                         if (ourfinisacked) {
13226                                 /*
13227                                  * If we can't receive any more data, then
13228                                  * closing user can proceed. Starting the
13229                                  * timer is contrary to the specification,
13230                                  * but if we don't get a FIN we'll hang
13231                                  * forever.
13232                                  *
13233                                  * XXXjl: we should release the tp also, and
13234                                  * use a compressed state.
13235                                  */
13236                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13237                                         soisdisconnected(so);
13238                                         tcp_timer_activate(tp, TT_2MSL,
13239                                             (tcp_fast_finwait2_recycle ?
13240                                             tcp_finwait2_timeout :
13241                                             TP_MAXIDLE(tp)));
13242                                 }
13243                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13244                         }
13245                 }
13246         }
13247         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13248            tiwin, thflags, nxt_pkt));
13249 }
13250
13251 /*
13252  * Return value of 1, the TCB is unlocked and most
13253  * likely gone, return value of 0, the TCP is still
13254  * locked.
13255  */
13256 static int
13257 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13258     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13259     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13260 {
13261         struct tcp_rack *rack;
13262         int32_t ret_val = 0;
13263         int32_t ourfinisacked = 0;
13264
13265         ctf_calc_rwin(so, tp);
13266         if ((thflags & TH_ACK) &&
13267             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13268             SEQ_GT(th->th_ack, tp->snd_max))) {
13269                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13270                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13271                 return (1);
13272         }
13273         rack = (struct tcp_rack *)tp->t_fb_ptr;
13274         if (IS_FASTOPEN(tp->t_flags)) {
13275                 /*
13276                  * When a TFO connection is in SYN_RECEIVED, the
13277                  * only valid packets are the initial SYN, a
13278                  * retransmit/copy of the initial SYN (possibly with
13279                  * a subset of the original data), a valid ACK, a
13280                  * FIN, or a RST.
13281                  */
13282                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13283                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13284                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13285                         return (1);
13286                 } else if (thflags & TH_SYN) {
13287                         /* non-initial SYN is ignored */
13288                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13289                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13290                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13291                                 ctf_do_drop(m, NULL);
13292                                 return (0);
13293                         }
13294                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13295                         ctf_do_drop(m, NULL);
13296                         return (0);
13297                 }
13298         }
13299
13300         if ((thflags & TH_RST) ||
13301             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13302                 return (__ctf_process_rst(m, th, so, tp,
13303                                           &rack->r_ctl.challenge_ack_ts,
13304                                           &rack->r_ctl.challenge_ack_cnt));
13305         /*
13306          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13307          * it's less than ts_recent, drop it.
13308          */
13309         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13310             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13311                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13312                         return (ret_val);
13313         }
13314         /*
13315          * In the SYN-RECEIVED state, validate that the packet belongs to
13316          * this connection before trimming the data to fit the receive
13317          * window.  Check the sequence number versus IRS since we know the
13318          * sequence numbers haven't wrapped.  This is a partial fix for the
13319          * "LAND" DoS attack.
13320          */
13321         if (SEQ_LT(th->th_seq, tp->irs)) {
13322                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13323                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13324                 return (1);
13325         }
13326         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13327                               &rack->r_ctl.challenge_ack_ts,
13328                               &rack->r_ctl.challenge_ack_cnt)) {
13329                 return (ret_val);
13330         }
13331         /*
13332          * If last ACK falls within this segment's sequence numbers, record
13333          * its timestamp. NOTE: 1) That the test incorporates suggestions
13334          * from the latest proposal of the tcplw@cray.com list (Braden
13335          * 1993/04/26). 2) That updating only on newer timestamps interferes
13336          * with our earlier PAWS tests, so this check should be solely
13337          * predicated on the sequence space of this segment. 3) That we
13338          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13339          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13340          * SEG.Len, This modified check allows us to overcome RFC1323's
13341          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13342          * p.869. In such cases, we can still calculate the RTT correctly
13343          * when RCV.NXT == Last.ACK.Sent.
13344          */
13345         if ((to->to_flags & TOF_TS) != 0 &&
13346             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13347             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13348             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13349                 tp->ts_recent_age = tcp_ts_getticks();
13350                 tp->ts_recent = to->to_tsval;
13351         }
13352         tp->snd_wnd = tiwin;
13353         rack_validate_fo_sendwin_up(tp, rack);
13354         /*
13355          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13356          * is on (half-synchronized state), then queue data for later
13357          * processing; else drop segment and return.
13358          */
13359         if ((thflags & TH_ACK) == 0) {
13360                 if (IS_FASTOPEN(tp->t_flags)) {
13361                         rack_cc_conn_init(tp);
13362                 }
13363                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13364                     tiwin, thflags, nxt_pkt));
13365         }
13366         KMOD_TCPSTAT_INC(tcps_connects);
13367         if (tp->t_flags & TF_SONOTCONN) {
13368                 tp->t_flags &= ~TF_SONOTCONN;
13369                 soisconnected(so);
13370         }
13371         /* Do window scaling? */
13372         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13373             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13374                 tp->rcv_scale = tp->request_r_scale;
13375         }
13376         /*
13377          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
13378          * FIN-WAIT-1
13379          */
13380         tp->t_starttime = ticks;
13381         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
13382                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13383                 tp->t_tfo_pending = NULL;
13384         }
13385         if (tp->t_flags & TF_NEEDFIN) {
13386                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
13387                 tp->t_flags &= ~TF_NEEDFIN;
13388         } else {
13389                 tcp_state_change(tp, TCPS_ESTABLISHED);
13390                 TCP_PROBE5(accept__established, NULL, tp,
13391                     mtod(m, const char *), tp, th);
13392                 /*
13393                  * TFO connections call cc_conn_init() during SYN
13394                  * processing.  Calling it again here for such connections
13395                  * is not harmless as it would undo the snd_cwnd reduction
13396                  * that occurs when a TFO SYN|ACK is retransmitted.
13397                  */
13398                 if (!IS_FASTOPEN(tp->t_flags))
13399                         rack_cc_conn_init(tp);
13400         }
13401         /*
13402          * Account for the ACK of our SYN prior to
13403          * regular ACK processing below, except for
13404          * simultaneous SYN, which is handled later.
13405          */
13406         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13407                 tp->snd_una++;
13408         /*
13409          * If segment contains data or ACK, will call tcp_reass() later; if
13410          * not, do so now to pass queued data to user.
13411          */
13412         if (tlen == 0 && (thflags & TH_FIN) == 0) {
13413                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13414                     (struct mbuf *)0);
13415                 if (tp->t_flags & TF_WAKESOR) {
13416                         tp->t_flags &= ~TF_WAKESOR;
13417                         /* NB: sorwakeup_locked() does an implicit unlock. */
13418                         sorwakeup_locked(so);
13419                 }
13420         }
13421         tp->snd_wl1 = th->th_seq - 1;
13422         /* For syn-recv we need to possibly update the rtt */
13423         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13424                 uint32_t t, mcts;
13425
13426                 mcts = tcp_ts_getticks();
13427                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13428                 if (!tp->t_rttlow || tp->t_rttlow > t)
13429                         tp->t_rttlow = t;
13430                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13431                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13432                 tcp_rack_xmit_timer_commit(rack, tp);
13433         }
13434         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13435                 return (ret_val);
13436         }
13437         if (tp->t_state == TCPS_FIN_WAIT_1) {
13438                 /* We could have went to FIN_WAIT_1 (or EST) above */
13439                 /*
13440                  * In FIN_WAIT_1 STATE in addition to the processing for the
13441                  * ESTABLISHED state if our FIN is now acknowledged then
13442                  * enter FIN_WAIT_2.
13443                  */
13444                 if (ourfinisacked) {
13445                         /*
13446                          * If we can't receive any more data, then closing
13447                          * user can proceed. Starting the timer is contrary
13448                          * to the specification, but if we don't get a FIN
13449                          * we'll hang forever.
13450                          *
13451                          * XXXjl: we should release the tp also, and use a
13452                          * compressed state.
13453                          */
13454                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13455                                 soisdisconnected(so);
13456                                 tcp_timer_activate(tp, TT_2MSL,
13457                                     (tcp_fast_finwait2_recycle ?
13458                                     tcp_finwait2_timeout :
13459                                     TP_MAXIDLE(tp)));
13460                         }
13461                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13462                 }
13463         }
13464         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13465             tiwin, thflags, nxt_pkt));
13466 }
13467
13468 /*
13469  * Return value of 1, the TCB is unlocked and most
13470  * likely gone, return value of 0, the TCP is still
13471  * locked.
13472  */
13473 static int
13474 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13475     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13476     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13477 {
13478         int32_t ret_val = 0;
13479         struct tcp_rack *rack;
13480
13481         /*
13482          * Header prediction: check for the two common cases of a
13483          * uni-directional data xfer.  If the packet has no control flags,
13484          * is in-sequence, the window didn't change and we're not
13485          * retransmitting, it's a candidate.  If the length is zero and the
13486          * ack moved forward, we're the sender side of the xfer.  Just free
13487          * the data acked & wake any higher level process that was blocked
13488          * waiting for space.  If the length is non-zero and the ack didn't
13489          * move, we're the receiver side.  If we're getting packets in-order
13490          * (the reassembly queue is empty), add the data toc The socket
13491          * buffer and note that we need a delayed ack. Make sure that the
13492          * hidden state-flags are also off. Since we check for
13493          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13494          */
13495         rack = (struct tcp_rack *)tp->t_fb_ptr;
13496         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13497             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13498             __predict_true(SEGQ_EMPTY(tp)) &&
13499             __predict_true(th->th_seq == tp->rcv_nxt)) {
13500                 if (tlen == 0) {
13501                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13502                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13503                                 return (0);
13504                         }
13505                 } else {
13506                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13507                             tiwin, nxt_pkt, iptos)) {
13508                                 return (0);
13509                         }
13510                 }
13511         }
13512         ctf_calc_rwin(so, tp);
13513
13514         if ((thflags & TH_RST) ||
13515             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13516                 return (__ctf_process_rst(m, th, so, tp,
13517                                           &rack->r_ctl.challenge_ack_ts,
13518                                           &rack->r_ctl.challenge_ack_cnt));
13519
13520         /*
13521          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13522          * synchronized state.
13523          */
13524         if (thflags & TH_SYN) {
13525                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13526                 return (ret_val);
13527         }
13528         /*
13529          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13530          * it's less than ts_recent, drop it.
13531          */
13532         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13533             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13534                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13535                         return (ret_val);
13536         }
13537         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13538                               &rack->r_ctl.challenge_ack_ts,
13539                               &rack->r_ctl.challenge_ack_cnt)) {
13540                 return (ret_val);
13541         }
13542         /*
13543          * If last ACK falls within this segment's sequence numbers, record
13544          * its timestamp. NOTE: 1) That the test incorporates suggestions
13545          * from the latest proposal of the tcplw@cray.com list (Braden
13546          * 1993/04/26). 2) That updating only on newer timestamps interferes
13547          * with our earlier PAWS tests, so this check should be solely
13548          * predicated on the sequence space of this segment. 3) That we
13549          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13550          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13551          * SEG.Len, This modified check allows us to overcome RFC1323's
13552          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13553          * p.869. In such cases, we can still calculate the RTT correctly
13554          * when RCV.NXT == Last.ACK.Sent.
13555          */
13556         if ((to->to_flags & TOF_TS) != 0 &&
13557             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13558             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13559             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13560                 tp->ts_recent_age = tcp_ts_getticks();
13561                 tp->ts_recent = to->to_tsval;
13562         }
13563         /*
13564          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13565          * is on (half-synchronized state), then queue data for later
13566          * processing; else drop segment and return.
13567          */
13568         if ((thflags & TH_ACK) == 0) {
13569                 if (tp->t_flags & TF_NEEDSYN) {
13570                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13571                             tiwin, thflags, nxt_pkt));
13572
13573                 } else if (tp->t_flags & TF_ACKNOW) {
13574                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13575                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13576                         return (ret_val);
13577                 } else {
13578                         ctf_do_drop(m, NULL);
13579                         return (0);
13580                 }
13581         }
13582         /*
13583          * Ack processing.
13584          */
13585         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
13586                 return (ret_val);
13587         }
13588         if (sbavail(&so->so_snd)) {
13589                 if (ctf_progress_timeout_check(tp, true)) {
13590                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
13591                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13592                         return (1);
13593                 }
13594         }
13595         /* State changes only happen in rack_process_data() */
13596         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13597             tiwin, thflags, nxt_pkt));
13598 }
13599
13600 /*
13601  * Return value of 1, the TCB is unlocked and most
13602  * likely gone, return value of 0, the TCP is still
13603  * locked.
13604  */
13605 static int
13606 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
13607     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13608     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13609 {
13610         int32_t ret_val = 0;
13611         struct tcp_rack *rack;
13612
13613         rack = (struct tcp_rack *)tp->t_fb_ptr;
13614         ctf_calc_rwin(so, tp);
13615         if ((thflags & TH_RST) ||
13616             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13617                 return (__ctf_process_rst(m, th, so, tp,
13618                                           &rack->r_ctl.challenge_ack_ts,
13619                                           &rack->r_ctl.challenge_ack_cnt));
13620         /*
13621          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13622          * synchronized state.
13623          */
13624         if (thflags & TH_SYN) {
13625                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13626                 return (ret_val);
13627         }
13628         /*
13629          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13630          * it's less than ts_recent, drop it.
13631          */
13632         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13633             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13634                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13635                         return (ret_val);
13636         }
13637         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13638                               &rack->r_ctl.challenge_ack_ts,
13639                               &rack->r_ctl.challenge_ack_cnt)) {
13640                 return (ret_val);
13641         }
13642         /*
13643          * If last ACK falls within this segment's sequence numbers, record
13644          * its timestamp. NOTE: 1) That the test incorporates suggestions
13645          * from the latest proposal of the tcplw@cray.com list (Braden
13646          * 1993/04/26). 2) That updating only on newer timestamps interferes
13647          * with our earlier PAWS tests, so this check should be solely
13648          * predicated on the sequence space of this segment. 3) That we
13649          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13650          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13651          * SEG.Len, This modified check allows us to overcome RFC1323's
13652          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13653          * p.869. In such cases, we can still calculate the RTT correctly
13654          * when RCV.NXT == Last.ACK.Sent.
13655          */
13656         if ((to->to_flags & TOF_TS) != 0 &&
13657             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13658             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13659             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13660                 tp->ts_recent_age = tcp_ts_getticks();
13661                 tp->ts_recent = to->to_tsval;
13662         }
13663         /*
13664          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13665          * is on (half-synchronized state), then queue data for later
13666          * processing; else drop segment and return.
13667          */
13668         if ((thflags & TH_ACK) == 0) {
13669                 if (tp->t_flags & TF_NEEDSYN) {
13670                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13671                             tiwin, thflags, nxt_pkt));
13672
13673                 } else if (tp->t_flags & TF_ACKNOW) {
13674                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13675                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13676                         return (ret_val);
13677                 } else {
13678                         ctf_do_drop(m, NULL);
13679                         return (0);
13680                 }
13681         }
13682         /*
13683          * Ack processing.
13684          */
13685         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
13686                 return (ret_val);
13687         }
13688         if (sbavail(&so->so_snd)) {
13689                 if (ctf_progress_timeout_check(tp, true)) {
13690                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13691                                                 tp, tick, PROGRESS_DROP, __LINE__);
13692                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13693                         return (1);
13694                 }
13695         }
13696         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13697             tiwin, thflags, nxt_pkt));
13698 }
13699
13700 static int
13701 rack_check_data_after_close(struct mbuf *m,
13702     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
13703 {
13704         struct tcp_rack *rack;
13705
13706         rack = (struct tcp_rack *)tp->t_fb_ptr;
13707         if (rack->rc_allow_data_af_clo == 0) {
13708         close_now:
13709                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13710                 /* tcp_close will kill the inp pre-log the Reset */
13711                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13712                 tp = tcp_close(tp);
13713                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
13714                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
13715                 return (1);
13716         }
13717         if (sbavail(&so->so_snd) == 0)
13718                 goto close_now;
13719         /* Ok we allow data that is ignored and a followup reset */
13720         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13721         tp->rcv_nxt = th->th_seq + *tlen;
13722         tp->t_flags2 |= TF2_DROP_AF_DATA;
13723         rack->r_wanted_output = 1;
13724         *tlen = 0;
13725         return (0);
13726 }
13727
13728 /*
13729  * Return value of 1, the TCB is unlocked and most
13730  * likely gone, return value of 0, the TCP is still
13731  * locked.
13732  */
13733 static int
13734 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
13735     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13736     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13737 {
13738         int32_t ret_val = 0;
13739         int32_t ourfinisacked = 0;
13740         struct tcp_rack *rack;
13741
13742         rack = (struct tcp_rack *)tp->t_fb_ptr;
13743         ctf_calc_rwin(so, tp);
13744
13745         if ((thflags & TH_RST) ||
13746             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13747                 return (__ctf_process_rst(m, th, so, tp,
13748                                           &rack->r_ctl.challenge_ack_ts,
13749                                           &rack->r_ctl.challenge_ack_cnt));
13750         /*
13751          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13752          * synchronized state.
13753          */
13754         if (thflags & TH_SYN) {
13755                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13756                 return (ret_val);
13757         }
13758         /*
13759          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13760          * it's less than ts_recent, drop it.
13761          */
13762         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13763             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13764                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13765                         return (ret_val);
13766         }
13767         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13768                               &rack->r_ctl.challenge_ack_ts,
13769                               &rack->r_ctl.challenge_ack_cnt)) {
13770                 return (ret_val);
13771         }
13772         /*
13773          * If new data are received on a connection after the user processes
13774          * are gone, then RST the other end.
13775          */
13776         if ((tp->t_flags & TF_CLOSED) && tlen &&
13777             rack_check_data_after_close(m, tp, &tlen, th, so))
13778                 return (1);
13779         /*
13780          * If last ACK falls within this segment's sequence numbers, record
13781          * its timestamp. NOTE: 1) That the test incorporates suggestions
13782          * from the latest proposal of the tcplw@cray.com list (Braden
13783          * 1993/04/26). 2) That updating only on newer timestamps interferes
13784          * with our earlier PAWS tests, so this check should be solely
13785          * predicated on the sequence space of this segment. 3) That we
13786          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13787          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13788          * SEG.Len, This modified check allows us to overcome RFC1323's
13789          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13790          * p.869. In such cases, we can still calculate the RTT correctly
13791          * when RCV.NXT == Last.ACK.Sent.
13792          */
13793         if ((to->to_flags & TOF_TS) != 0 &&
13794             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13795             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13796             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13797                 tp->ts_recent_age = tcp_ts_getticks();
13798                 tp->ts_recent = to->to_tsval;
13799         }
13800         /*
13801          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13802          * is on (half-synchronized state), then queue data for later
13803          * processing; else drop segment and return.
13804          */
13805         if ((thflags & TH_ACK) == 0) {
13806                 if (tp->t_flags & TF_NEEDSYN) {
13807                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13808                             tiwin, thflags, nxt_pkt));
13809                 } else if (tp->t_flags & TF_ACKNOW) {
13810                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13811                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13812                         return (ret_val);
13813                 } else {
13814                         ctf_do_drop(m, NULL);
13815                         return (0);
13816                 }
13817         }
13818         /*
13819          * Ack processing.
13820          */
13821         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13822                 return (ret_val);
13823         }
13824         if (ourfinisacked) {
13825                 /*
13826                  * If we can't receive any more data, then closing user can
13827                  * proceed. Starting the timer is contrary to the
13828                  * specification, but if we don't get a FIN we'll hang
13829                  * forever.
13830                  *
13831                  * XXXjl: we should release the tp also, and use a
13832                  * compressed state.
13833                  */
13834                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13835                         soisdisconnected(so);
13836                         tcp_timer_activate(tp, TT_2MSL,
13837                             (tcp_fast_finwait2_recycle ?
13838                             tcp_finwait2_timeout :
13839                             TP_MAXIDLE(tp)));
13840                 }
13841                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13842         }
13843         if (sbavail(&so->so_snd)) {
13844                 if (ctf_progress_timeout_check(tp, true)) {
13845                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13846                                                 tp, tick, PROGRESS_DROP, __LINE__);
13847                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13848                         return (1);
13849                 }
13850         }
13851         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13852             tiwin, thflags, nxt_pkt));
13853 }
13854
13855 /*
13856  * Return value of 1, the TCB is unlocked and most
13857  * likely gone, return value of 0, the TCP is still
13858  * locked.
13859  */
13860 static int
13861 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
13862     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13863     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13864 {
13865         int32_t ret_val = 0;
13866         int32_t ourfinisacked = 0;
13867         struct tcp_rack *rack;
13868
13869         rack = (struct tcp_rack *)tp->t_fb_ptr;
13870         ctf_calc_rwin(so, tp);
13871
13872         if ((thflags & TH_RST) ||
13873             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13874                 return (__ctf_process_rst(m, th, so, tp,
13875                                           &rack->r_ctl.challenge_ack_ts,
13876                                           &rack->r_ctl.challenge_ack_cnt));
13877         /*
13878          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13879          * synchronized state.
13880          */
13881         if (thflags & TH_SYN) {
13882                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13883                 return (ret_val);
13884         }
13885         /*
13886          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13887          * it's less than ts_recent, drop it.
13888          */
13889         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13890             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13891                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13892                         return (ret_val);
13893         }
13894         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13895                               &rack->r_ctl.challenge_ack_ts,
13896                               &rack->r_ctl.challenge_ack_cnt)) {
13897                 return (ret_val);
13898         }
13899         /*
13900          * If new data are received on a connection after the user processes
13901          * are gone, then RST the other end.
13902          */
13903         if ((tp->t_flags & TF_CLOSED) && tlen &&
13904             rack_check_data_after_close(m, tp, &tlen, th, so))
13905                 return (1);
13906         /*
13907          * If last ACK falls within this segment's sequence numbers, record
13908          * its timestamp. NOTE: 1) That the test incorporates suggestions
13909          * from the latest proposal of the tcplw@cray.com list (Braden
13910          * 1993/04/26). 2) That updating only on newer timestamps interferes
13911          * with our earlier PAWS tests, so this check should be solely
13912          * predicated on the sequence space of this segment. 3) That we
13913          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13914          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13915          * SEG.Len, This modified check allows us to overcome RFC1323's
13916          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13917          * p.869. In such cases, we can still calculate the RTT correctly
13918          * when RCV.NXT == Last.ACK.Sent.
13919          */
13920         if ((to->to_flags & TOF_TS) != 0 &&
13921             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13922             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13923             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13924                 tp->ts_recent_age = tcp_ts_getticks();
13925                 tp->ts_recent = to->to_tsval;
13926         }
13927         /*
13928          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13929          * is on (half-synchronized state), then queue data for later
13930          * processing; else drop segment and return.
13931          */
13932         if ((thflags & TH_ACK) == 0) {
13933                 if (tp->t_flags & TF_NEEDSYN) {
13934                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13935                             tiwin, thflags, nxt_pkt));
13936                 } else if (tp->t_flags & TF_ACKNOW) {
13937                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13938                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13939                         return (ret_val);
13940                 } else {
13941                         ctf_do_drop(m, NULL);
13942                         return (0);
13943                 }
13944         }
13945         /*
13946          * Ack processing.
13947          */
13948         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13949                 return (ret_val);
13950         }
13951         if (ourfinisacked) {
13952                 tcp_twstart(tp);
13953                 m_freem(m);
13954                 return (1);
13955         }
13956         if (sbavail(&so->so_snd)) {
13957                 if (ctf_progress_timeout_check(tp, true)) {
13958                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13959                                                 tp, tick, PROGRESS_DROP, __LINE__);
13960                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13961                         return (1);
13962                 }
13963         }
13964         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13965             tiwin, thflags, nxt_pkt));
13966 }
13967
13968 /*
13969  * Return value of 1, the TCB is unlocked and most
13970  * likely gone, return value of 0, the TCP is still
13971  * locked.
13972  */
13973 static int
13974 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
13975     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13976     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13977 {
13978         int32_t ret_val = 0;
13979         int32_t ourfinisacked = 0;
13980         struct tcp_rack *rack;
13981
13982         rack = (struct tcp_rack *)tp->t_fb_ptr;
13983         ctf_calc_rwin(so, tp);
13984
13985         if ((thflags & TH_RST) ||
13986             (tp->t_fin_is_rst && (thflags & TH_FIN)))
13987                 return (__ctf_process_rst(m, th, so, tp,
13988                                           &rack->r_ctl.challenge_ack_ts,
13989                                           &rack->r_ctl.challenge_ack_cnt));
13990         /*
13991          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13992          * synchronized state.
13993          */
13994         if (thflags & TH_SYN) {
13995                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13996                 return (ret_val);
13997         }
13998         /*
13999          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14000          * it's less than ts_recent, drop it.
14001          */
14002         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14003             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14004                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14005                         return (ret_val);
14006         }
14007         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14008                               &rack->r_ctl.challenge_ack_ts,
14009                               &rack->r_ctl.challenge_ack_cnt)) {
14010                 return (ret_val);
14011         }
14012         /*
14013          * If new data are received on a connection after the user processes
14014          * are gone, then RST the other end.
14015          */
14016         if ((tp->t_flags & TF_CLOSED) && tlen &&
14017             rack_check_data_after_close(m, tp, &tlen, th, so))
14018                 return (1);
14019         /*
14020          * If last ACK falls within this segment's sequence numbers, record
14021          * its timestamp. NOTE: 1) That the test incorporates suggestions
14022          * from the latest proposal of the tcplw@cray.com list (Braden
14023          * 1993/04/26). 2) That updating only on newer timestamps interferes
14024          * with our earlier PAWS tests, so this check should be solely
14025          * predicated on the sequence space of this segment. 3) That we
14026          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14027          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14028          * SEG.Len, This modified check allows us to overcome RFC1323's
14029          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14030          * p.869. In such cases, we can still calculate the RTT correctly
14031          * when RCV.NXT == Last.ACK.Sent.
14032          */
14033         if ((to->to_flags & TOF_TS) != 0 &&
14034             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14035             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14036             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14037                 tp->ts_recent_age = tcp_ts_getticks();
14038                 tp->ts_recent = to->to_tsval;
14039         }
14040         /*
14041          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14042          * is on (half-synchronized state), then queue data for later
14043          * processing; else drop segment and return.
14044          */
14045         if ((thflags & TH_ACK) == 0) {
14046                 if (tp->t_flags & TF_NEEDSYN) {
14047                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14048                             tiwin, thflags, nxt_pkt));
14049                 } else if (tp->t_flags & TF_ACKNOW) {
14050                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14051                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14052                         return (ret_val);
14053                 } else {
14054                         ctf_do_drop(m, NULL);
14055                         return (0);
14056                 }
14057         }
14058         /*
14059          * case TCPS_LAST_ACK: Ack processing.
14060          */
14061         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14062                 return (ret_val);
14063         }
14064         if (ourfinisacked) {
14065                 tp = tcp_close(tp);
14066                 ctf_do_drop(m, tp);
14067                 return (1);
14068         }
14069         if (sbavail(&so->so_snd)) {
14070                 if (ctf_progress_timeout_check(tp, true)) {
14071                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14072                                                 tp, tick, PROGRESS_DROP, __LINE__);
14073                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14074                         return (1);
14075                 }
14076         }
14077         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14078             tiwin, thflags, nxt_pkt));
14079 }
14080
14081 /*
14082  * Return value of 1, the TCB is unlocked and most
14083  * likely gone, return value of 0, the TCP is still
14084  * locked.
14085  */
14086 static int
14087 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
14088     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14089     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14090 {
14091         int32_t ret_val = 0;
14092         int32_t ourfinisacked = 0;
14093         struct tcp_rack *rack;
14094
14095         rack = (struct tcp_rack *)tp->t_fb_ptr;
14096         ctf_calc_rwin(so, tp);
14097
14098         /* Reset receive buffer auto scaling when not in bulk receive mode. */
14099         if ((thflags & TH_RST) ||
14100             (tp->t_fin_is_rst && (thflags & TH_FIN)))
14101                 return (__ctf_process_rst(m, th, so, tp,
14102                                           &rack->r_ctl.challenge_ack_ts,
14103                                           &rack->r_ctl.challenge_ack_cnt));
14104         /*
14105          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14106          * synchronized state.
14107          */
14108         if (thflags & TH_SYN) {
14109                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14110                 return (ret_val);
14111         }
14112         /*
14113          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14114          * it's less than ts_recent, drop it.
14115          */
14116         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14117             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14118                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14119                         return (ret_val);
14120         }
14121         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14122                               &rack->r_ctl.challenge_ack_ts,
14123                               &rack->r_ctl.challenge_ack_cnt)) {
14124                 return (ret_val);
14125         }
14126         /*
14127          * If new data are received on a connection after the user processes
14128          * are gone, then RST the other end.
14129          */
14130         if ((tp->t_flags & TF_CLOSED) && tlen &&
14131             rack_check_data_after_close(m, tp, &tlen, th, so))
14132                 return (1);
14133         /*
14134          * If last ACK falls within this segment's sequence numbers, record
14135          * its timestamp. NOTE: 1) That the test incorporates suggestions
14136          * from the latest proposal of the tcplw@cray.com list (Braden
14137          * 1993/04/26). 2) That updating only on newer timestamps interferes
14138          * with our earlier PAWS tests, so this check should be solely
14139          * predicated on the sequence space of this segment. 3) That we
14140          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14141          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14142          * SEG.Len, This modified check allows us to overcome RFC1323's
14143          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14144          * p.869. In such cases, we can still calculate the RTT correctly
14145          * when RCV.NXT == Last.ACK.Sent.
14146          */
14147         if ((to->to_flags & TOF_TS) != 0 &&
14148             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14149             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14150             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14151                 tp->ts_recent_age = tcp_ts_getticks();
14152                 tp->ts_recent = to->to_tsval;
14153         }
14154         /*
14155          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14156          * is on (half-synchronized state), then queue data for later
14157          * processing; else drop segment and return.
14158          */
14159         if ((thflags & TH_ACK) == 0) {
14160                 if (tp->t_flags & TF_NEEDSYN) {
14161                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14162                             tiwin, thflags, nxt_pkt));
14163                 } else if (tp->t_flags & TF_ACKNOW) {
14164                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14165                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14166                         return (ret_val);
14167                 } else {
14168                         ctf_do_drop(m, NULL);
14169                         return (0);
14170                 }
14171         }
14172         /*
14173          * Ack processing.
14174          */
14175         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14176                 return (ret_val);
14177         }
14178         if (sbavail(&so->so_snd)) {
14179                 if (ctf_progress_timeout_check(tp, true)) {
14180                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14181                                                 tp, tick, PROGRESS_DROP, __LINE__);
14182                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14183                         return (1);
14184                 }
14185         }
14186         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14187             tiwin, thflags, nxt_pkt));
14188 }
14189
14190 static void inline
14191 rack_clear_rate_sample(struct tcp_rack *rack)
14192 {
14193         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
14194         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
14195         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
14196 }
14197
14198 static void
14199 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
14200 {
14201         uint64_t bw_est, rate_wanted;
14202         int chged = 0;
14203         uint32_t user_max, orig_min, orig_max;
14204
14205 #ifdef TCP_REQUEST_TRK
14206         if (rack->rc_hybrid_mode &&
14207             (rack->r_ctl.rc_pace_max_segs != 0) &&
14208             (rack_hybrid_allow_set_maxseg == 1) &&
14209             (rack->r_ctl.rc_last_sft != NULL)) {
14210                 rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
14211                 return;
14212         }
14213 #endif
14214         orig_min = rack->r_ctl.rc_pace_min_segs;
14215         orig_max = rack->r_ctl.rc_pace_max_segs;
14216         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
14217         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
14218                 chged = 1;
14219         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
14220         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
14221                 if (user_max != rack->r_ctl.rc_pace_max_segs)
14222                         chged = 1;
14223         }
14224         if (rack->rc_force_max_seg) {
14225                 rack->r_ctl.rc_pace_max_segs = user_max;
14226         } else if (rack->use_fixed_rate) {
14227                 bw_est = rack_get_bw(rack);
14228                 if ((rack->r_ctl.crte == NULL) ||
14229                     (bw_est != rack->r_ctl.crte->rate)) {
14230                         rack->r_ctl.rc_pace_max_segs = user_max;
14231                 } else {
14232                         /* We are pacing right at the hardware rate */
14233                         uint32_t segsiz, pace_one;
14234
14235                         if (rack_pace_one_seg ||
14236                             (rack->r_ctl.rc_user_set_min_segs == 1))
14237                                 pace_one = 1;
14238                         else
14239                                 pace_one = 0;
14240                         segsiz = min(ctf_fixed_maxseg(tp),
14241                                      rack->r_ctl.rc_pace_min_segs);
14242                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14243                                 tp, bw_est, segsiz, pace_one,
14244                                 rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14245                 }
14246         } else if (rack->rc_always_pace) {
14247                 if (rack->r_ctl.gp_bw ||
14248                     rack->r_ctl.init_rate) {
14249                         /* We have a rate of some sort set */
14250                         uint32_t  orig;
14251
14252                         bw_est = rack_get_bw(rack);
14253                         orig = rack->r_ctl.rc_pace_max_segs;
14254                         if (fill_override)
14255                                 rate_wanted = *fill_override;
14256                         else
14257                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
14258                         if (rate_wanted) {
14259                                 /* We have something */
14260                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14261                                                                                    rate_wanted,
14262                                                                                    ctf_fixed_maxseg(rack->rc_tp));
14263                         } else
14264                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14265                         if (orig != rack->r_ctl.rc_pace_max_segs)
14266                                 chged = 1;
14267                 } else if ((rack->r_ctl.gp_bw == 0) &&
14268                            (rack->r_ctl.rc_pace_max_segs == 0)) {
14269                         /*
14270                          * If we have nothing limit us to bursting
14271                          * out IW sized pieces.
14272                          */
14273                         chged = 1;
14274                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14275                 }
14276         }
14277         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14278                 chged = 1;
14279                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14280         }
14281         if (chged)
14282                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14283 }
14284
14285
14286 static void
14287 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14288 {
14289 #ifdef INET6
14290         struct ip6_hdr *ip6 = NULL;
14291 #endif
14292 #ifdef INET
14293         struct ip *ip = NULL;
14294 #endif
14295         struct udphdr *udp = NULL;
14296
14297         /* Ok lets fill in the fast block, it can only be used with no IP options! */
14298 #ifdef INET6
14299         if (rack->r_is_v6) {
14300                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14301                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14302                 if (tp->t_port) {
14303                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14304                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14305                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14306                         udp->uh_dport = tp->t_port;
14307                         rack->r_ctl.fsb.udp = udp;
14308                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14309                 } else
14310                 {
14311                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14312                         rack->r_ctl.fsb.udp = NULL;
14313                 }
14314                 tcpip_fillheaders(rack->rc_inp,
14315                                   tp->t_port,
14316                                   ip6, rack->r_ctl.fsb.th);
14317                 rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14318         } else
14319 #endif                          /* INET6 */
14320 #ifdef INET
14321         {
14322                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14323                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14324                 if (tp->t_port) {
14325                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14326                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14327                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14328                         udp->uh_dport = tp->t_port;
14329                         rack->r_ctl.fsb.udp = udp;
14330                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14331                 } else
14332                 {
14333                         rack->r_ctl.fsb.udp = NULL;
14334                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14335                 }
14336                 tcpip_fillheaders(rack->rc_inp,
14337                                   tp->t_port,
14338                                   ip, rack->r_ctl.fsb.th);
14339                 rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14340         }
14341 #endif
14342         rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14343             (long)TCP_MAXWIN << tp->rcv_scale);
14344         rack->r_fsb_inited = 1;
14345 }
14346
14347 static int
14348 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14349 {
14350         /*
14351          * Allocate the larger of spaces V6 if available else just
14352          * V4 and include udphdr (overbook)
14353          */
14354 #ifdef INET6
14355         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14356 #else
14357         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14358 #endif
14359         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14360                                             M_TCPFSB, M_NOWAIT|M_ZERO);
14361         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14362                 return (ENOMEM);
14363         }
14364         rack->r_fsb_inited = 0;
14365         return (0);
14366 }
14367
14368 static void
14369 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14370 {
14371         /*
14372          * Types of logs (mod value)
14373          * 20 - Initial round setup
14374          * 21 - Rack declares a new round.
14375          */
14376         struct tcpcb *tp;
14377
14378         tp = rack->rc_tp;
14379         if (tcp_bblogging_on(tp)) {
14380                 union tcp_log_stackspecific log;
14381                 struct timeval tv;
14382
14383                 memset(&log, 0, sizeof(log));
14384                 log.u_bbr.flex1 = rack->r_ctl.current_round;
14385                 log.u_bbr.flex2 = rack->r_ctl.roundends;
14386                 log.u_bbr.flex3 = high_seq;
14387                 log.u_bbr.flex4 = tp->snd_max;
14388                 log.u_bbr.flex8 = mod;
14389                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14390                 log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14391                 log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14392                 TCP_LOG_EVENTP(tp, NULL,
14393                     &tptosocket(tp)->so_rcv,
14394                     &tptosocket(tp)->so_snd,
14395                     TCP_HYSTART, 0,
14396                     0, &log, false, &tv);
14397         }
14398 }
14399
14400 static void
14401 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14402 {
14403         rack->rack_deferred_inited = 1;
14404         rack->r_ctl.roundends = tp->snd_max;
14405         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14406         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14407 }
14408
14409 static void
14410 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14411 {
14412         /* Retransmit bit controls.
14413          *
14414          * The setting of these values control one of
14415          * three settings you can have and dictate
14416          * how rack does retransmissions. Note this
14417          * is in *any* mode i.e. pacing on or off DGP
14418          * fixed rate pacing, or just bursting rack.
14419          *
14420          * 1 - Use full sized retransmits i.e. limit
14421          *     the size to whatever the pace_max_segments
14422          *     size is.
14423          *
14424          * 2 - Use pacer min granularity as a guide to
14425          *     the size combined with the current calculated
14426          *     goodput b/w measurement. So for example if
14427          *     the goodput is measured at 20Mbps we would
14428          *     calculate 8125 (pacer minimum 250usec in
14429          *     that b/w) and then round it up to the next
14430          *     MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14431          *
14432          * 0 - The rack default 1 MSS (anything not 0/1/2
14433          *     fall here too if we are setting via rack_init()).
14434          *
14435          */
14436         if (ctl == 1) {
14437                 rack->full_size_rxt = 1;
14438                 rack->shape_rxt_to_pacing_min  = 0;
14439         } else if (ctl == 2) {
14440                 rack->full_size_rxt = 0;
14441                 rack->shape_rxt_to_pacing_min  = 1;
14442         } else {
14443                 rack->full_size_rxt = 0;
14444                 rack->shape_rxt_to_pacing_min  = 0;
14445         }
14446 }
14447
14448 static void
14449 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14450                   uint32_t flex1,
14451                   uint32_t flex2,
14452                   uint32_t flex3)
14453 {
14454         if (tcp_bblogging_on(rack->rc_tp)) {
14455                 union tcp_log_stackspecific log;
14456                 struct timeval tv;
14457
14458                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14459                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14460                 log.u_bbr.flex8 = mod;
14461                 log.u_bbr.flex1 = flex1;
14462                 log.u_bbr.flex2 = flex2;
14463                 log.u_bbr.flex3 = flex3;
14464                 tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14465                                0, &log, false, NULL, __func__, __LINE__, &tv);
14466         }
14467 }
14468
14469 static int
14470 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14471 {
14472         struct tcp_rack *rack;
14473         struct rack_sendmap *rsm;
14474         int i;
14475
14476
14477         rack = (struct tcp_rack *)tp->t_fb_ptr;
14478         switch (reqr->req) {
14479         case TCP_QUERY_SENDMAP:
14480                 if ((reqr->req_param == tp->snd_max) ||
14481                     (tp->snd_max == tp->snd_una)){
14482                         /* Unlikely */
14483                         return (0);
14484                 }
14485                 rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14486                 if (rsm == NULL) {
14487                         /* Can't find that seq -- unlikely */
14488                         return (0);
14489                 }
14490                 reqr->sendmap_start = rsm->r_start;
14491                 reqr->sendmap_end = rsm->r_end;
14492                 reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14493                 reqr->sendmap_fas = rsm->r_fas;
14494                 if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14495                         reqr->sendmap_send_cnt = SNDMAP_NRTX;
14496                 for(i=0; i<reqr->sendmap_send_cnt; i++)
14497                         reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14498                 reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14499                 reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14500                 reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14501                 reqr->sendmap_dupacks = rsm->r_dupack;
14502                 rack_log_chg_info(tp, rack, 1,
14503                                   rsm->r_start,
14504                                   rsm->r_end,
14505                                   rsm->r_flags);
14506                 return(1);
14507                 break;
14508         case TCP_QUERY_TIMERS_UP:
14509                 if (rack->r_ctl.rc_hpts_flags == 0) {
14510                         /* no timers up */
14511                         return (0);
14512                 }
14513                 reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14514                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14515                         reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14516                 }
14517                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14518                         reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
14519                 }
14520                 rack_log_chg_info(tp, rack, 2,
14521                                   rack->r_ctl.rc_hpts_flags,
14522                                   rack->r_ctl.rc_last_output_to,
14523                                   rack->r_ctl.rc_timer_exp);
14524                 return (1);
14525                 break;
14526         case TCP_QUERY_RACK_TIMES:
14527                 /* Reordering items */
14528                 reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
14529                 reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
14530                 /* Timerstamps and timers */
14531                 reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
14532                 reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
14533                 reqr->rack_rtt = rack->rc_rack_rtt;
14534                 reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
14535                 reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
14536                 /* PRR data */
14537                 reqr->rack_sacked = rack->r_ctl.rc_sacked;
14538                 reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
14539                 reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
14540                 reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
14541                 reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
14542                 reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
14543                 /* TLP and persists info */
14544                 reqr->rack_tlp_out = rack->rc_tlp_in_progress;
14545                 reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
14546                 if (rack->rc_in_persist) {
14547                         reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
14548                         reqr->rack_in_persist = 1;
14549                 } else {
14550                         reqr->rack_time_went_idle = 0;
14551                         reqr->rack_in_persist = 0;
14552                 }
14553                 if (rack->r_wanted_output)
14554                         reqr->rack_wanted_output = 1;
14555                 else
14556                         reqr->rack_wanted_output = 0;
14557                 return (1);
14558                 break;
14559         default:
14560                 return (-EINVAL);
14561         }
14562 }
14563
14564 static void
14565 rack_switch_failed(struct tcpcb *tp)
14566 {
14567         /*
14568          * This method gets called if a stack switch was
14569          * attempted and it failed. We are left
14570          * but our hpts timers were stopped and we
14571          * need to validate time units and t_flags2.
14572          */
14573         struct tcp_rack *rack;
14574         struct timeval tv;
14575         uint32_t cts;
14576         uint32_t toval;
14577         struct hpts_diag diag;
14578
14579         rack = (struct tcp_rack *)tp->t_fb_ptr;
14580         tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
14581         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14582                 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14583         else
14584                 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14585         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14586                 tp->t_flags2 |= TF2_MBUF_ACKCMP;
14587         if (tp->t_in_hpts > IHPTS_NONE) {
14588                 /* Strange */
14589                 return;
14590         }
14591         cts = tcp_get_usecs(&tv);
14592         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14593                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
14594                         toval = rack->r_ctl.rc_last_output_to - cts;
14595                 } else {
14596                         /* one slot please */
14597                         toval = HPTS_TICKS_PER_SLOT;
14598                 }
14599         } else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14600                 if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
14601                         toval = rack->r_ctl.rc_timer_exp - cts;
14602                 } else {
14603                         /* one slot please */
14604                         toval = HPTS_TICKS_PER_SLOT;
14605                 }
14606         } else
14607                 toval = HPTS_TICKS_PER_SLOT;
14608         (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
14609                                    __LINE__, &diag);
14610         rack_log_hpts_diag(rack, cts, &diag, &tv);
14611 }
14612
14613 static int
14614 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
14615 {
14616         struct rack_sendmap *rsm, *ersm;
14617         int insret __diagused;
14618         /*
14619          * When initing outstanding, we must be quite careful
14620          * to not refer to tp->t_fb_ptr. This has the old rack
14621          * pointer in it, not the "new" one (when we are doing
14622          * a stack switch).
14623          */
14624
14625
14626         if (tp->t_fb->tfb_chg_query == NULL) {
14627                 /* Create a send map for the current outstanding data */
14628
14629                 rsm = rack_alloc(rack);
14630                 if (rsm == NULL) {
14631                         uma_zfree(rack_pcb_zone, ptr);
14632                         return (ENOMEM);
14633                 }
14634                 rsm->r_no_rtt_allowed = 1;
14635                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
14636                 rsm->r_rtr_cnt = 1;
14637                 rsm->r_rtr_bytes = 0;
14638                 if (tp->t_flags & TF_SENTFIN)
14639                         rsm->r_flags |= RACK_HAS_FIN;
14640                 rsm->r_end = tp->snd_max;
14641                 if (tp->snd_una == tp->iss) {
14642                         /* The data space is one beyond snd_una */
14643                         rsm->r_flags |= RACK_HAS_SYN;
14644                         rsm->r_start = tp->iss;
14645                         rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
14646                 } else
14647                         rsm->r_start = tp->snd_una;
14648                 rsm->r_dupack = 0;
14649                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
14650                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
14651                         if (rsm->m) {
14652                                 rsm->orig_m_len = rsm->m->m_len;
14653                                 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14654                         } else {
14655                                 rsm->orig_m_len = 0;
14656                                 rsm->orig_t_space = 0;
14657                         }
14658                 } else {
14659                         /*
14660                          * This can happen if we have a stand-alone FIN or
14661                          *  SYN.
14662                          */
14663                         rsm->m = NULL;
14664                         rsm->orig_m_len = 0;
14665                         rsm->orig_t_space = 0;
14666                         rsm->soff = 0;
14667                 }
14668 #ifdef INVARIANTS
14669                 if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14670                         panic("Insert in rb tree fails ret:%d rack:%p rsm:%p",
14671                               insret, rack, rsm);
14672                 }
14673 #else
14674                 (void)tqhash_insert(rack->r_ctl.tqh, rsm);
14675 #endif
14676                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14677                 rsm->r_in_tmap = 1;
14678         } else {
14679                 /* We have a query mechanism, lets use it */
14680                 struct tcp_query_resp qr;
14681                 int i;
14682                 tcp_seq at;
14683
14684                 at = tp->snd_una;
14685                 while (at != tp->snd_max) {
14686                         memset(&qr, 0, sizeof(qr));
14687                         qr.req = TCP_QUERY_SENDMAP;
14688                         qr.req_param = at;
14689                         if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
14690                                 break;
14691                         /* Move forward */
14692                         at = qr.sendmap_end;
14693                         /* Now lets build the entry for this one */
14694                         rsm = rack_alloc(rack);
14695                         if (rsm == NULL) {
14696                                 uma_zfree(rack_pcb_zone, ptr);
14697                                 return (ENOMEM);
14698                         }
14699                         memset(rsm, 0, sizeof(struct rack_sendmap));
14700                         /* Now configure the rsm and insert it */
14701                         rsm->r_dupack = qr.sendmap_dupacks;
14702                         rsm->r_start = qr.sendmap_start;
14703                         rsm->r_end = qr.sendmap_end;
14704                         if (qr.sendmap_fas)
14705                                 rsm->r_fas = qr.sendmap_end;
14706                         else
14707                                 rsm->r_fas = rsm->r_start - tp->snd_una;
14708                         /*
14709                          * We have carefully aligned the bits
14710                          * so that all we have to do is copy over
14711                          * the bits with the mask.
14712                          */
14713                         rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
14714                         rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
14715                         rsm->r_rtr_cnt = qr.sendmap_send_cnt;
14716                         rsm->r_ack_arrival = qr.sendmap_ack_arrival;
14717                         for (i=0 ; i<rsm->r_rtr_cnt; i++)
14718                                 rsm->r_tim_lastsent[i]  = qr.sendmap_time[i];
14719                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
14720                                            (rsm->r_start - tp->snd_una), &rsm->soff);
14721                         if (rsm->m) {
14722                                 rsm->orig_m_len = rsm->m->m_len;
14723                                 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14724                         } else {
14725                                 rsm->orig_m_len = 0;
14726                                 rsm->orig_t_space = 0;
14727                         }
14728 #ifdef INVARIANTS
14729                         if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14730                                 panic("Insert in rb tree fails ret:%d rack:%p rsm:%p",
14731                                       insret, rack, rsm);
14732                         }
14733 #else
14734                         (void)tqhash_insert(rack->r_ctl.tqh, rsm);
14735 #endif
14736                         if ((rsm->r_flags & RACK_ACKED) == 0)  {
14737                                 TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
14738                                         if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
14739                                             rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
14740                                                 /*
14741                                                  * If the existing ersm was sent at
14742                                                  * a later time than the new one, then
14743                                                  * the new one should appear ahead of this
14744                                                  * ersm.
14745                                                  */
14746                                                 rsm->r_in_tmap = 1;
14747                                                 TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
14748                                                 break;
14749                                         }
14750                                 }
14751                                 if (rsm->r_in_tmap == 0) {
14752                                         /*
14753                                          * Not found so shove it on the tail.
14754                                          */
14755                                         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14756                                         rsm->r_in_tmap = 1;
14757                                 }
14758                         } else {
14759                                 if ((rack->r_ctl.rc_sacklast == NULL) ||
14760                                     (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
14761                                         rack->r_ctl.rc_sacklast = rsm;
14762                                 }
14763                         }
14764                         rack_log_chg_info(tp, rack, 3,
14765                                           rsm->r_start,
14766                                           rsm->r_end,
14767                                           rsm->r_flags);
14768                 }
14769         }
14770         return (0);
14771 }
14772
14773 static void
14774 rack_translate_clamp_value(struct tcp_rack *rack, uint32_t optval)
14775 {
14776         /*
14777          * P = percent bits
14778          * F = fill cw bit -- Toggle fillcw if this bit is set.
14779          * S = Segment bits
14780          * M = set max segment bit
14781          * U = Unclamined
14782          * C = If set to non-zero override the max number of clamps.
14783          * L = Bit to indicate if clamped gets lower.
14784          *
14785          * CCCC CCCCC UUUU UULF PPPP PPPP PPPP PPPP
14786          *
14787          * The lowest 3 nibbles is the perentage .1 - 6553.5%
14788          * where 10.1 = 101, max 6553.5
14789          * The upper 16 bits  holds some options.
14790          * The F bit will turn on fill-cw on if you are
14791          * not pacing, it will turn it off if dgp is on.
14792          * The L bit will change it so when clamped we get
14793          * the min(gp, lt-bw) for dgp.
14794          */
14795         uint16_t per;
14796
14797         rack->r_ctl.saved_rxt_clamp_val = optval;
14798         per = optval & 0x0000ffff;
14799         rack->r_ctl.rxt_threshold = (uint64_t)(per & 0xffff);
14800         if (optval > 0) {
14801                 uint16_t clamp_opt;
14802
14803                 rack->excess_rxt_on = 1;
14804                 clamp_opt = ((optval & 0xffff0000) >> 16);
14805                 rack->r_ctl.clamp_options = clamp_opt & 0x00ff;
14806                 if (clamp_opt & 0xff00) {
14807                         /* A max clamps is also present */
14808                         rack->r_ctl.max_clamps = (clamp_opt >> 8);
14809                 } else {
14810                         /* No specified clamps means no limit */
14811                         rack->r_ctl.max_clamps = 0;
14812                 }
14813                 if (rack->r_ctl.clamp_options & 0x0002) {
14814                         rack->r_clamped_gets_lower  = 1;
14815                 } else {
14816                         rack->r_clamped_gets_lower  = 0;
14817                 }
14818         } else {
14819                 /* Turn it off back to default */
14820                 rack->excess_rxt_on = 0;
14821                 rack->r_clamped_gets_lower  = 0;
14822         }
14823
14824 }
14825
14826
14827 static int32_t
14828 rack_init(struct tcpcb *tp, void **ptr)
14829 {
14830         struct inpcb *inp = tptoinpcb(tp);
14831         struct tcp_rack *rack = NULL;
14832         uint32_t iwin, snt, us_cts;
14833         int err, no_query;
14834
14835         /*
14836          * First are we the initial or are we a switched stack?
14837          * If we are initing via tcp_newtcppcb the ptr passed
14838          * will be tp->t_fb_ptr. If its a stack switch that
14839          * has a previous stack we can query it will be a local
14840          * var that will in the end be set into t_fb_ptr.
14841          */
14842         if (ptr == &tp->t_fb_ptr)
14843                 no_query = 1;
14844         else
14845                 no_query = 0;
14846         *ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
14847         if (*ptr == NULL) {
14848                 /*
14849                  * We need to allocate memory but cant. The INP and INP_INFO
14850                  * locks and they are recursive (happens during setup. So a
14851                  * scheme to drop the locks fails :(
14852                  *
14853                  */
14854                 return(ENOMEM);
14855         }
14856         memset(*ptr, 0, sizeof(struct tcp_rack));
14857         rack = (struct tcp_rack *)*ptr;
14858         rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
14859         if (rack->r_ctl.tqh == NULL) {
14860                 uma_zfree(rack_pcb_zone, rack);
14861                 return(ENOMEM);
14862         }
14863         tqhash_init(rack->r_ctl.tqh);
14864         TAILQ_INIT(&rack->r_ctl.rc_free);
14865         TAILQ_INIT(&rack->r_ctl.rc_tmap);
14866         rack->rc_tp = tp;
14867         rack->rc_inp = inp;
14868         /* Set the flag */
14869         rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
14870         /* Probably not needed but lets be sure */
14871         rack_clear_rate_sample(rack);
14872         /*
14873          * Save off the default values, socket options will poke
14874          * at these if pacing is not on or we have not yet
14875          * reached where pacing is on (gp_ready/fixed enabled).
14876          * When they get set into the CC module (when gp_ready
14877          * is enabled or we enable fixed) then we will set these
14878          * values into the CC and place in here the old values
14879          * so we have a restoral. Then we will set the flag
14880          * rc_pacing_cc_set. That way whenever we turn off pacing
14881          * or switch off this stack, we will know to go restore
14882          * the saved values.
14883          *
14884          * We specifically put into the beta the ecn value for pacing.
14885          */
14886         rack->rc_new_rnd_needed = 1;
14887         rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
14888         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
14889         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
14890         /* We want abe like behavior as well */
14891         rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
14892         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
14893         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
14894         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
14895         if (rack_rxt_clamp_thresh) {
14896                 rack_translate_clamp_value(rack, rack_rxt_clamp_thresh);
14897                 rack->excess_rxt_on = 1;
14898         }
14899         if (rack_uses_full_dgp_in_rec)
14900                 rack->r_ctl.full_dgp_in_rec = 1;
14901         if (rack_fill_cw_state)
14902                 rack->rc_pace_to_cwnd = 1;
14903         if (rack_pacing_min_seg)
14904                 rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
14905         if (use_rack_rr)
14906                 rack->use_rack_rr = 1;
14907         if (rack_dnd_default) {
14908                 rack->rc_pace_dnd = 1;
14909         }
14910         if (V_tcp_delack_enabled)
14911                 tp->t_delayed_ack = 1;
14912         else
14913                 tp->t_delayed_ack = 0;
14914 #ifdef TCP_ACCOUNTING
14915         if (rack_tcp_accounting) {
14916                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
14917         }
14918 #endif
14919         rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
14920         rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
14921         if (rack_enable_shared_cwnd)
14922                 rack->rack_enable_scwnd = 1;
14923         rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
14924         rack->rc_user_set_max_segs = rack_hptsi_segments;
14925         rack->rc_force_max_seg = 0;
14926         TAILQ_INIT(&rack->r_ctl.opt_list);
14927         if (rack_hibeta_setting)
14928                 rack->rack_hibeta = 1;
14929         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
14930         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
14931         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
14932         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
14933         rack->r_ctl.rc_highest_us_rtt = 0;
14934         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
14935         rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
14936         if (rack_use_cmp_acks)
14937                 rack->r_use_cmp_ack = 1;
14938         if (rack_disable_prr)
14939                 rack->rack_no_prr = 1;
14940         if (rack_gp_no_rec_chg)
14941                 rack->rc_gp_no_rec_chg = 1;
14942         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
14943                 rack->rc_always_pace = 1;
14944                 if ((rack->gp_ready) && (rack->rc_always_pace && (rack->use_fixed_rate == 0)))
14945                         rack_set_cc_pacing(rack);
14946         } else
14947                 rack->rc_always_pace = 0;
14948         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
14949                 rack->r_mbuf_queue = 1;
14950         else
14951                 rack->r_mbuf_queue = 0;
14952         rack_set_pace_segments(tp, rack, __LINE__, NULL);
14953         if (rack_limits_scwnd)
14954                 rack->r_limit_scw = 1;
14955         else
14956                 rack->r_limit_scw = 0;
14957         rack_init_retransmit_value(rack, rack_rxt_controls);
14958         rack->rc_labc = V_tcp_abc_l_var;
14959         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
14960         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
14961         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
14962         rack->r_ctl.rc_min_to = rack_min_to;
14963         microuptime(&rack->r_ctl.act_rcv_time);
14964         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
14965         rack->rc_init_win = rack_default_init_window;
14966         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
14967         if (rack_hw_up_only)
14968                 rack->r_up_only = 1;
14969         if (rack_do_dyn_mul) {
14970                 /* When dynamic adjustment is on CA needs to start at 100% */
14971                 rack->rc_gp_dyn_mul = 1;
14972                 if (rack_do_dyn_mul >= 100)
14973                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
14974         } else
14975                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
14976         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
14977         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
14978         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
14979         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
14980                                 rack_probertt_filter_life);
14981         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14982         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
14983         rack->r_ctl.rc_time_of_last_probertt = us_cts;
14984         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
14985         rack->r_ctl.rc_time_probertt_starts = 0;
14986         if (rack_dsack_std_based & 0x1) {
14987                 /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
14988                 rack->rc_rack_tmr_std_based = 1;
14989         }
14990         if (rack_dsack_std_based & 0x2) {
14991                 /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
14992                 rack->rc_rack_use_dsack = 1;
14993         }
14994         /* We require at least one measurement, even if the sysctl is 0 */
14995         if (rack_req_measurements)
14996                 rack->r_ctl.req_measurements = rack_req_measurements;
14997         else
14998                 rack->r_ctl.req_measurements = 1;
14999         if (rack_enable_hw_pacing)
15000                 rack->rack_hdw_pace_ena = 1;
15001         if (rack_hw_rate_caps)
15002                 rack->r_rack_hw_rate_caps = 1;
15003 #ifdef TCP_SAD_DETECTION
15004         rack->do_detection = 1;
15005 #else
15006         rack->do_detection = 0;
15007 #endif
15008         if (rack_non_rxt_use_cr)
15009                 rack->rack_rec_nonrxt_use_cr = 1;
15010         /* Lets setup the fsb block */
15011         err = rack_init_fsb(tp, rack);
15012         if (err) {
15013                 uma_zfree(rack_pcb_zone, *ptr);
15014                 *ptr = NULL;
15015                 return (err);
15016         }
15017         if (rack_do_hystart) {
15018                 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
15019                 if (rack_do_hystart > 1)
15020                         tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
15021                 if (rack_do_hystart > 2)
15022                         tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
15023         }
15024         /* Log what we will do with queries */
15025         rack_log_chg_info(tp, rack, 7,
15026                           no_query, 0, 0);
15027         if (rack_def_profile)
15028                 rack_set_profile(rack, rack_def_profile);
15029         /* Cancel the GP measurement in progress */
15030         tp->t_flags &= ~TF_GPUTINPROG;
15031         if ((tp->t_state != TCPS_CLOSED) &&
15032             (tp->t_state != TCPS_TIME_WAIT)) {
15033                 /*
15034                  * We are already open, we may
15035                  * need to adjust a few things.
15036                  */
15037                 if (SEQ_GT(tp->snd_max, tp->iss))
15038                         snt = tp->snd_max - tp->iss;
15039                 else
15040                         snt = 0;
15041                 iwin = rc_init_window(rack);
15042                 if ((snt < iwin) &&
15043                     (no_query == 1)) {
15044                         /* We are not past the initial window
15045                          * on the first init (i.e. a stack switch
15046                          * has not yet occured) so we need to make
15047                          * sure cwnd and ssthresh is correct.
15048                          */
15049                         if (tp->snd_cwnd < iwin)
15050                                 tp->snd_cwnd = iwin;
15051                         /*
15052                          * If we are within the initial window
15053                          * we want ssthresh to be unlimited. Setting
15054                          * it to the rwnd (which the default stack does
15055                          * and older racks) is not really a good idea
15056                          * since we want to be in SS and grow both the
15057                          * cwnd and the rwnd (via dynamic rwnd growth). If
15058                          * we set it to the rwnd then as the peer grows its
15059                          * rwnd we will be stuck in CA and never hit SS.
15060                          *
15061                          * Its far better to raise it up high (this takes the
15062                          * risk that there as been a loss already, probably
15063                          * we should have an indicator in all stacks of loss
15064                          * but we don't), but considering the normal use this
15065                          * is a risk worth taking. The consequences of not
15066                          * hitting SS are far worse than going one more time
15067                          * into it early on (before we have sent even a IW).
15068                          * It is highly unlikely that we will have had a loss
15069                          * before getting the IW out.
15070                          */
15071                         tp->snd_ssthresh = 0xffffffff;
15072                 }
15073                 /*
15074                  * Any init based on sequence numbers
15075                  * should be done in the deferred init path
15076                  * since we can be CLOSED and not have them
15077                  * inited when rack_init() is called. We
15078                  * are not closed so lets call it.
15079                  */
15080                 rack_deferred_init(tp, rack);
15081         }
15082         if ((tp->t_state != TCPS_CLOSED) &&
15083             (tp->t_state != TCPS_TIME_WAIT) &&
15084             (no_query == 0) &&
15085             (tp->snd_una != tp->snd_max))  {
15086                 err = rack_init_outstanding(tp, rack, us_cts, *ptr);
15087                 if (err) {
15088                         *ptr = NULL;
15089                         return(err);
15090                 }
15091         }
15092         rack_stop_all_timers(tp, rack);
15093         /* Setup all the t_flags2 */
15094         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
15095                 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
15096         else
15097                 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
15098         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15099                 tp->t_flags2 |= TF2_MBUF_ACKCMP;
15100         /*
15101          * Timers in Rack are kept in microseconds so lets
15102          * convert any initial incoming variables
15103          * from ticks into usecs. Note that we
15104          * also change the values of t_srtt and t_rttvar, if
15105          * they are non-zero. They are kept with a 5
15106          * bit decimal so we have to carefully convert
15107          * these to get the full precision.
15108          */
15109         rack_convert_rtts(tp);
15110         rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
15111         if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
15112                 /* We do not start any timers on DROPPED connections */
15113                 if (tp->t_fb->tfb_chg_query == NULL) {
15114                         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15115                 } else {
15116                         struct tcp_query_resp qr;
15117                         int ret;
15118
15119                         memset(&qr, 0, sizeof(qr));
15120
15121                         /* Get the misc time stamps and such for rack */
15122                         qr.req = TCP_QUERY_RACK_TIMES;
15123                         ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15124                         if (ret == 1) {
15125                                 rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
15126                                 rack->r_ctl.num_dsack  = qr.rack_num_dsacks;
15127                                 rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
15128                                 rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
15129                                 rack->rc_rack_rtt = qr.rack_rtt;
15130                                 rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
15131                                 rack->r_ctl.rc_sacked = qr.rack_sacked;
15132                                 rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
15133                                 rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
15134                                 rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
15135                                 rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
15136                                 rack->r_ctl.rc_prr_out = qr.rack_prr_out;
15137                                 if (qr.rack_tlp_out) {
15138                                         rack->rc_tlp_in_progress = 1;
15139                                         rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
15140                                 } else {
15141                                         rack->rc_tlp_in_progress = 0;
15142                                         rack->r_ctl.rc_tlp_cnt_out = 0;
15143                                 }
15144                                 if (qr.rack_srtt_measured)
15145                                         rack->rc_srtt_measure_made = 1;
15146                                 if (qr.rack_in_persist == 1) {
15147                                         rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
15148 #ifdef NETFLIX_SHARED_CWND
15149                                         if (rack->r_ctl.rc_scw) {
15150                                                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
15151                                                 rack->rack_scwnd_is_idle = 1;
15152                                         }
15153 #endif
15154                                         rack->r_ctl.persist_lost_ends = 0;
15155                                         rack->probe_not_answered = 0;
15156                                         rack->forced_ack = 0;
15157                                         tp->t_rxtshift = 0;
15158                                         rack->rc_in_persist = 1;
15159                                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
15160                                                            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
15161                                 }
15162                                 if (qr.rack_wanted_output)
15163                                         rack->r_wanted_output = 1;
15164                                 rack_log_chg_info(tp, rack, 6,
15165                                                   qr.rack_min_rtt,
15166                                                   qr.rack_rtt,
15167                                                   qr.rack_reorder_ts);
15168                         }
15169                         /* Get the old stack timers */
15170                         qr.req_param = 0;
15171                         qr.req = TCP_QUERY_TIMERS_UP;
15172                         ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15173                         if (ret) {
15174                                 /*
15175                                  * non-zero return means we have a timer('s)
15176                                  * to start. Zero means no timer (no keepalive
15177                                  * I suppose).
15178                                  */
15179                                 uint32_t tov = 0;
15180
15181                                 rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
15182                                 if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
15183                                         rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
15184                                         if (TSTMP_GT(qr.timer_pacing_to, us_cts))
15185                                                 tov = qr.timer_pacing_to - us_cts;
15186                                         else
15187                                                 tov = HPTS_TICKS_PER_SLOT;
15188                                 }
15189                                 if (qr.timer_hpts_flags & PACE_TMR_MASK) {
15190                                         rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
15191                                         if (tov == 0) {
15192                                                 if (TSTMP_GT(qr.timer_timer_exp, us_cts))
15193                                                         tov = qr.timer_timer_exp - us_cts;
15194                                                 else
15195                                                         tov = HPTS_TICKS_PER_SLOT;
15196                                         }
15197                                 }
15198                                 rack_log_chg_info(tp, rack, 4,
15199                                                   rack->r_ctl.rc_hpts_flags,
15200                                                   rack->r_ctl.rc_last_output_to,
15201                                                   rack->r_ctl.rc_timer_exp);
15202                                 if (tov) {
15203                                         struct hpts_diag diag;
15204
15205                                         (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
15206                                                                    __LINE__, &diag);
15207                                         rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
15208                                 }
15209                         }
15210                 }
15211                 rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
15212                                      __LINE__, RACK_RTTS_INIT);
15213         }
15214         return (0);
15215 }
15216
15217 static int
15218 rack_handoff_ok(struct tcpcb *tp)
15219 {
15220         if ((tp->t_state == TCPS_CLOSED) ||
15221             (tp->t_state == TCPS_LISTEN)) {
15222                 /* Sure no problem though it may not stick */
15223                 return (0);
15224         }
15225         if ((tp->t_state == TCPS_SYN_SENT) ||
15226             (tp->t_state == TCPS_SYN_RECEIVED)) {
15227                 /*
15228                  * We really don't know if you support sack,
15229                  * you have to get to ESTAB or beyond to tell.
15230                  */
15231                 return (EAGAIN);
15232         }
15233         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15234                 /*
15235                  * Rack will only send a FIN after all data is acknowledged.
15236                  * So in this case we have more data outstanding. We can't
15237                  * switch stacks until either all data and only the FIN
15238                  * is left (in which case rack_init() now knows how
15239                  * to deal with that) <or> all is acknowledged and we
15240                  * are only left with incoming data, though why you
15241                  * would want to switch to rack after all data is acknowledged
15242                  * I have no idea (rrs)!
15243                  */
15244                 return (EAGAIN);
15245         }
15246         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15247                 return (0);
15248         }
15249         /*
15250          * If we reach here we don't do SACK on this connection so we can
15251          * never do rack.
15252          */
15253         return (EINVAL);
15254 }
15255
15256 static void
15257 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15258 {
15259
15260         if (tp->t_fb_ptr) {
15261                 uint32_t cnt_free = 0;
15262                 struct tcp_rack *rack;
15263                 struct rack_sendmap *rsm;
15264
15265                 tcp_handle_orphaned_packets(tp);
15266                 tp->t_flags &= ~TF_FORCEDATA;
15267                 rack = (struct tcp_rack *)tp->t_fb_ptr;
15268                 rack_log_pacing_delay_calc(rack,
15269                                            0,
15270                                            0,
15271                                            0,
15272                                            rack_get_gp_est(rack), /* delRate */
15273                                            rack_get_lt_bw(rack), /* rttProp */
15274                                            20, __LINE__, NULL, 0);
15275 #ifdef NETFLIX_SHARED_CWND
15276                 if (rack->r_ctl.rc_scw) {
15277                         uint32_t limit;
15278
15279                         if (rack->r_limit_scw)
15280                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15281                         else
15282                                 limit = 0;
15283                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15284                                                   rack->r_ctl.rc_scw_index,
15285                                                   limit);
15286                         rack->r_ctl.rc_scw = NULL;
15287                 }
15288 #endif
15289                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
15290                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15291                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15292                         rack->r_ctl.fsb.th = NULL;
15293                 }
15294                 if (rack->rc_always_pace) {
15295                         tcp_decrement_paced_conn();
15296                         rack_undo_cc_pacing(rack);
15297                         rack->rc_always_pace = 0;
15298                 }
15299                 /* Clean up any options if they were not applied */
15300                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15301                         struct deferred_opt_list *dol;
15302
15303                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15304                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15305                         free(dol, M_TCPDO);
15306                 }
15307                 /* rack does not use force data but other stacks may clear it */
15308                 if (rack->r_ctl.crte != NULL) {
15309                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15310                         rack->rack_hdrw_pacing = 0;
15311                         rack->r_ctl.crte = NULL;
15312                 }
15313 #ifdef TCP_BLACKBOX
15314                 tcp_log_flowend(tp);
15315 #endif
15316                 /*
15317                  * Lets take a different approach to purging just
15318                  * get each one and free it like a cum-ack would and
15319                  * not use a foreach loop.
15320                  */
15321                 rsm = tqhash_min(rack->r_ctl.tqh);
15322                 while (rsm) {
15323                         tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15324                         rack->r_ctl.rc_num_maps_alloced--;
15325                         uma_zfree(rack_zone, rsm);
15326                         rsm = tqhash_min(rack->r_ctl.tqh);
15327                 }
15328                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15329                 while (rsm) {
15330                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15331                         rack->r_ctl.rc_num_maps_alloced--;
15332                         rack->rc_free_cnt--;
15333                         cnt_free++;
15334                         uma_zfree(rack_zone, rsm);
15335                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15336                 }
15337                 if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15338                     (tcp_bblogging_on(tp))) {
15339                         union tcp_log_stackspecific log;
15340                         struct timeval tv;
15341
15342                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15343                         log.u_bbr.flex8 = 10;
15344                         log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15345                         log.u_bbr.flex2 = rack->rc_free_cnt;
15346                         log.u_bbr.flex3 = cnt_free;
15347                         log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15348                         rsm = tqhash_min(rack->r_ctl.tqh);
15349                         log.u_bbr.delRate = (uint64_t)rsm;
15350                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15351                         log.u_bbr.cur_del_rate = (uint64_t)rsm;
15352                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15353                         log.u_bbr.pkt_epoch = __LINE__;
15354                         (void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15355                                              0, &log, false, NULL, NULL, 0, &tv);
15356                 }
15357                 KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15358                         ("rack:%p num_aloc:%u after freeing all?",
15359                          rack,
15360                          rack->r_ctl.rc_num_maps_alloced));
15361                 rack->rc_free_cnt = 0;
15362                 free(rack->r_ctl.tqh, M_TCPFSB);
15363                 rack->r_ctl.tqh = NULL;
15364                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15365                 tp->t_fb_ptr = NULL;
15366         }
15367         /* Make sure snd_nxt is correctly set */
15368         tp->snd_nxt = tp->snd_max;
15369 }
15370
15371 static void
15372 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15373 {
15374         if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15375                 rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15376         }
15377         switch (tp->t_state) {
15378         case TCPS_SYN_SENT:
15379                 rack->r_state = TCPS_SYN_SENT;
15380                 rack->r_substate = rack_do_syn_sent;
15381                 break;
15382         case TCPS_SYN_RECEIVED:
15383                 rack->r_state = TCPS_SYN_RECEIVED;
15384                 rack->r_substate = rack_do_syn_recv;
15385                 break;
15386         case TCPS_ESTABLISHED:
15387                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15388                 rack->r_state = TCPS_ESTABLISHED;
15389                 rack->r_substate = rack_do_established;
15390                 break;
15391         case TCPS_CLOSE_WAIT:
15392                 rack->r_state = TCPS_CLOSE_WAIT;
15393                 rack->r_substate = rack_do_close_wait;
15394                 break;
15395         case TCPS_FIN_WAIT_1:
15396                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15397                 rack->r_state = TCPS_FIN_WAIT_1;
15398                 rack->r_substate = rack_do_fin_wait_1;
15399                 break;
15400         case TCPS_CLOSING:
15401                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15402                 rack->r_state = TCPS_CLOSING;
15403                 rack->r_substate = rack_do_closing;
15404                 break;
15405         case TCPS_LAST_ACK:
15406                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15407                 rack->r_state = TCPS_LAST_ACK;
15408                 rack->r_substate = rack_do_lastack;
15409                 break;
15410         case TCPS_FIN_WAIT_2:
15411                 rack->r_state = TCPS_FIN_WAIT_2;
15412                 rack->r_substate = rack_do_fin_wait_2;
15413                 break;
15414         case TCPS_LISTEN:
15415         case TCPS_CLOSED:
15416         case TCPS_TIME_WAIT:
15417         default:
15418                 break;
15419         };
15420         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15421                 rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
15422
15423 }
15424
15425 static void
15426 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15427 {
15428         /*
15429          * We received an ack, and then did not
15430          * call send or were bounced out due to the
15431          * hpts was running. Now a timer is up as well, is
15432          * it the right timer?
15433          */
15434         struct rack_sendmap *rsm;
15435         int tmr_up;
15436
15437         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15438         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15439                 return;
15440         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15441         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15442             (tmr_up == PACE_TMR_RXT)) {
15443                 /* Should be an RXT */
15444                 return;
15445         }
15446         if (rsm == NULL) {
15447                 /* Nothing outstanding? */
15448                 if (tp->t_flags & TF_DELACK) {
15449                         if (tmr_up == PACE_TMR_DELACK)
15450                                 /* We are supposed to have delayed ack up and we do */
15451                                 return;
15452                 } else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
15453                         /*
15454                          * if we hit enobufs then we would expect the possibility
15455                          * of nothing outstanding and the RXT up (and the hptsi timer).
15456                          */
15457                         return;
15458                 } else if (((V_tcp_always_keepalive ||
15459                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
15460                             (tp->t_state <= TCPS_CLOSING)) &&
15461                            (tmr_up == PACE_TMR_KEEP) &&
15462                            (tp->snd_max == tp->snd_una)) {
15463                         /* We should have keep alive up and we do */
15464                         return;
15465                 }
15466         }
15467         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
15468                    ((tmr_up == PACE_TMR_TLP) ||
15469                     (tmr_up == PACE_TMR_RACK) ||
15470                     (tmr_up == PACE_TMR_RXT))) {
15471                 /*
15472                  * Either a Rack, TLP or RXT is fine if  we
15473                  * have outstanding data.
15474                  */
15475                 return;
15476         } else if (tmr_up == PACE_TMR_DELACK) {
15477                 /*
15478                  * If the delayed ack was going to go off
15479                  * before the rtx/tlp/rack timer were going to
15480                  * expire, then that would be the timer in control.
15481                  * Note we don't check the time here trusting the
15482                  * code is correct.
15483                  */
15484                 return;
15485         }
15486         /*
15487          * Ok the timer originally started is not what we want now.
15488          * We will force the hpts to be stopped if any, and restart
15489          * with the slot set to what was in the saved slot.
15490          */
15491         if (tcp_in_hpts(rack->rc_tp)) {
15492                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15493                         uint32_t us_cts;
15494
15495                         us_cts = tcp_get_usecs(NULL);
15496                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
15497                                 rack->r_early = 1;
15498                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
15499                         }
15500                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
15501                 }
15502                 tcp_hpts_remove(rack->rc_tp);
15503         }
15504         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15505         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15506 }
15507
15508
15509 static void
15510 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
15511 {
15512         if ((SEQ_LT(tp->snd_wl1, seq) ||
15513             (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
15514             (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
15515                 /* keep track of pure window updates */
15516                 if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
15517                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
15518                 tp->snd_wnd = tiwin;
15519                 rack_validate_fo_sendwin_up(tp, rack);
15520                 tp->snd_wl1 = seq;
15521                 tp->snd_wl2 = ack;
15522                 if (tp->snd_wnd > tp->max_sndwnd)
15523                         tp->max_sndwnd = tp->snd_wnd;
15524             rack->r_wanted_output = 1;
15525         } else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
15526                 tp->snd_wnd = tiwin;
15527                 rack_validate_fo_sendwin_up(tp, rack);
15528                 tp->snd_wl1 = seq;
15529                 tp->snd_wl2 = ack;
15530         } else {
15531                 /* Not a valid win update */
15532                 return;
15533         }
15534         if (tp->snd_wnd > tp->max_sndwnd)
15535                 tp->max_sndwnd = tp->snd_wnd;
15536         /* Do we exit persists? */
15537         if ((rack->rc_in_persist != 0) &&
15538             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
15539                                 rack->r_ctl.rc_pace_min_segs))) {
15540                 rack_exit_persist(tp, rack, cts);
15541         }
15542         /* Do we enter persists? */
15543         if ((rack->rc_in_persist == 0) &&
15544             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
15545             TCPS_HAVEESTABLISHED(tp->t_state) &&
15546             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
15547             sbavail(&tptosocket(tp)->so_snd) &&
15548             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
15549                 /*
15550                  * Here the rwnd is less than
15551                  * the pacing size, we are established,
15552                  * nothing is outstanding, and there is
15553                  * data to send. Enter persists.
15554                  */
15555                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
15556         }
15557 }
15558
15559 static void
15560 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
15561 {
15562
15563         if (tcp_bblogging_on(rack->rc_tp)) {
15564                 struct inpcb *inp = tptoinpcb(tp);
15565                 union tcp_log_stackspecific log;
15566                 struct timeval ltv;
15567                 char tcp_hdr_buf[60];
15568                 struct tcphdr *th;
15569                 struct timespec ts;
15570                 uint32_t orig_snd_una;
15571                 uint8_t xx = 0;
15572
15573 #ifdef TCP_REQUEST_TRK
15574                 struct tcp_sendfile_track *tcp_req;
15575
15576                 if (SEQ_GT(ae->ack, tp->snd_una)) {
15577                         tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
15578                 } else {
15579                         tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
15580                 }
15581 #endif
15582                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15583                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
15584                 if (rack->rack_no_prr == 0)
15585                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15586                 else
15587                         log.u_bbr.flex1 = 0;
15588                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
15589                 log.u_bbr.use_lt_bw <<= 1;
15590                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
15591                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
15592                 log.u_bbr.bbr_state = rack->rc_free_cnt;
15593                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15594                 log.u_bbr.pkts_out = tp->t_maxseg;
15595                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
15596                 log.u_bbr.flex7 = 1;
15597                 log.u_bbr.lost = ae->flags;
15598                 log.u_bbr.cwnd_gain = ackval;
15599                 log.u_bbr.pacing_gain = 0x2;
15600                 if (ae->flags & TSTMP_HDWR) {
15601                         /* Record the hardware timestamp if present */
15602                         log.u_bbr.flex3 = M_TSTMP;
15603                         ts.tv_sec = ae->timestamp / 1000000000;
15604                         ts.tv_nsec = ae->timestamp % 1000000000;
15605                         ltv.tv_sec = ts.tv_sec;
15606                         ltv.tv_usec = ts.tv_nsec / 1000;
15607                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
15608                 } else if (ae->flags & TSTMP_LRO) {
15609                         /* Record the LRO the arrival timestamp */
15610                         log.u_bbr.flex3 = M_TSTMP_LRO;
15611                         ts.tv_sec = ae->timestamp / 1000000000;
15612                         ts.tv_nsec = ae->timestamp % 1000000000;
15613                         ltv.tv_sec = ts.tv_sec;
15614                         ltv.tv_usec = ts.tv_nsec / 1000;
15615                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
15616                 }
15617                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
15618                 /* Log the rcv time */
15619                 log.u_bbr.delRate = ae->timestamp;
15620 #ifdef TCP_REQUEST_TRK
15621                 log.u_bbr.applimited = tp->t_tcpreq_closed;
15622                 log.u_bbr.applimited <<= 8;
15623                 log.u_bbr.applimited |= tp->t_tcpreq_open;
15624                 log.u_bbr.applimited <<= 8;
15625                 log.u_bbr.applimited |= tp->t_tcpreq_req;
15626                 if (tcp_req) {
15627                         /* Copy out any client req info */
15628                         /* seconds */
15629                         log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
15630                         /* useconds */
15631                         log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
15632                         log.u_bbr.rttProp = tcp_req->timestamp;
15633                         log.u_bbr.cur_del_rate = tcp_req->start;
15634                         if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
15635                                 log.u_bbr.flex8 |= 1;
15636                         } else {
15637                                 log.u_bbr.flex8 |= 2;
15638                                 log.u_bbr.bw_inuse = tcp_req->end;
15639                         }
15640                         log.u_bbr.flex6 = tcp_req->start_seq;
15641                         if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
15642                                 log.u_bbr.flex8 |= 4;
15643                                 log.u_bbr.epoch = tcp_req->end_seq;
15644                         }
15645                 }
15646 #endif
15647                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
15648                 th = (struct tcphdr *)tcp_hdr_buf;
15649                 th->th_seq = ae->seq;
15650                 th->th_ack = ae->ack;
15651                 th->th_win = ae->win;
15652                 /* Now fill in the ports */
15653                 th->th_sport = inp->inp_fport;
15654                 th->th_dport = inp->inp_lport;
15655                 tcp_set_flags(th, ae->flags);
15656                 /* Now do we have a timestamp option? */
15657                 if (ae->flags & HAS_TSTMP) {
15658                         u_char *cp;
15659                         uint32_t val;
15660
15661                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
15662                         cp = (u_char *)(th + 1);
15663                         *cp = TCPOPT_NOP;
15664                         cp++;
15665                         *cp = TCPOPT_NOP;
15666                         cp++;
15667                         *cp = TCPOPT_TIMESTAMP;
15668                         cp++;
15669                         *cp = TCPOLEN_TIMESTAMP;
15670                         cp++;
15671                         val = htonl(ae->ts_value);
15672                         bcopy((char *)&val,
15673                               (char *)cp, sizeof(uint32_t));
15674                         val = htonl(ae->ts_echo);
15675                         bcopy((char *)&val,
15676                               (char *)(cp + 4), sizeof(uint32_t));
15677                 } else
15678                         th->th_off = (sizeof(struct tcphdr) >> 2);
15679
15680                 /*
15681                  * For sane logging we need to play a little trick.
15682                  * If the ack were fully processed we would have moved
15683                  * snd_una to high_seq, but since compressed acks are
15684                  * processed in two phases, at this point (logging) snd_una
15685                  * won't be advanced. So we would see multiple acks showing
15686                  * the advancement. We can prevent that by "pretending" that
15687                  * snd_una was advanced and then un-advancing it so that the
15688                  * logging code has the right value for tlb_snd_una.
15689                  */
15690                 if (tp->snd_una != high_seq) {
15691                         orig_snd_una = tp->snd_una;
15692                         tp->snd_una = high_seq;
15693                         xx = 1;
15694                 } else
15695                         xx = 0;
15696                 TCP_LOG_EVENTP(tp, th,
15697                                &tptosocket(tp)->so_rcv,
15698                                &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
15699                                0, &log, true, &ltv);
15700                 if (xx) {
15701                         tp->snd_una = orig_snd_una;
15702                 }
15703         }
15704
15705 }
15706
15707 static void
15708 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
15709 {
15710         uint32_t us_rtt;
15711         /*
15712          * A persist or keep-alive was forced out, update our
15713          * min rtt time. Note now worry about lost responses.
15714          * When a subsequent keep-alive or persist times out
15715          * and forced_ack is still on, then the last probe
15716          * was not responded to. In such cases we have a
15717          * sysctl that controls the behavior. Either we apply
15718          * the rtt but with reduced confidence (0). Or we just
15719          * plain don't apply the rtt estimate. Having data flow
15720          * will clear the probe_not_answered flag i.e. cum-ack
15721          * move forward <or> exiting and reentering persists.
15722          */
15723
15724         rack->forced_ack = 0;
15725         rack->rc_tp->t_rxtshift = 0;
15726         if ((rack->rc_in_persist &&
15727              (tiwin == rack->rc_tp->snd_wnd)) ||
15728             (rack->rc_in_persist == 0)) {
15729                 /*
15730                  * In persists only apply the RTT update if this is
15731                  * a response to our window probe. And that
15732                  * means the rwnd sent must match the current
15733                  * snd_wnd. If it does not, then we got a
15734                  * window update ack instead. For keepalive
15735                  * we allow the answer no matter what the window.
15736                  *
15737                  * Note that if the probe_not_answered is set then
15738                  * the forced_ack_ts is the oldest one i.e. the first
15739                  * probe sent that might have been lost. This assures
15740                  * us that if we do calculate an RTT it is longer not
15741                  * some short thing.
15742                  */
15743                 if (rack->rc_in_persist)
15744                         counter_u64_add(rack_persists_acks, 1);
15745                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
15746                 if (us_rtt == 0)
15747                         us_rtt = 1;
15748                 if (rack->probe_not_answered == 0) {
15749                         rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15750                         tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
15751                 } else {
15752                         /* We have a retransmitted probe here too */
15753                         if (rack_apply_rtt_with_reduced_conf) {
15754                                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15755                                 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
15756                         }
15757                 }
15758         }
15759 }
15760
15761 static int
15762 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
15763 {
15764         /*
15765          * Handle a "special" compressed ack mbuf. Each incoming
15766          * ack has only four possible dispositions:
15767          *
15768          * A) It moves the cum-ack forward
15769          * B) It is behind the cum-ack.
15770          * C) It is a window-update ack.
15771          * D) It is a dup-ack.
15772          *
15773          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
15774          * in the incoming mbuf. We also need to still pay attention
15775          * to nxt_pkt since there may be another packet after this
15776          * one.
15777          */
15778 #ifdef TCP_ACCOUNTING
15779         uint64_t ts_val;
15780         uint64_t rdstc;
15781 #endif
15782         int segsiz;
15783         struct timespec ts;
15784         struct tcp_rack *rack;
15785         struct tcp_ackent *ae;
15786         uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
15787         int cnt, i, did_out, ourfinisacked = 0;
15788         struct tcpopt to_holder, *to = NULL;
15789 #ifdef TCP_ACCOUNTING
15790         int win_up_req = 0;
15791 #endif
15792         int nsegs = 0;
15793         int under_pacing = 0;
15794         int recovery = 0;
15795 #ifdef TCP_ACCOUNTING
15796         sched_pin();
15797 #endif
15798         rack = (struct tcp_rack *)tp->t_fb_ptr;
15799         if (rack->gp_ready &&
15800             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
15801                 under_pacing = 1;
15802
15803         if (rack->r_state != tp->t_state)
15804                 rack_set_state(tp, rack);
15805         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
15806             (tp->t_flags & TF_GPUTINPROG)) {
15807                 /*
15808                  * We have a goodput in progress
15809                  * and we have entered a late state.
15810                  * Do we have enough data in the sb
15811                  * to handle the GPUT request?
15812                  */
15813                 uint32_t bytes;
15814
15815                 bytes = tp->gput_ack - tp->gput_seq;
15816                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
15817                         bytes += tp->gput_seq - tp->snd_una;
15818                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
15819                         /*
15820                          * There are not enough bytes in the socket
15821                          * buffer that have been sent to cover this
15822                          * measurement. Cancel it.
15823                          */
15824                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
15825                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
15826                                                    tp->gput_seq,
15827                                                    0, 0, 18, __LINE__, NULL, 0);
15828                         tp->t_flags &= ~TF_GPUTINPROG;
15829                 }
15830         }
15831         to = &to_holder;
15832         to->to_flags = 0;
15833         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
15834                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
15835         cnt = m->m_len / sizeof(struct tcp_ackent);
15836         counter_u64_add(rack_multi_single_eq, cnt);
15837         high_seq = tp->snd_una;
15838         the_win = tp->snd_wnd;
15839         win_seq = tp->snd_wl1;
15840         win_upd_ack = tp->snd_wl2;
15841         cts = tcp_tv_to_usectick(tv);
15842         ms_cts = tcp_tv_to_mssectick(tv);
15843         rack->r_ctl.rc_rcvtime = cts;
15844         segsiz = ctf_fixed_maxseg(tp);
15845         if ((rack->rc_gp_dyn_mul) &&
15846             (rack->use_fixed_rate == 0) &&
15847             (rack->rc_always_pace)) {
15848                 /* Check in on probertt */
15849                 rack_check_probe_rtt(rack, cts);
15850         }
15851         for (i = 0; i < cnt; i++) {
15852 #ifdef TCP_ACCOUNTING
15853                 ts_val = get_cyclecount();
15854 #endif
15855                 rack_clear_rate_sample(rack);
15856                 ae = ((mtod(m, struct tcp_ackent *)) + i);
15857                 if (ae->flags & TH_FIN)
15858                         rack_log_pacing_delay_calc(rack,
15859                                                    0,
15860                                                    0,
15861                                                    0,
15862                                                    rack_get_gp_est(rack), /* delRate */
15863                                                    rack_get_lt_bw(rack), /* rttProp */
15864                                                    20, __LINE__, NULL, 0);
15865                 /* Setup the window */
15866                 tiwin = ae->win << tp->snd_scale;
15867                 if (tiwin > rack->r_ctl.rc_high_rwnd)
15868                         rack->r_ctl.rc_high_rwnd = tiwin;
15869                 /* figure out the type of ack */
15870                 if (SEQ_LT(ae->ack, high_seq)) {
15871                         /* Case B*/
15872                         ae->ack_val_set = ACK_BEHIND;
15873                 } else if (SEQ_GT(ae->ack, high_seq)) {
15874                         /* Case A */
15875                         ae->ack_val_set = ACK_CUMACK;
15876                 } else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
15877                         /* Case D */
15878                         ae->ack_val_set = ACK_DUPACK;
15879                 } else {
15880                         /* Case C */
15881                         ae->ack_val_set = ACK_RWND;
15882                 }
15883                 if (rack->sack_attack_disable > 0) {
15884                         rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
15885                         rack->r_ctl.ack_during_sd++;
15886                 }
15887                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
15888                 /* Validate timestamp */
15889                 if (ae->flags & HAS_TSTMP) {
15890                         /* Setup for a timestamp */
15891                         to->to_flags = TOF_TS;
15892                         ae->ts_echo -= tp->ts_offset;
15893                         to->to_tsecr = ae->ts_echo;
15894                         to->to_tsval = ae->ts_value;
15895                         /*
15896                          * If echoed timestamp is later than the current time, fall back to
15897                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
15898                          * were used when this connection was established.
15899                          */
15900                         if (TSTMP_GT(ae->ts_echo, ms_cts))
15901                                 to->to_tsecr = 0;
15902                         if (tp->ts_recent &&
15903                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
15904                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
15905 #ifdef TCP_ACCOUNTING
15906                                         rdstc = get_cyclecount();
15907                                         if (rdstc > ts_val) {
15908                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15909                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15910                                                 }
15911                                         }
15912 #endif
15913                                         continue;
15914                                 }
15915                         }
15916                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
15917                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
15918                                 tp->ts_recent_age = tcp_ts_getticks();
15919                                 tp->ts_recent = ae->ts_value;
15920                         }
15921                 } else {
15922                         /* Setup for a no options */
15923                         to->to_flags = 0;
15924                 }
15925                 /* Update the rcv time and perform idle reduction possibly */
15926                 if  (tp->t_idle_reduce &&
15927                      (tp->snd_max == tp->snd_una) &&
15928                      (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
15929                         counter_u64_add(rack_input_idle_reduces, 1);
15930                         rack_cc_after_idle(rack, tp);
15931                 }
15932                 tp->t_rcvtime = ticks;
15933                 /* Now what about ECN of a chain of pure ACKs? */
15934                 if (tcp_ecn_input_segment(tp, ae->flags, 0,
15935                         tcp_packets_this_ack(tp, ae->ack),
15936                         ae->codepoint))
15937                         rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
15938 #ifdef TCP_ACCOUNTING
15939                 /* Count for the specific type of ack in */
15940                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15941                         tp->tcp_cnt_counters[ae->ack_val_set]++;
15942                 }
15943 #endif
15944                 /*
15945                  * Note how we could move up these in the determination
15946                  * above, but we don't so that way the timestamp checks (and ECN)
15947                  * is done first before we do any processing on the ACK.
15948                  * The non-compressed path through the code has this
15949                  * weakness (noted by @jtl) that it actually does some
15950                  * processing before verifying the timestamp information.
15951                  * We don't take that path here which is why we set
15952                  * the ack_val_set first, do the timestamp and ecn
15953                  * processing, and then look at what we have setup.
15954                  */
15955                 if (ae->ack_val_set == ACK_BEHIND) {
15956                         /*
15957                          * Case B flag reordering, if window is not closed
15958                          * or it could be a keep-alive or persists
15959                          */
15960                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
15961                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15962                                 if (rack->r_ctl.rc_reorder_ts == 0)
15963                                         rack->r_ctl.rc_reorder_ts = 1;
15964                         }
15965                 } else if (ae->ack_val_set == ACK_DUPACK) {
15966                         /* Case D */
15967                         rack_strike_dupack(rack);
15968                 } else if (ae->ack_val_set == ACK_RWND) {
15969                         /* Case C */
15970                         if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15971                                 ts.tv_sec = ae->timestamp / 1000000000;
15972                                 ts.tv_nsec = ae->timestamp % 1000000000;
15973                                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15974                                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15975                         } else {
15976                                 rack->r_ctl.act_rcv_time = *tv;
15977                         }
15978                         if (rack->forced_ack) {
15979                                 rack_handle_probe_response(rack, tiwin,
15980                                                            tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
15981                         }
15982 #ifdef TCP_ACCOUNTING
15983                         win_up_req = 1;
15984 #endif
15985                         win_upd_ack = ae->ack;
15986                         win_seq = ae->seq;
15987                         the_win = tiwin;
15988                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15989                 } else {
15990                         /* Case A */
15991                         if (SEQ_GT(ae->ack, tp->snd_max)) {
15992                                 /*
15993                                  * We just send an ack since the incoming
15994                                  * ack is beyond the largest seq we sent.
15995                                  */
15996                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
15997                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
15998                                         if (tp->t_flags && TF_ACKNOW)
15999                                                 rack->r_wanted_output = 1;
16000                                 }
16001                         } else {
16002                                 nsegs++;
16003                                 /* If the window changed setup to update */
16004                                 if (tiwin != tp->snd_wnd) {
16005                                         win_upd_ack = ae->ack;
16006                                         win_seq = ae->seq;
16007                                         the_win = tiwin;
16008                                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
16009                                 }
16010 #ifdef TCP_ACCOUNTING
16011                                 /* Account for the acks */
16012                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16013                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
16014                                 }
16015 #endif
16016                                 high_seq = ae->ack;
16017                                 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp))
16018                                         rack_log_hystart_event(rack, high_seq, 8);
16019                                 /* Setup our act_rcv_time */
16020                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
16021                                         ts.tv_sec = ae->timestamp / 1000000000;
16022                                         ts.tv_nsec = ae->timestamp % 1000000000;
16023                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16024                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16025                                 } else {
16026                                         rack->r_ctl.act_rcv_time = *tv;
16027                                 }
16028                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to,
16029                                                        tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
16030 #ifdef TCP_REQUEST_TRK
16031                                 rack_req_check_for_comp(rack, high_seq);
16032 #endif
16033                                 if (rack->rc_dsack_round_seen) {
16034                                         /* Is the dsack round over? */
16035                                         if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
16036                                                 /* Yes it is */
16037                                                 rack->rc_dsack_round_seen = 0;
16038                                                 rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
16039                                         }
16040                                 }
16041                         }
16042                 }
16043                 /* And lets be sure to commit the rtt measurements for this ack */
16044                 tcp_rack_xmit_timer_commit(rack, tp);
16045 #ifdef TCP_ACCOUNTING
16046                 rdstc = get_cyclecount();
16047                 if (rdstc > ts_val) {
16048                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16049                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
16050                                 if (ae->ack_val_set == ACK_CUMACK)
16051                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
16052                         }
16053                 }
16054 #endif
16055         }
16056 #ifdef TCP_ACCOUNTING
16057         ts_val = get_cyclecount();
16058 #endif
16059         /* Tend to any collapsed window */
16060         if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
16061                 /* The peer collapsed the window */
16062                 rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
16063         } else if (rack->rc_has_collapsed)
16064                 rack_un_collapse_window(rack, __LINE__);
16065         if ((rack->r_collapse_point_valid) &&
16066             (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
16067                 rack->r_collapse_point_valid = 0;
16068         acked_amount = acked = (high_seq - tp->snd_una);
16069         if (acked) {
16070                 /*
16071                  * The draft (v3) calls for us to use SEQ_GEQ, but that
16072                  * causes issues when we are just going app limited. Lets
16073                  * instead use SEQ_GT <or> where its equal but more data
16074                  * is outstanding.
16075                  *
16076                  * Also make sure we are on the last ack of a series. We
16077                  * have to have all the ack's processed in queue to know
16078                  * if there is something left outstanding.
16079                  *
16080                  */
16081                 if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
16082                     (rack->rc_new_rnd_needed == 0) &&
16083                     (nxt_pkt == 0)) {
16084                         rack_log_hystart_event(rack, high_seq, 21);
16085                         rack->r_ctl.current_round++;
16086                         /* Force the next send to setup the next round */
16087                         rack->rc_new_rnd_needed = 1;
16088                         if (CC_ALGO(tp)->newround != NULL) {
16089                                 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
16090                         }
16091                 }
16092                 /*
16093                  * Clear the probe not answered flag
16094                  * since cum-ack moved forward.
16095                  */
16096                 rack->probe_not_answered = 0;
16097                 if (rack->sack_attack_disable == 0)
16098                         rack_do_decay(rack);
16099                 if (acked >= segsiz) {
16100                         /*
16101                          * You only get credit for
16102                          * MSS and greater (and you get extra
16103                          * credit for larger cum-ack moves).
16104                          */
16105                         int ac;
16106
16107                         ac = acked / segsiz;
16108                         rack->r_ctl.ack_count += ac;
16109                         counter_u64_add(rack_ack_total, ac);
16110                 }
16111                 if (rack->r_ctl.ack_count > 0xfff00000) {
16112                         /*
16113                          * reduce the number to keep us under
16114                          * a uint32_t.
16115                          */
16116                         rack->r_ctl.ack_count /= 2;
16117                         rack->r_ctl.sack_count /= 2;
16118                 }
16119                 if (tp->t_flags & TF_NEEDSYN) {
16120                         /*
16121                          * T/TCP: Connection was half-synchronized, and our SYN has
16122                          * been ACK'd (so connection is now fully synchronized).  Go
16123                          * to non-starred state, increment snd_una for ACK of SYN,
16124                          * and check if we can do window scaling.
16125                          */
16126                         tp->t_flags &= ~TF_NEEDSYN;
16127                         tp->snd_una++;
16128                         acked_amount = acked = (high_seq - tp->snd_una);
16129                 }
16130                 if (acked > sbavail(&so->so_snd))
16131                         acked_amount = sbavail(&so->so_snd);
16132 #ifdef TCP_SAD_DETECTION
16133                 /*
16134                  * We only care on a cum-ack move if we are in a sack-disabled
16135                  * state. We have already added in to the ack_count, and we never
16136                  * would disable on a cum-ack move, so we only care to do the
16137                  * detection if it may "undo" it, i.e. we were in disabled already.
16138                  */
16139                 if (rack->sack_attack_disable)
16140                         rack_do_detection(tp, rack, acked_amount, segsiz);
16141 #endif
16142                 if (IN_FASTRECOVERY(tp->t_flags) &&
16143                     (rack->rack_no_prr == 0))
16144                         rack_update_prr(tp, rack, acked_amount, high_seq);
16145                 if (IN_RECOVERY(tp->t_flags)) {
16146                         if (SEQ_LT(high_seq, tp->snd_recover) &&
16147                             (SEQ_LT(high_seq, tp->snd_max))) {
16148                                 tcp_rack_partialack(tp);
16149                         } else {
16150                                 rack_post_recovery(tp, high_seq);
16151                                 recovery = 1;
16152                         }
16153                 }
16154                 /* Handle the rack-log-ack part (sendmap) */
16155                 if ((sbused(&so->so_snd) == 0) &&
16156                     (acked > acked_amount) &&
16157                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
16158                     (tp->t_flags & TF_SENTFIN)) {
16159                         /*
16160                          * We must be sure our fin
16161                          * was sent and acked (we can be
16162                          * in FIN_WAIT_1 without having
16163                          * sent the fin).
16164                          */
16165                         ourfinisacked = 1;
16166                         /*
16167                          * Lets make sure snd_una is updated
16168                          * since most likely acked_amount = 0 (it
16169                          * should be).
16170                          */
16171                         tp->snd_una = high_seq;
16172                 }
16173                 /* Did we make a RTO error? */
16174                 if ((tp->t_flags & TF_PREVVALID) &&
16175                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16176                         tp->t_flags &= ~TF_PREVVALID;
16177                         if (tp->t_rxtshift == 1 &&
16178                             (int)(ticks - tp->t_badrxtwin) < 0)
16179                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16180                 }
16181                 /* Handle the data in the socket buffer */
16182                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16183                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16184                 if (acked_amount > 0) {
16185                         struct mbuf *mfree;
16186
16187                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
16188                         SOCKBUF_LOCK(&so->so_snd);
16189                         mfree = sbcut_locked(&so->so_snd, acked_amount);
16190                         tp->snd_una = high_seq;
16191                         /* Note we want to hold the sb lock through the sendmap adjust */
16192                         rack_adjust_sendmap_head(rack, &so->so_snd);
16193                         /* Wake up the socket if we have room to write more */
16194                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16195                         sowwakeup_locked(so);
16196                         if ((recovery == 1) &&
16197                             (rack->excess_rxt_on) &&
16198                             (rack->r_cwnd_was_clamped == 0)) {
16199                                 do_rack_excess_rxt(tp, rack);
16200                         } else if (rack->r_cwnd_was_clamped)
16201                                 do_rack_check_for_unclamp(tp, rack);
16202                         m_freem(mfree);
16203                 }
16204                 /* update progress */
16205                 tp->t_acktime = ticks;
16206                 rack_log_progress_event(rack, tp, tp->t_acktime,
16207                                         PROGRESS_UPDATE, __LINE__);
16208                 /* Clear out shifts and such */
16209                 tp->t_rxtshift = 0;
16210                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16211                                    rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16212                 rack->rc_tlp_in_progress = 0;
16213                 rack->r_ctl.rc_tlp_cnt_out = 0;
16214                 /* Send recover and snd_nxt must be dragged along */
16215                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
16216                         tp->snd_recover = tp->snd_una;
16217                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
16218                         tp->snd_nxt = tp->snd_una;
16219                 /*
16220                  * If the RXT timer is running we want to
16221                  * stop it, so we can restart a TLP (or new RXT).
16222                  */
16223                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16224                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16225                 tp->snd_wl2 = high_seq;
16226                 tp->t_dupacks = 0;
16227                 if (under_pacing &&
16228                     (rack->use_fixed_rate == 0) &&
16229                     (rack->in_probe_rtt == 0) &&
16230                     rack->rc_gp_dyn_mul &&
16231                     rack->rc_always_pace) {
16232                         /* Check if we are dragging bottom */
16233                         rack_check_bottom_drag(tp, rack, so);
16234                 }
16235                 if (tp->snd_una == tp->snd_max) {
16236                         tp->t_flags &= ~TF_PREVVALID;
16237                         rack->r_ctl.retran_during_recovery = 0;
16238                         rack->rc_suspicious = 0;
16239                         rack->r_ctl.dsack_byte_cnt = 0;
16240                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16241                         if (rack->r_ctl.rc_went_idle_time == 0)
16242                                 rack->r_ctl.rc_went_idle_time = 1;
16243                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16244                         if (sbavail(&tptosocket(tp)->so_snd) == 0)
16245                                 tp->t_acktime = 0;
16246                         /* Set so we might enter persists... */
16247                         rack->r_wanted_output = 1;
16248                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16249                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16250                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16251                             (sbavail(&so->so_snd) == 0) &&
16252                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16253                                 /*
16254                                  * The socket was gone and the
16255                                  * peer sent data (not now in the past), time to
16256                                  * reset him.
16257                                  */
16258                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16259                                 /* tcp_close will kill the inp pre-log the Reset */
16260                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16261 #ifdef TCP_ACCOUNTING
16262                                 rdstc = get_cyclecount();
16263                                 if (rdstc > ts_val) {
16264                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16265                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16266                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16267                                         }
16268                                 }
16269 #endif
16270                                 m_freem(m);
16271                                 tp = tcp_close(tp);
16272                                 if (tp == NULL) {
16273 #ifdef TCP_ACCOUNTING
16274                                         sched_unpin();
16275 #endif
16276                                         return (1);
16277                                 }
16278                                 /*
16279                                  * We would normally do drop-with-reset which would
16280                                  * send back a reset. We can't since we don't have
16281                                  * all the needed bits. Instead lets arrange for
16282                                  * a call to tcp_output(). That way since we
16283                                  * are in the closed state we will generate a reset.
16284                                  *
16285                                  * Note if tcp_accounting is on we don't unpin since
16286                                  * we do that after the goto label.
16287                                  */
16288                                 goto send_out_a_rst;
16289                         }
16290                         if ((sbused(&so->so_snd) == 0) &&
16291                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
16292                             (tp->t_flags & TF_SENTFIN)) {
16293                                 /*
16294                                  * If we can't receive any more data, then closing user can
16295                                  * proceed. Starting the timer is contrary to the
16296                                  * specification, but if we don't get a FIN we'll hang
16297                                  * forever.
16298                                  *
16299                                  */
16300                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16301                                         soisdisconnected(so);
16302                                         tcp_timer_activate(tp, TT_2MSL,
16303                                                            (tcp_fast_finwait2_recycle ?
16304                                                             tcp_finwait2_timeout :
16305                                                             TP_MAXIDLE(tp)));
16306                                 }
16307                                 if (ourfinisacked == 0) {
16308                                         /*
16309                                          * We don't change to fin-wait-2 if we have our fin acked
16310                                          * which means we are probably in TCPS_CLOSING.
16311                                          */
16312                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
16313                                 }
16314                         }
16315                 }
16316                 /* Wake up the socket if we have room to write more */
16317                 if (sbavail(&so->so_snd)) {
16318                         rack->r_wanted_output = 1;
16319                         if (ctf_progress_timeout_check(tp, true)) {
16320                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16321                                                         tp, tick, PROGRESS_DROP, __LINE__);
16322                                 /*
16323                                  * We cheat here and don't send a RST, we should send one
16324                                  * when the pacer drops the connection.
16325                                  */
16326 #ifdef TCP_ACCOUNTING
16327                                 rdstc = get_cyclecount();
16328                                 if (rdstc > ts_val) {
16329                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16330                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16331                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16332                                         }
16333                                 }
16334                                 sched_unpin();
16335 #endif
16336                                 (void)tcp_drop(tp, ETIMEDOUT);
16337                                 m_freem(m);
16338                                 return (1);
16339                         }
16340                 }
16341                 if (ourfinisacked) {
16342                         switch(tp->t_state) {
16343                         case TCPS_CLOSING:
16344 #ifdef TCP_ACCOUNTING
16345                                 rdstc = get_cyclecount();
16346                                 if (rdstc > ts_val) {
16347                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16348                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16349                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16350                                         }
16351                                 }
16352                                 sched_unpin();
16353 #endif
16354                                 tcp_twstart(tp);
16355                                 m_freem(m);
16356                                 return (1);
16357                                 break;
16358                         case TCPS_LAST_ACK:
16359 #ifdef TCP_ACCOUNTING
16360                                 rdstc = get_cyclecount();
16361                                 if (rdstc > ts_val) {
16362                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16363                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16364                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16365                                         }
16366                                 }
16367                                 sched_unpin();
16368 #endif
16369                                 tp = tcp_close(tp);
16370                                 ctf_do_drop(m, tp);
16371                                 return (1);
16372                                 break;
16373                         case TCPS_FIN_WAIT_1:
16374 #ifdef TCP_ACCOUNTING
16375                                 rdstc = get_cyclecount();
16376                                 if (rdstc > ts_val) {
16377                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16378                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16379                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16380                                         }
16381                                 }
16382 #endif
16383                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16384                                         soisdisconnected(so);
16385                                         tcp_timer_activate(tp, TT_2MSL,
16386                                                            (tcp_fast_finwait2_recycle ?
16387                                                             tcp_finwait2_timeout :
16388                                                             TP_MAXIDLE(tp)));
16389                                 }
16390                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
16391                                 break;
16392                         default:
16393                                 break;
16394                         }
16395                 }
16396                 if (rack->r_fast_output) {
16397                         /*
16398                          * We re doing fast output.. can we expand that?
16399                          */
16400                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
16401                 }
16402 #ifdef TCP_ACCOUNTING
16403                 rdstc = get_cyclecount();
16404                 if (rdstc > ts_val) {
16405                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16406                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16407                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16408                         }
16409                 }
16410
16411         } else if (win_up_req) {
16412                 rdstc = get_cyclecount();
16413                 if (rdstc > ts_val) {
16414                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16415                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
16416                         }
16417                 }
16418 #endif
16419         }
16420         /* Now is there a next packet, if so we are done */
16421         m_freem(m);
16422         did_out = 0;
16423         if (nxt_pkt) {
16424 #ifdef TCP_ACCOUNTING
16425                 sched_unpin();
16426 #endif
16427                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
16428                 return (0);
16429         }
16430         rack_handle_might_revert(tp, rack);
16431         ctf_calc_rwin(so, tp);
16432         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
16433         send_out_a_rst:
16434                 if (tcp_output(tp) < 0) {
16435 #ifdef TCP_ACCOUNTING
16436                         sched_unpin();
16437 #endif
16438                         return (1);
16439                 }
16440                 did_out = 1;
16441         }
16442         if (tp->t_flags2 & TF2_HPTS_CALLS)
16443                 tp->t_flags2 &= ~TF2_HPTS_CALLS;
16444         rack_free_trim(rack);
16445 #ifdef TCP_ACCOUNTING
16446         sched_unpin();
16447 #endif
16448         rack_timer_audit(tp, rack, &so->so_snd);
16449         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
16450         return (0);
16451 }
16452
16453 #define TCP_LRO_TS_OPTION \
16454     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
16455           (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
16456
16457 static int
16458 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
16459     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
16460     struct timeval *tv)
16461 {
16462         struct inpcb *inp = tptoinpcb(tp);
16463         struct socket *so = tptosocket(tp);
16464 #ifdef TCP_ACCOUNTING
16465         uint64_t ts_val;
16466 #endif
16467         int32_t thflags, retval, did_out = 0;
16468         int32_t way_out = 0;
16469         /*
16470          * cts - is the current time from tv (caller gets ts) in microseconds.
16471          * ms_cts - is the current time from tv in milliseconds.
16472          * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
16473          */
16474         uint32_t cts, us_cts, ms_cts;
16475         uint32_t tiwin, high_seq;
16476         struct timespec ts;
16477         struct tcpopt to;
16478         struct tcp_rack *rack;
16479         struct rack_sendmap *rsm;
16480         int32_t prev_state = 0;
16481         int no_output = 0;
16482         int slot_remaining = 0;
16483 #ifdef TCP_ACCOUNTING
16484         int ack_val_set = 0xf;
16485 #endif
16486         int nsegs;
16487
16488         NET_EPOCH_ASSERT();
16489         INP_WLOCK_ASSERT(inp);
16490
16491         /*
16492          * tv passed from common code is from either M_TSTMP_LRO or
16493          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
16494          */
16495         rack = (struct tcp_rack *)tp->t_fb_ptr;
16496         if (rack->rack_deferred_inited == 0) {
16497                 /*
16498                  * If we are the connecting socket we will
16499                  * hit rack_init() when no sequence numbers
16500                  * are setup. This makes it so we must defer
16501                  * some initialization. Call that now.
16502                  */
16503                 rack_deferred_init(tp, rack);
16504         }
16505         /*
16506          * Check to see if we need to skip any output plans. This
16507          * can happen in the non-LRO path where we are pacing and
16508          * must process the ack coming in but need to defer sending
16509          * anything becase a pacing timer is running.
16510          */
16511         us_cts = tcp_tv_to_usectick(tv);
16512         if (m->m_flags & M_ACKCMP) {
16513                 /*
16514                  * All compressed ack's are ack's by definition so
16515                  * remove any ack required flag and then do the processing.
16516                  */
16517                 rack->rc_ack_required = 0;
16518                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
16519         }
16520         thflags = tcp_get_flags(th);
16521         if ((rack->rc_always_pace == 1) &&
16522             (rack->rc_ack_can_sendout_data == 0) &&
16523             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16524             (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
16525                 /*
16526                  * Ok conditions are right for queuing the packets
16527                  * but we do have to check the flags in the inp, it
16528                  * could be, if a sack is present, we want to be awoken and
16529                  * so should process the packets.
16530                  */
16531                 slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
16532                 if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
16533                         no_output = 1;
16534                 } else {
16535                         /*
16536                          * If there is no options, or just a
16537                          * timestamp option, we will want to queue
16538                          * the packets. This is the same that LRO does
16539                          * and will need to change with accurate ECN.
16540                          */
16541                         uint32_t *ts_ptr;
16542                         int optlen;
16543
16544                         optlen = (th->th_off << 2) - sizeof(struct tcphdr);
16545                         ts_ptr = (uint32_t *)(th + 1);
16546                         if ((optlen == 0) ||
16547                             ((optlen == TCPOLEN_TSTAMP_APPA) &&
16548                              (*ts_ptr == TCP_LRO_TS_OPTION)))
16549                                 no_output = 1;
16550                 }
16551                 if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
16552                         /*
16553                          * It is unrealistic to think we can pace in less than
16554                          * the minimum granularity of the pacer (def:250usec). So
16555                          * if we have less than that time remaining we should go
16556                          * ahead and allow output to be "early". We will attempt to
16557                          * make up for it in any pacing time we try to apply on
16558                          * the outbound packet.
16559                          */
16560                         no_output = 0;
16561                 }
16562         }
16563         /*
16564          * If there is a RST or FIN lets dump out the bw
16565          * with a FIN the connection may go on but we
16566          * may not.
16567          */
16568         if ((thflags & TH_FIN) || (thflags & TH_RST))
16569                 rack_log_pacing_delay_calc(rack,
16570                                            rack->r_ctl.gp_bw,
16571                                            0,
16572                                            0,
16573                                            rack_get_gp_est(rack), /* delRate */
16574                                            rack_get_lt_bw(rack), /* rttProp */
16575                                            20, __LINE__, NULL, 0);
16576         if (m->m_flags & M_ACKCMP) {
16577                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
16578         }
16579         cts = tcp_tv_to_usectick(tv);
16580         ms_cts =  tcp_tv_to_mssectick(tv);
16581         nsegs = m->m_pkthdr.lro_nsegs;
16582         counter_u64_add(rack_proc_non_comp_ack, 1);
16583 #ifdef TCP_ACCOUNTING
16584         sched_pin();
16585         if (thflags & TH_ACK)
16586                 ts_val = get_cyclecount();
16587 #endif
16588         if ((m->m_flags & M_TSTMP) ||
16589             (m->m_flags & M_TSTMP_LRO)) {
16590                 mbuf_tstmp2timespec(m, &ts);
16591                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16592                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16593         } else
16594                 rack->r_ctl.act_rcv_time = *tv;
16595         kern_prefetch(rack, &prev_state);
16596         prev_state = 0;
16597         /*
16598          * Unscale the window into a 32-bit value. For the SYN_SENT state
16599          * the scale is zero.
16600          */
16601         tiwin = th->th_win << tp->snd_scale;
16602 #ifdef TCP_ACCOUNTING
16603         if (thflags & TH_ACK) {
16604                 /*
16605                  * We have a tradeoff here. We can either do what we are
16606                  * doing i.e. pinning to this CPU and then doing the accounting
16607                  * <or> we could do a critical enter, setup the rdtsc and cpu
16608                  * as in below, and then validate we are on the same CPU on
16609                  * exit. I have choosen to not do the critical enter since
16610                  * that often will gain you a context switch, and instead lock
16611                  * us (line above this if) to the same CPU with sched_pin(). This
16612                  * means we may be context switched out for a higher priority
16613                  * interupt but we won't be moved to another CPU.
16614                  *
16615                  * If this occurs (which it won't very often since we most likely
16616                  * are running this code in interupt context and only a higher
16617                  * priority will bump us ... clock?) we will falsely add in
16618                  * to the time the interupt processing time plus the ack processing
16619                  * time. This is ok since its a rare event.
16620                  */
16621                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
16622                                                     ctf_fixed_maxseg(tp));
16623         }
16624 #endif
16625         /*
16626          * Parse options on any incoming segment.
16627          */
16628         memset(&to, 0, sizeof(to));
16629         tcp_dooptions(&to, (u_char *)(th + 1),
16630             (th->th_off << 2) - sizeof(struct tcphdr),
16631             (thflags & TH_SYN) ? TO_SYN : 0);
16632         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
16633             __func__));
16634         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
16635             __func__));
16636
16637         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16638             (tp->t_flags & TF_GPUTINPROG)) {
16639                 /*
16640                  * We have a goodput in progress
16641                  * and we have entered a late state.
16642                  * Do we have enough data in the sb
16643                  * to handle the GPUT request?
16644                  */
16645                 uint32_t bytes;
16646
16647                 bytes = tp->gput_ack - tp->gput_seq;
16648                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
16649                         bytes += tp->gput_seq - tp->snd_una;
16650                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16651                         /*
16652                          * There are not enough bytes in the socket
16653                          * buffer that have been sent to cover this
16654                          * measurement. Cancel it.
16655                          */
16656                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16657                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
16658                                                    tp->gput_seq,
16659                                                    0, 0, 18, __LINE__, NULL, 0);
16660                         tp->t_flags &= ~TF_GPUTINPROG;
16661                 }
16662         }
16663         high_seq = th->th_ack;
16664         if (tcp_bblogging_on(rack->rc_tp)) {
16665                 union tcp_log_stackspecific log;
16666                 struct timeval ltv;
16667 #ifdef TCP_REQUEST_TRK
16668                 struct tcp_sendfile_track *tcp_req;
16669
16670                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
16671                         tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
16672                 } else {
16673                         tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
16674                 }
16675 #endif
16676                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16677                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16678                 if (rack->rack_no_prr == 0)
16679                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16680                 else
16681                         log.u_bbr.flex1 = 0;
16682                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16683                 log.u_bbr.use_lt_bw <<= 1;
16684                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
16685                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16686                 log.u_bbr.bbr_state = rack->rc_free_cnt;
16687                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16688                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
16689                 log.u_bbr.flex3 = m->m_flags;
16690                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16691                 log.u_bbr.lost = thflags;
16692                 log.u_bbr.pacing_gain = 0x1;
16693 #ifdef TCP_ACCOUNTING
16694                 log.u_bbr.cwnd_gain = ack_val_set;
16695 #endif
16696                 log.u_bbr.flex7 = 2;
16697                 if (m->m_flags & M_TSTMP) {
16698                         /* Record the hardware timestamp if present */
16699                         mbuf_tstmp2timespec(m, &ts);
16700                         ltv.tv_sec = ts.tv_sec;
16701                         ltv.tv_usec = ts.tv_nsec / 1000;
16702                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
16703                 } else if (m->m_flags & M_TSTMP_LRO) {
16704                         /* Record the LRO the arrival timestamp */
16705                         mbuf_tstmp2timespec(m, &ts);
16706                         ltv.tv_sec = ts.tv_sec;
16707                         ltv.tv_usec = ts.tv_nsec / 1000;
16708                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
16709                 }
16710                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
16711                 /* Log the rcv time */
16712                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
16713 #ifdef TCP_REQUEST_TRK
16714                 log.u_bbr.applimited = tp->t_tcpreq_closed;
16715                 log.u_bbr.applimited <<= 8;
16716                 log.u_bbr.applimited |= tp->t_tcpreq_open;
16717                 log.u_bbr.applimited <<= 8;
16718                 log.u_bbr.applimited |= tp->t_tcpreq_req;
16719                 if (tcp_req) {
16720                         /* Copy out any client req info */
16721                         /* seconds */
16722                         log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16723                         /* useconds */
16724                         log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16725                         log.u_bbr.rttProp = tcp_req->timestamp;
16726                         log.u_bbr.cur_del_rate = tcp_req->start;
16727                         if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16728                                 log.u_bbr.flex8 |= 1;
16729                         } else {
16730                                 log.u_bbr.flex8 |= 2;
16731                                 log.u_bbr.bw_inuse = tcp_req->end;
16732                         }
16733                         log.u_bbr.flex6 = tcp_req->start_seq;
16734                         if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16735                                 log.u_bbr.flex8 |= 4;
16736                                 log.u_bbr.epoch = tcp_req->end_seq;
16737                         }
16738                 }
16739 #endif
16740                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
16741                     tlen, &log, true, &ltv);
16742         }
16743         /* Remove ack required flag if set, we have one  */
16744         if (thflags & TH_ACK)
16745                 rack->rc_ack_required = 0;
16746         if (rack->sack_attack_disable > 0) {
16747                 rack->r_ctl.ack_during_sd++;
16748                 rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16749         }
16750         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
16751                 way_out = 4;
16752                 retval = 0;
16753                 m_freem(m);
16754                 goto done_with_input;
16755         }
16756         /*
16757          * If a segment with the ACK-bit set arrives in the SYN-SENT state
16758          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
16759          */
16760         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
16761             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
16762                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
16763                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
16764 #ifdef TCP_ACCOUNTING
16765                 sched_unpin();
16766 #endif
16767                 return (1);
16768         }
16769         /*
16770          * If timestamps were negotiated during SYN/ACK and a
16771          * segment without a timestamp is received, silently drop
16772          * the segment, unless it is a RST segment or missing timestamps are
16773          * tolerated.
16774          * See section 3.2 of RFC 7323.
16775          */
16776         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
16777             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
16778                 way_out = 5;
16779                 retval = 0;
16780                 m_freem(m);
16781                 goto done_with_input;
16782         }
16783
16784         /*
16785          * Segment received on connection. Reset idle time and keep-alive
16786          * timer. XXX: This should be done after segment validation to
16787          * ignore broken/spoofed segs.
16788          */
16789         if  (tp->t_idle_reduce &&
16790              (tp->snd_max == tp->snd_una) &&
16791              (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16792                 counter_u64_add(rack_input_idle_reduces, 1);
16793                 rack_cc_after_idle(rack, tp);
16794         }
16795         tp->t_rcvtime = ticks;
16796 #ifdef STATS
16797         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
16798 #endif
16799         if (tiwin > rack->r_ctl.rc_high_rwnd)
16800                 rack->r_ctl.rc_high_rwnd = tiwin;
16801         /*
16802          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
16803          * this to occur after we've validated the segment.
16804          */
16805         if (tcp_ecn_input_segment(tp, thflags, tlen,
16806             tcp_packets_this_ack(tp, th->th_ack),
16807             iptos))
16808                 rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
16809
16810         /*
16811          * If echoed timestamp is later than the current time, fall back to
16812          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16813          * were used when this connection was established.
16814          */
16815         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
16816                 to.to_tsecr -= tp->ts_offset;
16817                 if (TSTMP_GT(to.to_tsecr, ms_cts))
16818                         to.to_tsecr = 0;
16819         }
16820
16821         /*
16822          * If its the first time in we need to take care of options and
16823          * verify we can do SACK for rack!
16824          */
16825         if (rack->r_state == 0) {
16826                 /* Should be init'd by rack_init() */
16827                 KASSERT(rack->rc_inp != NULL,
16828                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
16829                 if (rack->rc_inp == NULL) {
16830                         rack->rc_inp = inp;
16831                 }
16832
16833                 /*
16834                  * Process options only when we get SYN/ACK back. The SYN
16835                  * case for incoming connections is handled in tcp_syncache.
16836                  * According to RFC1323 the window field in a SYN (i.e., a
16837                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
16838                  * this is traditional behavior, may need to be cleaned up.
16839                  */
16840                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
16841                         /* Handle parallel SYN for ECN */
16842                         tcp_ecn_input_parallel_syn(tp, thflags, iptos);
16843                         if ((to.to_flags & TOF_SCALE) &&
16844                             (tp->t_flags & TF_REQ_SCALE)) {
16845                                 tp->t_flags |= TF_RCVD_SCALE;
16846                                 tp->snd_scale = to.to_wscale;
16847                         } else
16848                                 tp->t_flags &= ~TF_REQ_SCALE;
16849                         /*
16850                          * Initial send window.  It will be updated with the
16851                          * next incoming segment to the scaled value.
16852                          */
16853                         tp->snd_wnd = th->th_win;
16854                         rack_validate_fo_sendwin_up(tp, rack);
16855                         if ((to.to_flags & TOF_TS) &&
16856                             (tp->t_flags & TF_REQ_TSTMP)) {
16857                                 tp->t_flags |= TF_RCVD_TSTMP;
16858                                 tp->ts_recent = to.to_tsval;
16859                                 tp->ts_recent_age = cts;
16860                         } else
16861                                 tp->t_flags &= ~TF_REQ_TSTMP;
16862                         if (to.to_flags & TOF_MSS) {
16863                                 tcp_mss(tp, to.to_mss);
16864                         }
16865                         if ((tp->t_flags & TF_SACK_PERMIT) &&
16866                             (to.to_flags & TOF_SACKPERM) == 0)
16867                                 tp->t_flags &= ~TF_SACK_PERMIT;
16868                         if (IS_FASTOPEN(tp->t_flags)) {
16869                                 if (to.to_flags & TOF_FASTOPEN) {
16870                                         uint16_t mss;
16871
16872                                         if (to.to_flags & TOF_MSS)
16873                                                 mss = to.to_mss;
16874                                         else
16875                                                 if ((inp->inp_vflag & INP_IPV6) != 0)
16876                                                         mss = TCP6_MSS;
16877                                                 else
16878                                                         mss = TCP_MSS;
16879                                         tcp_fastopen_update_cache(tp, mss,
16880                                             to.to_tfo_len, to.to_tfo_cookie);
16881                                 } else
16882                                         tcp_fastopen_disable_path(tp);
16883                         }
16884                 }
16885                 /*
16886                  * At this point we are at the initial call. Here we decide
16887                  * if we are doing RACK or not. We do this by seeing if
16888                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
16889                  * The code now does do dup-ack counting so if you don't
16890                  * switch back you won't get rack & TLP, but you will still
16891                  * get this stack.
16892                  */
16893
16894                 if ((rack_sack_not_required == 0) &&
16895                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
16896                         tcp_switch_back_to_default(tp);
16897                         (*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
16898                             tlen, iptos);
16899 #ifdef TCP_ACCOUNTING
16900                         sched_unpin();
16901 #endif
16902                         return (1);
16903                 }
16904                 tcp_set_hpts(tp);
16905                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
16906         }
16907         if (thflags & TH_FIN)
16908                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
16909         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
16910         if ((rack->rc_gp_dyn_mul) &&
16911             (rack->use_fixed_rate == 0) &&
16912             (rack->rc_always_pace)) {
16913                 /* Check in on probertt */
16914                 rack_check_probe_rtt(rack, us_cts);
16915         }
16916         rack_clear_rate_sample(rack);
16917         if ((rack->forced_ack) &&
16918             ((tcp_get_flags(th) & TH_RST) == 0)) {
16919                 rack_handle_probe_response(rack, tiwin, us_cts);
16920         }
16921         /*
16922          * This is the one exception case where we set the rack state
16923          * always. All other times (timers etc) we must have a rack-state
16924          * set (so we assure we have done the checks above for SACK).
16925          */
16926         rack->r_ctl.rc_rcvtime = cts;
16927         if (rack->r_state != tp->t_state)
16928                 rack_set_state(tp, rack);
16929         if (SEQ_GT(th->th_ack, tp->snd_una) &&
16930             (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
16931                 kern_prefetch(rsm, &prev_state);
16932         prev_state = rack->r_state;
16933         if ((thflags & TH_RST) &&
16934             ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
16935               SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
16936              (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
16937                 /* The connection will be killed by a reset check the tracepoint */
16938                 tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
16939         }
16940         retval = (*rack->r_substate) (m, th, so,
16941             tp, &to, drop_hdrlen,
16942             tlen, tiwin, thflags, nxt_pkt, iptos);
16943         if (retval == 0) {
16944                 /*
16945                  * If retval is 1 the tcb is unlocked and most likely the tp
16946                  * is gone.
16947                  */
16948                 INP_WLOCK_ASSERT(inp);
16949                 if ((rack->rc_gp_dyn_mul) &&
16950                     (rack->rc_always_pace) &&
16951                     (rack->use_fixed_rate == 0) &&
16952                     rack->in_probe_rtt &&
16953                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
16954                         /*
16955                          * If we are going for target, lets recheck before
16956                          * we output.
16957                          */
16958                         rack_check_probe_rtt(rack, us_cts);
16959                 }
16960                 if (rack->set_pacing_done_a_iw == 0) {
16961                         /* How much has been acked? */
16962                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
16963                                 /* We have enough to set in the pacing segment size */
16964                                 rack->set_pacing_done_a_iw = 1;
16965                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
16966                         }
16967                 }
16968                 tcp_rack_xmit_timer_commit(rack, tp);
16969 #ifdef TCP_ACCOUNTING
16970                 /*
16971                  * If we set the ack_val_se to what ack processing we are doing
16972                  * we also want to track how many cycles we burned. Note
16973                  * the bits after tcp_output we let be "free". This is because
16974                  * we are also tracking the tcp_output times as well. Note the
16975                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
16976                  * 0xf cannot be returned and is what we initialize it too to
16977                  * indicate we are not doing the tabulations.
16978                  */
16979                 if (ack_val_set != 0xf) {
16980                         uint64_t crtsc;
16981
16982                         crtsc = get_cyclecount();
16983                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16984                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
16985                         }
16986                 }
16987 #endif
16988                 if ((nxt_pkt == 0) && (no_output == 0)) {
16989                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
16990 do_output_now:
16991                                 if (tcp_output(tp) < 0) {
16992 #ifdef TCP_ACCOUNTING
16993                                         sched_unpin();
16994 #endif
16995                                         return (1);
16996                                 }
16997                                 did_out = 1;
16998                         }
16999                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
17000                         rack_free_trim(rack);
17001                 } else if ((no_output == 1) &&
17002                            (nxt_pkt == 0)  &&
17003                            (tcp_in_hpts(rack->rc_tp) == 0)) {
17004                         /*
17005                          * We are not in hpts and we had a pacing timer up. Use
17006                          * the remaining time (slot_remaining) to restart the timer.
17007                          */
17008                         KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
17009                         rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
17010                         rack_free_trim(rack);
17011                 }
17012                 /* Clear the flag, it may have been cleared by output but we may not have  */
17013                 if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
17014                         tp->t_flags2 &= ~TF2_HPTS_CALLS;
17015                 /* Update any rounds needed */
17016                 if (rack_verbose_logging &&  tcp_bblogging_on(rack->rc_tp))
17017                         rack_log_hystart_event(rack, high_seq, 8);
17018                 /*
17019                  * The draft (v3) calls for us to use SEQ_GEQ, but that
17020                  * causes issues when we are just going app limited. Lets
17021                  * instead use SEQ_GT <or> where its equal but more data
17022                  * is outstanding.
17023                  *
17024                  * Also make sure we are on the last ack of a series. We
17025                  * have to have all the ack's processed in queue to know
17026                  * if there is something left outstanding.
17027                  */
17028                 if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
17029                     (rack->rc_new_rnd_needed == 0) &&
17030                     (nxt_pkt == 0)) {
17031                         rack_log_hystart_event(rack, tp->snd_una, 21);
17032                         rack->r_ctl.current_round++;
17033                         /* Force the next send to setup the next round */
17034                         rack->rc_new_rnd_needed = 1;
17035                         if (CC_ALGO(tp)->newround != NULL) {
17036                                 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
17037                         }
17038                 }
17039                 if ((nxt_pkt == 0) &&
17040                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
17041                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
17042                      (tp->t_flags & TF_DELACK) ||
17043                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
17044                       (tp->t_state <= TCPS_CLOSING)))) {
17045                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
17046                         if ((tp->snd_max == tp->snd_una) &&
17047                             ((tp->t_flags & TF_DELACK) == 0) &&
17048                             (tcp_in_hpts(rack->rc_tp)) &&
17049                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
17050                                 /* keep alive not needed if we are hptsi output yet */
17051                                 ;
17052                         } else {
17053                                 int late = 0;
17054                                 if (tcp_in_hpts(tp)) {
17055                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
17056                                                 us_cts = tcp_get_usecs(NULL);
17057                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
17058                                                         rack->r_early = 1;
17059                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
17060                                                 } else
17061                                                         late = 1;
17062                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
17063                                         }
17064                                         tcp_hpts_remove(tp);
17065                                 }
17066                                 if (late && (did_out == 0)) {
17067                                         /*
17068                                          * We are late in the sending
17069                                          * and we did not call the output
17070                                          * (this probably should not happen).
17071                                          */
17072                                         goto do_output_now;
17073                                 }
17074                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
17075                         }
17076                         way_out = 1;
17077                 } else if (nxt_pkt == 0) {
17078                         /* Do we have the correct timer running? */
17079                         rack_timer_audit(tp, rack, &so->so_snd);
17080                         way_out = 2;
17081                 }
17082         done_with_input:
17083                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
17084                 if (did_out)
17085                         rack->r_wanted_output = 0;
17086         }
17087 #ifdef TCP_ACCOUNTING
17088         sched_unpin();
17089 #endif
17090         return (retval);
17091 }
17092
17093 static void
17094 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17095     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17096 {
17097         struct timeval tv;
17098
17099         /* First lets see if we have old packets */
17100         if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17101                 if (ctf_do_queued_segments(tp, 1)) {
17102                         m_freem(m);
17103                         return;
17104                 }
17105         }
17106         if (m->m_flags & M_TSTMP_LRO) {
17107                 mbuf_tstmp2timeval(m, &tv);
17108         } else {
17109                 /* Should not be should we kassert instead? */
17110                 tcp_get_usecs(&tv);
17111         }
17112         if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17113             &tv) == 0) {
17114                 INP_WUNLOCK(tptoinpcb(tp));
17115         }
17116 }
17117
17118 struct rack_sendmap *
17119 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17120 {
17121         struct rack_sendmap *rsm = NULL;
17122         int32_t idx;
17123         uint32_t srtt = 0, thresh = 0, ts_low = 0;
17124         int no_sack = 0;
17125
17126         /* Return the next guy to be re-transmitted */
17127         if (tqhash_empty(rack->r_ctl.tqh)) {
17128                 return (NULL);
17129         }
17130         if (tp->t_flags & TF_SENTFIN) {
17131                 /* retran the end FIN? */
17132                 return (NULL);
17133         }
17134         /* ok lets look at this one */
17135         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17136         if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17137                 return (rsm);
17138         }
17139         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17140                 goto check_it;
17141         }
17142         rsm = rack_find_lowest_rsm(rack);
17143         if (rsm == NULL) {
17144                 return (NULL);
17145         }
17146 check_it:
17147         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) ||
17148             (rack->sack_attack_disable > 0)) {
17149                 no_sack = 1;
17150         }
17151         if ((no_sack > 0) &&
17152             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17153                 /*
17154                  * No sack so we automatically do the 3 strikes and
17155                  * retransmit (no rack timer would be started).
17156                  */
17157                 return (rsm);
17158         }
17159         if (rsm->r_flags & RACK_ACKED) {
17160                 return (NULL);
17161         }
17162         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17163             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17164                 /* Its not yet ready */
17165                 return (NULL);
17166         }
17167         srtt = rack_grab_rtt(tp, rack);
17168         idx = rsm->r_rtr_cnt - 1;
17169         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17170         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
17171         if ((tsused == ts_low) ||
17172             (TSTMP_LT(tsused, ts_low))) {
17173                 /* No time since sending */
17174                 return (NULL);
17175         }
17176         if ((tsused - ts_low) < thresh) {
17177                 /* It has not been long enough yet */
17178                 return (NULL);
17179         }
17180         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17181             ((rsm->r_flags & RACK_SACK_PASSED) &&
17182              (rack->sack_attack_disable == 0))) {
17183                 /*
17184                  * We have passed the dup-ack threshold <or>
17185                  * a SACK has indicated this is missing.
17186                  * Note that if you are a declared attacker
17187                  * it is only the dup-ack threshold that
17188                  * will cause retransmits.
17189                  */
17190                 /* log retransmit reason */
17191                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17192                 rack->r_fast_output = 0;
17193                 return (rsm);
17194         }
17195         return (NULL);
17196 }
17197
17198 static void
17199 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
17200                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17201                            int line, struct rack_sendmap *rsm, uint8_t quality)
17202 {
17203         if (tcp_bblogging_on(rack->rc_tp)) {
17204                 union tcp_log_stackspecific log;
17205                 struct timeval tv;
17206
17207                 memset(&log, 0, sizeof(log));
17208                 log.u_bbr.flex1 = slot;
17209                 log.u_bbr.flex2 = len;
17210                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17211                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17212                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17213                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17214                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17215                 log.u_bbr.use_lt_bw <<= 1;
17216                 log.u_bbr.use_lt_bw |= rack->r_late;
17217                 log.u_bbr.use_lt_bw <<= 1;
17218                 log.u_bbr.use_lt_bw |= rack->r_early;
17219                 log.u_bbr.use_lt_bw <<= 1;
17220                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17221                 log.u_bbr.use_lt_bw <<= 1;
17222                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17223                 log.u_bbr.use_lt_bw <<= 1;
17224                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17225                 log.u_bbr.use_lt_bw <<= 1;
17226                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17227                 log.u_bbr.use_lt_bw <<= 1;
17228                 log.u_bbr.use_lt_bw |= rack->gp_ready;
17229                 log.u_bbr.pkt_epoch = line;
17230                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17231                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17232                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17233                 log.u_bbr.bw_inuse = bw_est;
17234                 log.u_bbr.delRate = bw;
17235                 if (rack->r_ctl.gp_bw == 0)
17236                         log.u_bbr.cur_del_rate = 0;
17237                 else
17238                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
17239                 log.u_bbr.rttProp = len_time;
17240                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17241                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17242                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17243                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17244                         /* We are in slow start */
17245                         log.u_bbr.flex7 = 1;
17246                 } else {
17247                         /* we are on congestion avoidance */
17248                         log.u_bbr.flex7 = 0;
17249                 }
17250                 log.u_bbr.flex8 = method;
17251                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17252                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17253                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17254                 log.u_bbr.cwnd_gain <<= 1;
17255                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17256                 log.u_bbr.cwnd_gain <<= 1;
17257                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17258                 log.u_bbr.bbr_substate = quality;
17259                 log.u_bbr.bbr_state = rack->dgp_on;
17260                 log.u_bbr.bbr_state <<= 1;
17261                 log.u_bbr.bbr_state |= rack->r_fill_less_agg;
17262                 log.u_bbr.bbr_state <<= 1;
17263                 log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17264                 log.u_bbr.bbr_state <<= 2;
17265                 log.u_bbr.bbr_state |= rack->r_pacing_discount;
17266                 log.u_bbr.flex7 = ((rack->r_ctl.pacing_discount_amm << 1) | log.u_bbr.flex7);
17267                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
17268                     &rack->rc_inp->inp_socket->so_rcv,
17269                     &rack->rc_inp->inp_socket->so_snd,
17270                     BBR_LOG_HPTSI_CALC, 0,
17271                     0, &log, false, &tv);
17272         }
17273 }
17274
17275 static uint32_t
17276 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17277 {
17278         uint32_t new_tso, user_max, pace_one;
17279
17280         user_max = rack->rc_user_set_max_segs * mss;
17281         if (rack->rc_force_max_seg) {
17282                 return (user_max);
17283         }
17284         if (rack->use_fixed_rate &&
17285             ((rack->r_ctl.crte == NULL) ||
17286              (bw != rack->r_ctl.crte->rate))) {
17287                 /* Use the user mss since we are not exactly matched */
17288                 return (user_max);
17289         }
17290         if (rack_pace_one_seg ||
17291             (rack->r_ctl.rc_user_set_min_segs == 1))
17292                 pace_one = 1;
17293         else
17294                 pace_one = 0;
17295
17296         new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17297                      pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17298         if (new_tso > user_max)
17299                 new_tso = user_max;
17300         if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17301                 if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17302                         new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17303         }
17304         if (rack->r_ctl.rc_user_set_min_segs &&
17305             ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17306             new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17307         return (new_tso);
17308 }
17309
17310 static int32_t
17311 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)
17312 {
17313         uint64_t lentim, fill_bw;
17314
17315         /* Lets first see if we are full, if so continue with normal rate */
17316         rack->r_via_fill_cw = 0;
17317         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
17318                 return (slot);
17319         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
17320                 return (slot);
17321         if (rack->r_ctl.rc_last_us_rtt == 0)
17322                 return (slot);
17323         if (rack->rc_pace_fill_if_rttin_range &&
17324             (rack->r_ctl.rc_last_us_rtt >=
17325              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
17326                 /* The rtt is huge, N * smallest, lets not fill */
17327                 return (slot);
17328         }
17329         /*
17330          * first lets calculate the b/w based on the last us-rtt
17331          * and the the smallest send window.
17332          */
17333         fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17334         /* Take the rwnd if its smaller */
17335         if (fill_bw > rack->rc_tp->snd_wnd)
17336                 fill_bw = rack->rc_tp->snd_wnd;
17337         /* Now lets make it into a b/w */
17338         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
17339         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17340         if (rack->r_fill_less_agg) {
17341                 /*
17342                  * We want the average of the rate_wanted
17343                  * and our fill-cw calculated bw. We also want
17344                  * to cap any increase to be no more than
17345                  * X times the lt_bw (where X is the rack_bw_multipler).
17346                  */
17347                 uint64_t lt_bw, rate;
17348
17349                 lt_bw = rack_get_lt_bw(rack);
17350                 if (lt_bw > *rate_wanted)
17351                         rate = lt_bw;
17352                 else
17353                         rate = *rate_wanted;
17354                 fill_bw += rate;
17355                 fill_bw /= 2;
17356                 if (rack_bw_multipler && (fill_bw > (rate * rack_bw_multipler))) {
17357                         fill_bw = rate * rack_bw_multipler;
17358                 }
17359         }
17360         /* We are below the min b/w */
17361         if (non_paced)
17362                 *rate_wanted = fill_bw;
17363         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
17364                 return (slot);
17365         rack->r_via_fill_cw = 1;
17366         if (rack->r_rack_hw_rate_caps &&
17367             (rack->r_ctl.crte != NULL)) {
17368                 uint64_t high_rate;
17369
17370                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
17371                 if (fill_bw > high_rate) {
17372                         /* We are capping bw at the highest rate table entry */
17373                         if (*rate_wanted > high_rate) {
17374                                 /* The original rate was also capped */
17375                                 rack->r_via_fill_cw = 0;
17376                         }
17377                         rack_log_hdwr_pacing(rack,
17378                                              fill_bw, high_rate, __LINE__,
17379                                              0, 3);
17380                         fill_bw = high_rate;
17381                         if (capped)
17382                                 *capped = 1;
17383                 }
17384         } else if ((rack->r_ctl.crte == NULL) &&
17385                    (rack->rack_hdrw_pacing == 0) &&
17386                    (rack->rack_hdw_pace_ena) &&
17387                    rack->r_rack_hw_rate_caps &&
17388                    (rack->rack_attempt_hdwr_pace == 0) &&
17389                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
17390                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17391                 /*
17392                  * Ok we may have a first attempt that is greater than our top rate
17393                  * lets check.
17394                  */
17395                 uint64_t high_rate;
17396
17397                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
17398                 if (high_rate) {
17399                         if (fill_bw > high_rate) {
17400                                 fill_bw = high_rate;
17401                                 if (capped)
17402                                         *capped = 1;
17403                         }
17404                 }
17405         }
17406         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
17407                 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
17408                                    fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL);
17409                 fill_bw = rack->r_ctl.bw_rate_cap;
17410         }
17411         /*
17412          * Ok fill_bw holds our mythical b/w to fill the cwnd
17413          * in an rtt (unless it was capped), what does that
17414          * time wise equate too?
17415          */
17416         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
17417         lentim /= fill_bw;
17418         *rate_wanted = fill_bw;
17419         if (non_paced || (lentim < slot)) {
17420                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
17421                                            0, lentim, 12, __LINE__, NULL, 0);
17422                 return ((int32_t)lentim);
17423         } else
17424                 return (slot);
17425 }
17426
17427 static int32_t
17428 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
17429 {
17430         uint64_t srtt;
17431         int32_t slot = 0;
17432         int32_t minslot = 0;
17433         int can_start_hw_pacing = 1;
17434         int err;
17435         int pace_one;
17436
17437         if (rack_pace_one_seg ||
17438             (rack->r_ctl.rc_user_set_min_segs == 1))
17439                 pace_one = 1;
17440         else
17441                 pace_one = 0;
17442         if (rack->rc_always_pace == 0) {
17443                 /*
17444                  * We use the most optimistic possible cwnd/srtt for
17445                  * sending calculations. This will make our
17446                  * calculation anticipate getting more through
17447                  * quicker then possible. But thats ok we don't want
17448                  * the peer to have a gap in data sending.
17449                  */
17450                 uint64_t cwnd, tr_perms = 0;
17451                 int32_t reduce = 0;
17452
17453         old_method:
17454                 /*
17455                  * We keep no precise pacing with the old method
17456                  * instead we use the pacer to mitigate bursts.
17457                  */
17458                 if (rack->r_ctl.rc_rack_min_rtt)
17459                         srtt = rack->r_ctl.rc_rack_min_rtt;
17460                 else
17461                         srtt = max(tp->t_srtt, 1);
17462                 if (rack->r_ctl.rc_rack_largest_cwnd)
17463                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
17464                 else
17465                         cwnd = rack->r_ctl.cwnd_to_use;
17466                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
17467                 tr_perms = (cwnd * 1000) / srtt;
17468                 if (tr_perms == 0) {
17469                         tr_perms = ctf_fixed_maxseg(tp);
17470                 }
17471                 /*
17472                  * Calculate how long this will take to drain, if
17473                  * the calculation comes out to zero, thats ok we
17474                  * will use send_a_lot to possibly spin around for
17475                  * more increasing tot_len_this_send to the point
17476                  * that its going to require a pace, or we hit the
17477                  * cwnd. Which in that case we are just waiting for
17478                  * a ACK.
17479                  */
17480                 slot = len / tr_perms;
17481                 /* Now do we reduce the time so we don't run dry? */
17482                 if (slot && rack_slot_reduction) {
17483                         reduce = (slot / rack_slot_reduction);
17484                         if (reduce < slot) {
17485                                 slot -= reduce;
17486                         } else
17487                                 slot = 0;
17488                 }
17489                 slot *= HPTS_USEC_IN_MSEC;
17490                 if (rack->rc_pace_to_cwnd) {
17491                         uint64_t rate_wanted = 0;
17492
17493                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
17494                         rack->rc_ack_can_sendout_data = 1;
17495                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
17496                 } else
17497                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
17498                 /*******************************************************/
17499                 /* RRS: We insert non-paced call to stats here for len */
17500                 /*******************************************************/
17501         } else {
17502                 uint64_t bw_est, res, lentim, rate_wanted;
17503                 uint32_t segs, oh;
17504                 int capped = 0;
17505                 int prev_fill;
17506
17507                 if ((rack->r_rr_config == 1) && rsm) {
17508                         return (rack->r_ctl.rc_min_to);
17509                 }
17510                 if (rack->use_fixed_rate) {
17511                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
17512                 } else if ((rack->r_ctl.init_rate == 0) &&
17513                            (rack->r_ctl.gp_bw == 0)) {
17514                         /* no way to yet do an estimate */
17515                         bw_est = rate_wanted = 0;
17516                 } else {
17517                         bw_est = rack_get_bw(rack);
17518                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
17519                 }
17520                 if ((bw_est == 0) || (rate_wanted == 0) ||
17521                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
17522                         /*
17523                          * No way yet to make a b/w estimate or
17524                          * our raise is set incorrectly.
17525                          */
17526                         goto old_method;
17527                 }
17528                 rack_rate_cap_bw(rack, &rate_wanted, &capped);
17529                 /* We need to account for all the overheads */
17530                 segs = (len + segsiz - 1) / segsiz;
17531                 /*
17532                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
17533                  * and how much data we put in each packet. Yes this
17534                  * means we may be off if we are larger than 1500 bytes
17535                  * or smaller. But this just makes us more conservative.
17536                  */
17537
17538                 oh =  (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
17539                 if (rack->r_is_v6) {
17540 #ifdef INET6
17541                         oh += sizeof(struct ip6_hdr);
17542 #endif
17543                 } else {
17544 #ifdef INET
17545                         oh += sizeof(struct ip);
17546 #endif
17547                 }
17548                 /* We add a fixed 14 for the ethernet header */
17549                 oh += 14;
17550                 segs *= oh;
17551                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
17552                 res = lentim / rate_wanted;
17553                 slot = (uint32_t)res;
17554                 if (rack_hw_rate_min &&
17555                     (rate_wanted < rack_hw_rate_min)) {
17556                         can_start_hw_pacing = 0;
17557                         if (rack->r_ctl.crte) {
17558                                 /*
17559                                  * Ok we need to release it, we
17560                                  * have fallen too low.
17561                                  */
17562                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17563                                 rack->r_ctl.crte = NULL;
17564                                 rack->rack_attempt_hdwr_pace = 0;
17565                                 rack->rack_hdrw_pacing = 0;
17566                         }
17567                 }
17568                 if (rack->r_ctl.crte &&
17569                     (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17570                         /*
17571                          * We want more than the hardware can give us,
17572                          * don't start any hw pacing.
17573                          */
17574                         can_start_hw_pacing = 0;
17575                         if (rack->r_rack_hw_rate_caps == 0) {
17576                                 /*
17577                                  * Ok we need to release it, we
17578                                  * want more than the card can give us and
17579                                  * no rate cap is in place. Set it up so
17580                                  * when we want less we can retry.
17581                                  */
17582                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17583                                 rack->r_ctl.crte = NULL;
17584                                 rack->rack_attempt_hdwr_pace = 0;
17585                                 rack->rack_hdrw_pacing = 0;
17586                         }
17587                 }
17588                 if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
17589                         /*
17590                          * We lost our rate somehow, this can happen
17591                          * if the interface changed underneath us.
17592                          */
17593                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17594                         rack->r_ctl.crte = NULL;
17595                         /* Lets re-allow attempting to setup pacing */
17596                         rack->rack_hdrw_pacing = 0;
17597                         rack->rack_attempt_hdwr_pace = 0;
17598                         rack_log_hdwr_pacing(rack,
17599                                              rate_wanted, bw_est, __LINE__,
17600                                              0, 6);
17601                 }
17602                 prev_fill = rack->r_via_fill_cw;
17603                 if ((rack->rc_pace_to_cwnd) &&
17604                     (capped == 0) &&
17605                     (rack->use_fixed_rate == 0) &&
17606                     (rack->in_probe_rtt == 0) &&
17607                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
17608                         /*
17609                          * We want to pace at our rate *or* faster to
17610                          * fill the cwnd to the max if its not full.
17611                          */
17612                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
17613                         /* Re-check to make sure we are not exceeding our max b/w */
17614                         if ((rack->r_ctl.crte != NULL) &&
17615                             (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17616                                 /*
17617                                  * We want more than the hardware can give us,
17618                                  * don't start any hw pacing.
17619                                  */
17620                                 can_start_hw_pacing = 0;
17621                                 if (rack->r_rack_hw_rate_caps == 0) {
17622                                         /*
17623                                          * Ok we need to release it, we
17624                                          * want more than the card can give us and
17625                                          * no rate cap is in place. Set it up so
17626                                          * when we want less we can retry.
17627                                          */
17628                                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17629                                         rack->r_ctl.crte = NULL;
17630                                         rack->rack_attempt_hdwr_pace = 0;
17631                                         rack->rack_hdrw_pacing = 0;
17632                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
17633                                 }
17634                         }
17635                 }
17636                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
17637                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17638                         if ((rack->rack_hdw_pace_ena) &&
17639                             (can_start_hw_pacing > 0) &&
17640                             (rack->rack_hdrw_pacing == 0) &&
17641                             (rack->rack_attempt_hdwr_pace == 0)) {
17642                                 /*
17643                                  * Lets attempt to turn on hardware pacing
17644                                  * if we can.
17645                                  */
17646                                 rack->rack_attempt_hdwr_pace = 1;
17647                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
17648                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
17649                                                                        rate_wanted,
17650                                                                        RS_PACING_GEQ,
17651                                                                        &err, &rack->r_ctl.crte_prev_rate);
17652                                 if (rack->r_ctl.crte) {
17653                                         rack->rack_hdrw_pacing = 1;
17654                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
17655                                                                           pace_one, rack->r_ctl.crte,
17656                                                                           NULL, rack->r_ctl.pace_len_divisor);
17657                                         rack_log_hdwr_pacing(rack,
17658                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17659                                                              err, 0);
17660                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
17661                                 } else {
17662                                         counter_u64_add(rack_hw_pace_init_fail, 1);
17663                                 }
17664                         } else if (rack->rack_hdrw_pacing &&
17665                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
17666                                 /* Do we need to adjust our rate? */
17667                                 const struct tcp_hwrate_limit_table *nrte;
17668
17669                                 if (rack->r_up_only &&
17670                                     (rate_wanted < rack->r_ctl.crte->rate)) {
17671                                         /**
17672                                          * We have four possible states here
17673                                          * having to do with the previous time
17674                                          * and this time.
17675                                          *   previous  |  this-time
17676                                          * A)     0      |     0   -- fill_cw not in the picture
17677                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
17678                                          * C)     1      |     1   -- all rates from fill_cw
17679                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
17680                                          *
17681                                          * For case A, C and D we don't allow a drop. But for
17682                                          * case B where we now our on our steady rate we do
17683                                          * allow a drop.
17684                                          *
17685                                          */
17686                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
17687                                                 goto done_w_hdwr;
17688                                 }
17689                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
17690                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
17691                                         if (rack_hw_rate_to_low &&
17692                                             (bw_est < rack_hw_rate_to_low)) {
17693                                                 /*
17694                                                  * The pacing rate is too low for hardware, but
17695                                                  * do allow hardware pacing to be restarted.
17696                                                  */
17697                                                 rack_log_hdwr_pacing(rack,
17698                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
17699                                                              0, 5);
17700                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17701                                                 rack->r_ctl.crte = NULL;
17702                                                 rack->rack_attempt_hdwr_pace = 0;
17703                                                 rack->rack_hdrw_pacing = 0;
17704                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17705                                                 goto done_w_hdwr;
17706                                         }
17707                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
17708                                                                             rack->rc_tp,
17709                                                                             rack->rc_inp->inp_route.ro_nh->nh_ifp,
17710                                                                             rate_wanted,
17711                                                                             RS_PACING_GEQ,
17712                                                                             &err, &rack->r_ctl.crte_prev_rate);
17713                                         if (nrte == NULL) {
17714                                                 /*
17715                                                  * Lost the rate, lets drop hardware pacing
17716                                                  * period.
17717                                                  */
17718                                                 rack->rack_hdrw_pacing = 0;
17719                                                 rack->r_ctl.crte = NULL;
17720                                                 rack_log_hdwr_pacing(rack,
17721                                                                      rate_wanted, 0, __LINE__,
17722                                                                      err, 1);
17723                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17724                                                 counter_u64_add(rack_hw_pace_lost, 1);
17725                                         } else if (nrte != rack->r_ctl.crte) {
17726                                                 rack->r_ctl.crte = nrte;
17727                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
17728                                                                                  segsiz, pace_one, rack->r_ctl.crte,
17729                                                                                  NULL, rack->r_ctl.pace_len_divisor);
17730                                                 rack_log_hdwr_pacing(rack,
17731                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17732                                                                      err, 2);
17733                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
17734                                         }
17735                                 } else {
17736                                         /* We just need to adjust the segment size */
17737                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17738                                         rack_log_hdwr_pacing(rack,
17739                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17740                                                              0, 4);
17741                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
17742                                 }
17743                         }
17744                 }
17745                 if (minslot && (minslot > slot)) {
17746                         rack_log_pacing_delay_calc(rack, minslot, slot, rack->r_ctl.crte->rate, bw_est, lentim,
17747                                                    98, __LINE__, NULL, 0);
17748                         slot = minslot;
17749                 }
17750 done_w_hdwr:
17751                 if (rack_limit_time_with_srtt &&
17752                     (rack->use_fixed_rate == 0) &&
17753                     (rack->rack_hdrw_pacing == 0)) {
17754                         /*
17755                          * Sanity check, we do not allow the pacing delay
17756                          * to be longer than the SRTT of the path. If it is
17757                          * a slow path, then adding a packet should increase
17758                          * the RTT and compensate for this i.e. the srtt will
17759                          * be greater so the allowed pacing time will be greater.
17760                          *
17761                          * Note this restriction is not for where a peak rate
17762                          * is set, we are doing fixed pacing or hardware pacing.
17763                          */
17764                         if (rack->rc_tp->t_srtt)
17765                                 srtt = rack->rc_tp->t_srtt;
17766                         else
17767                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
17768                         if (srtt < (uint64_t)slot) {
17769                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
17770                                 slot = srtt;
17771                         }
17772                 }
17773                 /*******************************************************************/
17774                 /* RRS: We insert paced call to stats here for len and rate_wanted */
17775                 /*******************************************************************/
17776                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
17777         }
17778         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
17779                 /*
17780                  * If this rate is seeing enobufs when it
17781                  * goes to send then either the nic is out
17782                  * of gas or we are mis-estimating the time
17783                  * somehow and not letting the queue empty
17784                  * completely. Lets add to the pacing time.
17785                  */
17786                 int hw_boost_delay;
17787
17788                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
17789                 if (hw_boost_delay > rack_enobuf_hw_max)
17790                         hw_boost_delay = rack_enobuf_hw_max;
17791                 else if (hw_boost_delay < rack_enobuf_hw_min)
17792                         hw_boost_delay = rack_enobuf_hw_min;
17793                 slot += hw_boost_delay;
17794         }
17795         return (slot);
17796 }
17797
17798 static void
17799 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
17800     tcp_seq startseq, uint32_t sb_offset)
17801 {
17802         struct rack_sendmap *my_rsm = NULL;
17803
17804         if (tp->t_state < TCPS_ESTABLISHED) {
17805                 /*
17806                  * We don't start any measurements if we are
17807                  * not at least established.
17808                  */
17809                 return;
17810         }
17811         if (tp->t_state >= TCPS_FIN_WAIT_1) {
17812                 /*
17813                  * We will get no more data into the SB
17814                  * this means we need to have the data available
17815                  * before we start a measurement.
17816                  */
17817
17818                 if (sbavail(&tptosocket(tp)->so_snd) <
17819                     max(rc_init_window(rack),
17820                         (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
17821                         /* Nope not enough data */
17822                         return;
17823                 }
17824         }
17825         tp->t_flags |= TF_GPUTINPROG;
17826         rack->r_ctl.rc_gp_cumack_ts = 0;
17827         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
17828         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
17829         tp->gput_seq = startseq;
17830         rack->app_limited_needs_set = 0;
17831         if (rack->in_probe_rtt)
17832                 rack->measure_saw_probe_rtt = 1;
17833         else if ((rack->measure_saw_probe_rtt) &&
17834                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
17835                 rack->measure_saw_probe_rtt = 0;
17836         if (rack->rc_gp_filled)
17837                 tp->gput_ts = rack->r_ctl.last_cumack_advance;
17838         else {
17839                 /* Special case initial measurement */
17840                 struct timeval tv;
17841
17842                 tp->gput_ts = tcp_get_usecs(&tv);
17843                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
17844         }
17845         /*
17846          * We take a guess out into the future,
17847          * if we have no measurement and no
17848          * initial rate, we measure the first
17849          * initial-windows worth of data to
17850          * speed up getting some GP measurement and
17851          * thus start pacing.
17852          */
17853         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
17854                 rack->app_limited_needs_set = 1;
17855                 tp->gput_ack = startseq + max(rc_init_window(rack),
17856                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
17857                 rack_log_pacing_delay_calc(rack,
17858                                            tp->gput_seq,
17859                                            tp->gput_ack,
17860                                            0,
17861                                            tp->gput_ts,
17862                                            (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
17863                                            9,
17864                                            __LINE__, NULL, 0);
17865                 rack_tend_gp_marks(tp, rack);
17866                 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
17867                 return;
17868         }
17869         if (sb_offset) {
17870                 /*
17871                  * We are out somewhere in the sb
17872                  * can we use the already outstanding data?
17873                  */
17874
17875                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
17876                         /*
17877                          * Yes first one is good and in this case
17878                          * the tp->gput_ts is correctly set based on
17879                          * the last ack that arrived (no need to
17880                          * set things up when an ack comes in).
17881                          */
17882                         my_rsm = tqhash_min(rack->r_ctl.tqh);
17883                         if ((my_rsm == NULL) ||
17884                             (my_rsm->r_rtr_cnt != 1)) {
17885                                 /* retransmission? */
17886                                 goto use_latest;
17887                         }
17888                 } else {
17889                         if (rack->r_ctl.rc_first_appl == NULL) {
17890                                 /*
17891                                  * If rc_first_appl is NULL
17892                                  * then the cnt should be 0.
17893                                  * This is probably an error, maybe
17894                                  * a KASSERT would be approprate.
17895                                  */
17896                                 goto use_latest;
17897                         }
17898                         /*
17899                          * If we have a marker pointer to the last one that is
17900                          * app limited we can use that, but we need to set
17901                          * things up so that when it gets ack'ed we record
17902                          * the ack time (if its not already acked).
17903                          */
17904                         rack->app_limited_needs_set = 1;
17905                         /*
17906                          * We want to get to the rsm that is either
17907                          * next with space i.e. over 1 MSS or the one
17908                          * after that (after the app-limited).
17909                          */
17910                         my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
17911                         if (my_rsm) {
17912                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
17913                                         /* Have to use the next one */
17914                                         my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17915                                 else {
17916                                         /* Use after the first MSS of it is acked */
17917                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
17918                                         goto start_set;
17919                                 }
17920                         }
17921                         if ((my_rsm == NULL) ||
17922                             (my_rsm->r_rtr_cnt != 1)) {
17923                                 /*
17924                                  * Either its a retransmit or
17925                                  * the last is the app-limited one.
17926                                  */
17927                                 goto use_latest;
17928                         }
17929                 }
17930                 tp->gput_seq = my_rsm->r_start;
17931 start_set:
17932                 if (my_rsm->r_flags & RACK_ACKED) {
17933                         /*
17934                          * This one has been acked use the arrival ack time
17935                          */
17936                         struct rack_sendmap *nrsm;
17937
17938                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
17939                         rack->app_limited_needs_set = 0;
17940                         /*
17941                          * Ok in this path we need to use the r_end now
17942                          * since this guy is the starting ack.
17943                          */
17944                         tp->gput_seq = my_rsm->r_end;
17945                         /*
17946                          * We also need to adjust up the sendtime
17947                          * to the send of the next data after my_rsm.
17948                          */
17949                         nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17950                         if (nrsm != NULL)
17951                                 my_rsm = nrsm;
17952                         else {
17953                                 /*
17954                                  * The next as not been sent, thats the
17955                                  * case for using the latest.
17956                                  */
17957                                 goto use_latest;
17958                         }
17959                 }
17960                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
17961                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
17962                 rack->r_ctl.rc_gp_cumack_ts = 0;
17963                 rack_log_pacing_delay_calc(rack,
17964                                            tp->gput_seq,
17965                                            tp->gput_ack,
17966                                            (uint64_t)my_rsm,
17967                                            tp->gput_ts,
17968                                            (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
17969                                            9,
17970                                            __LINE__, my_rsm, 0);
17971                 /* Now lets make sure all are marked as they should be */
17972                 rack_tend_gp_marks(tp, rack);
17973                 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
17974                 return;
17975         }
17976
17977 use_latest:
17978         /*
17979          * We don't know how long we may have been
17980          * idle or if this is the first-send. Lets
17981          * setup the flag so we will trim off
17982          * the first ack'd data so we get a true
17983          * measurement.
17984          */
17985         rack->app_limited_needs_set = 1;
17986         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
17987         rack->r_ctl.rc_gp_cumack_ts = 0;
17988         /* Find this guy so we can pull the send time */
17989         my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
17990         if (my_rsm) {
17991                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
17992                 if (my_rsm->r_flags & RACK_ACKED) {
17993                         /*
17994                          * Unlikely since its probably what was
17995                          * just transmitted (but I am paranoid).
17996                          */
17997                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
17998                         rack->app_limited_needs_set = 0;
17999                 }
18000                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18001                         /* This also is unlikely */
18002                         tp->gput_seq = my_rsm->r_start;
18003                 }
18004         } else {
18005                 /*
18006                  * TSNH unless we have some send-map limit,
18007                  * and even at that it should not be hitting
18008                  * that limit (we should have stopped sending).
18009                  */
18010                 struct timeval tv;
18011
18012                 microuptime(&tv);
18013                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18014         }
18015         rack_tend_gp_marks(tp, rack);
18016         rack_log_pacing_delay_calc(rack,
18017                                    tp->gput_seq,
18018                                    tp->gput_ack,
18019                                    (uint64_t)my_rsm,
18020                                    tp->gput_ts,
18021                                    (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18022                                    9, __LINE__, NULL, 0);
18023         rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18024 }
18025
18026 static inline uint32_t
18027 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
18028     uint32_t avail, int32_t sb_offset)
18029 {
18030         uint32_t len;
18031         uint32_t sendwin;
18032
18033         if (tp->snd_wnd > cwnd_to_use)
18034                 sendwin = cwnd_to_use;
18035         else
18036                 sendwin = tp->snd_wnd;
18037         if (ctf_outstanding(tp) >= tp->snd_wnd) {
18038                 /* We never want to go over our peers rcv-window */
18039                 len = 0;
18040         } else {
18041                 uint32_t flight;
18042
18043                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18044                 if (flight >= sendwin) {
18045                         /*
18046                          * We have in flight what we are allowed by cwnd (if
18047                          * it was rwnd blocking it would have hit above out
18048                          * >= tp->snd_wnd).
18049                          */
18050                         return (0);
18051                 }
18052                 len = sendwin - flight;
18053                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18054                         /* We would send too much (beyond the rwnd) */
18055                         len = tp->snd_wnd - ctf_outstanding(tp);
18056                 }
18057                 if ((len + sb_offset) > avail) {
18058                         /*
18059                          * We don't have that much in the SB, how much is
18060                          * there?
18061                          */
18062                         len = avail - sb_offset;
18063                 }
18064         }
18065         return (len);
18066 }
18067
18068 static void
18069 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18070              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18071              int rsm_is_null, int optlen, int line, uint16_t mode)
18072 {
18073         if (tcp_bblogging_on(rack->rc_tp)) {
18074                 union tcp_log_stackspecific log;
18075                 struct timeval tv;
18076
18077                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18078                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18079                 log.u_bbr.flex1 = error;
18080                 log.u_bbr.flex2 = flags;
18081                 log.u_bbr.flex3 = rsm_is_null;
18082                 log.u_bbr.flex4 = ipoptlen;
18083                 log.u_bbr.flex5 = tp->rcv_numsacks;
18084                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18085                 log.u_bbr.flex7 = optlen;
18086                 log.u_bbr.flex8 = rack->r_fsb_inited;
18087                 log.u_bbr.applimited = rack->r_fast_output;
18088                 log.u_bbr.bw_inuse = rack_get_bw(rack);
18089                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18090                 log.u_bbr.cwnd_gain = mode;
18091                 log.u_bbr.pkts_out = orig_len;
18092                 log.u_bbr.lt_epoch = len;
18093                 log.u_bbr.delivered = line;
18094                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18095                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18096                 tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
18097                                len, &log, false, NULL, __func__, __LINE__, &tv);
18098         }
18099 }
18100
18101
18102 static struct mbuf *
18103 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
18104                    struct rack_fast_send_blk *fsb,
18105                    int32_t seglimit, int32_t segsize, int hw_tls)
18106 {
18107 #ifdef KERN_TLS
18108         struct ktls_session *tls, *ntls;
18109 #ifdef INVARIANTS
18110         struct mbuf *start;
18111 #endif
18112 #endif
18113         struct mbuf *m, *n, **np, *smb;
18114         struct mbuf *top;
18115         int32_t off, soff;
18116         int32_t len = *plen;
18117         int32_t fragsize;
18118         int32_t len_cp = 0;
18119         uint32_t mlen, frags;
18120
18121         soff = off = the_off;
18122         smb = m = the_m;
18123         np = &top;
18124         top = NULL;
18125 #ifdef KERN_TLS
18126         if (hw_tls && (m->m_flags & M_EXTPG))
18127                 tls = m->m_epg_tls;
18128         else
18129                 tls = NULL;
18130 #ifdef INVARIANTS
18131         start = m;
18132 #endif
18133 #endif
18134         while (len > 0) {
18135                 if (m == NULL) {
18136                         *plen = len_cp;
18137                         break;
18138                 }
18139 #ifdef KERN_TLS
18140                 if (hw_tls) {
18141                         if (m->m_flags & M_EXTPG)
18142                                 ntls = m->m_epg_tls;
18143                         else
18144                                 ntls = NULL;
18145
18146                         /*
18147                          * Avoid mixing TLS records with handshake
18148                          * data or TLS records from different
18149                          * sessions.
18150                          */
18151                         if (tls != ntls) {
18152                                 MPASS(m != start);
18153                                 *plen = len_cp;
18154                                 break;
18155                         }
18156                 }
18157 #endif
18158                 mlen = min(len, m->m_len - off);
18159                 if (seglimit) {
18160                         /*
18161                          * For M_EXTPG mbufs, add 3 segments
18162                          * + 1 in case we are crossing page boundaries
18163                          * + 2 in case the TLS hdr/trailer are used
18164                          * It is cheaper to just add the segments
18165                          * than it is to take the cache miss to look
18166                          * at the mbuf ext_pgs state in detail.
18167                          */
18168                         if (m->m_flags & M_EXTPG) {
18169                                 fragsize = min(segsize, PAGE_SIZE);
18170                                 frags = 3;
18171                         } else {
18172                                 fragsize = segsize;
18173                                 frags = 0;
18174                         }
18175
18176                         /* Break if we really can't fit anymore. */
18177                         if ((frags + 1) >= seglimit) {
18178                                 *plen = len_cp;
18179                                 break;
18180                         }
18181
18182                         /*
18183                          * Reduce size if you can't copy the whole
18184                          * mbuf. If we can't copy the whole mbuf, also
18185                          * adjust len so the loop will end after this
18186                          * mbuf.
18187                          */
18188                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
18189                                 mlen = (seglimit - frags - 1) * fragsize;
18190                                 len = mlen;
18191                                 *plen = len_cp + len;
18192                         }
18193                         frags += howmany(mlen, fragsize);
18194                         if (frags == 0)
18195                                 frags++;
18196                         seglimit -= frags;
18197                         KASSERT(seglimit > 0,
18198                             ("%s: seglimit went too low", __func__));
18199                 }
18200                 n = m_get(M_NOWAIT, m->m_type);
18201                 *np = n;
18202                 if (n == NULL)
18203                         goto nospace;
18204                 n->m_len = mlen;
18205                 soff += mlen;
18206                 len_cp += n->m_len;
18207                 if (m->m_flags & (M_EXT|M_EXTPG)) {
18208                         n->m_data = m->m_data + off;
18209                         mb_dupcl(n, m);
18210                 } else {
18211                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
18212                             (u_int)n->m_len);
18213                 }
18214                 len -= n->m_len;
18215                 off = 0;
18216                 m = m->m_next;
18217                 np = &n->m_next;
18218                 if (len || (soff == smb->m_len)) {
18219                         /*
18220                          * We have more so we move forward  or
18221                          * we have consumed the entire mbuf and
18222                          * len has fell to 0.
18223                          */
18224                         soff = 0;
18225                         smb = m;
18226                 }
18227
18228         }
18229         if (fsb != NULL) {
18230                 fsb->m = smb;
18231                 fsb->off = soff;
18232                 if (smb) {
18233                         /*
18234                          * Save off the size of the mbuf. We do
18235                          * this so that we can recognize when it
18236                          * has been trimmed by sbcut() as acks
18237                          * come in.
18238                          */
18239                         fsb->o_m_len = smb->m_len;
18240                         fsb->o_t_len = M_TRAILINGROOM(smb);
18241                 } else {
18242                         /*
18243                          * This is the case where the next mbuf went to NULL. This
18244                          * means with this copy we have sent everything in the sb.
18245                          * In theory we could clear the fast_output flag, but lets
18246                          * not since its possible that we could get more added
18247                          * and acks that call the extend function which would let
18248                          * us send more.
18249                          */
18250                         fsb->o_m_len = 0;
18251                         fsb->o_t_len = 0;
18252                 }
18253         }
18254         return (top);
18255 nospace:
18256         if (top)
18257                 m_freem(top);
18258         return (NULL);
18259
18260 }
18261
18262 /*
18263  * This is a copy of m_copym(), taking the TSO segment size/limit
18264  * constraints into account, and advancing the sndptr as it goes.
18265  */
18266 static struct mbuf *
18267 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
18268                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
18269 {
18270         struct mbuf *m, *n;
18271         int32_t soff;
18272
18273         m = rack->r_ctl.fsb.m;
18274         if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
18275                 /*
18276                  * The trailing space changed, mbufs can grow
18277                  * at the tail but they can't shrink from
18278                  * it, KASSERT that. Adjust the orig_m_len to
18279                  * compensate for this change.
18280                  */
18281                 KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
18282                         ("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
18283                          m,
18284                          rack,
18285                          (intmax_t)M_TRAILINGROOM(m),
18286                          rack->r_ctl.fsb.o_t_len,
18287                          rack->r_ctl.fsb.o_m_len,
18288                          m->m_len));
18289                 rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
18290                 rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
18291         }
18292         if (m->m_len < rack->r_ctl.fsb.o_m_len) {
18293                 /*
18294                  * Mbuf shrank, trimmed off the top by an ack, our
18295                  * offset changes.
18296                  */
18297                 KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
18298                         ("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
18299                          m, m->m_len,
18300                          rack, rack->r_ctl.fsb.o_m_len,
18301                          rack->r_ctl.fsb.off));
18302
18303                 if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
18304                         rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
18305                 else
18306                         rack->r_ctl.fsb.off = 0;
18307                 rack->r_ctl.fsb.o_m_len = m->m_len;
18308 #ifdef INVARIANTS
18309         } else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
18310                 panic("rack:%p m:%p m_len grew outside of t_space compensation",
18311                       rack, m);
18312 #endif
18313         }
18314         soff = rack->r_ctl.fsb.off;
18315         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
18316         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
18317         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
18318                                  __FUNCTION__,
18319                                  rack, *plen, m, m->m_len));
18320         /* Save off the right location before we copy and advance */
18321         *s_soff = soff;
18322         *s_mb = rack->r_ctl.fsb.m;
18323         n = rack_fo_base_copym(m, soff, plen,
18324                                &rack->r_ctl.fsb,
18325                                seglimit, segsize, rack->r_ctl.fsb.hw_tls);
18326         return (n);
18327 }
18328
18329 /* Log the buffer level */
18330 static void
18331 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
18332                      int len, struct timeval *tv,
18333                      uint32_t cts)
18334 {
18335         uint32_t p_rate = 0, p_queue = 0, err = 0;
18336         union tcp_log_stackspecific log;
18337
18338 #ifdef RATELIMIT
18339         err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18340         err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18341 #endif
18342         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18343         log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18344         log.u_bbr.flex1 = p_rate;
18345         log.u_bbr.flex2 = p_queue;
18346         log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18347         log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18348         log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18349         log.u_bbr.flex7 = 99;
18350         log.u_bbr.flex8 = 0;
18351         log.u_bbr.pkts_out = err;
18352         log.u_bbr.delRate = rack->r_ctl.crte->rate;
18353         log.u_bbr.timeStamp = cts;
18354         log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18355         tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18356                        len, &log, false, NULL, __func__, __LINE__, tv);
18357
18358 }
18359
18360 static uint32_t
18361 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
18362                        struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
18363 {
18364         uint64_t lentime = 0;
18365 #ifdef RATELIMIT
18366         uint32_t p_rate = 0, p_queue = 0, err;
18367         union tcp_log_stackspecific log;
18368         uint64_t bw;
18369
18370         err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18371         /* Failed or queue is zero */
18372         if (err || (p_queue == 0)) {
18373                 lentime = 0;
18374                 goto out;
18375         }
18376         err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18377         if (err) {
18378                 lentime = 0;
18379                 goto out;
18380         }
18381         /*
18382          * If we reach here we have some bytes in
18383          * the queue. The number returned is a value
18384          * between 0 and 0xffff where ffff is full
18385          * and 0 is empty. So how best to make this into
18386          * something usable?
18387          *
18388          * The "safer" way is lets take the b/w gotten
18389          * from the query (which should be our b/w rate)
18390          * and pretend that a full send (our rc_pace_max_segs)
18391          * is outstanding. We factor it so its as if a full
18392          * number of our MSS segment is terms of full
18393          * ethernet segments are outstanding.
18394          */
18395         bw = p_rate / 8;
18396         if (bw) {
18397                 lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
18398                 lentime *= ETHERNET_SEGMENT_SIZE;
18399                 lentime *= (uint64_t)HPTS_USEC_IN_SEC;
18400                 lentime /= bw;
18401         } else {
18402                 /* TSNH -- KASSERT? */
18403                 lentime = 0;
18404         }
18405 out:
18406         if (tcp_bblogging_on(tp)) {
18407                 memset(&log, 0, sizeof(log));
18408                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18409                 log.u_bbr.flex1 = p_rate;
18410                 log.u_bbr.flex2 = p_queue;
18411                 log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18412                 log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18413                 log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18414                 log.u_bbr.flex7 = 99;
18415                 log.u_bbr.flex8 = 0;
18416                 log.u_bbr.pkts_out = err;
18417                 log.u_bbr.delRate = rack->r_ctl.crte->rate;
18418                 log.u_bbr.cur_del_rate = lentime;
18419                 log.u_bbr.timeStamp = cts;
18420                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18421                 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18422                                len, &log, false, NULL, __func__, __LINE__,tv);
18423         }
18424 #endif
18425         return ((uint32_t)lentime);
18426 }
18427
18428 static int
18429 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
18430                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
18431 {
18432         /*
18433          * Enter the fast retransmit path. We are given that a sched_pin is
18434          * in place (if accounting is compliled in) and the cycle count taken
18435          * at the entry is in the ts_val. The concept her is that the rsm
18436          * now holds the mbuf offsets and such so we can directly transmit
18437          * without a lot of overhead, the len field is already set for
18438          * us to prohibit us from sending too much (usually its 1MSS).
18439          */
18440         struct ip *ip = NULL;
18441         struct udphdr *udp = NULL;
18442         struct tcphdr *th = NULL;
18443         struct mbuf *m = NULL;
18444         struct inpcb *inp;
18445         uint8_t *cpto;
18446         struct tcp_log_buffer *lgb;
18447 #ifdef TCP_ACCOUNTING
18448         uint64_t crtsc;
18449         int cnt_thru = 1;
18450 #endif
18451         struct tcpopt to;
18452         u_char opt[TCP_MAXOLEN];
18453         uint32_t hdrlen, optlen;
18454         int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
18455         uint16_t flags;
18456         uint32_t if_hw_tsomaxsegcount = 0, startseq;
18457         uint32_t if_hw_tsomaxsegsize;
18458         int32_t ip_sendflag = IP_NO_SND_TAG_RL;
18459
18460 #ifdef INET6
18461         struct ip6_hdr *ip6 = NULL;
18462
18463         if (rack->r_is_v6) {
18464                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18465                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18466         } else
18467 #endif                          /* INET6 */
18468         {
18469                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18470                 hdrlen = sizeof(struct tcpiphdr);
18471         }
18472         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
18473                 goto failed;
18474         }
18475         if (doing_tlp) {
18476                 /* Its a TLP add the flag, it may already be there but be sure */
18477                 rsm->r_flags |= RACK_TLP;
18478         } else {
18479                 /* If it was a TLP it is not not on this retransmit */
18480                 rsm->r_flags &= ~RACK_TLP;
18481         }
18482         startseq = rsm->r_start;
18483         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
18484         inp = rack->rc_inp;
18485         to.to_flags = 0;
18486         flags = tcp_outflags[tp->t_state];
18487         if (flags & (TH_SYN|TH_RST)) {
18488                 goto failed;
18489         }
18490         if (rsm->r_flags & RACK_HAS_FIN) {
18491                 /* We can't send a FIN here */
18492                 goto failed;
18493         }
18494         if (flags & TH_FIN) {
18495                 /* We never send a FIN */
18496                 flags &= ~TH_FIN;
18497         }
18498         if (tp->t_flags & TF_RCVD_TSTMP) {
18499                 to.to_tsval = ms_cts + tp->ts_offset;
18500                 to.to_tsecr = tp->ts_recent;
18501                 to.to_flags = TOF_TS;
18502         }
18503         optlen = tcp_addoptions(&to, opt);
18504         hdrlen += optlen;
18505         udp = rack->r_ctl.fsb.udp;
18506         if (udp)
18507                 hdrlen += sizeof(struct udphdr);
18508         if (rack->r_ctl.rc_pace_max_segs)
18509                 max_val = rack->r_ctl.rc_pace_max_segs;
18510         else if (rack->rc_user_set_max_segs)
18511                 max_val = rack->rc_user_set_max_segs * segsiz;
18512         else
18513                 max_val = len;
18514         if ((tp->t_flags & TF_TSO) &&
18515             V_tcp_do_tso &&
18516             (len > segsiz) &&
18517             (tp->t_port == 0))
18518                 tso = 1;
18519 #ifdef INET6
18520         if (MHLEN < hdrlen + max_linkhdr)
18521                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18522         else
18523 #endif
18524                 m = m_gethdr(M_NOWAIT, MT_DATA);
18525         if (m == NULL)
18526                 goto failed;
18527         m->m_data += max_linkhdr;
18528         m->m_len = hdrlen;
18529         th = rack->r_ctl.fsb.th;
18530         /* Establish the len to send */
18531         if (len > max_val)
18532                 len = max_val;
18533         if ((tso) && (len + optlen > segsiz)) {
18534                 uint32_t if_hw_tsomax;
18535                 int32_t max_len;
18536
18537                 /* extract TSO information */
18538                 if_hw_tsomax = tp->t_tsomax;
18539                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18540                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18541                 /*
18542                  * Check if we should limit by maximum payload
18543                  * length:
18544                  */
18545                 if (if_hw_tsomax != 0) {
18546                         /* compute maximum TSO length */
18547                         max_len = (if_hw_tsomax - hdrlen -
18548                                    max_linkhdr);
18549                         if (max_len <= 0) {
18550                                 goto failed;
18551                         } else if (len > max_len) {
18552                                 len = max_len;
18553                         }
18554                 }
18555                 if (len <= segsiz) {
18556                         /*
18557                          * In case there are too many small fragments don't
18558                          * use TSO:
18559                          */
18560                         tso = 0;
18561                 }
18562         } else {
18563                 tso = 0;
18564         }
18565         if ((tso == 0) && (len > segsiz))
18566                 len = segsiz;
18567         (void)tcp_get_usecs(tv);
18568         if ((len == 0) ||
18569             (len <= MHLEN - hdrlen - max_linkhdr)) {
18570                 goto failed;
18571         }
18572         th->th_seq = htonl(rsm->r_start);
18573         th->th_ack = htonl(tp->rcv_nxt);
18574         /*
18575          * The PUSH bit should only be applied
18576          * if the full retransmission is made. If
18577          * we are sending less than this is the
18578          * left hand edge and should not have
18579          * the PUSH bit.
18580          */
18581         if ((rsm->r_flags & RACK_HAD_PUSH) &&
18582             (len == (rsm->r_end - rsm->r_start)))
18583                 flags |= TH_PUSH;
18584         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
18585         if (th->th_win == 0) {
18586                 tp->t_sndzerowin++;
18587                 tp->t_flags |= TF_RXWIN0SENT;
18588         } else
18589                 tp->t_flags &= ~TF_RXWIN0SENT;
18590         if (rsm->r_flags & RACK_TLP) {
18591                 /*
18592                  * TLP should not count in retran count, but
18593                  * in its own bin
18594                  */
18595                 counter_u64_add(rack_tlp_retran, 1);
18596                 counter_u64_add(rack_tlp_retran_bytes, len);
18597         } else {
18598                 tp->t_sndrexmitpack++;
18599                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18600                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18601         }
18602 #ifdef STATS
18603         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18604                                  len);
18605 #endif
18606         if (rsm->m == NULL)
18607                 goto failed;
18608         if (rsm->m &&
18609             ((rsm->orig_m_len != rsm->m->m_len) ||
18610              (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
18611                 /* Fix up the orig_m_len and possibly the mbuf offset */
18612                 rack_adjust_orig_mlen(rsm);
18613         }
18614         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
18615         if (len <= segsiz) {
18616                 /*
18617                  * Must have ran out of mbufs for the copy
18618                  * shorten it to no longer need tso. Lets
18619                  * not put on sendalot since we are low on
18620                  * mbufs.
18621                  */
18622                 tso = 0;
18623         }
18624         if ((m->m_next == NULL) || (len <= 0)){
18625                 goto failed;
18626         }
18627         if (udp) {
18628                 if (rack->r_is_v6)
18629                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
18630                 else
18631                         ulen = hdrlen + len - sizeof(struct ip);
18632                 udp->uh_ulen = htons(ulen);
18633         }
18634         m->m_pkthdr.rcvif = (struct ifnet *)0;
18635         if (TCPS_HAVERCVDSYN(tp->t_state) &&
18636             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18637                 int ect = tcp_ecn_output_established(tp, &flags, len, true);
18638                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18639                     (tp->t_flags2 & TF2_ECN_SND_ECE))
18640                     tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18641 #ifdef INET6
18642                 if (rack->r_is_v6) {
18643                     ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18644                     ip6->ip6_flow |= htonl(ect << 20);
18645                 }
18646                 else
18647 #endif
18648                 {
18649                     ip->ip_tos &= ~IPTOS_ECN_MASK;
18650                     ip->ip_tos |= ect;
18651                 }
18652         }
18653         if (rack->r_ctl.crte != NULL) {
18654                 /* See if we can send via the hw queue */
18655                 slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
18656                 /* If there is nothing in queue (no pacing time) we can send via the hw queue */
18657                 if (slot == 0)
18658                         ip_sendflag = 0;
18659         }
18660         tcp_set_flags(th, flags);
18661         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
18662 #ifdef INET6
18663         if (rack->r_is_v6) {
18664                 if (tp->t_port) {
18665                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18666                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18667                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18668                         th->th_sum = htons(0);
18669                         UDPSTAT_INC(udps_opackets);
18670                 } else {
18671                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18672                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18673                         th->th_sum = in6_cksum_pseudo(ip6,
18674                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18675                                                       0);
18676                 }
18677         }
18678 #endif
18679 #if defined(INET6) && defined(INET)
18680         else
18681 #endif
18682 #ifdef INET
18683         {
18684                 if (tp->t_port) {
18685                         m->m_pkthdr.csum_flags = CSUM_UDP;
18686                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18687                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18688                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18689                         th->th_sum = htons(0);
18690                         UDPSTAT_INC(udps_opackets);
18691                 } else {
18692                         m->m_pkthdr.csum_flags = CSUM_TCP;
18693                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18694                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
18695                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18696                                                                         IPPROTO_TCP + len + optlen));
18697                 }
18698                 /* IP version must be set here for ipv4/ipv6 checking later */
18699                 KASSERT(ip->ip_v == IPVERSION,
18700                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
18701         }
18702 #endif
18703         if (tso) {
18704                 /*
18705                  * Here we use segsiz since we have no added options besides
18706                  * any standard timestamp options (no DSACKs or SACKS are sent
18707                  * via either fast-path).
18708                  */
18709                 KASSERT(len > segsiz,
18710                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
18711                 m->m_pkthdr.csum_flags |= CSUM_TSO;
18712                 m->m_pkthdr.tso_segsz = segsiz;
18713         }
18714 #ifdef INET6
18715         if (rack->r_is_v6) {
18716                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
18717                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18718                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18719                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18720                 else
18721                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18722         }
18723 #endif
18724 #if defined(INET) && defined(INET6)
18725         else
18726 #endif
18727 #ifdef INET
18728         {
18729                 ip->ip_len = htons(m->m_pkthdr.len);
18730                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
18731                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18732                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18733                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18734                                 ip->ip_off |= htons(IP_DF);
18735                         }
18736                 } else {
18737                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18738                 }
18739         }
18740 #endif
18741         if (doing_tlp == 0) {
18742                 /* Set we retransmitted */
18743                 rack->rc_gp_saw_rec = 1;
18744         } else {
18745                 /* Its a TLP set ca or ss */
18746                 if (tp->snd_cwnd > tp->snd_ssthresh) {
18747                         /* Set we sent in CA */
18748                         rack->rc_gp_saw_ca = 1;
18749                 } else {
18750                         /* Set we sent in SS */
18751                         rack->rc_gp_saw_ss = 1;
18752                 }
18753         }
18754         /* Time to copy in our header */
18755         cpto = mtod(m, uint8_t *);
18756         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18757         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18758         if (optlen) {
18759                 bcopy(opt, th + 1, optlen);
18760                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18761         } else {
18762                 th->th_off = sizeof(struct tcphdr) >> 2;
18763         }
18764         if (tcp_bblogging_on(rack->rc_tp)) {
18765                 union tcp_log_stackspecific log;
18766
18767                 if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18768                         rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18769                         counter_u64_add(rack_collapsed_win_rxt, 1);
18770                         counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18771                 }
18772                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18773                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18774                 if (rack->rack_no_prr)
18775                         log.u_bbr.flex1 = 0;
18776                 else
18777                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18778                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18779                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18780                 log.u_bbr.flex4 = max_val;
18781                 /* Save off the early/late values */
18782                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18783                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18784                 log.u_bbr.bw_inuse = rack_get_bw(rack);
18785                 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
18786                 if (doing_tlp == 0)
18787                         log.u_bbr.flex8 = 1;
18788                 else
18789                         log.u_bbr.flex8 = 2;
18790                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18791                 log.u_bbr.flex7 = 55;
18792                 log.u_bbr.pkts_out = tp->t_maxseg;
18793                 log.u_bbr.timeStamp = cts;
18794                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18795                 if (rsm && (rsm->r_rtr_cnt > 0)) {
18796                         /*
18797                          * When we have a retransmit we want to log the
18798                          * burst at send and flight at send from before.
18799                          */
18800                         log.u_bbr.flex5 = rsm->r_fas;
18801                         log.u_bbr.bbr_substate = rsm->r_bas;
18802                 } else {
18803                         /*
18804                          * This is currently unlikely until we do the
18805                          * packet pair probes but I will add it for completeness.
18806                          */
18807                         log.u_bbr.flex5 = log.u_bbr.inflight;
18808                         log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
18809                 }
18810                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
18811                 log.u_bbr.delivered = 0;
18812                 log.u_bbr.rttProp = (uint64_t)rsm;
18813                 log.u_bbr.delRate = rsm->r_flags;
18814                 log.u_bbr.delRate <<= 31;
18815                 log.u_bbr.delRate |= rack->r_must_retran;
18816                 log.u_bbr.delRate <<= 1;
18817                 log.u_bbr.delRate |= 1;
18818                 log.u_bbr.pkt_epoch = __LINE__;
18819                 lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
18820                                      len, &log, false, NULL, __func__, __LINE__, tv);
18821         } else
18822                 lgb = NULL;
18823         if ((rack->r_ctl.crte != NULL) &&
18824             tcp_bblogging_on(tp)) {
18825                 rack_log_queue_level(tp, rack, len, tv, cts);
18826         }
18827 #ifdef INET6
18828         if (rack->r_is_v6) {
18829                 error = ip6_output(m, inp->in6p_outputopts,
18830                                    &inp->inp_route6,
18831                                    ip_sendflag, NULL, NULL, inp);
18832         }
18833         else
18834 #endif
18835 #ifdef INET
18836         {
18837                 error = ip_output(m, NULL,
18838                                   &inp->inp_route,
18839                                   ip_sendflag, 0, inp);
18840         }
18841 #endif
18842         m = NULL;
18843         if (lgb) {
18844                 lgb->tlb_errno = error;
18845                 lgb = NULL;
18846         }
18847         if (error) {
18848                 goto failed;
18849         } else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
18850                 rack->rc_hw_nobuf = 0;
18851                 rack->r_ctl.rc_agg_delayed = 0;
18852                 rack->r_early = 0;
18853                 rack->r_late = 0;
18854                 rack->r_ctl.rc_agg_early = 0;
18855         }
18856
18857         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
18858                         rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
18859         if (doing_tlp) {
18860                 rack->rc_tlp_in_progress = 1;
18861                 rack->r_ctl.rc_tlp_cnt_out++;
18862         }
18863         if (error == 0) {
18864                 counter_u64_add(rack_total_bytes, len);
18865                 tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
18866                 if (doing_tlp) {
18867                         rack->rc_last_sent_tlp_past_cumack = 0;
18868                         rack->rc_last_sent_tlp_seq_valid = 1;
18869                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18870                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18871                 }
18872         }
18873         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18874         rack->forced_ack = 0;   /* If we send something zap the FA flag */
18875         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18876                 rack->r_ctl.retran_during_recovery += len;
18877         {
18878                 int idx;
18879
18880                 idx = (len / segsiz) + 3;
18881                 if (idx >= TCP_MSS_ACCT_ATIMER)
18882                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18883                 else
18884                         counter_u64_add(rack_out_size[idx], 1);
18885         }
18886         if (tp->t_rtttime == 0) {
18887                 tp->t_rtttime = ticks;
18888                 tp->t_rtseq = startseq;
18889                 KMOD_TCPSTAT_INC(tcps_segstimed);
18890         }
18891         counter_u64_add(rack_fto_rsm_send, 1);
18892         if (error && (error == ENOBUFS)) {
18893                 if (rack->r_ctl.crte != NULL) {
18894                         tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
18895                         if (tcp_bblogging_on(rack->rc_tp))
18896                                 rack_log_queue_level(tp, rack, len, tv, cts);
18897                 } else
18898                         tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
18899                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18900                 if (rack->rc_enobuf < 0x7f)
18901                         rack->rc_enobuf++;
18902                 if (slot < (10 * HPTS_USEC_IN_MSEC))
18903                         slot = 10 * HPTS_USEC_IN_MSEC;
18904                 if (rack->r_ctl.crte != NULL) {
18905                         counter_u64_add(rack_saw_enobuf_hw, 1);
18906                         tcp_rl_log_enobuf(rack->r_ctl.crte);
18907                 }
18908                 counter_u64_add(rack_saw_enobuf, 1);
18909         } else
18910                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
18911         if ((slot == 0) ||
18912             (rack->rc_always_pace == 0) ||
18913             (rack->r_rr_config == 1)) {
18914                 /*
18915                  * We have no pacing set or we
18916                  * are using old-style rack or
18917                  * we are overridden to use the old 1ms pacing.
18918                  */
18919                 slot = rack->r_ctl.rc_min_to;
18920         }
18921         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
18922 #ifdef TCP_ACCOUNTING
18923         crtsc = get_cyclecount();
18924         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18925                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
18926         }
18927         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18928                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
18929         }
18930         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18931                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
18932         }
18933         sched_unpin();
18934 #endif
18935         return (0);
18936 failed:
18937         if (m)
18938                 m_free(m);
18939         return (-1);
18940 }
18941
18942 static void
18943 rack_sndbuf_autoscale(struct tcp_rack *rack)
18944 {
18945         /*
18946          * Automatic sizing of send socket buffer.  Often the send buffer
18947          * size is not optimally adjusted to the actual network conditions
18948          * at hand (delay bandwidth product).  Setting the buffer size too
18949          * small limits throughput on links with high bandwidth and high
18950          * delay (eg. trans-continental/oceanic links).  Setting the
18951          * buffer size too big consumes too much real kernel memory,
18952          * especially with many connections on busy servers.
18953          *
18954          * The criteria to step up the send buffer one notch are:
18955          *  1. receive window of remote host is larger than send buffer
18956          *     (with a fudge factor of 5/4th);
18957          *  2. send buffer is filled to 7/8th with data (so we actually
18958          *     have data to make use of it);
18959          *  3. send buffer fill has not hit maximal automatic size;
18960          *  4. our send window (slow start and cogestion controlled) is
18961          *     larger than sent but unacknowledged data in send buffer.
18962          *
18963          * Note that the rack version moves things much faster since
18964          * we want to avoid hitting cache lines in the rack_fast_output()
18965          * path so this is called much less often and thus moves
18966          * the SB forward by a percentage.
18967          */
18968         struct socket *so;
18969         struct tcpcb *tp;
18970         uint32_t sendwin, scaleup;
18971
18972         tp = rack->rc_tp;
18973         so = rack->rc_inp->inp_socket;
18974         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
18975         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
18976                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
18977                     sbused(&so->so_snd) >=
18978                     (so->so_snd.sb_hiwat / 8 * 7) &&
18979                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
18980                     sendwin >= (sbused(&so->so_snd) -
18981                     (tp->snd_nxt - tp->snd_una))) {
18982                         if (rack_autosndbuf_inc)
18983                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
18984                         else
18985                                 scaleup = V_tcp_autosndbuf_inc;
18986                         if (scaleup < V_tcp_autosndbuf_inc)
18987                                 scaleup = V_tcp_autosndbuf_inc;
18988                         scaleup += so->so_snd.sb_hiwat;
18989                         if (scaleup > V_tcp_autosndbuf_max)
18990                                 scaleup = V_tcp_autosndbuf_max;
18991                         if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
18992                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
18993                 }
18994         }
18995 }
18996
18997 static int
18998 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
18999                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
19000 {
19001         /*
19002          * Enter to do fast output. We are given that the sched_pin is
19003          * in place (if accounting is compiled in) and the cycle count taken
19004          * at entry is in place in ts_val. The idea here is that
19005          * we know how many more bytes needs to be sent (presumably either
19006          * during pacing or to fill the cwnd and that was greater than
19007          * the max-burst). We have how much to send and all the info we
19008          * need to just send.
19009          */
19010 #ifdef INET
19011         struct ip *ip = NULL;
19012 #endif
19013         struct udphdr *udp = NULL;
19014         struct tcphdr *th = NULL;
19015         struct mbuf *m, *s_mb;
19016         struct inpcb *inp;
19017         uint8_t *cpto;
19018         struct tcp_log_buffer *lgb;
19019 #ifdef TCP_ACCOUNTING
19020         uint64_t crtsc;
19021 #endif
19022         struct tcpopt to;
19023         u_char opt[TCP_MAXOLEN];
19024         uint32_t hdrlen, optlen;
19025 #ifdef TCP_ACCOUNTING
19026         int cnt_thru = 1;
19027 #endif
19028         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19029         uint16_t flags;
19030         uint32_t s_soff;
19031         uint32_t if_hw_tsomaxsegcount = 0, startseq;
19032         uint32_t if_hw_tsomaxsegsize;
19033         uint16_t add_flag = RACK_SENT_FP;
19034 #ifdef INET6
19035         struct ip6_hdr *ip6 = NULL;
19036
19037         if (rack->r_is_v6) {
19038                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19039                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19040         } else
19041 #endif                          /* INET6 */
19042         {
19043 #ifdef INET
19044                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19045                 hdrlen = sizeof(struct tcpiphdr);
19046 #endif
19047         }
19048         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19049                 m = NULL;
19050                 goto failed;
19051         }
19052         startseq = tp->snd_max;
19053         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19054         inp = rack->rc_inp;
19055         len = rack->r_ctl.fsb.left_to_send;
19056         to.to_flags = 0;
19057         flags = rack->r_ctl.fsb.tcp_flags;
19058         if (tp->t_flags & TF_RCVD_TSTMP) {
19059                 to.to_tsval = ms_cts + tp->ts_offset;
19060                 to.to_tsecr = tp->ts_recent;
19061                 to.to_flags = TOF_TS;
19062         }
19063         optlen = tcp_addoptions(&to, opt);
19064         hdrlen += optlen;
19065         udp = rack->r_ctl.fsb.udp;
19066         if (udp)
19067                 hdrlen += sizeof(struct udphdr);
19068         if (rack->r_ctl.rc_pace_max_segs)
19069                 max_val = rack->r_ctl.rc_pace_max_segs;
19070         else if (rack->rc_user_set_max_segs)
19071                 max_val = rack->rc_user_set_max_segs * segsiz;
19072         else
19073                 max_val = len;
19074         if ((tp->t_flags & TF_TSO) &&
19075             V_tcp_do_tso &&
19076             (len > segsiz) &&
19077             (tp->t_port == 0))
19078                 tso = 1;
19079 again:
19080 #ifdef INET6
19081         if (MHLEN < hdrlen + max_linkhdr)
19082                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19083         else
19084 #endif
19085                 m = m_gethdr(M_NOWAIT, MT_DATA);
19086         if (m == NULL)
19087                 goto failed;
19088         m->m_data += max_linkhdr;
19089         m->m_len = hdrlen;
19090         th = rack->r_ctl.fsb.th;
19091         /* Establish the len to send */
19092         if (len > max_val)
19093                 len = max_val;
19094         if ((tso) && (len + optlen > segsiz)) {
19095                 uint32_t if_hw_tsomax;
19096                 int32_t max_len;
19097
19098                 /* extract TSO information */
19099                 if_hw_tsomax = tp->t_tsomax;
19100                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19101                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19102                 /*
19103                  * Check if we should limit by maximum payload
19104                  * length:
19105                  */
19106                 if (if_hw_tsomax != 0) {
19107                         /* compute maximum TSO length */
19108                         max_len = (if_hw_tsomax - hdrlen -
19109                                    max_linkhdr);
19110                         if (max_len <= 0) {
19111                                 goto failed;
19112                         } else if (len > max_len) {
19113                                 len = max_len;
19114                         }
19115                 }
19116                 if (len <= segsiz) {
19117                         /*
19118                          * In case there are too many small fragments don't
19119                          * use TSO:
19120                          */
19121                         tso = 0;
19122                 }
19123         } else {
19124                 tso = 0;
19125         }
19126         if ((tso == 0) && (len > segsiz))
19127                 len = segsiz;
19128         (void)tcp_get_usecs(tv);
19129         if ((len == 0) ||
19130             (len <= MHLEN - hdrlen - max_linkhdr)) {
19131                 goto failed;
19132         }
19133         sb_offset = tp->snd_max - tp->snd_una;
19134         th->th_seq = htonl(tp->snd_max);
19135         th->th_ack = htonl(tp->rcv_nxt);
19136         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19137         if (th->th_win == 0) {
19138                 tp->t_sndzerowin++;
19139                 tp->t_flags |= TF_RXWIN0SENT;
19140         } else
19141                 tp->t_flags &= ~TF_RXWIN0SENT;
19142         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
19143         KMOD_TCPSTAT_INC(tcps_sndpack);
19144         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
19145 #ifdef STATS
19146         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
19147                                  len);
19148 #endif
19149         if (rack->r_ctl.fsb.m == NULL)
19150                 goto failed;
19151
19152         /* s_mb and s_soff are saved for rack_log_output */
19153         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
19154                                     &s_mb, &s_soff);
19155         if (len <= segsiz) {
19156                 /*
19157                  * Must have ran out of mbufs for the copy
19158                  * shorten it to no longer need tso. Lets
19159                  * not put on sendalot since we are low on
19160                  * mbufs.
19161                  */
19162                 tso = 0;
19163         }
19164         if (rack->r_ctl.fsb.rfo_apply_push &&
19165             (len == rack->r_ctl.fsb.left_to_send)) {
19166                 tcp_set_flags(th, flags | TH_PUSH);
19167                 add_flag |= RACK_HAD_PUSH;
19168         }
19169         if ((m->m_next == NULL) || (len <= 0)){
19170                 goto failed;
19171         }
19172         if (udp) {
19173                 if (rack->r_is_v6)
19174                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
19175                 else
19176                         ulen = hdrlen + len - sizeof(struct ip);
19177                 udp->uh_ulen = htons(ulen);
19178         }
19179         m->m_pkthdr.rcvif = (struct ifnet *)0;
19180         if (TCPS_HAVERCVDSYN(tp->t_state) &&
19181             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19182                 int ect = tcp_ecn_output_established(tp, &flags, len, false);
19183                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19184                     (tp->t_flags2 & TF2_ECN_SND_ECE))
19185                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19186 #ifdef INET6
19187                 if (rack->r_is_v6) {
19188                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19189                         ip6->ip6_flow |= htonl(ect << 20);
19190                 }
19191                 else
19192 #endif
19193                 {
19194 #ifdef INET
19195                         ip->ip_tos &= ~IPTOS_ECN_MASK;
19196                         ip->ip_tos |= ect;
19197 #endif
19198                 }
19199         }
19200         tcp_set_flags(th, flags);
19201         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
19202 #ifdef INET6
19203         if (rack->r_is_v6) {
19204                 if (tp->t_port) {
19205                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19206                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19207                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19208                         th->th_sum = htons(0);
19209                         UDPSTAT_INC(udps_opackets);
19210                 } else {
19211                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19212                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19213                         th->th_sum = in6_cksum_pseudo(ip6,
19214                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19215                                                       0);
19216                 }
19217         }
19218 #endif
19219 #if defined(INET6) && defined(INET)
19220         else
19221 #endif
19222 #ifdef INET
19223         {
19224                 if (tp->t_port) {
19225                         m->m_pkthdr.csum_flags = CSUM_UDP;
19226                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19227                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19228                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19229                         th->th_sum = htons(0);
19230                         UDPSTAT_INC(udps_opackets);
19231                 } else {
19232                         m->m_pkthdr.csum_flags = CSUM_TCP;
19233                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19234                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
19235                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19236                                                                         IPPROTO_TCP + len + optlen));
19237                 }
19238                 /* IP version must be set here for ipv4/ipv6 checking later */
19239                 KASSERT(ip->ip_v == IPVERSION,
19240                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
19241         }
19242 #endif
19243         if (tso) {
19244                 /*
19245                  * Here we use segsiz since we have no added options besides
19246                  * any standard timestamp options (no DSACKs or SACKS are sent
19247                  * via either fast-path).
19248                  */
19249                 KASSERT(len > segsiz,
19250                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
19251                 m->m_pkthdr.csum_flags |= CSUM_TSO;
19252                 m->m_pkthdr.tso_segsz = segsiz;
19253         }
19254 #ifdef INET6
19255         if (rack->r_is_v6) {
19256                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19257                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19258                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19259                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19260                 else
19261                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19262         }
19263 #endif
19264 #if defined(INET) && defined(INET6)
19265         else
19266 #endif
19267 #ifdef INET
19268         {
19269                 ip->ip_len = htons(m->m_pkthdr.len);
19270                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19271                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19272                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19273                         if (tp->t_port == 0 || len < V_tcp_minmss) {
19274                                 ip->ip_off |= htons(IP_DF);
19275                         }
19276                 } else {
19277                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19278                 }
19279         }
19280 #endif
19281         if (tp->snd_cwnd > tp->snd_ssthresh) {
19282                 /* Set we sent in CA */
19283                 rack->rc_gp_saw_ca = 1;
19284         } else {
19285                 /* Set we sent in SS */
19286                 rack->rc_gp_saw_ss = 1;
19287         }
19288         /* Time to copy in our header */
19289         cpto = mtod(m, uint8_t *);
19290         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19291         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19292         if (optlen) {
19293                 bcopy(opt, th + 1, optlen);
19294                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19295         } else {
19296                 th->th_off = sizeof(struct tcphdr) >> 2;
19297         }
19298         if ((rack->r_ctl.crte != NULL) &&
19299             tcp_bblogging_on(tp)) {
19300                 rack_log_queue_level(tp, rack, len, tv, cts);
19301         }
19302         if (tcp_bblogging_on(rack->rc_tp)) {
19303                 union tcp_log_stackspecific log;
19304
19305                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19306                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19307                 if (rack->rack_no_prr)
19308                         log.u_bbr.flex1 = 0;
19309                 else
19310                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19311                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19312                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19313                 log.u_bbr.flex4 = max_val;
19314                 /* Save off the early/late values */
19315                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19316                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19317                 log.u_bbr.bw_inuse = rack_get_bw(rack);
19318                 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19319                 log.u_bbr.flex8 = 0;
19320                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19321                 log.u_bbr.flex7 = 44;
19322                 log.u_bbr.pkts_out = tp->t_maxseg;
19323                 log.u_bbr.timeStamp = cts;
19324                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19325                 log.u_bbr.flex5 = log.u_bbr.inflight;
19326                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19327                 log.u_bbr.delivered = 0;
19328                 log.u_bbr.rttProp = 0;
19329                 log.u_bbr.delRate = rack->r_must_retran;
19330                 log.u_bbr.delRate <<= 1;
19331                 log.u_bbr.pkt_epoch = __LINE__;
19332                 /* For fast output no retrans so just inflight and how many mss we send */
19333                 log.u_bbr.flex5 = log.u_bbr.inflight;
19334                 log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19335                 lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19336                                      len, &log, false, NULL, __func__, __LINE__, tv);
19337         } else
19338                 lgb = NULL;
19339 #ifdef INET6
19340         if (rack->r_is_v6) {
19341                 error = ip6_output(m, inp->in6p_outputopts,
19342                                    &inp->inp_route6,
19343                                    0, NULL, NULL, inp);
19344         }
19345 #endif
19346 #if defined(INET) && defined(INET6)
19347         else
19348 #endif
19349 #ifdef INET
19350         {
19351                 error = ip_output(m, NULL,
19352                                   &inp->inp_route,
19353                                   0, 0, inp);
19354         }
19355 #endif
19356         if (lgb) {
19357                 lgb->tlb_errno = error;
19358                 lgb = NULL;
19359         }
19360         if (error) {
19361                 *send_err = error;
19362                 m = NULL;
19363                 goto failed;
19364         } else if (rack->rc_hw_nobuf) {
19365                 rack->rc_hw_nobuf = 0;
19366                 rack->r_ctl.rc_agg_delayed = 0;
19367                 rack->r_early = 0;
19368                 rack->r_late = 0;
19369                 rack->r_ctl.rc_agg_early = 0;
19370         }
19371         if ((error == 0) && (rack->lt_bw_up == 0)) {
19372                 /* Unlikely */
19373                 rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(tv);
19374                 rack->r_ctl.lt_seq = tp->snd_una;
19375                 rack->lt_bw_up = 1;
19376         }
19377         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
19378                         NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
19379         m = NULL;
19380         if (tp->snd_una == tp->snd_max) {
19381                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
19382                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19383                 tp->t_acktime = ticks;
19384         }
19385         counter_u64_add(rack_total_bytes, len);
19386         tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
19387
19388         rack->forced_ack = 0;   /* If we send something zap the FA flag */
19389         tot_len += len;
19390         if ((tp->t_flags & TF_GPUTINPROG) == 0)
19391                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
19392         tp->snd_max += len;
19393         tp->snd_nxt = tp->snd_max;
19394         if (rack->rc_new_rnd_needed) {
19395                 /*
19396                  * Update the rnd to start ticking not
19397                  * that from a time perspective all of
19398                  * the preceding idle time is "in the round"
19399                  */
19400                 rack->rc_new_rnd_needed = 0;
19401                 rack->r_ctl.roundends = tp->snd_max;
19402         }
19403         {
19404                 int idx;
19405
19406                 idx = (len / segsiz) + 3;
19407                 if (idx >= TCP_MSS_ACCT_ATIMER)
19408                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19409                 else
19410                         counter_u64_add(rack_out_size[idx], 1);
19411         }
19412         if (len <= rack->r_ctl.fsb.left_to_send)
19413                 rack->r_ctl.fsb.left_to_send -= len;
19414         else
19415                 rack->r_ctl.fsb.left_to_send = 0;
19416         if (rack->r_ctl.fsb.left_to_send < segsiz) {
19417                 rack->r_fast_output = 0;
19418                 rack->r_ctl.fsb.left_to_send = 0;
19419                 /* At the end of fast_output scale up the sb */
19420                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
19421                 rack_sndbuf_autoscale(rack);
19422                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
19423         }
19424         if (tp->t_rtttime == 0) {
19425                 tp->t_rtttime = ticks;
19426                 tp->t_rtseq = startseq;
19427                 KMOD_TCPSTAT_INC(tcps_segstimed);
19428         }
19429         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
19430             (max_val > len) &&
19431             (tso == 0)) {
19432                 max_val -= len;
19433                 len = segsiz;
19434                 th = rack->r_ctl.fsb.th;
19435 #ifdef TCP_ACCOUNTING
19436                 cnt_thru++;
19437 #endif
19438                 goto again;
19439         }
19440         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19441         counter_u64_add(rack_fto_send, 1);
19442         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
19443         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
19444 #ifdef TCP_ACCOUNTING
19445         crtsc = get_cyclecount();
19446         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19447                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19448         }
19449         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19450                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19451         }
19452         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19453                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
19454         }
19455         sched_unpin();
19456 #endif
19457         return (0);
19458 failed:
19459         if (m)
19460                 m_free(m);
19461         rack->r_fast_output = 0;
19462         return (-1);
19463 }
19464
19465 static inline void
19466 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
19467                        struct sockbuf *sb,
19468                        int len, int orig_len, int segsiz, uint32_t pace_max_seg,
19469                        bool hw_tls,
19470                        uint16_t flags)
19471 {
19472         rack->r_fast_output = 1;
19473         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19474         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19475         rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
19476         rack->r_ctl.fsb.tcp_flags = flags;
19477         rack->r_ctl.fsb.left_to_send = orig_len - len;
19478         if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
19479                 /* Less than a full sized pace, lets not  */
19480                 rack->r_fast_output = 0;
19481                 return;
19482         } else {
19483                 /* Round down to the nearest pace_max_seg */
19484                 rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
19485         }
19486         if (hw_tls)
19487                 rack->r_ctl.fsb.hw_tls = 1;
19488         else
19489                 rack->r_ctl.fsb.hw_tls = 0;
19490         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19491                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19492                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19493                  (tp->snd_max - tp->snd_una)));
19494         if (rack->r_ctl.fsb.left_to_send < segsiz)
19495                 rack->r_fast_output = 0;
19496         else {
19497                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19498                         rack->r_ctl.fsb.rfo_apply_push = 1;
19499                 else
19500                         rack->r_ctl.fsb.rfo_apply_push = 0;
19501         }
19502 }
19503
19504 static uint32_t
19505 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
19506 {
19507         uint64_t min_time;
19508         uint32_t maxlen;
19509
19510         min_time = (uint64_t)get_hpts_min_sleep_time();
19511         maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
19512         maxlen = roundup(maxlen, segsiz);
19513         return (maxlen);
19514 }
19515
19516 static struct rack_sendmap *
19517 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
19518 {
19519         struct rack_sendmap *rsm = NULL;
19520         int thresh;
19521
19522 restart:
19523         rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
19524         if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
19525                 /* Nothing, strange turn off validity  */
19526                 rack->r_collapse_point_valid = 0;
19527                 return (NULL);
19528         }
19529         /* Can we send it yet? */
19530         if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
19531                 /*
19532                  * Receiver window has not grown enough for
19533                  * the segment to be put on the wire.
19534                  */
19535                 return (NULL);
19536         }
19537         if (rsm->r_flags & RACK_ACKED) {
19538                 /*
19539                  * It has been sacked, lets move to the
19540                  * next one if possible.
19541                  */
19542                 rack->r_ctl.last_collapse_point = rsm->r_end;
19543                 /* Are we done? */
19544                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19545                             rack->r_ctl.high_collapse_point)) {
19546                         rack->r_collapse_point_valid = 0;
19547                         return (NULL);
19548                 }
19549                 goto restart;
19550         }
19551         /* Now has it been long enough ? */
19552         thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
19553         if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
19554                 rack_log_collapse(rack, rsm->r_start,
19555                                   (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19556                                   thresh, __LINE__, 6, rsm->r_flags, rsm);
19557                 return (rsm);
19558         }
19559         /* Not enough time */
19560         rack_log_collapse(rack, rsm->r_start,
19561                           (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19562                           thresh, __LINE__, 7, rsm->r_flags, rsm);
19563         return (NULL);
19564 }
19565
19566 static inline void
19567 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
19568 {
19569         if ((rack->full_size_rxt == 0) &&
19570             (rack->shape_rxt_to_pacing_min == 0) &&
19571             (*len >= segsiz)) {
19572                 *len = segsiz;
19573         } else if (rack->shape_rxt_to_pacing_min &&
19574                  rack->gp_ready) {
19575                 /* We use pacing min as shaping len req */
19576                 uint32_t maxlen;
19577
19578                 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
19579                 if (*len > maxlen)
19580                         *len = maxlen;
19581         } else {
19582                 /*
19583                  * The else is full_size_rxt is on so send it all
19584                  * note we do need to check this for exceeding
19585                  * our max segment size due to the fact that
19586                  * we do sometimes merge chunks together i.e.
19587                  * we cannot just assume that we will never have
19588                  * a chunk greater than pace_max_seg
19589                  */
19590                 if (*len > pace_max_seg)
19591                         *len = pace_max_seg;
19592         }
19593 }
19594
19595 static int
19596 rack_output(struct tcpcb *tp)
19597 {
19598         struct socket *so;
19599         uint32_t recwin;
19600         uint32_t sb_offset, s_moff = 0;
19601         int32_t len, error = 0;
19602         uint16_t flags;
19603         struct mbuf *m, *s_mb = NULL;
19604         struct mbuf *mb;
19605         uint32_t if_hw_tsomaxsegcount = 0;
19606         uint32_t if_hw_tsomaxsegsize;
19607         int32_t segsiz, minseg;
19608         long tot_len_this_send = 0;
19609 #ifdef INET
19610         struct ip *ip = NULL;
19611 #endif
19612         struct udphdr *udp = NULL;
19613         struct tcp_rack *rack;
19614         struct tcphdr *th;
19615         uint8_t pass = 0;
19616         uint8_t mark = 0;
19617         uint8_t check_done = 0;
19618         uint8_t wanted_cookie = 0;
19619         u_char opt[TCP_MAXOLEN];
19620         unsigned ipoptlen, optlen, hdrlen, ulen=0;
19621         uint32_t rack_seq;
19622
19623 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
19624         unsigned ipsec_optlen = 0;
19625
19626 #endif
19627         int32_t idle, sendalot;
19628         int32_t sub_from_prr = 0;
19629         volatile int32_t sack_rxmit;
19630         struct rack_sendmap *rsm = NULL;
19631         int32_t tso, mtu;
19632         struct tcpopt to;
19633         int32_t slot = 0;
19634         int32_t sup_rack = 0;
19635         uint32_t cts, ms_cts, delayed, early;
19636         uint16_t add_flag = RACK_SENT_SP;
19637         /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
19638         uint8_t doing_tlp = 0;
19639         uint32_t cwnd_to_use, pace_max_seg;
19640         int32_t do_a_prefetch = 0;
19641         int32_t prefetch_rsm = 0;
19642         int32_t orig_len = 0;
19643         struct timeval tv;
19644         int32_t prefetch_so_done = 0;
19645         struct tcp_log_buffer *lgb;
19646         struct inpcb *inp = tptoinpcb(tp);
19647         struct sockbuf *sb;
19648         uint64_t ts_val = 0;
19649 #ifdef TCP_ACCOUNTING
19650         uint64_t crtsc;
19651 #endif
19652 #ifdef INET6
19653         struct ip6_hdr *ip6 = NULL;
19654         int32_t isipv6;
19655 #endif
19656         bool hpts_calling, hw_tls = false;
19657
19658         NET_EPOCH_ASSERT();
19659         INP_WLOCK_ASSERT(inp);
19660
19661         /* setup and take the cache hits here */
19662         rack = (struct tcp_rack *)tp->t_fb_ptr;
19663 #ifdef TCP_ACCOUNTING
19664         sched_pin();
19665         ts_val = get_cyclecount();
19666 #endif
19667         hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
19668         tp->t_flags2 &= ~TF2_HPTS_CALLS;
19669 #ifdef TCP_OFFLOAD
19670         if (tp->t_flags & TF_TOE) {
19671 #ifdef TCP_ACCOUNTING
19672                 sched_unpin();
19673 #endif
19674                 return (tcp_offload_output(tp));
19675         }
19676 #endif
19677         if (rack->rack_deferred_inited == 0) {
19678                 /*
19679                  * If we are the connecting socket we will
19680                  * hit rack_init() when no sequence numbers
19681                  * are setup. This makes it so we must defer
19682                  * some initialization. Call that now.
19683                  */
19684                 rack_deferred_init(tp, rack);
19685         }
19686         /*
19687          * For TFO connections in SYN_RECEIVED, only allow the initial
19688          * SYN|ACK and those sent by the retransmit timer.
19689          */
19690         if (IS_FASTOPEN(tp->t_flags) &&
19691             (tp->t_state == TCPS_SYN_RECEIVED) &&
19692             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
19693             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
19694 #ifdef TCP_ACCOUNTING
19695                 sched_unpin();
19696 #endif
19697                 return (0);
19698         }
19699 #ifdef INET6
19700         if (rack->r_state) {
19701                 /* Use the cache line loaded if possible */
19702                 isipv6 = rack->r_is_v6;
19703         } else {
19704                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
19705         }
19706 #endif
19707         early = 0;
19708         cts = tcp_get_usecs(&tv);
19709         ms_cts = tcp_tv_to_mssectick(&tv);
19710         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
19711             tcp_in_hpts(rack->rc_tp)) {
19712                 /*
19713                  * We are on the hpts for some timer but not hptsi output.
19714                  * Remove from the hpts unconditionally.
19715                  */
19716                 rack_timer_cancel(tp, rack, cts, __LINE__);
19717         }
19718         /* Are we pacing and late? */
19719         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19720             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
19721                 /* We are delayed */
19722                 delayed = cts - rack->r_ctl.rc_last_output_to;
19723         } else {
19724                 delayed = 0;
19725         }
19726         /* Do the timers, which may override the pacer */
19727         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
19728                 int retval;
19729
19730                 retval = rack_process_timers(tp, rack, cts, hpts_calling,
19731                                              &doing_tlp);
19732                 if (retval != 0) {
19733                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
19734 #ifdef TCP_ACCOUNTING
19735                         sched_unpin();
19736 #endif
19737                         /*
19738                          * If timers want tcp_drop(), then pass error out,
19739                          * otherwise suppress it.
19740                          */
19741                         return (retval < 0 ? retval : 0);
19742                 }
19743         }
19744         if (rack->rc_in_persist) {
19745                 if (tcp_in_hpts(rack->rc_tp) == 0) {
19746                         /* Timer is not running */
19747                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19748                 }
19749 #ifdef TCP_ACCOUNTING
19750                 sched_unpin();
19751 #endif
19752                 return (0);
19753         }
19754         if ((rack->rc_ack_required == 1) &&
19755             (rack->r_timer_override == 0)){
19756                 /* A timeout occurred and no ack has arrived */
19757                 if (tcp_in_hpts(rack->rc_tp) == 0) {
19758                         /* Timer is not running */
19759                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19760                 }
19761 #ifdef TCP_ACCOUNTING
19762                 sched_unpin();
19763 #endif
19764                 return (0);
19765         }
19766         if ((rack->r_timer_override) ||
19767             (rack->rc_ack_can_sendout_data) ||
19768             (delayed) ||
19769             (tp->t_state < TCPS_ESTABLISHED)) {
19770                 rack->rc_ack_can_sendout_data = 0;
19771                 if (tcp_in_hpts(rack->rc_tp))
19772                         tcp_hpts_remove(rack->rc_tp);
19773         } else if (tcp_in_hpts(rack->rc_tp)) {
19774                 /*
19775                  * On the hpts you can't pass even if ACKNOW is on, we will
19776                  * when the hpts fires.
19777                  */
19778 #ifdef TCP_ACCOUNTING
19779                 crtsc = get_cyclecount();
19780                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19781                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
19782                 }
19783                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19784                         tp->tcp_cnt_counters[SND_BLOCKED]++;
19785                 }
19786                 sched_unpin();
19787 #endif
19788                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
19789                 return (0);
19790         }
19791         /* Finish out both pacing early and late accounting */
19792         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19793             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
19794                 early = rack->r_ctl.rc_last_output_to - cts;
19795         } else
19796                 early = 0;
19797         if (delayed) {
19798                 rack->r_ctl.rc_agg_delayed += delayed;
19799                 rack->r_late = 1;
19800         } else if (early) {
19801                 rack->r_ctl.rc_agg_early += early;
19802                 rack->r_early = 1;
19803         }
19804         /* Now that early/late accounting is done turn off the flag */
19805         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
19806         rack->r_wanted_output = 0;
19807         rack->r_timer_override = 0;
19808         if ((tp->t_state != rack->r_state) &&
19809             TCPS_HAVEESTABLISHED(tp->t_state)) {
19810                 rack_set_state(tp, rack);
19811         }
19812         if ((rack->r_fast_output) &&
19813             (doing_tlp == 0) &&
19814             (tp->rcv_numsacks == 0)) {
19815                 int ret;
19816
19817                 error = 0;
19818                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19819                 if (ret >= 0)
19820                         return(ret);
19821                 else if (error) {
19822                         inp = rack->rc_inp;
19823                         so = inp->inp_socket;
19824                         sb = &so->so_snd;
19825                         goto nomore;
19826                 }
19827         }
19828         inp = rack->rc_inp;
19829         /*
19830          * For TFO connections in SYN_SENT or SYN_RECEIVED,
19831          * only allow the initial SYN or SYN|ACK and those sent
19832          * by the retransmit timer.
19833          */
19834         if (IS_FASTOPEN(tp->t_flags) &&
19835             ((tp->t_state == TCPS_SYN_RECEIVED) ||
19836              (tp->t_state == TCPS_SYN_SENT)) &&
19837             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
19838             (tp->t_rxtshift == 0)) {              /* not a retransmit */
19839                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19840                 so = inp->inp_socket;
19841                 sb = &so->so_snd;
19842                 goto just_return_nolock;
19843         }
19844         /*
19845          * Determine length of data that should be transmitted, and flags
19846          * that will be used. If there is some data or critical controls
19847          * (SYN, RST) to send, then transmit; otherwise, investigate
19848          * further.
19849          */
19850         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
19851         if (tp->t_idle_reduce) {
19852                 if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
19853                         rack_cc_after_idle(rack, tp);
19854         }
19855         tp->t_flags &= ~TF_LASTIDLE;
19856         if (idle) {
19857                 if (tp->t_flags & TF_MORETOCOME) {
19858                         tp->t_flags |= TF_LASTIDLE;
19859                         idle = 0;
19860                 }
19861         }
19862         if ((tp->snd_una == tp->snd_max) &&
19863             rack->r_ctl.rc_went_idle_time &&
19864             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
19865                 idle = cts - rack->r_ctl.rc_went_idle_time;
19866                 if (idle > rack_min_probertt_hold) {
19867                         /* Count as a probe rtt */
19868                         if (rack->in_probe_rtt == 0) {
19869                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
19870                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
19871                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
19872                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
19873                         } else {
19874                                 rack_exit_probertt(rack, cts);
19875                         }
19876                 }
19877                 idle = 0;
19878         }
19879         if (rack_use_fsb &&
19880             (rack->r_ctl.fsb.tcp_ip_hdr) &&
19881             (rack->r_fsb_inited == 0) &&
19882             (rack->r_state != TCPS_CLOSED))
19883                 rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
19884 again:
19885         /*
19886          * If we've recently taken a timeout, snd_max will be greater than
19887          * snd_nxt.  There may be SACK information that allows us to avoid
19888          * resending already delivered data.  Adjust snd_nxt accordingly.
19889          */
19890         sendalot = 0;
19891         cts = tcp_get_usecs(&tv);
19892         ms_cts = tcp_tv_to_mssectick(&tv);
19893         tso = 0;
19894         mtu = 0;
19895         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19896         minseg = segsiz;
19897         if (rack->r_ctl.rc_pace_max_segs == 0)
19898                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
19899         else
19900                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
19901         sb_offset = tp->snd_max - tp->snd_una;
19902         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19903         flags = tcp_outflags[tp->t_state];
19904         while (rack->rc_free_cnt < rack_free_cache) {
19905                 rsm = rack_alloc(rack);
19906                 if (rsm == NULL) {
19907                         if (hpts_calling)
19908                                 /* Retry in a ms */
19909                                 slot = (1 * HPTS_USEC_IN_MSEC);
19910                         so = inp->inp_socket;
19911                         sb = &so->so_snd;
19912                         goto just_return_nolock;
19913                 }
19914                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
19915                 rack->rc_free_cnt++;
19916                 rsm = NULL;
19917         }
19918         sack_rxmit = 0;
19919         len = 0;
19920         rsm = NULL;
19921         if (flags & TH_RST) {
19922                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
19923                 so = inp->inp_socket;
19924                 sb = &so->so_snd;
19925                 goto send;
19926         }
19927         if (rack->r_ctl.rc_resend) {
19928                 /* Retransmit timer */
19929                 rsm = rack->r_ctl.rc_resend;
19930                 rack->r_ctl.rc_resend = NULL;
19931                 len = rsm->r_end - rsm->r_start;
19932                 sack_rxmit = 1;
19933                 sendalot = 0;
19934                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
19935                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
19936                          __func__, __LINE__,
19937                          rsm->r_start, tp->snd_una, tp, rack, rsm));
19938                 sb_offset = rsm->r_start - tp->snd_una;
19939                 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
19940         } else if (rack->r_collapse_point_valid &&
19941                    ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
19942                 /*
19943                  * If an RSM is returned then enough time has passed
19944                  * for us to retransmit it. Move up the collapse point,
19945                  * since this rsm has its chance to retransmit now.
19946                  */
19947                 tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
19948                 rack->r_ctl.last_collapse_point = rsm->r_end;
19949                 /* Are we done? */
19950                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19951                             rack->r_ctl.high_collapse_point))
19952                         rack->r_collapse_point_valid = 0;
19953                 sack_rxmit = 1;
19954                 /* We are not doing a TLP */
19955                 doing_tlp = 0;
19956                 len = rsm->r_end - rsm->r_start;
19957                 sb_offset = rsm->r_start - tp->snd_una;
19958                 sendalot = 0;
19959                 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
19960         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
19961                 /* We have a retransmit that takes precedence */
19962                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
19963                     ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
19964                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
19965                         /* Enter recovery if not induced by a time-out */
19966                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
19967                 }
19968 #ifdef INVARIANTS
19969                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
19970                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
19971                               tp, rack, rsm, rsm->r_start, tp->snd_una);
19972                 }
19973 #endif
19974                 len = rsm->r_end - rsm->r_start;
19975                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
19976                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
19977                          __func__, __LINE__,
19978                          rsm->r_start, tp->snd_una, tp, rack, rsm));
19979                 sb_offset = rsm->r_start - tp->snd_una;
19980                 sendalot = 0;
19981                 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
19982                 if (len > 0) {
19983                         sack_rxmit = 1;
19984                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
19985                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
19986                                          min(len, segsiz));
19987                 }
19988         } else if (rack->r_ctl.rc_tlpsend) {
19989                 /* Tail loss probe */
19990                 long cwin;
19991                 long tlen;
19992
19993                 /*
19994                  * Check if we can do a TLP with a RACK'd packet
19995                  * this can happen if we are not doing the rack
19996                  * cheat and we skipped to a TLP and it
19997                  * went off.
19998                  */
19999                 rsm = rack->r_ctl.rc_tlpsend;
20000                 /* We are doing a TLP make sure the flag is preent */
20001                 rsm->r_flags |= RACK_TLP;
20002                 rack->r_ctl.rc_tlpsend = NULL;
20003                 sack_rxmit = 1;
20004                 tlen = rsm->r_end - rsm->r_start;
20005                 if (tlen > segsiz)
20006                         tlen = segsiz;
20007                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20008                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20009                          __func__, __LINE__,
20010                          rsm->r_start, tp->snd_una, tp, rack, rsm));
20011                 sb_offset = rsm->r_start - tp->snd_una;
20012                 cwin = min(tp->snd_wnd, tlen);
20013                 len = cwin;
20014         }
20015         if (rack->r_must_retran &&
20016             (doing_tlp == 0) &&
20017             (SEQ_GT(tp->snd_max, tp->snd_una)) &&
20018             (rsm == NULL)) {
20019                 /*
20020                  * There are two different ways that we
20021                  * can get into this block:
20022                  * a) This is a non-sack connection, we had a time-out
20023                  *    and thus r_must_retran was set and everything
20024                  *    left outstanding as been marked for retransmit.
20025                  * b) The MTU of the path shrank, so that everything
20026                  *    was marked to be retransmitted with the smaller
20027                  *    mtu and r_must_retran was set.
20028                  *
20029                  * This means that we expect the sendmap (outstanding)
20030                  * to all be marked must. We can use the tmap to
20031                  * look at them.
20032                  *
20033                  */
20034                 int sendwin, flight;
20035
20036                 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
20037                 flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
20038                 if (flight >= sendwin) {
20039                         /*
20040                          * We can't send yet.
20041                          */
20042                         so = inp->inp_socket;
20043                         sb = &so->so_snd;
20044                         goto just_return_nolock;
20045                 }
20046                 /*
20047                  * This is the case a/b mentioned above. All
20048                  * outstanding/not-acked should be marked.
20049                  * We can use the tmap to find them.
20050                  */
20051                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
20052                 if (rsm == NULL) {
20053                         /* TSNH */
20054                         rack->r_must_retran = 0;
20055                         rack->r_ctl.rc_out_at_rto = 0;
20056                         so = inp->inp_socket;
20057                         sb = &so->so_snd;
20058                         goto just_return_nolock;
20059                 }
20060                 if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
20061                         /*
20062                          * The first one does not have the flag, did we collapse
20063                          * further up in our list?
20064                          */
20065                         rack->r_must_retran = 0;
20066                         rack->r_ctl.rc_out_at_rto = 0;
20067                         rsm = NULL;
20068                         sack_rxmit = 0;
20069                 } else {
20070                         sack_rxmit = 1;
20071                         len = rsm->r_end - rsm->r_start;
20072                         sb_offset = rsm->r_start - tp->snd_una;
20073                         sendalot = 0;
20074                         if ((rack->full_size_rxt == 0) &&
20075                             (rack->shape_rxt_to_pacing_min == 0) &&
20076                             (len >= segsiz))
20077                                 len = segsiz;
20078                         else if (rack->shape_rxt_to_pacing_min &&
20079                                  rack->gp_ready) {
20080                                 /* We use pacing min as shaping len req */
20081                                 uint32_t maxlen;
20082
20083                                 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20084                                 if (len > maxlen)
20085                                         len = maxlen;
20086                         }
20087                         /*
20088                          * Delay removing the flag RACK_MUST_RXT so
20089                          * that the fastpath for retransmit will
20090                          * work with this rsm.
20091                          */
20092                 }
20093         }
20094         /*
20095          * Enforce a connection sendmap count limit if set
20096          * as long as we are not retransmiting.
20097          */
20098         if ((rsm == NULL) &&
20099             (rack->do_detection == 0) &&
20100             (V_tcp_map_entries_limit > 0) &&
20101             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
20102                 counter_u64_add(rack_to_alloc_limited, 1);
20103                 if (!rack->alloc_limit_reported) {
20104                         rack->alloc_limit_reported = 1;
20105                         counter_u64_add(rack_alloc_limited_conns, 1);
20106                 }
20107                 so = inp->inp_socket;
20108                 sb = &so->so_snd;
20109                 goto just_return_nolock;
20110         }
20111         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
20112                 /* we are retransmitting the fin */
20113                 len--;
20114                 if (len) {
20115                         /*
20116                          * When retransmitting data do *not* include the
20117                          * FIN. This could happen from a TLP probe.
20118                          */
20119                         flags &= ~TH_FIN;
20120                 }
20121         }
20122         if (rsm && rack->r_fsb_inited &&
20123             rack_use_rsm_rfo &&
20124             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
20125                 int ret;
20126
20127                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
20128                 if (ret == 0)
20129                         return (0);
20130         }
20131         so = inp->inp_socket;
20132         sb = &so->so_snd;
20133         if (do_a_prefetch == 0) {
20134                 kern_prefetch(sb, &do_a_prefetch);
20135                 do_a_prefetch = 1;
20136         }
20137 #ifdef NETFLIX_SHARED_CWND
20138         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
20139             rack->rack_enable_scwnd) {
20140                 /* We are doing cwnd sharing */
20141                 if (rack->gp_ready &&
20142                     (rack->rack_attempted_scwnd == 0) &&
20143                     (rack->r_ctl.rc_scw == NULL) &&
20144                     tp->t_lib) {
20145                         /* The pcbid is in, lets make an attempt */
20146                         counter_u64_add(rack_try_scwnd, 1);
20147                         rack->rack_attempted_scwnd = 1;
20148                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
20149                                                                    &rack->r_ctl.rc_scw_index,
20150                                                                    segsiz);
20151                 }
20152                 if (rack->r_ctl.rc_scw &&
20153                     (rack->rack_scwnd_is_idle == 1) &&
20154                     sbavail(&so->so_snd)) {
20155                         /* we are no longer out of data */
20156                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20157                         rack->rack_scwnd_is_idle = 0;
20158                 }
20159                 if (rack->r_ctl.rc_scw) {
20160                         /* First lets update and get the cwnd */
20161                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
20162                                                                                        rack->r_ctl.rc_scw_index,
20163                                                                                        tp->snd_cwnd, tp->snd_wnd, segsiz);
20164                 }
20165         }
20166 #endif
20167         /*
20168          * Get standard flags, and add SYN or FIN if requested by 'hidden'
20169          * state flags.
20170          */
20171         if (tp->t_flags & TF_NEEDFIN)
20172                 flags |= TH_FIN;
20173         if (tp->t_flags & TF_NEEDSYN)
20174                 flags |= TH_SYN;
20175         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
20176                 void *end_rsm;
20177                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
20178                 if (end_rsm)
20179                         kern_prefetch(end_rsm, &prefetch_rsm);
20180                 prefetch_rsm = 1;
20181         }
20182         SOCKBUF_LOCK(sb);
20183         /*
20184          * If snd_nxt == snd_max and we have transmitted a FIN, the
20185          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
20186          * negative length.  This can also occur when TCP opens up its
20187          * congestion window while receiving additional duplicate acks after
20188          * fast-retransmit because TCP will reset snd_nxt to snd_max after
20189          * the fast-retransmit.
20190          *
20191          * In the normal retransmit-FIN-only case, however, snd_nxt will be
20192          * set to snd_una, the sb_offset will be 0, and the length may wind
20193          * up 0.
20194          *
20195          * If sack_rxmit is true we are retransmitting from the scoreboard
20196          * in which case len is already set.
20197          */
20198         if ((sack_rxmit == 0) &&
20199             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
20200                 uint32_t avail;
20201
20202                 avail = sbavail(sb);
20203                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
20204                         sb_offset = tp->snd_nxt - tp->snd_una;
20205                 else
20206                         sb_offset = 0;
20207                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
20208                         if (rack->r_ctl.rc_tlp_new_data) {
20209                                 /* TLP is forcing out new data */
20210                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
20211                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
20212                                 }
20213                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
20214                                         if (tp->snd_wnd > sb_offset)
20215                                                 len = tp->snd_wnd - sb_offset;
20216                                         else
20217                                                 len = 0;
20218                                 } else {
20219                                         len = rack->r_ctl.rc_tlp_new_data;
20220                                 }
20221                                 rack->r_ctl.rc_tlp_new_data = 0;
20222                         }  else {
20223                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
20224                         }
20225                         if ((rack->r_ctl.crte == NULL) &&
20226                             IN_FASTRECOVERY(tp->t_flags) &&
20227                             (rack->full_size_rxt == 0) &&
20228                             (rack->shape_rxt_to_pacing_min == 0) &&
20229                             (len > segsiz)) {
20230                                 /*
20231                                  * For prr=off, we need to send only 1 MSS
20232                                  * at a time. We do this because another sack could
20233                                  * be arriving that causes us to send retransmits and
20234                                  * we don't want to be on a long pace due to a larger send
20235                                  * that keeps us from sending out the retransmit.
20236                                  */
20237                                 len = segsiz;
20238                         } else if (rack->shape_rxt_to_pacing_min &&
20239                                    rack->gp_ready) {
20240                                 /* We use pacing min as shaping len req */
20241                                 uint32_t maxlen;
20242
20243                                 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20244                                 if (len > maxlen)
20245                                         len = maxlen;
20246                         }/* The else is full_size_rxt is on so send it all */
20247                 } else {
20248                         uint32_t outstanding;
20249                         /*
20250                          * We are inside of a Fast recovery episode, this
20251                          * is caused by a SACK or 3 dup acks. At this point
20252                          * we have sent all the retransmissions and we rely
20253                          * on PRR to dictate what we will send in the form of
20254                          * new data.
20255                          */
20256
20257                         outstanding = tp->snd_max - tp->snd_una;
20258                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
20259                                 if (tp->snd_wnd > outstanding) {
20260                                         len = tp->snd_wnd - outstanding;
20261                                         /* Check to see if we have the data */
20262                                         if ((sb_offset + len) > avail) {
20263                                                 /* It does not all fit */
20264                                                 if (avail > sb_offset)
20265                                                         len = avail - sb_offset;
20266                                                 else
20267                                                         len = 0;
20268                                         }
20269                                 } else {
20270                                         len = 0;
20271                                 }
20272                         } else if (avail > sb_offset) {
20273                                 len = avail - sb_offset;
20274                         } else {
20275                                 len = 0;
20276                         }
20277                         if (len > 0) {
20278                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
20279                                         len = rack->r_ctl.rc_prr_sndcnt;
20280                                 }
20281                                 if (len > 0) {
20282                                         sub_from_prr = 1;
20283                                 }
20284                         }
20285                         if (len > segsiz) {
20286                                 /*
20287                                  * We should never send more than a MSS when
20288                                  * retransmitting or sending new data in prr
20289                                  * mode unless the override flag is on. Most
20290                                  * likely the PRR algorithm is not going to
20291                                  * let us send a lot as well :-)
20292                                  */
20293                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
20294                                         len = segsiz;
20295                                 }
20296                         } else if (len < segsiz) {
20297                                 /*
20298                                  * Do we send any? The idea here is if the
20299                                  * send empty's the socket buffer we want to
20300                                  * do it. However if not then lets just wait
20301                                  * for our prr_sndcnt to get bigger.
20302                                  */
20303                                 long leftinsb;
20304
20305                                 leftinsb = sbavail(sb) - sb_offset;
20306                                 if (leftinsb > len) {
20307                                         /* This send does not empty the sb */
20308                                         len = 0;
20309                                 }
20310                         }
20311                 }
20312         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
20313                 /*
20314                  * If you have not established
20315                  * and are not doing FAST OPEN
20316                  * no data please.
20317                  */
20318                 if ((sack_rxmit == 0) &&
20319                     (!IS_FASTOPEN(tp->t_flags))){
20320                         len = 0;
20321                         sb_offset = 0;
20322                 }
20323         }
20324         if (prefetch_so_done == 0) {
20325                 kern_prefetch(so, &prefetch_so_done);
20326                 prefetch_so_done = 1;
20327         }
20328         /*
20329          * Lop off SYN bit if it has already been sent.  However, if this is
20330          * SYN-SENT state and if segment contains data and if we don't know
20331          * that foreign host supports TAO, suppress sending segment.
20332          */
20333         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
20334             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
20335                 /*
20336                  * When sending additional segments following a TFO SYN|ACK,
20337                  * do not include the SYN bit.
20338                  */
20339                 if (IS_FASTOPEN(tp->t_flags) &&
20340                     (tp->t_state == TCPS_SYN_RECEIVED))
20341                         flags &= ~TH_SYN;
20342         }
20343         /*
20344          * Be careful not to send data and/or FIN on SYN segments. This
20345          * measure is needed to prevent interoperability problems with not
20346          * fully conformant TCP implementations.
20347          */
20348         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
20349                 len = 0;
20350                 flags &= ~TH_FIN;
20351         }
20352         /*
20353          * On TFO sockets, ensure no data is sent in the following cases:
20354          *
20355          *  - When retransmitting SYN|ACK on a passively-created socket
20356          *
20357          *  - When retransmitting SYN on an actively created socket
20358          *
20359          *  - When sending a zero-length cookie (cookie request) on an
20360          *    actively created socket
20361          *
20362          *  - When the socket is in the CLOSED state (RST is being sent)
20363          */
20364         if (IS_FASTOPEN(tp->t_flags) &&
20365             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
20366              ((tp->t_state == TCPS_SYN_SENT) &&
20367               (tp->t_tfo_client_cookie_len == 0)) ||
20368              (flags & TH_RST))) {
20369                 sack_rxmit = 0;
20370                 len = 0;
20371         }
20372         /* Without fast-open there should never be data sent on a SYN */
20373         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
20374                 tp->snd_nxt = tp->iss;
20375                 len = 0;
20376         }
20377         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
20378                 /* We only send 1 MSS if we have a DSACK block */
20379                 add_flag |= RACK_SENT_W_DSACK;
20380                 len = segsiz;
20381         }
20382         orig_len = len;
20383         if (len <= 0) {
20384                 /*
20385                  * If FIN has been sent but not acked, but we haven't been
20386                  * called to retransmit, len will be < 0.  Otherwise, window
20387                  * shrank after we sent into it.  If window shrank to 0,
20388                  * cancel pending retransmit, pull snd_nxt back to (closed)
20389                  * window, and set the persist timer if it isn't already
20390                  * going.  If the window didn't close completely, just wait
20391                  * for an ACK.
20392                  *
20393                  * We also do a general check here to ensure that we will
20394                  * set the persist timer when we have data to send, but a
20395                  * 0-byte window. This makes sure the persist timer is set
20396                  * even if the packet hits one of the "goto send" lines
20397                  * below.
20398                  */
20399                 len = 0;
20400                 if ((tp->snd_wnd == 0) &&
20401                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20402                     (tp->snd_una == tp->snd_max) &&
20403                     (sb_offset < (int)sbavail(sb))) {
20404                         rack_enter_persist(tp, rack, cts, tp->snd_una);
20405                 }
20406         } else if ((rsm == NULL) &&
20407                    (doing_tlp == 0) &&
20408                    (len < pace_max_seg)) {
20409                 /*
20410                  * We are not sending a maximum sized segment for
20411                  * some reason. Should we not send anything (think
20412                  * sws or persists)?
20413                  */
20414                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20415                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20416                     (len < minseg) &&
20417                     (len < (int)(sbavail(sb) - sb_offset))) {
20418                         /*
20419                          * Here the rwnd is less than
20420                          * the minimum pacing size, this is not a retransmit,
20421                          * we are established and
20422                          * the send is not the last in the socket buffer
20423                          * we send nothing, and we may enter persists
20424                          * if nothing is outstanding.
20425                          */
20426                         len = 0;
20427                         if (tp->snd_max == tp->snd_una) {
20428                                 /*
20429                                  * Nothing out we can
20430                                  * go into persists.
20431                                  */
20432                                 rack_enter_persist(tp, rack, cts, tp->snd_una);
20433                         }
20434                 } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
20435                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20436                            (len < (int)(sbavail(sb) - sb_offset)) &&
20437                            (len < minseg)) {
20438                         /*
20439                          * Here we are not retransmitting, and
20440                          * the cwnd is not so small that we could
20441                          * not send at least a min size (rxt timer
20442                          * not having gone off), We have 2 segments or
20443                          * more already in flight, its not the tail end
20444                          * of the socket buffer  and the cwnd is blocking
20445                          * us from sending out a minimum pacing segment size.
20446                          * Lets not send anything.
20447                          */
20448                         len = 0;
20449                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
20450                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20451                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20452                            (len < (int)(sbavail(sb) - sb_offset)) &&
20453                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
20454                         /*
20455                          * Here we have a send window but we have
20456                          * filled it up and we can't send another pacing segment.
20457                          * We also have in flight more than 2 segments
20458                          * and we are not completing the sb i.e. we allow
20459                          * the last bytes of the sb to go out even if
20460                          * its not a full pacing segment.
20461                          */
20462                         len = 0;
20463                 } else if ((rack->r_ctl.crte != NULL) &&
20464                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
20465                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
20466                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
20467                            (len < (int)(sbavail(sb) - sb_offset))) {
20468                         /*
20469                          * Here we are doing hardware pacing, this is not a TLP,
20470                          * we are not sending a pace max segment size, there is rwnd
20471                          * room to send at least N pace_max_seg, the cwnd is greater
20472                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
20473                          * more segments in flight and its not the tail of the socket buffer.
20474                          *
20475                          * We don't want to send instead we need to get more ack's in to
20476                          * allow us to send a full pacing segment. Normally, if we are pacing
20477                          * about the right speed, we should have finished our pacing
20478                          * send as most of the acks have come back if we are at the
20479                          * right rate. This is a bit fuzzy since return path delay
20480                          * can delay the acks, which is why we want to make sure we
20481                          * have cwnd space to have a bit more than a max pace segments in flight.
20482                          *
20483                          * If we have not gotten our acks back we are pacing at too high a
20484                          * rate delaying will not hurt and will bring our GP estimate down by
20485                          * injecting the delay. If we don't do this we will send
20486                          * 2 MSS out in response to the acks being clocked in which
20487                          * defeats the point of hw-pacing (i.e. to help us get
20488                          * larger TSO's out).
20489                          */
20490                         len = 0;
20491                 }
20492
20493         }
20494         /* len will be >= 0 after this point. */
20495         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
20496         rack_sndbuf_autoscale(rack);
20497         /*
20498          * Decide if we can use TCP Segmentation Offloading (if supported by
20499          * hardware).
20500          *
20501          * TSO may only be used if we are in a pure bulk sending state.  The
20502          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
20503          * options prevent using TSO.  With TSO the TCP header is the same
20504          * (except for the sequence number) for all generated packets.  This
20505          * makes it impossible to transmit any options which vary per
20506          * generated segment or packet.
20507          *
20508          * IPv4 handling has a clear separation of ip options and ip header
20509          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
20510          * the right thing below to provide length of just ip options and thus
20511          * checking for ipoptlen is enough to decide if ip options are present.
20512          */
20513         ipoptlen = 0;
20514 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20515         /*
20516          * Pre-calculate here as we save another lookup into the darknesses
20517          * of IPsec that way and can actually decide if TSO is ok.
20518          */
20519 #ifdef INET6
20520         if (isipv6 && IPSEC_ENABLED(ipv6))
20521                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
20522 #ifdef INET
20523         else
20524 #endif
20525 #endif                          /* INET6 */
20526 #ifdef INET
20527                 if (IPSEC_ENABLED(ipv4))
20528                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
20529 #endif                          /* INET */
20530 #endif
20531
20532 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20533         ipoptlen += ipsec_optlen;
20534 #endif
20535         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
20536             (tp->t_port == 0) &&
20537             ((tp->t_flags & TF_SIGNATURE) == 0) &&
20538             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
20539             ipoptlen == 0)
20540                 tso = 1;
20541         {
20542                 uint32_t outstanding __unused;
20543
20544                 outstanding = tp->snd_max - tp->snd_una;
20545                 if (tp->t_flags & TF_SENTFIN) {
20546                         /*
20547                          * If we sent a fin, snd_max is 1 higher than
20548                          * snd_una
20549                          */
20550                         outstanding--;
20551                 }
20552                 if (sack_rxmit) {
20553                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
20554                                 flags &= ~TH_FIN;
20555                 } else {
20556                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
20557                                    sbused(sb)))
20558                                 flags &= ~TH_FIN;
20559                 }
20560         }
20561         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
20562                       (long)TCP_MAXWIN << tp->rcv_scale);
20563
20564         /*
20565          * Sender silly window avoidance.   We transmit under the following
20566          * conditions when len is non-zero:
20567          *
20568          * - We have a full segment (or more with TSO) - This is the last
20569          * buffer in a write()/send() and we are either idle or running
20570          * NODELAY - we've timed out (e.g. persist timer) - we have more
20571          * then 1/2 the maximum send window's worth of data (receiver may be
20572          * limited the window size) - we need to retransmit
20573          */
20574         if (len) {
20575                 if (len >= segsiz) {
20576                         goto send;
20577                 }
20578                 /*
20579                  * NOTE! on localhost connections an 'ack' from the remote
20580                  * end may occur synchronously with the output and cause us
20581                  * to flush a buffer queued with moretocome.  XXX
20582                  *
20583                  */
20584                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
20585                     (idle || (tp->t_flags & TF_NODELAY)) &&
20586                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20587                     (tp->t_flags & TF_NOPUSH) == 0) {
20588                         pass = 2;
20589                         goto send;
20590                 }
20591                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
20592                         pass = 22;
20593                         goto send;
20594                 }
20595                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
20596                         pass = 4;
20597                         goto send;
20598                 }
20599                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
20600                         pass = 5;
20601                         goto send;
20602                 }
20603                 if (sack_rxmit) {
20604                         pass = 6;
20605                         goto send;
20606                 }
20607                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
20608                     (ctf_outstanding(tp) < (segsiz * 2))) {
20609                         /*
20610                          * We have less than two MSS outstanding (delayed ack)
20611                          * and our rwnd will not let us send a full sized
20612                          * MSS. Lets go ahead and let this small segment
20613                          * out because we want to try to have at least two
20614                          * packets inflight to not be caught by delayed ack.
20615                          */
20616                         pass = 12;
20617                         goto send;
20618                 }
20619         }
20620         /*
20621          * Sending of standalone window updates.
20622          *
20623          * Window updates are important when we close our window due to a
20624          * full socket buffer and are opening it again after the application
20625          * reads data from it.  Once the window has opened again and the
20626          * remote end starts to send again the ACK clock takes over and
20627          * provides the most current window information.
20628          *
20629          * We must avoid the silly window syndrome whereas every read from
20630          * the receive buffer, no matter how small, causes a window update
20631          * to be sent.  We also should avoid sending a flurry of window
20632          * updates when the socket buffer had queued a lot of data and the
20633          * application is doing small reads.
20634          *
20635          * Prevent a flurry of pointless window updates by only sending an
20636          * update when we can increase the advertized window by more than
20637          * 1/4th of the socket buffer capacity.  When the buffer is getting
20638          * full or is very small be more aggressive and send an update
20639          * whenever we can increase by two mss sized segments. In all other
20640          * situations the ACK's to new incoming data will carry further
20641          * window increases.
20642          *
20643          * Don't send an independent window update if a delayed ACK is
20644          * pending (it will get piggy-backed on it) or the remote side
20645          * already has done a half-close and won't send more data.  Skip
20646          * this if the connection is in T/TCP half-open state.
20647          */
20648         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
20649             !(tp->t_flags & TF_DELACK) &&
20650             !TCPS_HAVERCVDFIN(tp->t_state)) {
20651                 /*
20652                  * "adv" is the amount we could increase the window, taking
20653                  * into account that we are limited by TCP_MAXWIN <<
20654                  * tp->rcv_scale.
20655                  */
20656                 int32_t adv;
20657                 int oldwin;
20658
20659                 adv = recwin;
20660                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
20661                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
20662                         if (adv > oldwin)
20663                                 adv -= oldwin;
20664                         else {
20665                                 /* We can't increase the window */
20666                                 adv = 0;
20667                         }
20668                 } else
20669                         oldwin = 0;
20670
20671                 /*
20672                  * If the new window size ends up being the same as or less
20673                  * than the old size when it is scaled, then don't force
20674                  * a window update.
20675                  */
20676                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
20677                         goto dontupdate;
20678
20679                 if (adv >= (int32_t)(2 * segsiz) &&
20680                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
20681                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
20682                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
20683                         pass = 7;
20684                         goto send;
20685                 }
20686                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
20687                         pass = 23;
20688                         goto send;
20689                 }
20690         }
20691 dontupdate:
20692
20693         /*
20694          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
20695          * is also a catch-all for the retransmit timer timeout case.
20696          */
20697         if (tp->t_flags & TF_ACKNOW) {
20698                 pass = 8;
20699                 goto send;
20700         }
20701         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
20702                 pass = 9;
20703                 goto send;
20704         }
20705         /*
20706          * If our state indicates that FIN should be sent and we have not
20707          * yet done so, then we need to send.
20708          */
20709         if ((flags & TH_FIN) &&
20710             (tp->snd_nxt == tp->snd_una)) {
20711                 pass = 11;
20712                 goto send;
20713         }
20714         /*
20715          * No reason to send a segment, just return.
20716          */
20717 just_return:
20718         SOCKBUF_UNLOCK(sb);
20719 just_return_nolock:
20720         {
20721                 int app_limited = CTF_JR_SENT_DATA;
20722
20723                 if (tot_len_this_send > 0) {
20724                         /* Make sure snd_nxt is up to max */
20725                         rack->r_ctl.fsb.recwin = recwin;
20726                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
20727                         if ((error == 0) &&
20728                             rack_use_rfo &&
20729                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
20730                             (ipoptlen == 0) &&
20731                             (tp->snd_nxt == tp->snd_max) &&
20732                             (tp->rcv_numsacks == 0) &&
20733                             rack->r_fsb_inited &&
20734                             TCPS_HAVEESTABLISHED(tp->t_state) &&
20735                             ((IN_RECOVERY(tp->t_flags)) == 0) &&
20736                             (rack->r_must_retran == 0) &&
20737                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
20738                             (len > 0) && (orig_len > 0) &&
20739                             (orig_len > len) &&
20740                             ((orig_len - len) >= segsiz) &&
20741                             ((optlen == 0) ||
20742                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
20743                                 /* We can send at least one more MSS using our fsb */
20744                                 rack_setup_fast_output(tp, rack, sb, len, orig_len,
20745                                                        segsiz, pace_max_seg, hw_tls, flags);
20746                         } else
20747                                 rack->r_fast_output = 0;
20748
20749
20750                         rack_log_fsb(rack, tp, so, flags,
20751                                      ipoptlen, orig_len, len, 0,
20752                                      1, optlen, __LINE__, 1);
20753                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
20754                                 tp->snd_nxt = tp->snd_max;
20755                 } else {
20756                         int end_window = 0;
20757                         uint32_t seq = tp->gput_ack;
20758
20759                         rsm = tqhash_max(rack->r_ctl.tqh);
20760                         if (rsm) {
20761                                 /*
20762                                  * Mark the last sent that we just-returned (hinting
20763                                  * that delayed ack may play a role in any rtt measurement).
20764                                  */
20765                                 rsm->r_just_ret = 1;
20766                         }
20767                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
20768                         rack->r_ctl.rc_agg_delayed = 0;
20769                         rack->r_early = 0;
20770                         rack->r_late = 0;
20771                         rack->r_ctl.rc_agg_early = 0;
20772                         if ((ctf_outstanding(tp) +
20773                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
20774                                  minseg)) >= tp->snd_wnd) {
20775                                 /* We are limited by the rwnd */
20776                                 app_limited = CTF_JR_RWND_LIMITED;
20777                                 if (IN_FASTRECOVERY(tp->t_flags))
20778                                         rack->r_ctl.rc_prr_sndcnt = 0;
20779                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
20780                                 /* We are limited by whats available -- app limited */
20781                                 app_limited = CTF_JR_APP_LIMITED;
20782                                 if (IN_FASTRECOVERY(tp->t_flags))
20783                                         rack->r_ctl.rc_prr_sndcnt = 0;
20784                         } else if ((idle == 0) &&
20785                                    ((tp->t_flags & TF_NODELAY) == 0) &&
20786                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20787                                    (len < segsiz)) {
20788                                 /*
20789                                  * No delay is not on and the
20790                                  * user is sending less than 1MSS. This
20791                                  * brings out SWS avoidance so we
20792                                  * don't send. Another app-limited case.
20793                                  */
20794                                 app_limited = CTF_JR_APP_LIMITED;
20795                         } else if (tp->t_flags & TF_NOPUSH) {
20796                                 /*
20797                                  * The user has requested no push of
20798                                  * the last segment and we are
20799                                  * at the last segment. Another app
20800                                  * limited case.
20801                                  */
20802                                 app_limited = CTF_JR_APP_LIMITED;
20803                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
20804                                 /* Its the cwnd */
20805                                 app_limited = CTF_JR_CWND_LIMITED;
20806                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
20807                                    (rack->rack_no_prr == 0) &&
20808                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
20809                                 app_limited = CTF_JR_PRR;
20810                         } else {
20811                                 /* Now why here are we not sending? */
20812 #ifdef NOW
20813 #ifdef INVARIANTS
20814                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
20815 #endif
20816 #endif
20817                                 app_limited = CTF_JR_ASSESSING;
20818                         }
20819                         /*
20820                          * App limited in some fashion, for our pacing GP
20821                          * measurements we don't want any gap (even cwnd).
20822                          * Close  down the measurement window.
20823                          */
20824                         if (rack_cwnd_block_ends_measure &&
20825                             ((app_limited == CTF_JR_CWND_LIMITED) ||
20826                              (app_limited == CTF_JR_PRR))) {
20827                                 /*
20828                                  * The reason we are not sending is
20829                                  * the cwnd (or prr). We have been configured
20830                                  * to end the measurement window in
20831                                  * this case.
20832                                  */
20833                                 end_window = 1;
20834                         } else if (rack_rwnd_block_ends_measure &&
20835                                    (app_limited == CTF_JR_RWND_LIMITED)) {
20836                                 /*
20837                                  * We are rwnd limited and have been
20838                                  * configured to end the measurement
20839                                  * window in this case.
20840                                  */
20841                                 end_window = 1;
20842                         } else if (app_limited == CTF_JR_APP_LIMITED) {
20843                                 /*
20844                                  * A true application limited period, we have
20845                                  * ran out of data.
20846                                  */
20847                                 end_window = 1;
20848                         } else if (app_limited == CTF_JR_ASSESSING) {
20849                                 /*
20850                                  * In the assessing case we hit the end of
20851                                  * the if/else and had no known reason
20852                                  * This will panic us under invariants..
20853                                  *
20854                                  * If we get this out in logs we need to
20855                                  * investagate which reason we missed.
20856                                  */
20857                                 end_window = 1;
20858                         }
20859                         if (end_window) {
20860                                 uint8_t log = 0;
20861
20862                                 /* Adjust the Gput measurement */
20863                                 if ((tp->t_flags & TF_GPUTINPROG) &&
20864                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
20865                                         tp->gput_ack = tp->snd_max;
20866                                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
20867                                                 /*
20868                                                  * There is not enough to measure.
20869                                                  */
20870                                                 tp->t_flags &= ~TF_GPUTINPROG;
20871                                                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
20872                                                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
20873                                                                            tp->gput_seq,
20874                                                                            0, 0, 18, __LINE__, NULL, 0);
20875                                         } else
20876                                                 log = 1;
20877                                 }
20878                                 /* Mark the last packet has app limited */
20879                                 rsm = tqhash_max(rack->r_ctl.tqh);
20880                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
20881                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
20882                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
20883                                         else {
20884                                                 /*
20885                                                  * Go out to the end app limited and mark
20886                                                  * this new one as next and move the end_appl up
20887                                                  * to this guy.
20888                                                  */
20889                                                 if (rack->r_ctl.rc_end_appl)
20890                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
20891                                                 rack->r_ctl.rc_end_appl = rsm;
20892                                         }
20893                                         rsm->r_flags |= RACK_APP_LIMITED;
20894                                         rack->r_ctl.rc_app_limited_cnt++;
20895                                 }
20896                                 if (log)
20897                                         rack_log_pacing_delay_calc(rack,
20898                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
20899                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
20900                         }
20901                 }
20902                 /* Check if we need to go into persists or not */
20903                 if ((tp->snd_max == tp->snd_una) &&
20904                     TCPS_HAVEESTABLISHED(tp->t_state) &&
20905                     sbavail(sb) &&
20906                     (sbavail(sb) > tp->snd_wnd) &&
20907                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
20908                         /* Yes lets make sure to move to persist before timer-start */
20909                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
20910                 }
20911                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
20912                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
20913         }
20914 #ifdef NETFLIX_SHARED_CWND
20915         if ((sbavail(sb) == 0) &&
20916             rack->r_ctl.rc_scw) {
20917                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20918                 rack->rack_scwnd_is_idle = 1;
20919         }
20920 #endif
20921 #ifdef TCP_ACCOUNTING
20922         if (tot_len_this_send > 0) {
20923                 crtsc = get_cyclecount();
20924                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20925                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
20926                 }
20927                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20928                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
20929                 }
20930                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20931                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
20932                 }
20933         } else {
20934                 crtsc = get_cyclecount();
20935                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20936                         tp->tcp_cnt_counters[SND_LIMITED]++;
20937                 }
20938                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20939                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
20940                 }
20941         }
20942         sched_unpin();
20943 #endif
20944         return (0);
20945
20946 send:
20947         if ((rack->r_ctl.crte != NULL) &&
20948             (rsm == NULL) &&
20949             ((rack->rc_hw_nobuf == 1) ||
20950              (rack_hw_check_queue && (check_done == 0)))) {
20951                 /*
20952                  * We only want to do this once with the hw_check_queue,
20953                  * for the enobuf case we would only do it once if
20954                  * we come around to again, the flag will be clear.
20955                  */
20956                 check_done = 1;
20957                 slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
20958                 if (slot) {
20959                         rack->r_ctl.rc_agg_delayed = 0;
20960                         rack->r_ctl.rc_agg_early = 0;
20961                         rack->r_early = 0;
20962                         rack->r_late = 0;
20963                         SOCKBUF_UNLOCK(&so->so_snd);
20964                         goto skip_all_send;
20965                 }
20966         }
20967         if (rsm || sack_rxmit)
20968                 counter_u64_add(rack_nfto_resend, 1);
20969         else
20970                 counter_u64_add(rack_non_fto_send, 1);
20971         if ((flags & TH_FIN) &&
20972             sbavail(sb)) {
20973                 /*
20974                  * We do not transmit a FIN
20975                  * with data outstanding. We
20976                  * need to make it so all data
20977                  * is acked first.
20978                  */
20979                 flags &= ~TH_FIN;
20980         }
20981         /* Enforce stack imposed max seg size if we have one */
20982         if (rack->r_ctl.rc_pace_max_segs &&
20983             (len > rack->r_ctl.rc_pace_max_segs)) {
20984                 mark = 1;
20985                 len = rack->r_ctl.rc_pace_max_segs;
20986         }
20987         SOCKBUF_LOCK_ASSERT(sb);
20988         if (len > 0) {
20989                 if (len >= segsiz)
20990                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
20991                 else
20992                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
20993         }
20994         /*
20995          * Before ESTABLISHED, force sending of initial options unless TCP
20996          * set not to do any options. NOTE: we assume that the IP/TCP header
20997          * plus TCP options always fit in a single mbuf, leaving room for a
20998          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
20999          * + optlen <= MCLBYTES
21000          */
21001         optlen = 0;
21002 #ifdef INET6
21003         if (isipv6)
21004                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
21005         else
21006 #endif
21007                 hdrlen = sizeof(struct tcpiphdr);
21008
21009         /*
21010          * Compute options for segment. We only have to care about SYN and
21011          * established connection segments.  Options for SYN-ACK segments
21012          * are handled in TCP syncache.
21013          */
21014         to.to_flags = 0;
21015         if ((tp->t_flags & TF_NOOPT) == 0) {
21016                 /* Maximum segment size. */
21017                 if (flags & TH_SYN) {
21018                         tp->snd_nxt = tp->iss;
21019                         to.to_mss = tcp_mssopt(&inp->inp_inc);
21020                         if (tp->t_port)
21021                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
21022                         to.to_flags |= TOF_MSS;
21023
21024                         /*
21025                          * On SYN or SYN|ACK transmits on TFO connections,
21026                          * only include the TFO option if it is not a
21027                          * retransmit, as the presence of the TFO option may
21028                          * have caused the original SYN or SYN|ACK to have
21029                          * been dropped by a middlebox.
21030                          */
21031                         if (IS_FASTOPEN(tp->t_flags) &&
21032                             (tp->t_rxtshift == 0)) {
21033                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
21034                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
21035                                         to.to_tfo_cookie =
21036                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
21037                                         to.to_flags |= TOF_FASTOPEN;
21038                                         wanted_cookie = 1;
21039                                 } else if (tp->t_state == TCPS_SYN_SENT) {
21040                                         to.to_tfo_len =
21041                                                 tp->t_tfo_client_cookie_len;
21042                                         to.to_tfo_cookie =
21043                                                 tp->t_tfo_cookie.client;
21044                                         to.to_flags |= TOF_FASTOPEN;
21045                                         wanted_cookie = 1;
21046                                         /*
21047                                          * If we wind up having more data to
21048                                          * send with the SYN than can fit in
21049                                          * one segment, don't send any more
21050                                          * until the SYN|ACK comes back from
21051                                          * the other end.
21052                                          */
21053                                         sendalot = 0;
21054                                 }
21055                         }
21056                 }
21057                 /* Window scaling. */
21058                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
21059                         to.to_wscale = tp->request_r_scale;
21060                         to.to_flags |= TOF_SCALE;
21061                 }
21062                 /* Timestamps. */
21063                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
21064                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
21065                         to.to_tsval = ms_cts + tp->ts_offset;
21066                         to.to_tsecr = tp->ts_recent;
21067                         to.to_flags |= TOF_TS;
21068                 }
21069                 /* Set receive buffer autosizing timestamp. */
21070                 if (tp->rfbuf_ts == 0 &&
21071                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
21072                         tp->rfbuf_ts = tcp_ts_getticks();
21073                 /* Selective ACK's. */
21074                 if (tp->t_flags & TF_SACK_PERMIT) {
21075                         if (flags & TH_SYN)
21076                                 to.to_flags |= TOF_SACKPERM;
21077                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21078                                  tp->rcv_numsacks > 0) {
21079                                 to.to_flags |= TOF_SACK;
21080                                 to.to_nsacks = tp->rcv_numsacks;
21081                                 to.to_sacks = (u_char *)tp->sackblks;
21082                         }
21083                 }
21084 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21085                 /* TCP-MD5 (RFC2385). */
21086                 if (tp->t_flags & TF_SIGNATURE)
21087                         to.to_flags |= TOF_SIGNATURE;
21088 #endif                          /* TCP_SIGNATURE */
21089
21090                 /* Processing the options. */
21091                 hdrlen += optlen = tcp_addoptions(&to, opt);
21092                 /*
21093                  * If we wanted a TFO option to be added, but it was unable
21094                  * to fit, ensure no data is sent.
21095                  */
21096                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
21097                     !(to.to_flags & TOF_FASTOPEN))
21098                         len = 0;
21099         }
21100         if (tp->t_port) {
21101                 if (V_tcp_udp_tunneling_port == 0) {
21102                         /* The port was removed?? */
21103                         SOCKBUF_UNLOCK(&so->so_snd);
21104 #ifdef TCP_ACCOUNTING
21105                         crtsc = get_cyclecount();
21106                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21107                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
21108                         }
21109                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21110                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
21111                         }
21112                         sched_unpin();
21113 #endif
21114                         return (EHOSTUNREACH);
21115                 }
21116                 hdrlen += sizeof(struct udphdr);
21117         }
21118 #ifdef INET6
21119         if (isipv6)
21120                 ipoptlen = ip6_optlen(inp);
21121         else
21122 #endif
21123                 if (inp->inp_options)
21124                         ipoptlen = inp->inp_options->m_len -
21125                                 offsetof(struct ipoption, ipopt_list);
21126                 else
21127                         ipoptlen = 0;
21128 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21129         ipoptlen += ipsec_optlen;
21130 #endif
21131
21132         /*
21133          * Adjust data length if insertion of options will bump the packet
21134          * length beyond the t_maxseg length. Clear the FIN bit because we
21135          * cut off the tail of the segment.
21136          */
21137         if (len + optlen + ipoptlen > tp->t_maxseg) {
21138                 if (tso) {
21139                         uint32_t if_hw_tsomax;
21140                         uint32_t moff;
21141                         int32_t max_len;
21142
21143                         /* extract TSO information */
21144                         if_hw_tsomax = tp->t_tsomax;
21145                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
21146                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
21147                         KASSERT(ipoptlen == 0,
21148                                 ("%s: TSO can't do IP options", __func__));
21149
21150                         /*
21151                          * Check if we should limit by maximum payload
21152                          * length:
21153                          */
21154                         if (if_hw_tsomax != 0) {
21155                                 /* compute maximum TSO length */
21156                                 max_len = (if_hw_tsomax - hdrlen -
21157                                            max_linkhdr);
21158                                 if (max_len <= 0) {
21159                                         len = 0;
21160                                 } else if (len > max_len) {
21161                                         sendalot = 1;
21162                                         len = max_len;
21163                                         mark = 2;
21164                                 }
21165                         }
21166                         /*
21167                          * Prevent the last segment from being fractional
21168                          * unless the send sockbuf can be emptied:
21169                          */
21170                         max_len = (tp->t_maxseg - optlen);
21171                         if ((sb_offset + len) < sbavail(sb)) {
21172                                 moff = len % (u_int)max_len;
21173                                 if (moff != 0) {
21174                                         mark = 3;
21175                                         len -= moff;
21176                                 }
21177                         }
21178                         /*
21179                          * In case there are too many small fragments don't
21180                          * use TSO:
21181                          */
21182                         if (len <= max_len) {
21183                                 mark = 4;
21184                                 tso = 0;
21185                         }
21186                         /*
21187                          * Send the FIN in a separate segment after the bulk
21188                          * sending is done. We don't trust the TSO
21189                          * implementations to clear the FIN flag on all but
21190                          * the last segment.
21191                          */
21192                         if (tp->t_flags & TF_NEEDFIN) {
21193                                 sendalot = 4;
21194                         }
21195                 } else {
21196                         mark = 5;
21197                         if (optlen + ipoptlen >= tp->t_maxseg) {
21198                                 /*
21199                                  * Since we don't have enough space to put
21200                                  * the IP header chain and the TCP header in
21201                                  * one packet as required by RFC 7112, don't
21202                                  * send it. Also ensure that at least one
21203                                  * byte of the payload can be put into the
21204                                  * TCP segment.
21205                                  */
21206                                 SOCKBUF_UNLOCK(&so->so_snd);
21207                                 error = EMSGSIZE;
21208                                 sack_rxmit = 0;
21209                                 goto out;
21210                         }
21211                         len = tp->t_maxseg - optlen - ipoptlen;
21212                         sendalot = 5;
21213                 }
21214         } else {
21215                 tso = 0;
21216                 mark = 6;
21217         }
21218         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
21219                 ("%s: len > IP_MAXPACKET", __func__));
21220 #ifdef DIAGNOSTIC
21221 #ifdef INET6
21222         if (max_linkhdr + hdrlen > MCLBYTES)
21223 #else
21224                 if (max_linkhdr + hdrlen > MHLEN)
21225 #endif
21226                         panic("tcphdr too big");
21227 #endif
21228
21229         /*
21230          * This KASSERT is here to catch edge cases at a well defined place.
21231          * Before, those had triggered (random) panic conditions further
21232          * down.
21233          */
21234         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21235         if ((len == 0) &&
21236             (flags & TH_FIN) &&
21237             (sbused(sb))) {
21238                 /*
21239                  * We have outstanding data, don't send a fin by itself!.
21240                  */
21241                 goto just_return;
21242         }
21243         /*
21244          * Grab a header mbuf, attaching a copy of data to be transmitted,
21245          * and initialize the header from the template for sends on this
21246          * connection.
21247          */
21248         hw_tls = tp->t_nic_ktls_xmit != 0;
21249         if (len) {
21250                 uint32_t max_val;
21251                 uint32_t moff;
21252
21253                 if (rack->r_ctl.rc_pace_max_segs)
21254                         max_val = rack->r_ctl.rc_pace_max_segs;
21255                 else if (rack->rc_user_set_max_segs)
21256                         max_val = rack->rc_user_set_max_segs * segsiz;
21257                 else
21258                         max_val = len;
21259                 /*
21260                  * We allow a limit on sending with hptsi.
21261                  */
21262                 if (len > max_val) {
21263                         mark = 7;
21264                         len = max_val;
21265                 }
21266 #ifdef INET6
21267                 if (MHLEN < hdrlen + max_linkhdr)
21268                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
21269                 else
21270 #endif
21271                         m = m_gethdr(M_NOWAIT, MT_DATA);
21272
21273                 if (m == NULL) {
21274                         SOCKBUF_UNLOCK(sb);
21275                         error = ENOBUFS;
21276                         sack_rxmit = 0;
21277                         goto out;
21278                 }
21279                 m->m_data += max_linkhdr;
21280                 m->m_len = hdrlen;
21281
21282                 /*
21283                  * Start the m_copy functions from the closest mbuf to the
21284                  * sb_offset in the socket buffer chain.
21285                  */
21286                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
21287                 s_mb = mb;
21288                 s_moff = moff;
21289                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
21290                         m_copydata(mb, moff, (int)len,
21291                                    mtod(m, caddr_t)+hdrlen);
21292                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
21293                                 sbsndptr_adv(sb, mb, len);
21294                         m->m_len += len;
21295                 } else {
21296                         struct sockbuf *msb;
21297
21298                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
21299                                 msb = NULL;
21300                         else
21301                                 msb = sb;
21302                         m->m_next = tcp_m_copym(
21303                                 mb, moff, &len,
21304                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
21305                                 ((rsm == NULL) ? hw_tls : 0)
21306 #ifdef NETFLIX_COPY_ARGS
21307                                 , &s_mb, &s_moff
21308 #endif
21309                                 );
21310                         if (len <= (tp->t_maxseg - optlen)) {
21311                                 /*
21312                                  * Must have ran out of mbufs for the copy
21313                                  * shorten it to no longer need tso. Lets
21314                                  * not put on sendalot since we are low on
21315                                  * mbufs.
21316                                  */
21317                                 tso = 0;
21318                         }
21319                         if (m->m_next == NULL) {
21320                                 SOCKBUF_UNLOCK(sb);
21321                                 (void)m_free(m);
21322                                 error = ENOBUFS;
21323                                 sack_rxmit = 0;
21324                                 goto out;
21325                         }
21326                 }
21327                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
21328                         if (rsm && (rsm->r_flags & RACK_TLP)) {
21329                                 /*
21330                                  * TLP should not count in retran count, but
21331                                  * in its own bin
21332                                  */
21333                                 counter_u64_add(rack_tlp_retran, 1);
21334                                 counter_u64_add(rack_tlp_retran_bytes, len);
21335                         } else {
21336                                 tp->t_sndrexmitpack++;
21337                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
21338                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
21339                         }
21340 #ifdef STATS
21341                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
21342                                                  len);
21343 #endif
21344                 } else {
21345                         KMOD_TCPSTAT_INC(tcps_sndpack);
21346                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
21347 #ifdef STATS
21348                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
21349                                                  len);
21350 #endif
21351                 }
21352                 /*
21353                  * If we're sending everything we've got, set PUSH. (This
21354                  * will keep happy those implementations which only give
21355                  * data to the user when a buffer fills or a PUSH comes in.)
21356                  */
21357                 if (sb_offset + len == sbused(sb) &&
21358                     sbused(sb) &&
21359                     !(flags & TH_SYN)) {
21360                         flags |= TH_PUSH;
21361                         add_flag |= RACK_HAD_PUSH;
21362                 }
21363
21364                 SOCKBUF_UNLOCK(sb);
21365         } else {
21366                 SOCKBUF_UNLOCK(sb);
21367                 if (tp->t_flags & TF_ACKNOW)
21368                         KMOD_TCPSTAT_INC(tcps_sndacks);
21369                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
21370                         KMOD_TCPSTAT_INC(tcps_sndctrl);
21371                 else
21372                         KMOD_TCPSTAT_INC(tcps_sndwinup);
21373
21374                 m = m_gethdr(M_NOWAIT, MT_DATA);
21375                 if (m == NULL) {
21376                         error = ENOBUFS;
21377                         sack_rxmit = 0;
21378                         goto out;
21379                 }
21380 #ifdef INET6
21381                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
21382                     MHLEN >= hdrlen) {
21383                         M_ALIGN(m, hdrlen);
21384                 } else
21385 #endif
21386                         m->m_data += max_linkhdr;
21387                 m->m_len = hdrlen;
21388         }
21389         SOCKBUF_UNLOCK_ASSERT(sb);
21390         m->m_pkthdr.rcvif = (struct ifnet *)0;
21391 #ifdef MAC
21392         mac_inpcb_create_mbuf(inp, m);
21393 #endif
21394         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
21395 #ifdef INET6
21396                 if (isipv6)
21397                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
21398                 else
21399 #endif                          /* INET6 */
21400 #ifdef INET
21401                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
21402 #endif
21403                 th = rack->r_ctl.fsb.th;
21404                 udp = rack->r_ctl.fsb.udp;
21405                 if (udp) {
21406 #ifdef INET6
21407                         if (isipv6)
21408                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
21409                         else
21410 #endif                          /* INET6 */
21411                                 ulen = hdrlen + len - sizeof(struct ip);
21412                         udp->uh_ulen = htons(ulen);
21413                 }
21414         } else {
21415 #ifdef INET6
21416                 if (isipv6) {
21417                         ip6 = mtod(m, struct ip6_hdr *);
21418                         if (tp->t_port) {
21419                                 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
21420                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21421                                 udp->uh_dport = tp->t_port;
21422                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
21423                                 udp->uh_ulen = htons(ulen);
21424                                 th = (struct tcphdr *)(udp + 1);
21425                         } else
21426                                 th = (struct tcphdr *)(ip6 + 1);
21427                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
21428                 } else
21429 #endif                          /* INET6 */
21430                 {
21431 #ifdef INET
21432                         ip = mtod(m, struct ip *);
21433                         if (tp->t_port) {
21434                                 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
21435                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21436                                 udp->uh_dport = tp->t_port;
21437                                 ulen = hdrlen + len - sizeof(struct ip);
21438                                 udp->uh_ulen = htons(ulen);
21439                                 th = (struct tcphdr *)(udp + 1);
21440                         } else
21441                                 th = (struct tcphdr *)(ip + 1);
21442                         tcpip_fillheaders(inp, tp->t_port, ip, th);
21443 #endif
21444                 }
21445         }
21446         /*
21447          * Fill in fields, remembering maximum advertised window for use in
21448          * delaying messages about window sizes. If resending a FIN, be sure
21449          * not to use a new sequence number.
21450          */
21451         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
21452             tp->snd_nxt == tp->snd_max)
21453                 tp->snd_nxt--;
21454         /*
21455          * If we are starting a connection, send ECN setup SYN packet. If we
21456          * are on a retransmit, we may resend those bits a number of times
21457          * as per RFC 3168.
21458          */
21459         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
21460                 flags |= tcp_ecn_output_syn_sent(tp);
21461         }
21462         /* Also handle parallel SYN for ECN */
21463         if (TCPS_HAVERCVDSYN(tp->t_state) &&
21464             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
21465                 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
21466                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
21467                     (tp->t_flags2 & TF2_ECN_SND_ECE))
21468                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
21469 #ifdef INET6
21470                 if (isipv6) {
21471                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
21472                         ip6->ip6_flow |= htonl(ect << 20);
21473                 }
21474                 else
21475 #endif
21476                 {
21477 #ifdef INET
21478                         ip->ip_tos &= ~IPTOS_ECN_MASK;
21479                         ip->ip_tos |= ect;
21480 #endif
21481                 }
21482         }
21483         /*
21484          * If we are doing retransmissions, then snd_nxt will not reflect
21485          * the first unsent octet.  For ACK only packets, we do not want the
21486          * sequence number of the retransmitted packet, we want the sequence
21487          * number of the next unsent octet.  So, if there is no data (and no
21488          * SYN or FIN), use snd_max instead of snd_nxt when filling in
21489          * ti_seq.  But if we are in persist state, snd_max might reflect
21490          * one byte beyond the right edge of the window, so use snd_nxt in
21491          * that case, since we know we aren't doing a retransmission.
21492          * (retransmit and persist are mutually exclusive...)
21493          */
21494         if (sack_rxmit == 0) {
21495                 if (len || (flags & (TH_SYN | TH_FIN))) {
21496                         th->th_seq = htonl(tp->snd_nxt);
21497                         rack_seq = tp->snd_nxt;
21498                 } else {
21499                         th->th_seq = htonl(tp->snd_max);
21500                         rack_seq = tp->snd_max;
21501                 }
21502         } else {
21503                 th->th_seq = htonl(rsm->r_start);
21504                 rack_seq = rsm->r_start;
21505         }
21506         th->th_ack = htonl(tp->rcv_nxt);
21507         tcp_set_flags(th, flags);
21508         /*
21509          * Calculate receive window.  Don't shrink window, but avoid silly
21510          * window syndrome.
21511          * If a RST segment is sent, advertise a window of zero.
21512          */
21513         if (flags & TH_RST) {
21514                 recwin = 0;
21515         } else {
21516                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
21517                     recwin < (long)segsiz) {
21518                         recwin = 0;
21519                 }
21520                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
21521                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
21522                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
21523         }
21524
21525         /*
21526          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
21527          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
21528          * handled in syncache.
21529          */
21530         if (flags & TH_SYN)
21531                 th->th_win = htons((u_short)
21532                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
21533         else {
21534                 /* Avoid shrinking window with window scaling. */
21535                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
21536                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
21537         }
21538         /*
21539          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
21540          * window.  This may cause the remote transmitter to stall.  This
21541          * flag tells soreceive() to disable delayed acknowledgements when
21542          * draining the buffer.  This can occur if the receiver is
21543          * attempting to read more data than can be buffered prior to
21544          * transmitting on the connection.
21545          */
21546         if (th->th_win == 0) {
21547                 tp->t_sndzerowin++;
21548                 tp->t_flags |= TF_RXWIN0SENT;
21549         } else
21550                 tp->t_flags &= ~TF_RXWIN0SENT;
21551         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
21552         /* Now are we using fsb?, if so copy the template data to the mbuf */
21553         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21554                 uint8_t *cpto;
21555
21556                 cpto = mtod(m, uint8_t *);
21557                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
21558                 /*
21559                  * We have just copied in:
21560                  * IP/IP6
21561                  * <optional udphdr>
21562                  * tcphdr (no options)
21563                  *
21564                  * We need to grab the correct pointers into the mbuf
21565                  * for both the tcp header, and possibly the udp header (if tunneling).
21566                  * We do this by using the offset in the copy buffer and adding it
21567                  * to the mbuf base pointer (cpto).
21568                  */
21569 #ifdef INET6
21570                 if (isipv6)
21571                         ip6 = mtod(m, struct ip6_hdr *);
21572                 else
21573 #endif                          /* INET6 */
21574 #ifdef INET
21575                         ip = mtod(m, struct ip *);
21576 #endif
21577                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
21578                 /* If we have a udp header lets set it into the mbuf as well */
21579                 if (udp)
21580                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
21581         }
21582 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21583         if (to.to_flags & TOF_SIGNATURE) {
21584                 /*
21585                  * Calculate MD5 signature and put it into the place
21586                  * determined before.
21587                  * NOTE: since TCP options buffer doesn't point into
21588                  * mbuf's data, calculate offset and use it.
21589                  */
21590                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
21591                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
21592                         /*
21593                          * Do not send segment if the calculation of MD5
21594                          * digest has failed.
21595                          */
21596                         goto out;
21597                 }
21598         }
21599 #endif
21600         if (optlen) {
21601                 bcopy(opt, th + 1, optlen);
21602                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
21603         }
21604         /*
21605          * Put TCP length in extended header, and then checksum extended
21606          * header and data.
21607          */
21608         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
21609 #ifdef INET6
21610         if (isipv6) {
21611                 /*
21612                  * ip6_plen is not need to be filled now, and will be filled
21613                  * in ip6_output.
21614                  */
21615                 if (tp->t_port) {
21616                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
21617                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21618                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
21619                         th->th_sum = htons(0);
21620                         UDPSTAT_INC(udps_opackets);
21621                 } else {
21622                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
21623                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21624                         th->th_sum = in6_cksum_pseudo(ip6,
21625                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
21626                                                       0);
21627                 }
21628         }
21629 #endif
21630 #if defined(INET6) && defined(INET)
21631         else
21632 #endif
21633 #ifdef INET
21634         {
21635                 if (tp->t_port) {
21636                         m->m_pkthdr.csum_flags = CSUM_UDP;
21637                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21638                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
21639                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
21640                         th->th_sum = htons(0);
21641                         UDPSTAT_INC(udps_opackets);
21642                 } else {
21643                         m->m_pkthdr.csum_flags = CSUM_TCP;
21644                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21645                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
21646                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
21647                                                                         IPPROTO_TCP + len + optlen));
21648                 }
21649                 /* IP version must be set here for ipv4/ipv6 checking later */
21650                 KASSERT(ip->ip_v == IPVERSION,
21651                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
21652         }
21653 #endif
21654         /*
21655          * Enable TSO and specify the size of the segments. The TCP pseudo
21656          * header checksum is always provided. XXX: Fixme: This is currently
21657          * not the case for IPv6.
21658          */
21659         if (tso) {
21660                 /*
21661                  * Here we must use t_maxseg and the optlen since
21662                  * the optlen may include SACK's (or DSACK).
21663                  */
21664                 KASSERT(len > tp->t_maxseg - optlen,
21665                         ("%s: len <= tso_segsz", __func__));
21666                 m->m_pkthdr.csum_flags |= CSUM_TSO;
21667                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
21668         }
21669         KASSERT(len + hdrlen == m_length(m, NULL),
21670                 ("%s: mbuf chain different than expected: %d + %u != %u",
21671                  __func__, len, hdrlen, m_length(m, NULL)));
21672
21673 #ifdef TCP_HHOOK
21674         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
21675         hhook_run_tcp_est_out(tp, th, &to, len, tso);
21676 #endif
21677         if ((rack->r_ctl.crte != NULL) &&
21678             (rack->rc_hw_nobuf == 0) &&
21679             tcp_bblogging_on(tp)) {
21680                 rack_log_queue_level(tp, rack, len, &tv, cts);
21681         }
21682         /* We're getting ready to send; log now. */
21683         if (tcp_bblogging_on(rack->rc_tp)) {
21684                 union tcp_log_stackspecific log;
21685
21686                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
21687                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
21688                 if (rack->rack_no_prr)
21689                         log.u_bbr.flex1 = 0;
21690                 else
21691                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
21692                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
21693                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
21694                 log.u_bbr.flex4 = orig_len;
21695                 /* Save off the early/late values */
21696                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
21697                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
21698                 log.u_bbr.bw_inuse = rack_get_bw(rack);
21699                 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
21700                 log.u_bbr.flex8 = 0;
21701                 if (rsm) {
21702                         if (rsm->r_flags & RACK_RWND_COLLAPSED) {
21703                                 rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
21704                                 counter_u64_add(rack_collapsed_win_rxt, 1);
21705                                 counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
21706                         }
21707                         if (doing_tlp)
21708                                 log.u_bbr.flex8 = 2;
21709                         else
21710                                 log.u_bbr.flex8 = 1;
21711                 } else {
21712                         if (doing_tlp)
21713                                 log.u_bbr.flex8 = 3;
21714                 }
21715                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
21716                 log.u_bbr.flex7 = mark;
21717                 log.u_bbr.flex7 <<= 8;
21718                 log.u_bbr.flex7 |= pass;
21719                 log.u_bbr.pkts_out = tp->t_maxseg;
21720                 log.u_bbr.timeStamp = cts;
21721                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
21722                 if (rsm && (rsm->r_rtr_cnt > 0)) {
21723                         /*
21724                          * When we have a retransmit we want to log the
21725                          * burst at send and flight at send from before.
21726                          */
21727                         log.u_bbr.flex5 = rsm->r_fas;
21728                         log.u_bbr.bbr_substate = rsm->r_bas;
21729                 } else {
21730                         /*
21731                          * New transmits we log in flex5 the inflight again as
21732                          * well as the number of segments in our send in the
21733                          * substate field.
21734                          */
21735                         log.u_bbr.flex5 = log.u_bbr.inflight;
21736                         log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
21737                 }
21738                 log.u_bbr.lt_epoch = cwnd_to_use;
21739                 log.u_bbr.delivered = sendalot;
21740                 log.u_bbr.rttProp = (uint64_t)rsm;
21741                 log.u_bbr.pkt_epoch = __LINE__;
21742                 if (rsm) {
21743                         log.u_bbr.delRate = rsm->r_flags;
21744                         log.u_bbr.delRate <<= 31;
21745                         log.u_bbr.delRate |= rack->r_must_retran;
21746                         log.u_bbr.delRate <<= 1;
21747                         log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21748                 } else {
21749                         log.u_bbr.delRate = rack->r_must_retran;
21750                         log.u_bbr.delRate <<= 1;
21751                         log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21752                 }
21753                 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
21754                                     len, &log, false, NULL, __func__, __LINE__, &tv);
21755         } else
21756                 lgb = NULL;
21757
21758         /*
21759          * Fill in IP length and desired time to live and send to IP level.
21760          * There should be a better way to handle ttl and tos; we could keep
21761          * them in the template, but need a way to checksum without them.
21762          */
21763         /*
21764          * m->m_pkthdr.len should have been set before cksum calcuration,
21765          * because in6_cksum() need it.
21766          */
21767 #ifdef INET6
21768         if (isipv6) {
21769                 /*
21770                  * we separately set hoplimit for every segment, since the
21771                  * user might want to change the value via setsockopt. Also,
21772                  * desired default hop limit might be changed via Neighbor
21773                  * Discovery.
21774                  */
21775                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
21776
21777                 /*
21778                  * Set the packet size here for the benefit of DTrace
21779                  * probes. ip6_output() will set it properly; it's supposed
21780                  * to include the option header lengths as well.
21781                  */
21782                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
21783
21784                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
21785                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
21786                 else
21787                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
21788
21789                 if (tp->t_state == TCPS_SYN_SENT)
21790                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
21791
21792                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
21793                 /* TODO: IPv6 IP6TOS_ECT bit on */
21794                 error = ip6_output(m,
21795                                    inp->in6p_outputopts,
21796                                    &inp->inp_route6,
21797                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
21798                                    NULL, NULL, inp);
21799
21800                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
21801                         mtu = inp->inp_route6.ro_nh->nh_mtu;
21802         }
21803 #endif                          /* INET6 */
21804 #if defined(INET) && defined(INET6)
21805         else
21806 #endif
21807 #ifdef INET
21808         {
21809                 ip->ip_len = htons(m->m_pkthdr.len);
21810 #ifdef INET6
21811                 if (inp->inp_vflag & INP_IPV6PROTO)
21812                         ip->ip_ttl = in6_selecthlim(inp, NULL);
21813 #endif                          /* INET6 */
21814                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
21815                 /*
21816                  * If we do path MTU discovery, then we set DF on every
21817                  * packet. This might not be the best thing to do according
21818                  * to RFC3390 Section 2. However the tcp hostcache migitates
21819                  * the problem so it affects only the first tcp connection
21820                  * with a host.
21821                  *
21822                  * NB: Don't set DF on small MTU/MSS to have a safe
21823                  * fallback.
21824                  */
21825                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
21826                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
21827                         if (tp->t_port == 0 || len < V_tcp_minmss) {
21828                                 ip->ip_off |= htons(IP_DF);
21829                         }
21830                 } else {
21831                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
21832                 }
21833
21834                 if (tp->t_state == TCPS_SYN_SENT)
21835                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
21836
21837                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
21838
21839                 error = ip_output(m,
21840 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21841                                   inp->inp_options,
21842 #else
21843                                   NULL,
21844 #endif
21845                                   &inp->inp_route,
21846                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
21847                                   inp);
21848                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
21849                         mtu = inp->inp_route.ro_nh->nh_mtu;
21850         }
21851 #endif                          /* INET */
21852
21853 out:
21854         if (lgb) {
21855                 lgb->tlb_errno = error;
21856                 lgb = NULL;
21857         }
21858         /*
21859          * In transmit state, time the transmission and arrange for the
21860          * retransmit.  In persist state, just set snd_max.
21861          */
21862         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
21863                         rack_to_usec_ts(&tv),
21864                         rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
21865         if (error == 0) {
21866                 if (rsm == NULL) {
21867                         if (rack->lt_bw_up == 0) {
21868                                 rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
21869                                 rack->r_ctl.lt_seq = tp->snd_una;
21870                                 rack->lt_bw_up = 1;
21871                         } else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
21872                                 /*
21873                                  * Need to record what we have since we are
21874                                  * approaching seq wrap.
21875                                  */
21876                                 uint64_t tmark;
21877
21878                                 rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
21879                                 rack->r_ctl.lt_seq = tp->snd_una;
21880                                 tmark = tcp_tv_to_lusectick(&tv);
21881                                 rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
21882                                 rack->r_ctl.lt_timemark = tmark;
21883                         }
21884                 }
21885                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
21886                 counter_u64_add(rack_total_bytes, len);
21887                 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
21888                 if (rsm && doing_tlp) {
21889                         rack->rc_last_sent_tlp_past_cumack = 0;
21890                         rack->rc_last_sent_tlp_seq_valid = 1;
21891                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
21892                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
21893                 }
21894                 if (rack->rc_hw_nobuf) {
21895                         rack->rc_hw_nobuf = 0;
21896                         rack->r_ctl.rc_agg_delayed = 0;
21897                         rack->r_early = 0;
21898                         rack->r_late = 0;
21899                         rack->r_ctl.rc_agg_early = 0;
21900                 }
21901                 if (rsm && (doing_tlp == 0)) {
21902                         /* Set we retransmitted */
21903                         rack->rc_gp_saw_rec = 1;
21904                 } else {
21905                         if (cwnd_to_use > tp->snd_ssthresh) {
21906                                 /* Set we sent in CA */
21907                                 rack->rc_gp_saw_ca = 1;
21908                         } else {
21909                                 /* Set we sent in SS */
21910                                 rack->rc_gp_saw_ss = 1;
21911                         }
21912                 }
21913                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21914                     (tp->t_flags & TF_SACK_PERMIT) &&
21915                     tp->rcv_numsacks > 0)
21916                         tcp_clean_dsack_blocks(tp);
21917                 tot_len_this_send += len;
21918                 if (len == 0) {
21919                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
21920                 } else {
21921                         int idx;
21922
21923                         idx = (len / segsiz) + 3;
21924                         if (idx >= TCP_MSS_ACCT_ATIMER)
21925                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
21926                         else
21927                                 counter_u64_add(rack_out_size[idx], 1);
21928                 }
21929         }
21930         if ((rack->rack_no_prr == 0) &&
21931             sub_from_prr &&
21932             (error == 0)) {
21933                 if (rack->r_ctl.rc_prr_sndcnt >= len)
21934                         rack->r_ctl.rc_prr_sndcnt -= len;
21935                 else
21936                         rack->r_ctl.rc_prr_sndcnt = 0;
21937         }
21938         sub_from_prr = 0;
21939         if (doing_tlp) {
21940                 /* Make sure the TLP is added */
21941                 add_flag |= RACK_TLP;
21942         } else if (rsm) {
21943                 /* If its a resend without TLP then it must not have the flag */
21944                 rsm->r_flags &= ~RACK_TLP;
21945         }
21946
21947
21948         if ((error == 0) &&
21949             (len > 0) &&
21950             (tp->snd_una == tp->snd_max))
21951                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
21952         {
21953                 tcp_seq startseq = tp->snd_nxt;
21954
21955                 /* Track our lost count */
21956                 if (rsm && (doing_tlp == 0))
21957                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
21958                 /*
21959                  * Advance snd_nxt over sequence space of this segment.
21960                  */
21961                 if (error)
21962                         /* We don't log or do anything with errors */
21963                         goto nomore;
21964                 if (doing_tlp == 0) {
21965                         if (rsm == NULL) {
21966                                 /*
21967                                  * Not a retransmission of some
21968                                  * sort, new data is going out so
21969                                  * clear our TLP count and flag.
21970                                  */
21971                                 rack->rc_tlp_in_progress = 0;
21972                                 rack->r_ctl.rc_tlp_cnt_out = 0;
21973                         }
21974                 } else {
21975                         /*
21976                          * We have just sent a TLP, mark that it is true
21977                          * and make sure our in progress is set so we
21978                          * continue to check the count.
21979                          */
21980                         rack->rc_tlp_in_progress = 1;
21981                         rack->r_ctl.rc_tlp_cnt_out++;
21982                 }
21983                 if (flags & (TH_SYN | TH_FIN)) {
21984                         if (flags & TH_SYN)
21985                                 tp->snd_nxt++;
21986                         if (flags & TH_FIN) {
21987                                 tp->snd_nxt++;
21988                                 tp->t_flags |= TF_SENTFIN;
21989                         }
21990                 }
21991                 /* In the ENOBUFS case we do *not* update snd_max */
21992                 if (sack_rxmit)
21993                         goto nomore;
21994
21995                 tp->snd_nxt += len;
21996                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
21997                         if (tp->snd_una == tp->snd_max) {
21998                                 /*
21999                                  * Update the time we just added data since
22000                                  * none was outstanding.
22001                                  */
22002                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
22003                                 tp->t_acktime = ticks;
22004                         }
22005                         tp->snd_max = tp->snd_nxt;
22006                         if (rack->rc_new_rnd_needed) {
22007                                 /*
22008                                  * Update the rnd to start ticking not
22009                                  * that from a time perspective all of
22010                                  * the preceding idle time is "in the round"
22011                                  */
22012                                 rack->rc_new_rnd_needed = 0;
22013                                 rack->r_ctl.roundends = tp->snd_max;
22014                         }
22015                         /*
22016                          * Time this transmission if not a retransmission and
22017                          * not currently timing anything.
22018                          * This is only relevant in case of switching back to
22019                          * the base stack.
22020                          */
22021                         if (tp->t_rtttime == 0) {
22022                                 tp->t_rtttime = ticks;
22023                                 tp->t_rtseq = startseq;
22024                                 KMOD_TCPSTAT_INC(tcps_segstimed);
22025                         }
22026                         if (len &&
22027                             ((tp->t_flags & TF_GPUTINPROG) == 0))
22028                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
22029                 }
22030                 /*
22031                  * If we are doing FO we need to update the mbuf position and subtract
22032                  * this happens when the peer sends us duplicate information and
22033                  * we thus want to send a DSACK.
22034                  *
22035                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
22036                  * turned off? If not then we are going to echo multiple DSACK blocks
22037                  * out (with the TSO), which we should not be doing.
22038                  */
22039                 if (rack->r_fast_output && len) {
22040                         if (rack->r_ctl.fsb.left_to_send > len)
22041                                 rack->r_ctl.fsb.left_to_send -= len;
22042                         else
22043                                 rack->r_ctl.fsb.left_to_send = 0;
22044                         if (rack->r_ctl.fsb.left_to_send < segsiz)
22045                                 rack->r_fast_output = 0;
22046                         if (rack->r_fast_output) {
22047                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
22048                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
22049                                 rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
22050                         }
22051                 }
22052         }
22053 nomore:
22054         if (error) {
22055                 rack->r_ctl.rc_agg_delayed = 0;
22056                 rack->r_early = 0;
22057                 rack->r_late = 0;
22058                 rack->r_ctl.rc_agg_early = 0;
22059                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
22060                 /*
22061                  * Failures do not advance the seq counter above. For the
22062                  * case of ENOBUFS we will fall out and retry in 1ms with
22063                  * the hpts. Everything else will just have to retransmit
22064                  * with the timer.
22065                  *
22066                  * In any case, we do not want to loop around for another
22067                  * send without a good reason.
22068                  */
22069                 sendalot = 0;
22070                 switch (error) {
22071                 case EPERM:
22072                         tp->t_softerror = error;
22073 #ifdef TCP_ACCOUNTING
22074                         crtsc = get_cyclecount();
22075                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22076                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22077                         }
22078                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22079                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22080                         }
22081                         sched_unpin();
22082 #endif
22083                         return (error);
22084                 case ENOBUFS:
22085                         /*
22086                          * Pace us right away to retry in a some
22087                          * time
22088                          */
22089                         if (rack->r_ctl.crte != NULL) {
22090                                 tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
22091                                 if (tcp_bblogging_on(rack->rc_tp))
22092                                         rack_log_queue_level(tp, rack, len, &tv, cts);
22093                         } else
22094                                 tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
22095                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
22096                         if (rack->rc_enobuf < 0x7f)
22097                                 rack->rc_enobuf++;
22098                         if (slot < (10 * HPTS_USEC_IN_MSEC))
22099                                 slot = 10 * HPTS_USEC_IN_MSEC;
22100                         if (rack->r_ctl.crte != NULL) {
22101                                 counter_u64_add(rack_saw_enobuf_hw, 1);
22102                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
22103                         }
22104                         counter_u64_add(rack_saw_enobuf, 1);
22105                         goto enobufs;
22106                 case EMSGSIZE:
22107                         /*
22108                          * For some reason the interface we used initially
22109                          * to send segments changed to another or lowered
22110                          * its MTU. If TSO was active we either got an
22111                          * interface without TSO capabilits or TSO was
22112                          * turned off. If we obtained mtu from ip_output()
22113                          * then update it and try again.
22114                          */
22115                         if (tso)
22116                                 tp->t_flags &= ~TF_TSO;
22117                         if (mtu != 0) {
22118                                 int saved_mtu;
22119
22120                                 saved_mtu = tp->t_maxseg;
22121                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
22122                                 if (saved_mtu > tp->t_maxseg) {
22123                                         goto again;
22124                                 }
22125                         }
22126                         slot = 10 * HPTS_USEC_IN_MSEC;
22127                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22128 #ifdef TCP_ACCOUNTING
22129                         crtsc = get_cyclecount();
22130                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22131                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22132                         }
22133                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22134                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22135                         }
22136                         sched_unpin();
22137 #endif
22138                         return (error);
22139                 case ENETUNREACH:
22140                         counter_u64_add(rack_saw_enetunreach, 1);
22141                 case EHOSTDOWN:
22142                 case EHOSTUNREACH:
22143                 case ENETDOWN:
22144                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
22145                                 tp->t_softerror = error;
22146                         }
22147                         /* FALLTHROUGH */
22148                 default:
22149                         slot = 10 * HPTS_USEC_IN_MSEC;
22150                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22151 #ifdef TCP_ACCOUNTING
22152                         crtsc = get_cyclecount();
22153                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22154                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22155                         }
22156                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22157                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22158                         }
22159                         sched_unpin();
22160 #endif
22161                         return (error);
22162                 }
22163         } else {
22164                 rack->rc_enobuf = 0;
22165                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
22166                         rack->r_ctl.retran_during_recovery += len;
22167         }
22168         KMOD_TCPSTAT_INC(tcps_sndtotal);
22169
22170         /*
22171          * Data sent (as far as we can tell). If this advertises a larger
22172          * window than any other segment, then remember the size of the
22173          * advertised window. Any pending ACK has now been sent.
22174          */
22175         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
22176                 tp->rcv_adv = tp->rcv_nxt + recwin;
22177
22178         tp->last_ack_sent = tp->rcv_nxt;
22179         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
22180 enobufs:
22181         if (sendalot) {
22182                 /* Do we need to turn off sendalot? */
22183                 if (rack->r_ctl.rc_pace_max_segs &&
22184                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
22185                         /* We hit our max. */
22186                         sendalot = 0;
22187                 } else if ((rack->rc_user_set_max_segs) &&
22188                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
22189                         /* We hit the user defined max */
22190                         sendalot = 0;
22191                 }
22192         }
22193         if ((error == 0) && (flags & TH_FIN))
22194                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
22195         if (flags & TH_RST) {
22196                 /*
22197                  * We don't send again after sending a RST.
22198                  */
22199                 slot = 0;
22200                 sendalot = 0;
22201                 if (error == 0)
22202                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
22203         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
22204                 /*
22205                  * Get our pacing rate, if an error
22206                  * occurred in sending (ENOBUF) we would
22207                  * hit the else if with slot preset. Other
22208                  * errors return.
22209                  */
22210                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
22211         }
22212         if (rsm &&
22213             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
22214             rack->use_rack_rr) {
22215                 /* Its a retransmit and we use the rack cheat? */
22216                 if ((slot == 0) ||
22217                     (rack->rc_always_pace == 0) ||
22218                     (rack->r_rr_config == 1)) {
22219                         /*
22220                          * We have no pacing set or we
22221                          * are using old-style rack or
22222                          * we are overridden to use the old 1ms pacing.
22223                          */
22224                         slot = rack->r_ctl.rc_min_to;
22225                 }
22226         }
22227         /* We have sent clear the flag */
22228         rack->r_ent_rec_ns = 0;
22229         if (rack->r_must_retran) {
22230                 if (rsm) {
22231                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
22232                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
22233                                 /*
22234                                  * We have retransmitted all.
22235                                  */
22236                                 rack->r_must_retran = 0;
22237                                 rack->r_ctl.rc_out_at_rto = 0;
22238                         }
22239                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22240                         /*
22241                          * Sending new data will also kill
22242                          * the loop.
22243                          */
22244                         rack->r_must_retran = 0;
22245                         rack->r_ctl.rc_out_at_rto = 0;
22246                 }
22247         }
22248         rack->r_ctl.fsb.recwin = recwin;
22249         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
22250             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22251                 /*
22252                  * We hit an RTO and now have past snd_max at the RTO
22253                  * clear all the WAS flags.
22254                  */
22255                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
22256         }
22257         if (slot) {
22258                 /* set the rack tcb into the slot N */
22259                 if ((error == 0) &&
22260                     rack_use_rfo &&
22261                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
22262                     (rsm == NULL) &&
22263                     (tp->snd_nxt == tp->snd_max) &&
22264                     (ipoptlen == 0) &&
22265                     (tp->rcv_numsacks == 0) &&
22266                     rack->r_fsb_inited &&
22267                     TCPS_HAVEESTABLISHED(tp->t_state) &&
22268                     ((IN_RECOVERY(tp->t_flags)) == 0) &&
22269                     (rack->r_must_retran == 0) &&
22270                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
22271                     (len > 0) && (orig_len > 0) &&
22272                     (orig_len > len) &&
22273                     ((orig_len - len) >= segsiz) &&
22274                     ((optlen == 0) ||
22275                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22276                         /* We can send at least one more MSS using our fsb */
22277                         rack_setup_fast_output(tp, rack, sb, len, orig_len,
22278                                                segsiz, pace_max_seg, hw_tls, flags);
22279                 } else
22280                         rack->r_fast_output = 0;
22281                 rack_log_fsb(rack, tp, so, flags,
22282                              ipoptlen, orig_len, len, error,
22283                              (rsm == NULL), optlen, __LINE__, 2);
22284         } else if (sendalot) {
22285                 int ret;
22286
22287                 sack_rxmit = 0;
22288                 if ((error == 0) &&
22289                     rack_use_rfo &&
22290                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
22291                     (rsm == NULL) &&
22292                     (ipoptlen == 0) &&
22293                     (tp->rcv_numsacks == 0) &&
22294                     (tp->snd_nxt == tp->snd_max) &&
22295                     (rack->r_must_retran == 0) &&
22296                     rack->r_fsb_inited &&
22297                     TCPS_HAVEESTABLISHED(tp->t_state) &&
22298                     ((IN_RECOVERY(tp->t_flags)) == 0) &&
22299                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
22300                     (len > 0) && (orig_len > 0) &&
22301                     (orig_len > len) &&
22302                     ((orig_len - len) >= segsiz) &&
22303                     ((optlen == 0) ||
22304                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22305                         /* we can use fast_output for more */
22306                         rack_setup_fast_output(tp, rack, sb, len, orig_len,
22307                                                segsiz, pace_max_seg, hw_tls, flags);
22308                         if (rack->r_fast_output) {
22309                                 error = 0;
22310                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
22311                                 if (ret >= 0)
22312                                         return (ret);
22313                                 else if (error)
22314                                         goto nomore;
22315
22316                         }
22317                 }
22318                 goto again;
22319         }
22320         /* Assure when we leave that snd_nxt will point to top */
22321 skip_all_send:
22322         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22323                 tp->snd_nxt = tp->snd_max;
22324         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
22325 #ifdef TCP_ACCOUNTING
22326         crtsc = get_cyclecount() - ts_val;
22327         if (tot_len_this_send) {
22328                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22329                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
22330                 }
22331                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22332                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
22333                 }
22334                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22335                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
22336                 }
22337         } else {
22338                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22339                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
22340                 }
22341                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22342                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
22343                 }
22344         }
22345         sched_unpin();
22346 #endif
22347         if (error == ENOBUFS)
22348                 error = 0;
22349         return (error);
22350 }
22351
22352 static void
22353 rack_update_seg(struct tcp_rack *rack)
22354 {
22355         uint32_t orig_val;
22356
22357         orig_val = rack->r_ctl.rc_pace_max_segs;
22358         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
22359         if (orig_val != rack->r_ctl.rc_pace_max_segs)
22360                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
22361 }
22362
22363 static void
22364 rack_mtu_change(struct tcpcb *tp)
22365 {
22366         /*
22367          * The MSS may have changed
22368          */
22369         struct tcp_rack *rack;
22370         struct rack_sendmap *rsm;
22371
22372         rack = (struct tcp_rack *)tp->t_fb_ptr;
22373         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
22374                 /*
22375                  * The MTU has changed we need to resend everything
22376                  * since all we have sent is lost. We first fix
22377                  * up the mtu though.
22378                  */
22379                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
22380                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
22381                 rack_remxt_tmr(tp);
22382                 rack->r_fast_output = 0;
22383                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
22384                                                 rack->r_ctl.rc_sacked);
22385                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
22386                 rack->r_must_retran = 1;
22387                 /* Mark all inflight to needing to be rxt'd */
22388                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
22389                         rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
22390                 }
22391         }
22392         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
22393         /* We don't use snd_nxt to retransmit */
22394         tp->snd_nxt = tp->snd_max;
22395 }
22396
22397 static int
22398 rack_set_dgp(struct tcp_rack *rack)
22399 {
22400         /* pace_always=1 */
22401         if (rack->rc_always_pace == 0) {
22402                 if (tcp_can_enable_pacing() == 0)
22403                         return (EBUSY);
22404         }
22405         rack->dgp_on = 1;
22406         rack->rc_always_pace = 1;
22407         rack->use_fixed_rate = 0;
22408         if (rack->gp_ready)
22409                 rack_set_cc_pacing(rack);
22410         rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22411         rack->rack_attempt_hdwr_pace = 0;
22412         /* rxt settings */
22413         rack->full_size_rxt = 1;
22414         rack->shape_rxt_to_pacing_min  = 0;
22415         /* cmpack=1 */
22416         rack->r_use_cmp_ack = 1;
22417         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
22418             rack->r_use_cmp_ack)
22419                 rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22420         /* scwnd=1 */
22421         rack->rack_enable_scwnd = 1;
22422         /* dynamic=100 */
22423         rack->rc_gp_dyn_mul = 1;
22424         /* gp_inc_ca */
22425         rack->r_ctl.rack_per_of_gp_ca = 100;
22426         /* rrr_conf=3 */
22427         rack->r_rr_config = 3;
22428         /* npush=2 */
22429         rack->r_ctl.rc_no_push_at_mrtt = 2;
22430         /* fillcw=1 */
22431         if (rack->r_cwnd_was_clamped == 0) {
22432                 rack->rc_pace_to_cwnd = 1;
22433         } else {
22434                 rack->rc_pace_to_cwnd = 0;
22435                 /* Reset all multipliers to 100.0 so just the measured bw */
22436                 rack->r_ctl.rack_per_of_gp_ss = 100;
22437                 rack->r_ctl.rack_per_of_gp_ca = 100;
22438         }
22439         rack->rc_pace_fill_if_rttin_range = 0;
22440         rack->rtt_limit_mul = 0;
22441         /* noprr=1 */
22442         rack->rack_no_prr = 1;
22443         /* lscwnd=1 */
22444         rack->r_limit_scw = 1;
22445         /* gp_inc_rec */
22446         rack->r_ctl.rack_per_of_gp_rec = 90;
22447         rack_client_buffer_level_set(rack);
22448         return (0);
22449 }
22450
22451
22452
22453 static int
22454 rack_set_profile(struct tcp_rack *rack, int prof)
22455 {
22456         int err = EINVAL;
22457         if (prof == 1) {
22458                 /*
22459                  * Profile 1 is "standard" DGP. It ignores
22460                  * client buffer level.
22461                  */
22462                 rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL0;
22463                 err = rack_set_dgp(rack);
22464                 if (err)
22465                         return (err);
22466         } else if (prof == 2) {
22467                 /*
22468                  * Profile 2 is DGP. Less aggressive with
22469                  * respect to client buffer level.
22470                  */
22471                 rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL1;
22472                 err = rack_set_dgp(rack);
22473                 if (err)
22474                         return (err);
22475         } else if (prof == 3) {
22476                 /*
22477                  * Profile 3 is DGP. Even Less aggressive with
22478                  * respect to client buffer level.
22479                  */
22480                 rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL2;
22481                 err = rack_set_dgp(rack);
22482                 if (err)
22483                         return (err);
22484         } else if (prof == 4) {
22485                 /*
22486                  * Profile 4 is DGP with the most responsiveness
22487                  * to client buffer level.
22488                  */
22489                 rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL3;
22490                 err = rack_set_dgp(rack);
22491                 if (err)
22492                         return (err);
22493         } else if (prof == 0) {
22494                 /* This changes things back to the default settings */
22495                 rack->dgp_on = 0;
22496                 rack->rc_hybrid_mode = 0;
22497                 err = 0;
22498                 if (rack_fill_cw_state)
22499                         rack->rc_pace_to_cwnd = 1;
22500                 else
22501                         rack->rc_pace_to_cwnd = 0;
22502                 if (rack->rc_always_pace) {
22503                         tcp_decrement_paced_conn();
22504                         rack_undo_cc_pacing(rack);
22505                         rack->rc_always_pace = 0;
22506                 }
22507                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
22508                         rack->rc_always_pace = 1;
22509                         if ((rack->gp_ready) && (rack->use_fixed_rate == 0))
22510                                 rack_set_cc_pacing(rack);
22511                 } else
22512                         rack->rc_always_pace = 0;
22513                 if (rack_dsack_std_based & 0x1) {
22514                         /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
22515                         rack->rc_rack_tmr_std_based = 1;
22516                 }
22517                 if (rack_dsack_std_based & 0x2) {
22518                         /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
22519                         rack->rc_rack_use_dsack = 1;
22520                 }
22521                 if (rack_use_cmp_acks)
22522                         rack->r_use_cmp_ack = 1;
22523                 else
22524                         rack->r_use_cmp_ack = 0;
22525                 if (rack_disable_prr)
22526                         rack->rack_no_prr = 1;
22527                 else
22528                         rack->rack_no_prr = 0;
22529                 if (rack_gp_no_rec_chg)
22530                         rack->rc_gp_no_rec_chg = 1;
22531                 else
22532                         rack->rc_gp_no_rec_chg = 0;
22533                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
22534                         rack->r_mbuf_queue = 1;
22535                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
22536                                 rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22537                         rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22538                 } else {
22539                         rack->r_mbuf_queue = 0;
22540                         rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
22541                 }
22542                 if (rack_enable_shared_cwnd)
22543                         rack->rack_enable_scwnd = 1;
22544                 else
22545                         rack->rack_enable_scwnd = 0;
22546                 if (rack_do_dyn_mul) {
22547                         /* When dynamic adjustment is on CA needs to start at 100% */
22548                         rack->rc_gp_dyn_mul = 1;
22549                         if (rack_do_dyn_mul >= 100)
22550                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
22551                 } else {
22552                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
22553                         rack->rc_gp_dyn_mul = 0;
22554                 }
22555                 rack->r_rr_config = 0;
22556                 rack->r_ctl.rc_no_push_at_mrtt = 0;
22557                 rack->rc_pace_to_cwnd = 0;
22558                 rack->rc_pace_fill_if_rttin_range = 0;
22559                 rack->rtt_limit_mul = 0;
22560
22561                 if (rack_enable_hw_pacing)
22562                         rack->rack_hdw_pace_ena = 1;
22563                 else
22564                         rack->rack_hdw_pace_ena = 0;
22565                 if (rack_disable_prr)
22566                         rack->rack_no_prr = 1;
22567                 else
22568                         rack->rack_no_prr = 0;
22569                 if (rack_limits_scwnd)
22570                         rack->r_limit_scw  = 1;
22571                 else
22572                         rack->r_limit_scw  = 0;
22573                 rack_init_retransmit_value(rack, rack_rxt_controls);
22574                 err = 0;
22575         }
22576         return (err);
22577 }
22578
22579 static int
22580 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
22581 {
22582         struct deferred_opt_list *dol;
22583
22584         dol = malloc(sizeof(struct deferred_opt_list),
22585                      M_TCPFSB, M_NOWAIT|M_ZERO);
22586         if (dol == NULL) {
22587                 /*
22588                  * No space yikes -- fail out..
22589                  */
22590                 return (0);
22591         }
22592         dol->optname = sopt_name;
22593         dol->optval = loptval;
22594         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
22595         return (1);
22596 }
22597
22598 static int
22599 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
22600 {
22601 #ifdef TCP_REQUEST_TRK
22602         struct tcp_sendfile_track *sft;
22603         struct timeval tv;
22604         tcp_seq seq;
22605         int err;
22606
22607         microuptime(&tv);
22608
22609         /*
22610          * If BB logging is not on we need to look at the DTL flag.
22611          * If its on already then those reasons override the DTL input.
22612          * We do this with any request, you can turn DTL on, but it does
22613          * not turn off at least from hybrid pacing requests.
22614          */
22615         if (tcp_bblogging_on(rack->rc_tp) == 0) {
22616                 if (hybrid->hybrid_flags & TCP_HYBRID_PACING_DTL) {
22617                         /* Turn on BB point logging  */
22618                         tcp_set_bblog_state(rack->rc_tp, TCP_LOG_VIA_BBPOINTS,
22619                                             TCP_BBPOINT_REQ_LEVEL_LOGGING);
22620                 }
22621         }
22622         /* Make sure no fixed rate is on */
22623         rack->use_fixed_rate = 0;
22624         rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
22625         rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
22626         rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
22627         /* Now allocate or find our entry that will have these settings */
22628         sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusectick(&tv), 0);
22629         if (sft == NULL) {
22630                 rack->rc_tp->tcp_hybrid_error++;
22631                 /* no space, where would it have gone? */
22632                 seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
22633                 rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
22634                 return (ENOSPC);
22635         }
22636         /* The seq will be snd_una + everything in the buffer */
22637         seq = sft->start_seq;
22638         if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
22639                 /* Disabling hybrid pacing */
22640                 if (rack->rc_hybrid_mode) {
22641                         rack_set_profile(rack, 0);
22642                         rack->rc_tp->tcp_hybrid_stop++;
22643                 }
22644                 rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
22645                 return (0);
22646         }
22647         if (rack->dgp_on == 0) {
22648                 /*
22649                  * If we have not yet turned DGP on, do so
22650                  * now setting pure DGP mode, no buffer level
22651                  * response.
22652                  */
22653                 if ((err = rack_set_profile(rack, 1)) != 0){
22654                         /* Failed to turn pacing on */
22655                         rack->rc_tp->tcp_hybrid_error++;
22656                         rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
22657                         return (err);
22658                 }
22659         }
22660         /* Now set in our flags */
22661         sft->hybrid_flags = hybrid->hybrid_flags;
22662         if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
22663                 sft->cspr = hybrid->cspr;
22664         else
22665                 sft->cspr = 0;
22666         if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
22667                 sft->hint_maxseg = hybrid->hint_maxseg;
22668         else
22669                 sft->hint_maxseg = 0;
22670         rack->rc_hybrid_mode = 1;
22671         rack->rc_tp->tcp_hybrid_start++;
22672         rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
22673         return (0);
22674 #else
22675         return (ENOTSUP);
22676 #endif
22677 }
22678
22679 static int
22680 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
22681                     uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
22682
22683 {
22684         struct epoch_tracker et;
22685         struct sockopt sopt;
22686         struct cc_newreno_opts opt;
22687         uint64_t val;
22688         int error = 0;
22689         uint16_t ca, ss;
22690
22691         switch (sopt_name) {
22692         case TCP_RACK_SET_RXT_OPTIONS:
22693                 if ((optval >= 0) && (optval <= 2)) {
22694                         rack_init_retransmit_value(rack, optval);
22695                 } else {
22696                         /*
22697                          * You must send in 0, 1 or 2 all else is
22698                          * invalid.
22699                          */
22700                         error = EINVAL;
22701                 }
22702                 break;
22703         case TCP_RACK_DSACK_OPT:
22704                 RACK_OPTS_INC(tcp_rack_dsack_opt);
22705                 if (optval & 0x1) {
22706                         rack->rc_rack_tmr_std_based = 1;
22707                 } else {
22708                         rack->rc_rack_tmr_std_based = 0;
22709                 }
22710                 if (optval & 0x2) {
22711                         rack->rc_rack_use_dsack = 1;
22712                 } else {
22713                         rack->rc_rack_use_dsack = 0;
22714                 }
22715                 rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
22716                 break;
22717         case TCP_RACK_PACING_DIVISOR:
22718                 RACK_OPTS_INC(tcp_rack_pacing_divisor);
22719                 if (optval == 0) {
22720                         rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
22721                 } else {
22722                         if (optval < RL_MIN_DIVISOR)
22723                                 rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
22724                         else
22725                                 rack->r_ctl.pace_len_divisor = optval;
22726                 }
22727                 break;
22728         case TCP_RACK_HI_BETA:
22729                 RACK_OPTS_INC(tcp_rack_hi_beta);
22730                 if (optval)
22731                         rack->rack_hibeta = 1;
22732                 else
22733                         rack->rack_hibeta = 0;
22734                 break;
22735         case TCP_RACK_PACING_BETA:
22736                 RACK_OPTS_INC(tcp_rack_beta);
22737                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
22738                         /* This only works for newreno. */
22739                         error = EINVAL;
22740                         break;
22741                 }
22742                 if (rack->rc_pacing_cc_set) {
22743                         /*
22744                          * Set them into the real CC module
22745                          * whats in the rack pcb is the old values
22746                          * to be used on restoral/
22747                          */
22748                         sopt.sopt_dir = SOPT_SET;
22749                         opt.name = CC_NEWRENO_BETA;
22750                         opt.val = optval;
22751                         if (CC_ALGO(tp)->ctl_output != NULL)
22752                                 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
22753                         else {
22754                                 error = ENOENT;
22755                                 break;
22756                         }
22757                 } else {
22758                         /*
22759                          * Not pacing yet so set it into our local
22760                          * rack pcb storage.
22761                          */
22762                         rack->r_ctl.rc_saved_beta.beta = optval;
22763                 }
22764                 break;
22765         case TCP_RACK_TIMER_SLOP:
22766                 RACK_OPTS_INC(tcp_rack_timer_slop);
22767                 rack->r_ctl.timer_slop = optval;
22768                 if (rack->rc_tp->t_srtt) {
22769                         /*
22770                          * If we have an SRTT lets update t_rxtcur
22771                          * to have the new slop.
22772                          */
22773                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
22774                                            rack_rto_min, rack_rto_max,
22775                                            rack->r_ctl.timer_slop);
22776                 }
22777                 break;
22778         case TCP_RACK_PACING_BETA_ECN:
22779                 RACK_OPTS_INC(tcp_rack_beta_ecn);
22780                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
22781                         /* This only works for newreno. */
22782                         error = EINVAL;
22783                         break;
22784                 }
22785                 if (rack->rc_pacing_cc_set) {
22786                         /*
22787                          * Set them into the real CC module
22788                          * whats in the rack pcb is the old values
22789                          * to be used on restoral/
22790                          */
22791                         sopt.sopt_dir = SOPT_SET;
22792                         opt.name = CC_NEWRENO_BETA_ECN;
22793                         opt.val = optval;
22794                         if (CC_ALGO(tp)->ctl_output != NULL)
22795                                 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
22796                         else
22797                                 error = ENOENT;
22798                 } else {
22799                         /*
22800                          * Not pacing yet so set it into our local
22801                          * rack pcb storage.
22802                          */
22803                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
22804                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
22805                 }
22806                 break;
22807         case TCP_DEFER_OPTIONS:
22808                 RACK_OPTS_INC(tcp_defer_opt);
22809                 if (optval) {
22810                         if (rack->gp_ready) {
22811                                 /* Too late */
22812                                 error = EINVAL;
22813                                 break;
22814                         }
22815                         rack->defer_options = 1;
22816                 } else
22817                         rack->defer_options = 0;
22818                 break;
22819         case TCP_RACK_MEASURE_CNT:
22820                 RACK_OPTS_INC(tcp_rack_measure_cnt);
22821                 if (optval && (optval <= 0xff)) {
22822                         rack->r_ctl.req_measurements = optval;
22823                 } else
22824                         error = EINVAL;
22825                 break;
22826         case TCP_REC_ABC_VAL:
22827                 RACK_OPTS_INC(tcp_rec_abc_val);
22828                 if (optval > 0)
22829                         rack->r_use_labc_for_rec = 1;
22830                 else
22831                         rack->r_use_labc_for_rec = 0;
22832                 break;
22833         case TCP_RACK_ABC_VAL:
22834                 RACK_OPTS_INC(tcp_rack_abc_val);
22835                 if ((optval > 0) && (optval < 255))
22836                         rack->rc_labc = optval;
22837                 else
22838                         error = EINVAL;
22839                 break;
22840         case TCP_HDWR_UP_ONLY:
22841                 RACK_OPTS_INC(tcp_pacing_up_only);
22842                 if (optval)
22843                         rack->r_up_only = 1;
22844                 else
22845                         rack->r_up_only = 0;
22846                 break;
22847         case TCP_PACING_RATE_CAP:
22848                 RACK_OPTS_INC(tcp_pacing_rate_cap);
22849                 rack->r_ctl.bw_rate_cap = loptval;
22850                 break;
22851         case TCP_HYBRID_PACING:
22852                 if (hybrid == NULL) {
22853                         error = EINVAL;
22854                         break;
22855                 }
22856                 error = process_hybrid_pacing(rack, hybrid);
22857                 break;
22858         case TCP_RACK_PROFILE:
22859                 RACK_OPTS_INC(tcp_profile);
22860                 error = rack_set_profile(rack, optval);
22861                 break;
22862         case TCP_USE_CMP_ACKS:
22863                 RACK_OPTS_INC(tcp_use_cmp_acks);
22864                 if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
22865                         /* You can't turn it off once its on! */
22866                         error = EINVAL;
22867                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
22868                         rack->r_use_cmp_ack = 1;
22869                         rack->r_mbuf_queue = 1;
22870                         tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22871                 }
22872                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
22873                         tp->t_flags2 |= TF2_MBUF_ACKCMP;
22874                 break;
22875         case TCP_SHARED_CWND_TIME_LIMIT:
22876                 RACK_OPTS_INC(tcp_lscwnd);
22877                 if (optval)
22878                         rack->r_limit_scw = 1;
22879                 else
22880                         rack->r_limit_scw = 0;
22881                 break;
22882         case TCP_RACK_DGP_IN_REC:
22883                 RACK_OPTS_INC(tcp_dgp_in_rec);
22884                 if (optval)
22885                         rack->r_ctl.full_dgp_in_rec = 1;
22886                 else
22887                         rack->r_ctl.full_dgp_in_rec = 0;
22888                 break;
22889         case TCP_RXT_CLAMP:
22890                 RACK_OPTS_INC(tcp_rxt_clamp);
22891                 rack_translate_clamp_value(rack, optval);
22892                 break;
22893         case TCP_RACK_PACE_TO_FILL:
22894                 RACK_OPTS_INC(tcp_fillcw);
22895                 if (optval == 0)
22896                         rack->rc_pace_to_cwnd = 0;
22897                 else {
22898                         rack->rc_pace_to_cwnd = 1;
22899                         if (optval > 1)
22900                                 rack->r_fill_less_agg = 1;
22901                 }
22902                 if ((optval >= rack_gp_rtt_maxmul) &&
22903                     rack_gp_rtt_maxmul &&
22904                     (optval < 0xf)) {
22905                         rack->rc_pace_fill_if_rttin_range = 1;
22906                         rack->rtt_limit_mul = optval;
22907                 } else {
22908                         rack->rc_pace_fill_if_rttin_range = 0;
22909                         rack->rtt_limit_mul = 0;
22910                 }
22911                 break;
22912         case TCP_RACK_NO_PUSH_AT_MAX:
22913                 RACK_OPTS_INC(tcp_npush);
22914                 if (optval == 0)
22915                         rack->r_ctl.rc_no_push_at_mrtt = 0;
22916                 else if (optval < 0xff)
22917                         rack->r_ctl.rc_no_push_at_mrtt = optval;
22918                 else
22919                         error = EINVAL;
22920                 break;
22921         case TCP_SHARED_CWND_ENABLE:
22922                 RACK_OPTS_INC(tcp_rack_scwnd);
22923                 if (optval == 0)
22924                         rack->rack_enable_scwnd = 0;
22925                 else
22926                         rack->rack_enable_scwnd = 1;
22927                 break;
22928         case TCP_RACK_MBUF_QUEUE:
22929                 /* Now do we use the LRO mbuf-queue feature */
22930                 RACK_OPTS_INC(tcp_rack_mbufq);
22931                 if (optval || rack->r_use_cmp_ack)
22932                         rack->r_mbuf_queue = 1;
22933                 else
22934                         rack->r_mbuf_queue = 0;
22935                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
22936                         tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22937                 else
22938                         tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
22939                 break;
22940         case TCP_RACK_NONRXT_CFG_RATE:
22941                 RACK_OPTS_INC(tcp_rack_cfg_rate);
22942                 if (optval == 0)
22943                         rack->rack_rec_nonrxt_use_cr = 0;
22944                 else
22945                         rack->rack_rec_nonrxt_use_cr = 1;
22946                 break;
22947         case TCP_NO_PRR:
22948                 RACK_OPTS_INC(tcp_rack_noprr);
22949                 if (optval == 0)
22950                         rack->rack_no_prr = 0;
22951                 else if (optval == 1)
22952                         rack->rack_no_prr = 1;
22953                 else if (optval == 2)
22954                         rack->no_prr_addback = 1;
22955                 else
22956                         error = EINVAL;
22957                 break;
22958         case TCP_TIMELY_DYN_ADJ:
22959                 RACK_OPTS_INC(tcp_timely_dyn);
22960                 if (optval == 0)
22961                         rack->rc_gp_dyn_mul = 0;
22962                 else {
22963                         rack->rc_gp_dyn_mul = 1;
22964                         if (optval >= 100) {
22965                                 /*
22966                                  * If the user sets something 100 or more
22967                                  * its the gp_ca value.
22968                                  */
22969                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
22970                         }
22971                 }
22972                 break;
22973         case TCP_RACK_DO_DETECTION:
22974                 RACK_OPTS_INC(tcp_rack_do_detection);
22975                 if (optval == 0)
22976                         rack->do_detection = 0;
22977                 else
22978                         rack->do_detection = 1;
22979                 break;
22980         case TCP_RACK_TLP_USE:
22981                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
22982                         error = EINVAL;
22983                         break;
22984                 }
22985                 RACK_OPTS_INC(tcp_tlp_use);
22986                 rack->rack_tlp_threshold_use = optval;
22987                 break;
22988         case TCP_RACK_TLP_REDUCE:
22989                 /* RACK TLP cwnd reduction (bool) */
22990                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
22991                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
22992                 break;
22993                 /*  Pacing related ones */
22994         case TCP_RACK_PACE_ALWAYS:
22995                 /*
22996                  * zero is old rack method, 1 is new
22997                  * method using a pacing rate.
22998                  */
22999                 RACK_OPTS_INC(tcp_rack_pace_always);
23000                 if (optval > 0) {
23001                         if (rack->rc_always_pace) {
23002                                 error = EALREADY;
23003                                 break;
23004                         } else if (tcp_can_enable_pacing()) {
23005                                 rack->rc_always_pace = 1;
23006                                 if ((rack->gp_ready) && (rack->use_fixed_rate == 0))
23007                                         rack_set_cc_pacing(rack);
23008                         }
23009                         else {
23010                                 error = ENOSPC;
23011                                 break;
23012                         }
23013                 } else {
23014                         if (rack->rc_always_pace) {
23015                                 tcp_decrement_paced_conn();
23016                                 rack->rc_always_pace = 0;
23017                                 rack_undo_cc_pacing(rack);
23018                         }
23019                 }
23020                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23021                         tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23022                 else
23023                         tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23024                 /* A rate may be set irate or other, if so set seg size */
23025                 rack_update_seg(rack);
23026                 break;
23027         case TCP_BBR_RACK_INIT_RATE:
23028                 RACK_OPTS_INC(tcp_initial_rate);
23029                 val = optval;
23030                 /* Change from kbits per second to bytes per second */
23031                 val *= 1000;
23032                 val /= 8;
23033                 rack->r_ctl.init_rate = val;
23034                 if (rack->rc_init_win != rack_default_init_window) {
23035                         uint32_t win, snt;
23036
23037                         /*
23038                          * Options don't always get applied
23039                          * in the order you think. So in order
23040                          * to assure we update a cwnd we need
23041                          * to check and see if we are still
23042                          * where we should raise the cwnd.
23043                          */
23044                         win = rc_init_window(rack);
23045                         if (SEQ_GT(tp->snd_max, tp->iss))
23046                                 snt = tp->snd_max - tp->iss;
23047                         else
23048                                 snt = 0;
23049                         if ((snt < win) &&
23050                             (tp->snd_cwnd < win))
23051                                 tp->snd_cwnd = win;
23052                 }
23053                 if (rack->rc_always_pace)
23054                         rack_update_seg(rack);
23055                 break;
23056         case TCP_BBR_IWINTSO:
23057                 RACK_OPTS_INC(tcp_initial_win);
23058                 if (optval && (optval <= 0xff)) {
23059                         uint32_t win, snt;
23060
23061                         rack->rc_init_win = optval;
23062                         win = rc_init_window(rack);
23063                         if (SEQ_GT(tp->snd_max, tp->iss))
23064                                 snt = tp->snd_max - tp->iss;
23065                         else
23066                                 snt = 0;
23067                         if ((snt < win) &&
23068                             (tp->t_srtt |
23069                              rack->r_ctl.init_rate)) {
23070                                 /*
23071                                  * We are not past the initial window
23072                                  * and we have some bases for pacing,
23073                                  * so we need to possibly adjust up
23074                                  * the cwnd. Note even if we don't set
23075                                  * the cwnd, its still ok to raise the rc_init_win
23076                                  * which can be used coming out of idle when we
23077                                  * would have a rate.
23078                                  */
23079                                 if (tp->snd_cwnd < win)
23080                                         tp->snd_cwnd = win;
23081                         }
23082                         if (rack->rc_always_pace)
23083                                 rack_update_seg(rack);
23084                 } else
23085                         error = EINVAL;
23086                 break;
23087         case TCP_RACK_FORCE_MSEG:
23088                 RACK_OPTS_INC(tcp_rack_force_max_seg);
23089                 if (optval)
23090                         rack->rc_force_max_seg = 1;
23091                 else
23092                         rack->rc_force_max_seg = 0;
23093                 break;
23094         case TCP_RACK_PACE_MIN_SEG:
23095                 RACK_OPTS_INC(tcp_rack_min_seg);
23096                 rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
23097                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
23098                 break;
23099         case TCP_RACK_PACE_MAX_SEG:
23100                 /* Max segments size in a pace in bytes */
23101                 RACK_OPTS_INC(tcp_rack_max_seg);
23102                 rack->rc_user_set_max_segs = optval;
23103                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
23104                 break;
23105         case TCP_RACK_PACE_RATE_REC:
23106                 /* Set the fixed pacing rate in Bytes per second ca */
23107                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
23108                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23109                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23110                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23111                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23112                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23113                 rack->use_fixed_rate = 1;
23114                 if (rack->rc_always_pace && rack->gp_ready && rack->rack_hibeta)
23115                         rack_set_cc_pacing(rack);
23116                 rack_log_pacing_delay_calc(rack,
23117                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
23118                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
23119                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23120                                            __LINE__, NULL,0);
23121                 break;
23122
23123         case TCP_RACK_PACE_RATE_SS:
23124                 /* Set the fixed pacing rate in Bytes per second ca */
23125                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
23126                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23127                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23128                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23129                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23130                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23131                 rack->use_fixed_rate = 1;
23132                 if (rack->rc_always_pace && rack->gp_ready && rack->rack_hibeta)
23133                         rack_set_cc_pacing(rack);
23134                 rack_log_pacing_delay_calc(rack,
23135                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
23136                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
23137                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23138                                            __LINE__, NULL, 0);
23139                 break;
23140
23141         case TCP_RACK_PACE_RATE_CA:
23142                 /* Set the fixed pacing rate in Bytes per second ca */
23143                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
23144                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23145                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23146                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23147                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23148                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23149                 rack->use_fixed_rate = 1;
23150                 if (rack->rc_always_pace && rack->gp_ready && rack->rack_hibeta)
23151                         rack_set_cc_pacing(rack);
23152                 rack_log_pacing_delay_calc(rack,
23153                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
23154                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
23155                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23156                                            __LINE__, NULL, 0);
23157                 break;
23158         case TCP_RACK_GP_INCREASE_REC:
23159                 RACK_OPTS_INC(tcp_gp_inc_rec);
23160                 rack->r_ctl.rack_per_of_gp_rec = optval;
23161                 rack_log_pacing_delay_calc(rack,
23162                                            rack->r_ctl.rack_per_of_gp_ss,
23163                                            rack->r_ctl.rack_per_of_gp_ca,
23164                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23165                                            __LINE__, NULL, 0);
23166                 break;
23167         case TCP_RACK_GP_INCREASE_CA:
23168                 RACK_OPTS_INC(tcp_gp_inc_ca);
23169                 ca = optval;
23170                 if (ca < 100) {
23171                         /*
23172                          * We don't allow any reduction
23173                          * over the GP b/w.
23174                          */
23175                         error = EINVAL;
23176                         break;
23177                 }
23178                 rack->r_ctl.rack_per_of_gp_ca = ca;
23179                 rack_log_pacing_delay_calc(rack,
23180                                            rack->r_ctl.rack_per_of_gp_ss,
23181                                            rack->r_ctl.rack_per_of_gp_ca,
23182                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23183                                            __LINE__, NULL, 0);
23184                 break;
23185         case TCP_RACK_GP_INCREASE_SS:
23186                 RACK_OPTS_INC(tcp_gp_inc_ss);
23187                 ss = optval;
23188                 if (ss < 100) {
23189                         /*
23190                          * We don't allow any reduction
23191                          * over the GP b/w.
23192                          */
23193                         error = EINVAL;
23194                         break;
23195                 }
23196                 rack->r_ctl.rack_per_of_gp_ss = ss;
23197                 rack_log_pacing_delay_calc(rack,
23198                                            rack->r_ctl.rack_per_of_gp_ss,
23199                                            rack->r_ctl.rack_per_of_gp_ca,
23200                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23201                                            __LINE__, NULL, 0);
23202                 break;
23203         case TCP_RACK_RR_CONF:
23204                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
23205                 if (optval && optval <= 3)
23206                         rack->r_rr_config = optval;
23207                 else
23208                         rack->r_rr_config = 0;
23209                 break;
23210         case TCP_PACING_DND:                    /*  URL:dnd */
23211                 if (optval > 0)
23212                         rack->rc_pace_dnd = 1;
23213                 else
23214                         rack->rc_pace_dnd = 0;
23215                 break;
23216         case TCP_HDWR_RATE_CAP:
23217                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
23218                 if (optval) {
23219                         if (rack->r_rack_hw_rate_caps == 0)
23220                                 rack->r_rack_hw_rate_caps = 1;
23221                         else
23222                                 error = EALREADY;
23223                 } else {
23224                         rack->r_rack_hw_rate_caps = 0;
23225                 }
23226                 break;
23227         case TCP_RACK_SPLIT_LIMIT:
23228                 RACK_OPTS_INC(tcp_split_limit);
23229                 rack->r_ctl.rc_split_limit = optval;
23230                 break;
23231         case TCP_BBR_HDWR_PACE:
23232                 RACK_OPTS_INC(tcp_hdwr_pacing);
23233                 if (optval){
23234                         if (rack->rack_hdrw_pacing == 0) {
23235                                 rack->rack_hdw_pace_ena = 1;
23236                                 rack->rack_attempt_hdwr_pace = 0;
23237                         } else
23238                                 error = EALREADY;
23239                 } else {
23240                         rack->rack_hdw_pace_ena = 0;
23241 #ifdef RATELIMIT
23242                         if (rack->r_ctl.crte != NULL) {
23243                                 rack->rack_hdrw_pacing = 0;
23244                                 rack->rack_attempt_hdwr_pace = 0;
23245                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
23246                                 rack->r_ctl.crte = NULL;
23247                         }
23248 #endif
23249                 }
23250                 break;
23251                 /*  End Pacing related ones */
23252         case TCP_RACK_PRR_SENDALOT:
23253                 /* Allow PRR to send more than one seg */
23254                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
23255                 rack->r_ctl.rc_prr_sendalot = optval;
23256                 break;
23257         case TCP_RACK_MIN_TO:
23258                 /* Minimum time between rack t-o's in ms */
23259                 RACK_OPTS_INC(tcp_rack_min_to);
23260                 rack->r_ctl.rc_min_to = optval;
23261                 break;
23262         case TCP_RACK_EARLY_SEG:
23263                 /* If early recovery max segments */
23264                 RACK_OPTS_INC(tcp_rack_early_seg);
23265                 rack->r_ctl.rc_early_recovery_segs = optval;
23266                 break;
23267         case TCP_RACK_ENABLE_HYSTART:
23268         {
23269                 if (optval) {
23270                         tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23271                         if (rack_do_hystart > RACK_HYSTART_ON)
23272                                 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23273                         if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23274                                 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23275                 } else {
23276                         tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23277                 }
23278         }
23279         break;
23280         case TCP_RACK_REORD_THRESH:
23281                 /* RACK reorder threshold (shift amount) */
23282                 RACK_OPTS_INC(tcp_rack_reord_thresh);
23283                 if ((optval > 0) && (optval < 31))
23284                         rack->r_ctl.rc_reorder_shift = optval;
23285                 else
23286                         error = EINVAL;
23287                 break;
23288         case TCP_RACK_REORD_FADE:
23289                 /* Does reordering fade after ms time */
23290                 RACK_OPTS_INC(tcp_rack_reord_fade);
23291                 rack->r_ctl.rc_reorder_fade = optval;
23292                 break;
23293         case TCP_RACK_TLP_THRESH:
23294                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
23295                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
23296                 if (optval)
23297                         rack->r_ctl.rc_tlp_threshold = optval;
23298                 else
23299                         error = EINVAL;
23300                 break;
23301         case TCP_BBR_USE_RACK_RR:
23302                 RACK_OPTS_INC(tcp_rack_rr);
23303                 if (optval)
23304                         rack->use_rack_rr = 1;
23305                 else
23306                         rack->use_rack_rr = 0;
23307                 break;
23308         case TCP_RACK_PKT_DELAY:
23309                 /* RACK added ms i.e. rack-rtt + reord + N */
23310                 RACK_OPTS_INC(tcp_rack_pkt_delay);
23311                 rack->r_ctl.rc_pkt_delay = optval;
23312                 break;
23313         case TCP_DELACK:
23314                 RACK_OPTS_INC(tcp_rack_delayed_ack);
23315                 if (optval == 0)
23316                         tp->t_delayed_ack = 0;
23317                 else
23318                         tp->t_delayed_ack = 1;
23319                 if (tp->t_flags & TF_DELACK) {
23320                         tp->t_flags &= ~TF_DELACK;
23321                         tp->t_flags |= TF_ACKNOW;
23322                         NET_EPOCH_ENTER(et);
23323                         rack_output(tp);
23324                         NET_EPOCH_EXIT(et);
23325                 }
23326                 break;
23327
23328         case TCP_BBR_RACK_RTT_USE:
23329                 RACK_OPTS_INC(tcp_rack_rtt_use);
23330                 if ((optval != USE_RTT_HIGH) &&
23331                     (optval != USE_RTT_LOW) &&
23332                     (optval != USE_RTT_AVG))
23333                         error = EINVAL;
23334                 else
23335                         rack->r_ctl.rc_rate_sample_method = optval;
23336                 break;
23337         case TCP_DATA_AFTER_CLOSE:
23338                 RACK_OPTS_INC(tcp_data_after_close);
23339                 if (optval)
23340                         rack->rc_allow_data_af_clo = 1;
23341                 else
23342                         rack->rc_allow_data_af_clo = 0;
23343                 break;
23344         default:
23345                 break;
23346         }
23347         tcp_log_socket_option(tp, sopt_name, optval, error);
23348         return (error);
23349 }
23350
23351
23352 static void
23353 rack_apply_deferred_options(struct tcp_rack *rack)
23354 {
23355         struct deferred_opt_list *dol, *sdol;
23356         uint32_t s_optval;
23357
23358         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
23359                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
23360                 /* Disadvantage of deferal is you loose the error return */
23361                 s_optval = (uint32_t)dol->optval;
23362                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
23363                 free(dol, M_TCPDO);
23364         }
23365 }
23366
23367 static void
23368 rack_hw_tls_change(struct tcpcb *tp, int chg)
23369 {
23370         /* Update HW tls state */
23371         struct tcp_rack *rack;
23372
23373         rack = (struct tcp_rack *)tp->t_fb_ptr;
23374         if (chg)
23375                 rack->r_ctl.fsb.hw_tls = 1;
23376         else
23377                 rack->r_ctl.fsb.hw_tls = 0;
23378 }
23379
23380 static int
23381 rack_pru_options(struct tcpcb *tp, int flags)
23382 {
23383         if (flags & PRUS_OOB)
23384                 return (EOPNOTSUPP);
23385         return (0);
23386 }
23387
23388 static bool
23389 rack_wake_check(struct tcpcb *tp)
23390 {
23391         struct tcp_rack *rack;
23392         struct timeval tv;
23393         uint32_t cts;
23394
23395         rack = (struct tcp_rack *)tp->t_fb_ptr;
23396         if (rack->r_ctl.rc_hpts_flags) {
23397                 cts = tcp_get_usecs(&tv);
23398                 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
23399                         /*
23400                          * Pacing timer is up, check if we are ready.
23401                          */
23402                         if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
23403                                 return (true);
23404                 } else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
23405                         /*
23406                          * A timer is up, check if we are ready.
23407                          */
23408                         if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
23409                                 return (true);
23410                 }
23411         }
23412         return (false);
23413 }
23414
23415 static struct tcp_function_block __tcp_rack = {
23416         .tfb_tcp_block_name = __XSTRING(STACKNAME),
23417         .tfb_tcp_output = rack_output,
23418         .tfb_do_queued_segments = ctf_do_queued_segments,
23419         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
23420         .tfb_tcp_do_segment = rack_do_segment,
23421         .tfb_tcp_ctloutput = rack_ctloutput,
23422         .tfb_tcp_fb_init = rack_init,
23423         .tfb_tcp_fb_fini = rack_fini,
23424         .tfb_tcp_timer_stop_all = rack_stopall,
23425         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
23426         .tfb_tcp_handoff_ok = rack_handoff_ok,
23427         .tfb_tcp_mtu_chg = rack_mtu_change,
23428         .tfb_pru_options = rack_pru_options,
23429         .tfb_hwtls_change = rack_hw_tls_change,
23430         .tfb_chg_query = rack_chg_query,
23431         .tfb_switch_failed = rack_switch_failed,
23432         .tfb_early_wake_check = rack_wake_check,
23433         .tfb_compute_pipe = rack_compute_pipe,
23434         .tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
23435 };
23436
23437 /*
23438  * rack_ctloutput() must drop the inpcb lock before performing copyin on
23439  * socket option arguments.  When it re-acquires the lock after the copy, it
23440  * has to revalidate that the connection is still valid for the socket
23441  * option.
23442  */
23443 static int
23444 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
23445 {
23446         struct inpcb *inp = tptoinpcb(tp);
23447 #ifdef INET
23448         struct ip *ip;
23449 #endif
23450         struct tcp_rack *rack;
23451         struct tcp_hybrid_req hybrid;
23452         uint64_t loptval;
23453         int32_t error = 0, optval;
23454
23455         rack = (struct tcp_rack *)tp->t_fb_ptr;
23456         if (rack == NULL) {
23457                 INP_WUNLOCK(inp);
23458                 return (EINVAL);
23459         }
23460 #ifdef INET
23461         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
23462 #endif
23463
23464         switch (sopt->sopt_level) {
23465 #ifdef INET6
23466         case IPPROTO_IPV6:
23467                 MPASS(inp->inp_vflag & INP_IPV6PROTO);
23468                 switch (sopt->sopt_name) {
23469                 case IPV6_USE_MIN_MTU:
23470                         tcp6_use_min_mtu(tp);
23471                         break;
23472                 }
23473                 INP_WUNLOCK(inp);
23474                 return (0);
23475 #endif
23476 #ifdef INET
23477         case IPPROTO_IP:
23478                 switch (sopt->sopt_name) {
23479                 case IP_TOS:
23480                         /*
23481                          * The DSCP codepoint has changed, update the fsb.
23482                          */
23483                         ip->ip_tos = rack->rc_inp->inp_ip_tos;
23484                         break;
23485                 case IP_TTL:
23486                         /*
23487                          * The TTL has changed, update the fsb.
23488                          */
23489                         ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
23490                         break;
23491                 }
23492                 INP_WUNLOCK(inp);
23493                 return (0);
23494 #endif
23495 #ifdef SO_PEERPRIO
23496         case SOL_SOCKET:
23497                 switch (sopt->sopt_name) {
23498                 case SO_PEERPRIO:                       /*  SC-URL:bs */
23499                         /* Already read in and sanity checked in sosetopt(). */
23500                         if (inp->inp_socket) {
23501                                 rack->client_bufferlvl = inp->inp_socket->so_peerprio;
23502                                 rack_client_buffer_level_set(rack);
23503                         }
23504                         break;
23505                 }
23506                 INP_WUNLOCK(inp);
23507                 return (0);
23508 #endif
23509         case IPPROTO_TCP:
23510                 switch (sopt->sopt_name) {
23511                 case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
23512                 /*  Pacing related ones */
23513                 case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
23514                 case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
23515                 case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
23516                 case TCP_RACK_PACE_MIN_SEG:             /*  URL:pace_min_seg */
23517                 case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
23518                 case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
23519                 case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
23520                 case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
23521                 case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
23522                 case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
23523                 case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
23524                 case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
23525                 case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
23526                 case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
23527                 case TCP_HDWR_RATE_CAP:                 /*  URL:hdwrcap boolean */
23528                 case TCP_PACING_RATE_CAP:               /*  URL:cap  -- used by side-channel */
23529                 case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
23530                 case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
23531                 case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
23532                 case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
23533                 case TCP_RACK_DGP_IN_REC:               /*  URL:dgpinrec */
23534                         /* End pacing related */
23535                 case TCP_RXT_CLAMP:                     /*  URL:rxtclamp */
23536                 case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
23537                 case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
23538                 case TCP_RACK_MIN_TO:                   /*  URL:min_to */
23539                 case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
23540                 case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
23541                 case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
23542                 case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
23543                 case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
23544                 case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
23545                 case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
23546                 case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
23547                 case TCP_RACK_DO_DETECTION:             /*  URL:detect */
23548                 case TCP_NO_PRR:                        /*  URL:noprr */
23549                 case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
23550                 case TCP_DATA_AFTER_CLOSE:              /*  no URL */
23551                 case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
23552                 case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
23553                 case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
23554                 case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
23555                 case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
23556                 case TCP_RACK_PROFILE:                  /*  URL:profile */
23557                 case TCP_HYBRID_PACING:                 /*  URL:hybrid */
23558                 case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
23559                 case TCP_RACK_ABC_VAL:                  /*  URL:labc */
23560                 case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
23561                 case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
23562                 case TCP_DEFER_OPTIONS:                 /*  URL:defer */
23563                 case TCP_RACK_DSACK_OPT:                /*  URL:dsack */
23564                 case TCP_RACK_TIMER_SLOP:               /*  URL:timer_slop */
23565                 case TCP_RACK_ENABLE_HYSTART:           /*  URL:hystart */
23566                 case TCP_RACK_SET_RXT_OPTIONS:          /*  URL:rxtsz */
23567                 case TCP_RACK_HI_BETA:                  /*  URL:hibeta */
23568                 case TCP_RACK_SPLIT_LIMIT:              /*  URL:split */
23569                 case TCP_RACK_PACING_DIVISOR:           /*  URL:divisor */
23570                 case TCP_PACING_DND:                    /*  URL:dnd */
23571                         goto process_opt;
23572                         break;
23573                 default:
23574                         /* Filter off all unknown options to the base stack */
23575                         return (tcp_default_ctloutput(tp, sopt));
23576                         break;
23577                 }
23578
23579         default:
23580                 INP_WUNLOCK(inp);
23581                 return (0);
23582         }
23583 process_opt:
23584         INP_WUNLOCK(inp);
23585         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
23586                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
23587                 /*
23588                  * We truncate it down to 32 bits for the socket-option trace this
23589                  * means rates > 34Gbps won't show right, but thats probably ok.
23590                  */
23591                 optval = (uint32_t)loptval;
23592         } else if (sopt->sopt_name == TCP_HYBRID_PACING) {
23593                 error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
23594         } else {
23595                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
23596                 /* Save it in 64 bit form too */
23597                 loptval = optval;
23598         }
23599         if (error)
23600                 return (error);
23601         INP_WLOCK(inp);
23602         if (tp->t_fb != &__tcp_rack) {
23603                 INP_WUNLOCK(inp);
23604                 return (ENOPROTOOPT);
23605         }
23606         if (rack->defer_options && (rack->gp_ready == 0) &&
23607             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
23608             (sopt->sopt_name != TCP_HYBRID_PACING) &&
23609             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
23610             (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
23611             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
23612             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
23613                 /* Options are beind deferred */
23614                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
23615                         INP_WUNLOCK(inp);
23616                         return (0);
23617                 } else {
23618                         /* No memory to defer, fail */
23619                         INP_WUNLOCK(inp);
23620                         return (ENOMEM);
23621                 }
23622         }
23623         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
23624         INP_WUNLOCK(inp);
23625         return (error);
23626 }
23627
23628 static void
23629 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
23630 {
23631
23632         INP_WLOCK_ASSERT(tptoinpcb(tp));
23633         bzero(ti, sizeof(*ti));
23634
23635         ti->tcpi_state = tp->t_state;
23636         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
23637                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
23638         if (tp->t_flags & TF_SACK_PERMIT)
23639                 ti->tcpi_options |= TCPI_OPT_SACK;
23640         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
23641                 ti->tcpi_options |= TCPI_OPT_WSCALE;
23642                 ti->tcpi_snd_wscale = tp->snd_scale;
23643                 ti->tcpi_rcv_wscale = tp->rcv_scale;
23644         }
23645         if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
23646                 ti->tcpi_options |= TCPI_OPT_ECN;
23647         if (tp->t_flags & TF_FASTOPEN)
23648                 ti->tcpi_options |= TCPI_OPT_TFO;
23649         /* still kept in ticks is t_rcvtime */
23650         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
23651         /* Since we hold everything in precise useconds this is easy */
23652         ti->tcpi_rtt = tp->t_srtt;
23653         ti->tcpi_rttvar = tp->t_rttvar;
23654         ti->tcpi_rto = tp->t_rxtcur;
23655         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
23656         ti->tcpi_snd_cwnd = tp->snd_cwnd;
23657         /*
23658          * FreeBSD-specific extension fields for tcp_info.
23659          */
23660         ti->tcpi_rcv_space = tp->rcv_wnd;
23661         ti->tcpi_rcv_nxt = tp->rcv_nxt;
23662         ti->tcpi_snd_wnd = tp->snd_wnd;
23663         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
23664         ti->tcpi_snd_nxt = tp->snd_nxt;
23665         ti->tcpi_snd_mss = tp->t_maxseg;
23666         ti->tcpi_rcv_mss = tp->t_maxseg;
23667         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
23668         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
23669         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
23670         ti->tcpi_total_tlp = tp->t_sndtlppack;
23671         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
23672 #ifdef NETFLIX_STATS
23673         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
23674 #endif
23675 #ifdef TCP_OFFLOAD
23676         if (tp->t_flags & TF_TOE) {
23677                 ti->tcpi_options |= TCPI_OPT_TOE;
23678                 tcp_offload_tcp_info(tp, ti);
23679         }
23680 #endif
23681 }
23682
23683 static int
23684 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
23685 {
23686         struct inpcb *inp = tptoinpcb(tp);
23687         struct tcp_rack *rack;
23688         int32_t error, optval;
23689         uint64_t val, loptval;
23690         struct  tcp_info ti;
23691         /*
23692          * Because all our options are either boolean or an int, we can just
23693          * pull everything into optval and then unlock and copy. If we ever
23694          * add a option that is not a int, then this will have quite an
23695          * impact to this routine.
23696          */
23697         error = 0;
23698         rack = (struct tcp_rack *)tp->t_fb_ptr;
23699         if (rack == NULL) {
23700                 INP_WUNLOCK(inp);
23701                 return (EINVAL);
23702         }
23703         switch (sopt->sopt_name) {
23704         case TCP_INFO:
23705                 /* First get the info filled */
23706                 rack_fill_info(tp, &ti);
23707                 /* Fix up the rtt related fields if needed */
23708                 INP_WUNLOCK(inp);
23709                 error = sooptcopyout(sopt, &ti, sizeof ti);
23710                 return (error);
23711         /*
23712          * Beta is the congestion control value for NewReno that influences how
23713          * much of a backoff happens when loss is detected. It is normally set
23714          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
23715          * when you exit recovery.
23716          */
23717         case TCP_RACK_PACING_BETA:
23718                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
23719                         error = EINVAL;
23720                 else if (rack->rc_pacing_cc_set == 0)
23721                         optval = rack->r_ctl.rc_saved_beta.beta;
23722                 else {
23723                         /*
23724                          * Reach out into the CC data and report back what
23725                          * I have previously set. Yeah it looks hackish but
23726                          * we don't want to report the saved values.
23727                          */
23728                         if (tp->t_ccv.cc_data)
23729                                 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
23730                         else
23731                                 error = EINVAL;
23732                 }
23733                 break;
23734                 /*
23735                  * Beta_ecn is the congestion control value for NewReno that influences how
23736                  * much of a backoff happens when a ECN mark is detected. It is normally set
23737                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
23738                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
23739                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
23740                  */
23741
23742         case TCP_RACK_PACING_BETA_ECN:
23743                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
23744                         error = EINVAL;
23745                 else if (rack->rc_pacing_cc_set == 0)
23746                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
23747                 else {
23748                         /*
23749                          * Reach out into the CC data and report back what
23750                          * I have previously set. Yeah it looks hackish but
23751                          * we don't want to report the saved values.
23752                          */
23753                         if (tp->t_ccv.cc_data)
23754                                 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
23755                         else
23756                                 error = EINVAL;
23757                 }
23758                 break;
23759         case TCP_RACK_DSACK_OPT:
23760                 optval = 0;
23761                 if (rack->rc_rack_tmr_std_based) {
23762                         optval |= 1;
23763                 }
23764                 if (rack->rc_rack_use_dsack) {
23765                         optval |= 2;
23766                 }
23767                 break;
23768         case TCP_RACK_ENABLE_HYSTART:
23769         {
23770                 if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
23771                         optval = RACK_HYSTART_ON;
23772                         if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
23773                                 optval = RACK_HYSTART_ON_W_SC;
23774                         if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
23775                                 optval = RACK_HYSTART_ON_W_SC_C;
23776                 } else {
23777                         optval = RACK_HYSTART_OFF;
23778                 }
23779         }
23780         break;
23781         case TCP_RACK_DGP_IN_REC:
23782                 optval = rack->r_ctl.full_dgp_in_rec;
23783                 break;
23784         case TCP_RACK_HI_BETA:
23785                 optval = rack->rack_hibeta;
23786                 break;
23787         case TCP_RXT_CLAMP:
23788                 optval = rack->r_ctl.saved_rxt_clamp_val;
23789                 break;
23790         case TCP_DEFER_OPTIONS:
23791                 optval = rack->defer_options;
23792                 break;
23793         case TCP_RACK_MEASURE_CNT:
23794                 optval = rack->r_ctl.req_measurements;
23795                 break;
23796         case TCP_REC_ABC_VAL:
23797                 optval = rack->r_use_labc_for_rec;
23798                 break;
23799         case TCP_RACK_ABC_VAL:
23800                 optval = rack->rc_labc;
23801                 break;
23802         case TCP_HDWR_UP_ONLY:
23803                 optval= rack->r_up_only;
23804                 break;
23805         case TCP_PACING_RATE_CAP:
23806                 loptval = rack->r_ctl.bw_rate_cap;
23807                 break;
23808         case TCP_RACK_PROFILE:
23809                 /* You cannot retrieve a profile, its write only */
23810                 error = EINVAL;
23811                 break;
23812         case TCP_HYBRID_PACING:
23813                 /* You cannot retrieve hybrid pacing information, its write only */
23814                 error = EINVAL;
23815                 break;
23816         case TCP_USE_CMP_ACKS:
23817                 optval = rack->r_use_cmp_ack;
23818                 break;
23819         case TCP_RACK_PACE_TO_FILL:
23820                 optval = rack->rc_pace_to_cwnd;
23821                 if (optval && rack->r_fill_less_agg)
23822                         optval++;
23823                 break;
23824         case TCP_RACK_NO_PUSH_AT_MAX:
23825                 optval = rack->r_ctl.rc_no_push_at_mrtt;
23826                 break;
23827         case TCP_SHARED_CWND_ENABLE:
23828                 optval = rack->rack_enable_scwnd;
23829                 break;
23830         case TCP_RACK_NONRXT_CFG_RATE:
23831                 optval = rack->rack_rec_nonrxt_use_cr;
23832                 break;
23833         case TCP_NO_PRR:
23834                 if (rack->rack_no_prr  == 1)
23835                         optval = 1;
23836                 else if (rack->no_prr_addback == 1)
23837                         optval = 2;
23838                 else
23839                         optval = 0;
23840                 break;
23841         case TCP_RACK_DO_DETECTION:
23842                 optval = rack->do_detection;
23843                 break;
23844         case TCP_RACK_MBUF_QUEUE:
23845                 /* Now do we use the LRO mbuf-queue feature */
23846                 optval = rack->r_mbuf_queue;
23847                 break;
23848         case TCP_TIMELY_DYN_ADJ:
23849                 optval = rack->rc_gp_dyn_mul;
23850                 break;
23851         case TCP_BBR_IWINTSO:
23852                 optval = rack->rc_init_win;
23853                 break;
23854         case TCP_RACK_TLP_REDUCE:
23855                 /* RACK TLP cwnd reduction (bool) */
23856                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
23857                 break;
23858         case TCP_BBR_RACK_INIT_RATE:
23859                 val = rack->r_ctl.init_rate;
23860                 /* convert to kbits per sec */
23861                 val *= 8;
23862                 val /= 1000;
23863                 optval = (uint32_t)val;
23864                 break;
23865         case TCP_RACK_FORCE_MSEG:
23866                 optval = rack->rc_force_max_seg;
23867                 break;
23868         case TCP_RACK_PACE_MIN_SEG:
23869                 optval = rack->r_ctl.rc_user_set_min_segs;
23870                 break;
23871         case TCP_RACK_PACE_MAX_SEG:
23872                 /* Max segments in a pace */
23873                 optval = rack->rc_user_set_max_segs;
23874                 break;
23875         case TCP_RACK_PACE_ALWAYS:
23876                 /* Use the always pace method */
23877                 optval = rack->rc_always_pace;
23878                 break;
23879         case TCP_RACK_PRR_SENDALOT:
23880                 /* Allow PRR to send more than one seg */
23881                 optval = rack->r_ctl.rc_prr_sendalot;
23882                 break;
23883         case TCP_RACK_MIN_TO:
23884                 /* Minimum time between rack t-o's in ms */
23885                 optval = rack->r_ctl.rc_min_to;
23886                 break;
23887         case TCP_RACK_SPLIT_LIMIT:
23888                 optval = rack->r_ctl.rc_split_limit;
23889                 break;
23890         case TCP_RACK_EARLY_SEG:
23891                 /* If early recovery max segments */
23892                 optval = rack->r_ctl.rc_early_recovery_segs;
23893                 break;
23894         case TCP_RACK_REORD_THRESH:
23895                 /* RACK reorder threshold (shift amount) */
23896                 optval = rack->r_ctl.rc_reorder_shift;
23897                 break;
23898         case TCP_RACK_REORD_FADE:
23899                 /* Does reordering fade after ms time */
23900                 optval = rack->r_ctl.rc_reorder_fade;
23901                 break;
23902         case TCP_BBR_USE_RACK_RR:
23903                 /* Do we use the rack cheat for rxt */
23904                 optval = rack->use_rack_rr;
23905                 break;
23906         case TCP_RACK_RR_CONF:
23907                 optval = rack->r_rr_config;
23908                 break;
23909         case TCP_HDWR_RATE_CAP:
23910                 optval = rack->r_rack_hw_rate_caps;
23911                 break;
23912         case TCP_BBR_HDWR_PACE:
23913                 optval = rack->rack_hdw_pace_ena;
23914                 break;
23915         case TCP_RACK_TLP_THRESH:
23916                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
23917                 optval = rack->r_ctl.rc_tlp_threshold;
23918                 break;
23919         case TCP_RACK_PKT_DELAY:
23920                 /* RACK added ms i.e. rack-rtt + reord + N */
23921                 optval = rack->r_ctl.rc_pkt_delay;
23922                 break;
23923         case TCP_RACK_TLP_USE:
23924                 optval = rack->rack_tlp_threshold_use;
23925                 break;
23926         case TCP_PACING_DND:
23927                 optval = rack->rc_pace_dnd;
23928                 break;
23929         case TCP_RACK_PACE_RATE_CA:
23930                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
23931                 break;
23932         case TCP_RACK_PACE_RATE_SS:
23933                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
23934                 break;
23935         case TCP_RACK_PACE_RATE_REC:
23936                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
23937                 break;
23938         case TCP_RACK_GP_INCREASE_SS:
23939                 optval = rack->r_ctl.rack_per_of_gp_ca;
23940                 break;
23941         case TCP_RACK_GP_INCREASE_CA:
23942                 optval = rack->r_ctl.rack_per_of_gp_ss;
23943                 break;
23944         case TCP_RACK_PACING_DIVISOR:
23945                 optval = rack->r_ctl.pace_len_divisor;
23946                 break;
23947         case TCP_BBR_RACK_RTT_USE:
23948                 optval = rack->r_ctl.rc_rate_sample_method;
23949                 break;
23950         case TCP_DELACK:
23951                 optval = tp->t_delayed_ack;
23952                 break;
23953         case TCP_DATA_AFTER_CLOSE:
23954                 optval = rack->rc_allow_data_af_clo;
23955                 break;
23956         case TCP_SHARED_CWND_TIME_LIMIT:
23957                 optval = rack->r_limit_scw;
23958                 break;
23959         case TCP_RACK_TIMER_SLOP:
23960                 optval = rack->r_ctl.timer_slop;
23961                 break;
23962         default:
23963                 return (tcp_default_ctloutput(tp, sopt));
23964                 break;
23965         }
23966         INP_WUNLOCK(inp);
23967         if (error == 0) {
23968                 if (TCP_PACING_RATE_CAP)
23969                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
23970                 else
23971                         error = sooptcopyout(sopt, &optval, sizeof optval);
23972         }
23973         return (error);
23974 }
23975
23976 static int
23977 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
23978 {
23979         if (sopt->sopt_dir == SOPT_SET) {
23980                 return (rack_set_sockopt(tp, sopt));
23981         } else if (sopt->sopt_dir == SOPT_GET) {
23982                 return (rack_get_sockopt(tp, sopt));
23983         } else {
23984                 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
23985         }
23986 }
23987
23988 static const char *rack_stack_names[] = {
23989         __XSTRING(STACKNAME),
23990 #ifdef STACKALIAS
23991         __XSTRING(STACKALIAS),
23992 #endif
23993 };
23994
23995 static int
23996 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
23997 {
23998         memset(mem, 0, size);
23999         return (0);
24000 }
24001
24002 static void
24003 rack_dtor(void *mem, int32_t size, void *arg)
24004 {
24005
24006 }
24007
24008 static bool rack_mod_inited = false;
24009
24010 static int
24011 tcp_addrack(module_t mod, int32_t type, void *data)
24012 {
24013         int32_t err = 0;
24014         int num_stacks;
24015
24016         switch (type) {
24017         case MOD_LOAD:
24018                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
24019                     sizeof(struct rack_sendmap),
24020                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
24021
24022                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
24023                     sizeof(struct tcp_rack),
24024                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
24025
24026                 sysctl_ctx_init(&rack_sysctl_ctx);
24027                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
24028                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
24029                     OID_AUTO,
24030 #ifdef STACKALIAS
24031                     __XSTRING(STACKALIAS),
24032 #else
24033                     __XSTRING(STACKNAME),
24034 #endif
24035                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
24036                     "");
24037                 if (rack_sysctl_root == NULL) {
24038                         printf("Failed to add sysctl node\n");
24039                         err = EFAULT;
24040                         goto free_uma;
24041                 }
24042                 rack_init_sysctls();
24043                 num_stacks = nitems(rack_stack_names);
24044                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
24045                     rack_stack_names, &num_stacks);
24046                 if (err) {
24047                         printf("Failed to register %s stack name for "
24048                             "%s module\n", rack_stack_names[num_stacks],
24049                             __XSTRING(MODNAME));
24050                         sysctl_ctx_free(&rack_sysctl_ctx);
24051 free_uma:
24052                         uma_zdestroy(rack_zone);
24053                         uma_zdestroy(rack_pcb_zone);
24054                         rack_counter_destroy();
24055                         printf("Failed to register rack module -- err:%d\n", err);
24056                         return (err);
24057                 }
24058                 tcp_lro_reg_mbufq();
24059                 rack_mod_inited = true;
24060                 break;
24061         case MOD_QUIESCE:
24062                 err = deregister_tcp_functions(&__tcp_rack, true, false);
24063                 break;
24064         case MOD_UNLOAD:
24065                 err = deregister_tcp_functions(&__tcp_rack, false, true);
24066                 if (err == EBUSY)
24067                         break;
24068                 if (rack_mod_inited) {
24069                         uma_zdestroy(rack_zone);
24070                         uma_zdestroy(rack_pcb_zone);
24071                         sysctl_ctx_free(&rack_sysctl_ctx);
24072                         rack_counter_destroy();
24073                         rack_mod_inited = false;
24074                 }
24075                 tcp_lro_dereg_mbufq();
24076                 err = 0;
24077                 break;
24078         default:
24079                 return (EOPNOTSUPP);
24080         }
24081         return (err);
24082 }
24083
24084 static moduledata_t tcp_rack = {
24085         .name = __XSTRING(MODNAME),
24086         .evhand = tcp_addrack,
24087         .priv = 0
24088 };
24089
24090 MODULE_VERSION(MODNAME, 1);
24091 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
24092 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
24093
24094 #endif /* #if !defined(INET) && !defined(INET6) */