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_tcpdebug.h"
  34 #include "opt_ratelimit.h"
  35 #include "opt_kern_tls.h"
  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_log_buf.h>
  93 #include <netinet/tcp_seq.h>
  94 #include <netinet/tcp_timer.h>
  95 #include <netinet/tcp_var.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 TCPDEBUG
 109 #include <netinet/tcp_debug.h>
 110 #endif                          /* TCPDEBUG */
 111 #ifdef TCP_OFFLOAD
 112 #include <netinet/tcp_offload.h>
 113 #endif
 114 #ifdef INET6
 115 #include <netinet6/tcp6_var.h>
 116 #endif
 117 #include <netinet/tcp_ecn.h>
 118
 119 #include <netipsec/ipsec_support.h>
 120
 121 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 122 #include <netipsec/ipsec.h>
 123 #include <netipsec/ipsec6.h>
 124 #endif                          /* IPSEC */
 125
 126 #include <netinet/udp.h>
 127 #include <netinet/udp_var.h>
 128 #include <machine/in_cksum.h>
 129
 130 #ifdef MAC
 131 #include <security/mac/mac_framework.h>
 132 #endif
 133 #include "sack_filter.h"
 134 #include "tcp_rack.h"
 135 #include "rack_bbr_common.h"
 136
 137 uma_zone_t rack_zone;
 138 uma_zone_t rack_pcb_zone;
 139
 140 #ifndef TICKS2SBT
 141 #define TICKS2SBT(__t)  (tick_sbt * ((sbintime_t)(__t)))
 142 #endif
 143
 144 VNET_DECLARE(uint32_t, newreno_beta);
 145 VNET_DECLARE(uint32_t, newreno_beta_ecn);
 146 #define V_newreno_beta VNET(newreno_beta)
 147 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
 148
 149
 150 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
 151 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
 152
 153 struct sysctl_ctx_list rack_sysctl_ctx;
 154 struct sysctl_oid *rack_sysctl_root;
 155
 156 #define CUM_ACKED 1
 157 #define SACKED 2
 158
 159 /*
 160  * The RACK module incorporates a number of
 161  * TCP ideas that have been put out into the IETF
 162  * over the last few years:
 163  * - Matt Mathis's Rate Halving which slowly drops
 164  *    the congestion window so that the ack clock can
 165  *    be maintained during a recovery.
 166  * - Yuchung Cheng's RACK TCP (for which its named) that
 167  *    will stop us using the number of dup acks and instead
 168  *    use time as the gage of when we retransmit.
 169  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
 170  *    of Dukkipati et.al.
 171  * RACK depends on SACK, so if an endpoint arrives that
 172  * cannot do SACK the state machine below will shuttle the
 173  * connection back to using the "default" TCP stack that is
 174  * in FreeBSD.
 175  *
 176  * To implement RACK the original TCP stack was first decomposed
 177  * into a functional state machine with individual states
 178  * for each of the possible TCP connection states. The do_segment
 179  * functions role in life is to mandate the connection supports SACK
 180  * initially and then assure that the RACK state matches the conenction
 181  * state before calling the states do_segment function. Each
 182  * state is simplified due to the fact that the original do_segment
 183  * has been decomposed and we *know* what state we are in (no
 184  * switches on the state) and all tests for SACK are gone. This
 185  * greatly simplifies what each state does.
 186  *
 187  * TCP output is also over-written with a new version since it
 188  * must maintain the new rack scoreboard.
 189  *
 190  */
 191 static int32_t rack_tlp_thresh = 1;
 192 static int32_t rack_tlp_limit = 2;      /* No more than 2 TLPs w-out new data */
 193 static int32_t rack_tlp_use_greater = 1;
 194 static int32_t rack_reorder_thresh = 2;
 195 static int32_t rack_reorder_fade = 60000000;    /* 0 - never fade, def 60,000,000
 196                                                  * - 60 seconds */
 197 static uint8_t rack_req_measurements = 1;
 198 /* Attack threshold detections */
 199 static uint32_t rack_highest_sack_thresh_seen = 0;
 200 static uint32_t rack_highest_move_thresh_seen = 0;
 201 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
 202 static int32_t rack_hw_pace_extra_slots = 2;    /* 2 extra MSS time betweens */
 203 static int32_t rack_hw_rate_caps = 1; /* 1; */
 204 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
 205 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
 206 static int32_t rack_hw_up_only = 1;
 207 static int32_t rack_stats_gets_ms_rtt = 1;
 208 static int32_t rack_prr_addbackmax = 2;
 209 static int32_t rack_do_hystart = 0;
 210 static int32_t rack_apply_rtt_with_reduced_conf = 0;
 211
 212 static int32_t rack_pkt_delay = 1000;
 213 static int32_t rack_send_a_lot_in_prr = 1;
 214 static int32_t rack_min_to = 1000;      /* Number of microsecond  min timeout */
 215 static int32_t rack_verbose_logging = 0;
 216 static int32_t rack_ignore_data_after_close = 1;
 217 static int32_t rack_enable_shared_cwnd = 1;
 218 static int32_t rack_use_cmp_acks = 1;
 219 static int32_t rack_use_fsb = 1;
 220 static int32_t rack_use_rfo = 1;
 221 static int32_t rack_use_rsm_rfo = 1;
 222 static int32_t rack_max_abc_post_recovery = 2;
 223 static int32_t rack_client_low_buf = 0;
 224 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 */
 225 #ifdef TCP_ACCOUNTING
 226 static int32_t rack_tcp_accounting = 0;
 227 #endif
 228 static int32_t rack_limits_scwnd = 1;
 229 static int32_t rack_enable_mqueue_for_nonpaced = 0;
 230 static int32_t rack_disable_prr = 0;
 231 static int32_t use_rack_rr = 1;
 232 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
 233 static int32_t rack_persist_min = 250000;       /* 250usec */
 234 static int32_t rack_persist_max = 2000000;      /* 2 Second in usec's */
 235 static int32_t rack_sack_not_required = 1;      /* set to one to allow non-sack to use rack */
 236 static int32_t rack_default_init_window = 0;    /* Use system default */
 237 static int32_t rack_limit_time_with_srtt = 0;
 238 static int32_t rack_autosndbuf_inc = 20;        /* In percentage form */
 239 static int32_t rack_enobuf_hw_boost_mult = 2;   /* How many times the hw rate we boost slot using time_between */
 240 static int32_t rack_enobuf_hw_max = 12000;      /* 12 ms in usecs */
 241 static int32_t rack_enobuf_hw_min = 10000;      /* 10 ms in usecs */
 242 static int32_t rack_hw_rwnd_factor = 2;         /* How many max_segs the rwnd must be before we hold off sending */
 243
 244 /*
 245  * Currently regular tcp has a rto_min of 30ms
 246  * the backoff goes 12 times so that ends up
 247  * being a total of 122.850 seconds before a
 248  * connection is killed.
 249  */
 250 static uint32_t rack_def_data_window = 20;
 251 static uint32_t rack_goal_bdp = 2;
 252 static uint32_t rack_min_srtts = 1;
 253 static uint32_t rack_min_measure_usec = 0;
 254 static int32_t rack_tlp_min = 10000;    /* 10ms */
 255 static int32_t rack_rto_min = 30000;    /* 30,000 usec same as main freebsd */
 256 static int32_t rack_rto_max = 4000000;  /* 4 seconds in usec's */
 257 static const int32_t rack_free_cache = 2;
 258 static int32_t rack_hptsi_segments = 40;
 259 static int32_t rack_rate_sample_method = USE_RTT_LOW;
 260 static int32_t rack_pace_every_seg = 0;
 261 static int32_t rack_delayed_ack_time = 40000;   /* 40ms in usecs */
 262 static int32_t rack_slot_reduction = 4;
 263 static int32_t rack_wma_divisor = 8;            /* For WMA calculation */
 264 static int32_t rack_cwnd_block_ends_measure = 0;
 265 static int32_t rack_rwnd_block_ends_measure = 0;
 266 static int32_t rack_def_profile = 0;
 267
 268 static int32_t rack_lower_cwnd_at_tlp = 0;
 269 static int32_t rack_limited_retran = 0;
 270 static int32_t rack_always_send_oldest = 0;
 271 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
 272
 273 static uint16_t rack_per_of_gp_ss = 250;        /* 250 % slow-start */
 274 static uint16_t rack_per_of_gp_ca = 200;        /* 200 % congestion-avoidance */
 275 static uint16_t rack_per_of_gp_rec = 200;       /* 200 % of bw */
 276
 277 /* Probertt */
 278 static uint16_t rack_per_of_gp_probertt = 60;   /* 60% of bw */
 279 static uint16_t rack_per_of_gp_lowthresh = 40;  /* 40% is bottom */
 280 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
 281 static uint16_t rack_atexit_prtt_hbp = 130;     /* Clamp to 130% on exit prtt if highly buffered path */
 282 static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */
 283
 284 static uint32_t rack_max_drain_wait = 2;        /* How man gp srtt's before we give up draining */
 285 static uint32_t rack_must_drain = 1;            /* How many GP srtt's we *must* wait */
 286 static uint32_t rack_probertt_use_min_rtt_entry = 1;    /* Use the min to calculate the goal else gp_srtt */
 287 static uint32_t rack_probertt_use_min_rtt_exit = 0;
 288 static uint32_t rack_probe_rtt_sets_cwnd = 0;
 289 static uint32_t rack_probe_rtt_safety_val = 2000000;    /* No more than 2 sec in probe-rtt */
 290 static uint32_t rack_time_between_probertt = 9600000;   /* 9.6 sec in usecs */
 291 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;       /* How many srtt periods does probe-rtt last top fraction */
 292 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;       /* How many srtt periods does probe-rtt last bottom fraction */
 293 static uint32_t rack_min_probertt_hold = 40000;         /* Equal to delayed ack time */
 294 static uint32_t rack_probertt_filter_life = 10000000;
 295 static uint32_t rack_probertt_lower_within = 10;
 296 static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds)  to count as a lowering */
 297 static int32_t rack_pace_one_seg = 0;           /* Shall we pace for less than 1.4Meg 1MSS at a time */
 298 static int32_t rack_probertt_clear_is = 1;
 299 static int32_t rack_max_drain_hbp = 1;          /* Extra drain times gpsrtt for highly buffered paths */
 300 static int32_t rack_hbp_thresh = 3;             /* what is the divisor max_rtt/min_rtt to decided a hbp */
 301
 302 /* Part of pacing */
 303 static int32_t rack_max_per_above = 30;         /* When we go to increment stop if above 100+this% */
 304
 305 /* Timely information */
 306 /* Combine these two gives the range of 'no change' to bw */
 307 /* ie the up/down provide the upper and lower bound */
 308 static int32_t rack_gp_per_bw_mul_up = 2;       /* 2% */
 309 static int32_t rack_gp_per_bw_mul_down = 4;     /* 4% */
 310 static int32_t rack_gp_rtt_maxmul = 3;          /* 3 x maxmin */
 311 static int32_t rack_gp_rtt_minmul = 1;          /* minrtt + (minrtt/mindiv) is lower rtt */
 312 static int32_t rack_gp_rtt_mindiv = 4;          /* minrtt + (minrtt * minmul/mindiv) is lower rtt */
 313 static int32_t rack_gp_decrease_per = 20;       /* 20% decrease in multiplier */
 314 static int32_t rack_gp_increase_per = 2;        /* 2% increase in multiplier */
 315 static int32_t rack_per_lower_bound = 50;       /* Don't allow to drop below this multiplier */
 316 static int32_t rack_per_upper_bound_ss = 0;     /* Don't allow SS to grow above this */
 317 static int32_t rack_per_upper_bound_ca = 0;     /* Don't allow CA to grow above this */
 318 static int32_t rack_do_dyn_mul = 0;             /* Are the rack gp multipliers dynamic */
 319 static int32_t rack_gp_no_rec_chg = 1;          /* Prohibit recovery from reducing it's multiplier */
 320 static int32_t rack_timely_dec_clear = 6;       /* Do we clear decrement count at a value (6)? */
 321 static int32_t rack_timely_max_push_rise = 3;   /* One round of pushing */
 322 static int32_t rack_timely_max_push_drop = 3;   /* Three round of pushing */
 323 static int32_t rack_timely_min_segs = 4;        /* 4 segment minimum */
 324 static int32_t rack_use_max_for_nobackoff = 0;
 325 static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */
 326 static int32_t rack_timely_no_stopping = 0;
 327 static int32_t rack_down_raise_thresh = 100;
 328 static int32_t rack_req_segs = 1;
 329 static uint64_t rack_bw_rate_cap = 0;
 330 static uint32_t rack_trace_point_config = 0;
 331 static uint32_t rack_trace_point_bb_mode = 4;
 332 static int32_t rack_trace_point_count = 0;
 333
 334
 335 /* Weird delayed ack mode */
 336 static int32_t rack_use_imac_dack = 0;
 337 /* Rack specific counters */
 338 counter_u64_t rack_saw_enobuf;
 339 counter_u64_t rack_saw_enobuf_hw;
 340 counter_u64_t rack_saw_enetunreach;
 341 counter_u64_t rack_persists_sends;
 342 counter_u64_t rack_persists_acks;
 343 counter_u64_t rack_persists_loss;
 344 counter_u64_t rack_persists_lost_ends;
 345 #ifdef INVARIANTS
 346 counter_u64_t rack_adjust_map_bw;
 347 #endif
 348 /* Tail loss probe counters */
 349 counter_u64_t rack_tlp_tot;
 350 counter_u64_t rack_tlp_newdata;
 351 counter_u64_t rack_tlp_retran;
 352 counter_u64_t rack_tlp_retran_bytes;
 353 counter_u64_t rack_to_tot;
 354 counter_u64_t rack_hot_alloc;
 355 counter_u64_t rack_to_alloc;
 356 counter_u64_t rack_to_alloc_hard;
 357 counter_u64_t rack_to_alloc_emerg;
 358 counter_u64_t rack_to_alloc_limited;
 359 counter_u64_t rack_alloc_limited_conns;
 360 counter_u64_t rack_split_limited;
 361
 362 counter_u64_t rack_multi_single_eq;
 363 counter_u64_t rack_proc_non_comp_ack;
 364
 365 counter_u64_t rack_fto_send;
 366 counter_u64_t rack_fto_rsm_send;
 367 counter_u64_t rack_nfto_resend;
 368 counter_u64_t rack_non_fto_send;
 369 counter_u64_t rack_extended_rfo;
 370
 371 counter_u64_t rack_sack_proc_all;
 372 counter_u64_t rack_sack_proc_short;
 373 counter_u64_t rack_sack_proc_restart;
 374 counter_u64_t rack_sack_attacks_detected;
 375 counter_u64_t rack_sack_attacks_reversed;
 376 counter_u64_t rack_sack_used_next_merge;
 377 counter_u64_t rack_sack_splits;
 378 counter_u64_t rack_sack_used_prev_merge;
 379 counter_u64_t rack_sack_skipped_acked;
 380 counter_u64_t rack_ack_total;
 381 counter_u64_t rack_express_sack;
 382 counter_u64_t rack_sack_total;
 383 counter_u64_t rack_move_none;
 384 counter_u64_t rack_move_some;
 385
 386 counter_u64_t rack_input_idle_reduces;
 387 counter_u64_t rack_collapsed_win;
 388 counter_u64_t rack_collapsed_win_seen;
 389 counter_u64_t rack_collapsed_win_rxt;
 390 counter_u64_t rack_collapsed_win_rxt_bytes;
 391 counter_u64_t rack_try_scwnd;
 392 counter_u64_t rack_hw_pace_init_fail;
 393 counter_u64_t rack_hw_pace_lost;
 394
 395 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
 396 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
 397
 398
 399 #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
 400
 401 #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {  \
 402         (tv) = (value) + slop;   \
 403         if ((u_long)(tv) < (u_long)(tvmin)) \
 404                 (tv) = (tvmin); \
 405         if ((u_long)(tv) > (u_long)(tvmax)) \
 406                 (tv) = (tvmax); \
 407 } while (0)
 408
 409 static void
 410 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
 411
 412 static int
 413 rack_process_ack(struct mbuf *m, struct tcphdr *th,
 414     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
 415     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
 416 static int
 417 rack_process_data(struct mbuf *m, struct tcphdr *th,
 418     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 419     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
 420 static void
 421 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
 422    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
 423 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
 424 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
 425     uint8_t limit_type);
 426 static struct rack_sendmap *
 427 rack_check_recovery_mode(struct tcpcb *tp,
 428     uint32_t tsused);
 429 static void
 430 rack_cong_signal(struct tcpcb *tp,
 431                  uint32_t type, uint32_t ack, int );
 432 static void rack_counter_destroy(void);
 433 static int
 434 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt);
 435 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
 436 static void
 437 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
 438 static void
 439 rack_do_segment(struct mbuf *m, struct tcphdr *th,
 440     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 441     uint8_t iptos);
 442 static void rack_dtor(void *mem, int32_t size, void *arg);
 443 static void
 444 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
 445     uint32_t flex1, uint32_t flex2,
 446     uint32_t flex3, uint32_t flex4,
 447     uint32_t flex5, uint32_t flex6,
 448     uint16_t flex7, uint8_t mod);
 449
 450 static void
 451 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
 452    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
 453    struct rack_sendmap *rsm, uint8_t quality);
 454 static struct rack_sendmap *
 455 rack_find_high_nonack(struct tcp_rack *rack,
 456     struct rack_sendmap *rsm);
 457 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
 458 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
 459 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
 460 static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt);
 461 static void
 462 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
 463                             tcp_seq th_ack, int line, uint8_t quality);
 464 static uint32_t
 465 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
 466 static int32_t rack_handoff_ok(struct tcpcb *tp);
 467 static int32_t rack_init(struct tcpcb *tp);
 468 static void rack_init_sysctls(void);
 469 static void
 470 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
 471     struct tcphdr *th, int entered_rec, int dup_ack_struck);
 472 static void
 473 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
 474     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
 475     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
 476
 477 static void
 478 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
 479     struct rack_sendmap *rsm);
 480 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
 481 static int32_t rack_output(struct tcpcb *tp);
 482
 483 static uint32_t
 484 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
 485     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
 486     uint32_t cts, int *moved_two);
 487 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
 488 static void rack_remxt_tmr(struct tcpcb *tp);
 489 static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt);
 490 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
 491 static int32_t rack_stopall(struct tcpcb *tp);
 492 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
 493 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
 494 static uint32_t
 495 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
 496     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
 497 static void
 498 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
 499     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
 500 static int
 501 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
 502     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
 503 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
 504 static int
 505 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
 506     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 507     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 508 static int
 509 rack_do_closing(struct mbuf *m, struct tcphdr *th,
 510     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 511     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 512 static int
 513 rack_do_established(struct mbuf *m, struct tcphdr *th,
 514     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 515     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 516 static int
 517 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
 518     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 519     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
 520 static int
 521 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
 522     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 523     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 524 static int
 525 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
 526     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 527     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 528 static int
 529 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
 530     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 531     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 532 static int
 533 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
 534     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 535     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 536 static int
 537 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
 538     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 539     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 540 struct rack_sendmap *
 541 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
 542     uint32_t tsused);
 543 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
 544     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
 545 static void
 546      tcp_rack_partialack(struct tcpcb *tp);
 547 static int
 548 rack_set_profile(struct tcp_rack *rack, int prof);
 549 static void
 550 rack_apply_deferred_options(struct tcp_rack *rack);
 551
 552 int32_t rack_clear_counter=0;
 553
 554 static inline void
 555 rack_trace_point(struct tcp_rack *rack, int num)
 556 {
 557         if (((rack_trace_point_config == num)  ||
 558              (rack_trace_point_config = 0xffffffff)) &&
 559             (rack_trace_point_bb_mode != 0) &&
 560             (rack_trace_point_count > 0) &&
 561             (rack->rc_tp->t_logstate == 0)) {
 562                 int res;
 563                 res = atomic_fetchadd_int(&rack_trace_point_count, -1);
 564                 if (res > 0) {
 565                         rack->rc_tp->t_logstate = rack_trace_point_bb_mode;
 566                 } else {
 567                         /* Loss a race assure its zero now */
 568                         rack_trace_point_count = 0;
 569                 }
 570         }
 571 }
 572
 573 static void
 574 rack_set_cc_pacing(struct tcp_rack *rack)
 575 {
 576         struct sockopt sopt;
 577         struct cc_newreno_opts opt;
 578         struct newreno old, *ptr;
 579         struct tcpcb *tp;
 580         int error;
 581
 582         if (rack->rc_pacing_cc_set)
 583                 return;
 584
 585         tp = rack->rc_tp;
 586         if (tp->t_cc == NULL) {
 587                 /* Tcb is leaving */
 588                 return;
 589         }
 590         rack->rc_pacing_cc_set = 1;
 591         if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
 592                 /* Not new-reno we can't play games with beta! */
 593                 goto out;
 594         }
 595         ptr = ((struct newreno *)tp->t_ccv.cc_data);
 596         if (CC_ALGO(tp)->ctl_output == NULL)  {
 597                 /* Huh, why does new_reno no longer have a set function? */
 598                 goto out;
 599         }
 600         if (ptr == NULL) {
 601                 /* Just the default values */
 602                 old.beta = V_newreno_beta_ecn;
 603                 old.beta_ecn = V_newreno_beta_ecn;
 604                 old.newreno_flags = 0;
 605         } else {
 606                 old.beta = ptr->beta;
 607                 old.beta_ecn = ptr->beta_ecn;
 608                 old.newreno_flags = ptr->newreno_flags;
 609         }
 610         sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
 611         sopt.sopt_dir = SOPT_SET;
 612         opt.name = CC_NEWRENO_BETA;
 613         opt.val = rack->r_ctl.rc_saved_beta.beta;
 614         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 615         if (error)  {
 616                 goto out;
 617         }
 618         /*
 619          * Hack alert we need to set in our newreno_flags
 620          * so that Abe behavior is also applied.
 621          */
 622         ((struct newreno *)tp->t_ccv.cc_data)->newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
 623         opt.name = CC_NEWRENO_BETA_ECN;
 624         opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
 625         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 626         if (error) {
 627                 goto out;
 628         }
 629         /* Save off the original values for restoral */
 630         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 631 out:
 632         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 633                 union tcp_log_stackspecific log;
 634                 struct timeval tv;
 635
 636                 ptr = ((struct newreno *)tp->t_ccv.cc_data);
 637                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 638                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 639                 if (ptr) {
 640                         log.u_bbr.flex1 = ptr->beta;
 641                         log.u_bbr.flex2 = ptr->beta_ecn;
 642                         log.u_bbr.flex3 = ptr->newreno_flags;
 643                 }
 644                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 645                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 646                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 647                 log.u_bbr.flex7 = rack->gp_ready;
 648                 log.u_bbr.flex7 <<= 1;
 649                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 650                 log.u_bbr.flex7 <<= 1;
 651                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 652                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 653                 log.u_bbr.flex8 = 3;
 654                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
 655                                0, &log, false, NULL, NULL, 0, &tv);
 656         }
 657 }
 658
 659 static void
 660 rack_undo_cc_pacing(struct tcp_rack *rack)
 661 {
 662         struct newreno old, *ptr;
 663         struct tcpcb *tp;
 664
 665         if (rack->rc_pacing_cc_set == 0)
 666                 return;
 667         tp = rack->rc_tp;
 668         rack->rc_pacing_cc_set = 0;
 669         if (tp->t_cc == NULL)
 670                 /* Tcb is leaving */
 671                 return;
 672         if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
 673                 /* Not new-reno nothing to do! */
 674                 return;
 675         }
 676         ptr = ((struct newreno *)tp->t_ccv.cc_data);
 677         if (ptr == NULL) {
 678                 /*
 679                  * This happens at rack_fini() if the
 680                  * cc module gets freed on us. In that
 681                  * case we loose our "new" settings but
 682                  * thats ok, since the tcb is going away anyway.
 683                  */
 684                 return;
 685         }
 686         /* Grab out our set values */
 687         memcpy(&old, ptr, sizeof(struct newreno));
 688         /* Copy back in the original values */
 689         memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
 690         /* Now save back the values we had set in (for when pacing is restored) */
 691         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 692         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 693                 union tcp_log_stackspecific log;
 694                 struct timeval tv;
 695
 696                 ptr = ((struct newreno *)tp->t_ccv.cc_data);
 697                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 698                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 699                 log.u_bbr.flex1 = ptr->beta;
 700                 log.u_bbr.flex2 = ptr->beta_ecn;
 701                 log.u_bbr.flex3 = ptr->newreno_flags;
 702                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 703                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 704                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 705                 log.u_bbr.flex7 = rack->gp_ready;
 706                 log.u_bbr.flex7 <<= 1;
 707                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 708                 log.u_bbr.flex7 <<= 1;
 709                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 710                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 711                 log.u_bbr.flex8 = 4;
 712                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
 713                                0, &log, false, NULL, NULL, 0, &tv);
 714         }
 715 }
 716
 717 #ifdef NETFLIX_PEAKRATE
 718 static inline void
 719 rack_update_peakrate_thr(struct tcpcb *tp)
 720 {
 721         /* Keep in mind that t_maxpeakrate is in B/s. */
 722         uint64_t peak;
 723         peak = uqmax((tp->t_maxseg * 2),
 724                      (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
 725         tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
 726 }
 727 #endif
 728
 729 static int
 730 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
 731 {
 732         uint32_t stat;
 733         int32_t error;
 734
 735         error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
 736         if (error || req->newptr == NULL)
 737                 return error;
 738
 739         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
 740         if (error)
 741                 return (error);
 742         if (stat == 1) {
 743 #ifdef INVARIANTS
 744                 printf("Clearing RACK counters\n");
 745 #endif
 746                 counter_u64_zero(rack_tlp_tot);
 747                 counter_u64_zero(rack_tlp_newdata);
 748                 counter_u64_zero(rack_tlp_retran);
 749                 counter_u64_zero(rack_tlp_retran_bytes);
 750                 counter_u64_zero(rack_to_tot);
 751                 counter_u64_zero(rack_saw_enobuf);
 752                 counter_u64_zero(rack_saw_enobuf_hw);
 753                 counter_u64_zero(rack_saw_enetunreach);
 754                 counter_u64_zero(rack_persists_sends);
 755                 counter_u64_zero(rack_persists_acks);
 756                 counter_u64_zero(rack_persists_loss);
 757                 counter_u64_zero(rack_persists_lost_ends);
 758 #ifdef INVARIANTS
 759                 counter_u64_zero(rack_adjust_map_bw);
 760 #endif
 761                 counter_u64_zero(rack_to_alloc_hard);
 762                 counter_u64_zero(rack_to_alloc_emerg);
 763                 counter_u64_zero(rack_sack_proc_all);
 764                 counter_u64_zero(rack_fto_send);
 765                 counter_u64_zero(rack_fto_rsm_send);
 766                 counter_u64_zero(rack_extended_rfo);
 767                 counter_u64_zero(rack_hw_pace_init_fail);
 768                 counter_u64_zero(rack_hw_pace_lost);
 769                 counter_u64_zero(rack_non_fto_send);
 770                 counter_u64_zero(rack_nfto_resend);
 771                 counter_u64_zero(rack_sack_proc_short);
 772                 counter_u64_zero(rack_sack_proc_restart);
 773                 counter_u64_zero(rack_to_alloc);
 774                 counter_u64_zero(rack_to_alloc_limited);
 775                 counter_u64_zero(rack_alloc_limited_conns);
 776                 counter_u64_zero(rack_split_limited);
 777                 counter_u64_zero(rack_multi_single_eq);
 778                 counter_u64_zero(rack_proc_non_comp_ack);
 779                 counter_u64_zero(rack_sack_attacks_detected);
 780                 counter_u64_zero(rack_sack_attacks_reversed);
 781                 counter_u64_zero(rack_sack_used_next_merge);
 782                 counter_u64_zero(rack_sack_used_prev_merge);
 783                 counter_u64_zero(rack_sack_splits);
 784                 counter_u64_zero(rack_sack_skipped_acked);
 785                 counter_u64_zero(rack_ack_total);
 786                 counter_u64_zero(rack_express_sack);
 787                 counter_u64_zero(rack_sack_total);
 788                 counter_u64_zero(rack_move_none);
 789                 counter_u64_zero(rack_move_some);
 790                 counter_u64_zero(rack_try_scwnd);
 791                 counter_u64_zero(rack_collapsed_win);
 792                 counter_u64_zero(rack_collapsed_win_rxt);
 793                 counter_u64_zero(rack_collapsed_win_seen);
 794                 counter_u64_zero(rack_collapsed_win_rxt_bytes);
 795         }
 796         rack_clear_counter = 0;
 797         return (0);
 798 }
 799
 800 static void
 801 rack_init_sysctls(void)
 802 {
 803         struct sysctl_oid *rack_counters;
 804         struct sysctl_oid *rack_attack;
 805         struct sysctl_oid *rack_pacing;
 806         struct sysctl_oid *rack_timely;
 807         struct sysctl_oid *rack_timers;
 808         struct sysctl_oid *rack_tlp;
 809         struct sysctl_oid *rack_misc;
 810         struct sysctl_oid *rack_features;
 811         struct sysctl_oid *rack_measure;
 812         struct sysctl_oid *rack_probertt;
 813         struct sysctl_oid *rack_hw_pacing;
 814         struct sysctl_oid *rack_tracepoint;
 815
 816         rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 817             SYSCTL_CHILDREN(rack_sysctl_root),
 818             OID_AUTO,
 819             "sack_attack",
 820             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 821             "Rack Sack Attack Counters and Controls");
 822         rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 823             SYSCTL_CHILDREN(rack_sysctl_root),
 824             OID_AUTO,
 825             "stats",
 826             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 827             "Rack Counters");
 828         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 829             SYSCTL_CHILDREN(rack_sysctl_root),
 830             OID_AUTO, "rate_sample_method", CTLFLAG_RW,
 831             &rack_rate_sample_method , USE_RTT_LOW,
 832             "What method should we use for rate sampling 0=high, 1=low ");
 833         /* Probe rtt related controls */
 834         rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 835             SYSCTL_CHILDREN(rack_sysctl_root),
 836             OID_AUTO,
 837             "probertt",
 838             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 839             "ProbeRTT related Controls");
 840         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 841             SYSCTL_CHILDREN(rack_probertt),
 842             OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
 843             &rack_atexit_prtt_hbp, 130,
 844             "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
 845         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 846             SYSCTL_CHILDREN(rack_probertt),
 847             OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
 848             &rack_atexit_prtt, 130,
 849             "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
 850         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 851             SYSCTL_CHILDREN(rack_probertt),
 852             OID_AUTO, "gp_per_mul", CTLFLAG_RW,
 853             &rack_per_of_gp_probertt, 60,
 854             "What percentage of goodput do we pace at in probertt");
 855         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 856             SYSCTL_CHILDREN(rack_probertt),
 857             OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
 858             &rack_per_of_gp_probertt_reduce, 10,
 859             "What percentage of goodput do we reduce every gp_srtt");
 860         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 861             SYSCTL_CHILDREN(rack_probertt),
 862             OID_AUTO, "gp_per_low", CTLFLAG_RW,
 863             &rack_per_of_gp_lowthresh, 40,
 864             "What percentage of goodput do we allow the multiplier to fall to");
 865         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 866             SYSCTL_CHILDREN(rack_probertt),
 867             OID_AUTO, "time_between", CTLFLAG_RW,
 868             & rack_time_between_probertt, 96000000,
 869             "How many useconds between the lowest rtt falling must past before we enter probertt");
 870         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 871             SYSCTL_CHILDREN(rack_probertt),
 872             OID_AUTO, "safety", CTLFLAG_RW,
 873             &rack_probe_rtt_safety_val, 2000000,
 874             "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
 875         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 876             SYSCTL_CHILDREN(rack_probertt),
 877             OID_AUTO, "sets_cwnd", CTLFLAG_RW,
 878             &rack_probe_rtt_sets_cwnd, 0,
 879             "Do we set the cwnd too (if always_lower is on)");
 880         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 881             SYSCTL_CHILDREN(rack_probertt),
 882             OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
 883             &rack_max_drain_wait, 2,
 884             "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
 885         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 886             SYSCTL_CHILDREN(rack_probertt),
 887             OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
 888             &rack_must_drain, 1,
 889             "We must drain this many gp_srtt's waiting for flight to reach goal");
 890         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 891             SYSCTL_CHILDREN(rack_probertt),
 892             OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
 893             &rack_probertt_use_min_rtt_entry, 1,
 894             "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
 895         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 896             SYSCTL_CHILDREN(rack_probertt),
 897             OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
 898             &rack_probertt_use_min_rtt_exit, 0,
 899             "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
 900         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 901             SYSCTL_CHILDREN(rack_probertt),
 902             OID_AUTO, "length_div", CTLFLAG_RW,
 903             &rack_probertt_gpsrtt_cnt_div, 0,
 904             "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
 905         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 906             SYSCTL_CHILDREN(rack_probertt),
 907             OID_AUTO, "length_mul", CTLFLAG_RW,
 908             &rack_probertt_gpsrtt_cnt_mul, 0,
 909             "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
 910         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 911             SYSCTL_CHILDREN(rack_probertt),
 912             OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
 913             &rack_min_probertt_hold, 200000,
 914             "What is the minimum time we hold probertt at target");
 915         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 916             SYSCTL_CHILDREN(rack_probertt),
 917             OID_AUTO, "filter_life", CTLFLAG_RW,
 918             &rack_probertt_filter_life, 10000000,
 919             "What is the time for the filters life in useconds");
 920         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 921             SYSCTL_CHILDREN(rack_probertt),
 922             OID_AUTO, "lower_within", CTLFLAG_RW,
 923             &rack_probertt_lower_within, 10,
 924             "If the rtt goes lower within this percentage of the time, go into probe-rtt");
 925         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 926             SYSCTL_CHILDREN(rack_probertt),
 927             OID_AUTO, "must_move", CTLFLAG_RW,
 928             &rack_min_rtt_movement, 250,
 929             "How much is the minimum movement in rtt to count as a drop for probertt purposes");
 930         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 931             SYSCTL_CHILDREN(rack_probertt),
 932             OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
 933             &rack_probertt_clear_is, 1,
 934             "Do we clear I/S counts on exiting probe-rtt");
 935         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 936             SYSCTL_CHILDREN(rack_probertt),
 937             OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
 938             &rack_max_drain_hbp, 1,
 939             "How many extra drain gpsrtt's do we get in highly buffered paths");
 940         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 941             SYSCTL_CHILDREN(rack_probertt),
 942             OID_AUTO, "hbp_threshold", CTLFLAG_RW,
 943             &rack_hbp_thresh, 3,
 944             "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
 945
 946         rack_tracepoint = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 947             SYSCTL_CHILDREN(rack_sysctl_root),
 948             OID_AUTO,
 949             "tp",
 950             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 951             "Rack tracepoint facility");
 952         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 953             SYSCTL_CHILDREN(rack_tracepoint),
 954             OID_AUTO, "number", CTLFLAG_RW,
 955             &rack_trace_point_config, 0,
 956             "What is the trace point number to activate (0=none, 0xffffffff = all)?");
 957         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 958             SYSCTL_CHILDREN(rack_tracepoint),
 959             OID_AUTO, "bbmode", CTLFLAG_RW,
 960             &rack_trace_point_bb_mode, 4,
 961             "What is BB logging mode that is activated?");
 962         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 963             SYSCTL_CHILDREN(rack_tracepoint),
 964             OID_AUTO, "count", CTLFLAG_RW,
 965             &rack_trace_point_count, 0,
 966             "How many connections will have BB logging turned on that hit the tracepoint?");
 967         /* Pacing related sysctls */
 968         rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 969             SYSCTL_CHILDREN(rack_sysctl_root),
 970             OID_AUTO,
 971             "pacing",
 972             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 973             "Pacing related Controls");
 974         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 975             SYSCTL_CHILDREN(rack_pacing),
 976             OID_AUTO, "max_pace_over", CTLFLAG_RW,
 977             &rack_max_per_above, 30,
 978             "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
 979         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 980             SYSCTL_CHILDREN(rack_pacing),
 981             OID_AUTO, "pace_to_one", CTLFLAG_RW,
 982             &rack_pace_one_seg, 0,
 983             "Do we allow low b/w pacing of 1MSS instead of two");
 984         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 985             SYSCTL_CHILDREN(rack_pacing),
 986             OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
 987             &rack_limit_time_with_srtt, 0,
 988             "Do we limit pacing time based on srtt");
 989         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 990             SYSCTL_CHILDREN(rack_pacing),
 991             OID_AUTO, "init_win", CTLFLAG_RW,
 992             &rack_default_init_window, 0,
 993             "Do we have a rack initial window 0 = system default");
 994         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 995             SYSCTL_CHILDREN(rack_pacing),
 996             OID_AUTO, "gp_per_ss", CTLFLAG_RW,
 997             &rack_per_of_gp_ss, 250,
 998             "If non zero, what percentage of goodput to pace at in slow start");
 999         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1000             SYSCTL_CHILDREN(rack_pacing),
1001             OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1002             &rack_per_of_gp_ca, 150,
1003             "If non zero, what percentage of goodput to pace at in congestion avoidance");
1004         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1005             SYSCTL_CHILDREN(rack_pacing),
1006             OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1007             &rack_per_of_gp_rec, 200,
1008             "If non zero, what percentage of goodput to pace at in recovery");
1009         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1010             SYSCTL_CHILDREN(rack_pacing),
1011             OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1012             &rack_hptsi_segments, 40,
1013             "What size is the max for TSO segments in pacing and burst mitigation");
1014         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1015             SYSCTL_CHILDREN(rack_pacing),
1016             OID_AUTO, "burst_reduces", CTLFLAG_RW,
1017             &rack_slot_reduction, 4,
1018             "When doing only burst mitigation what is the reduce divisor");
1019         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1020             SYSCTL_CHILDREN(rack_sysctl_root),
1021             OID_AUTO, "use_pacing", CTLFLAG_RW,
1022             &rack_pace_every_seg, 0,
1023             "If set we use pacing, if clear we use only the original burst mitigation");
1024         SYSCTL_ADD_U64(&rack_sysctl_ctx,
1025             SYSCTL_CHILDREN(rack_pacing),
1026             OID_AUTO, "rate_cap", CTLFLAG_RW,
1027             &rack_bw_rate_cap, 0,
1028             "If set we apply this value to the absolute rate cap used by pacing");
1029         SYSCTL_ADD_U8(&rack_sysctl_ctx,
1030             SYSCTL_CHILDREN(rack_sysctl_root),
1031             OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1032             &rack_req_measurements, 1,
1033             "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1034         /* Hardware pacing */
1035         rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1036             SYSCTL_CHILDREN(rack_sysctl_root),
1037             OID_AUTO,
1038             "hdwr_pacing",
1039             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1040             "Pacing related Controls");
1041         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1042             SYSCTL_CHILDREN(rack_hw_pacing),
1043             OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1044             &rack_hw_rwnd_factor, 2,
1045             "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1046         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1047             SYSCTL_CHILDREN(rack_hw_pacing),
1048             OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1049             &rack_enobuf_hw_boost_mult, 2,
1050             "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1051         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1052             SYSCTL_CHILDREN(rack_hw_pacing),
1053             OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1054             &rack_enobuf_hw_max, 2,
1055             "What is the max boost the pacing time if we see a ENOBUFS?");
1056         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1057             SYSCTL_CHILDREN(rack_hw_pacing),
1058             OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1059             &rack_enobuf_hw_min, 2,
1060             "What is the min boost the pacing time if we see a ENOBUFS?");
1061         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1062             SYSCTL_CHILDREN(rack_hw_pacing),
1063             OID_AUTO, "enable", CTLFLAG_RW,
1064             &rack_enable_hw_pacing, 0,
1065             "Should RACK attempt to use hw pacing?");
1066         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1067             SYSCTL_CHILDREN(rack_hw_pacing),
1068             OID_AUTO, "rate_cap", CTLFLAG_RW,
1069             &rack_hw_rate_caps, 1,
1070             "Does the highest hardware pacing rate cap the rate we will send at??");
1071         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1072             SYSCTL_CHILDREN(rack_hw_pacing),
1073             OID_AUTO, "rate_min", CTLFLAG_RW,
1074             &rack_hw_rate_min, 0,
1075             "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1076         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1077             SYSCTL_CHILDREN(rack_hw_pacing),
1078             OID_AUTO, "rate_to_low", CTLFLAG_RW,
1079             &rack_hw_rate_to_low, 0,
1080             "If we fall below this rate, dis-engage hw pacing?");
1081         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1082             SYSCTL_CHILDREN(rack_hw_pacing),
1083             OID_AUTO, "up_only", CTLFLAG_RW,
1084             &rack_hw_up_only, 1,
1085             "Do we allow hw pacing to lower the rate selected?");
1086         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1087             SYSCTL_CHILDREN(rack_hw_pacing),
1088             OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1089             &rack_hw_pace_extra_slots, 2,
1090             "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1091         rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1092             SYSCTL_CHILDREN(rack_sysctl_root),
1093             OID_AUTO,
1094             "timely",
1095             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1096             "Rack Timely RTT Controls");
1097         /* Timely based GP dynmics */
1098         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1099             SYSCTL_CHILDREN(rack_timely),
1100             OID_AUTO, "upper", CTLFLAG_RW,
1101             &rack_gp_per_bw_mul_up, 2,
1102             "Rack timely upper range for equal b/w (in percentage)");
1103         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1104             SYSCTL_CHILDREN(rack_timely),
1105             OID_AUTO, "lower", CTLFLAG_RW,
1106             &rack_gp_per_bw_mul_down, 4,
1107             "Rack timely lower range for equal b/w (in percentage)");
1108         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1109             SYSCTL_CHILDREN(rack_timely),
1110             OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1111             &rack_gp_rtt_maxmul, 3,
1112             "Rack timely multiplier of lowest rtt for rtt_max");
1113         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1114             SYSCTL_CHILDREN(rack_timely),
1115             OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1116             &rack_gp_rtt_mindiv, 4,
1117             "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1118         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1119             SYSCTL_CHILDREN(rack_timely),
1120             OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1121             &rack_gp_rtt_minmul, 1,
1122             "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1123         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1124             SYSCTL_CHILDREN(rack_timely),
1125             OID_AUTO, "decrease", CTLFLAG_RW,
1126             &rack_gp_decrease_per, 20,
1127             "Rack timely decrease percentage of our GP multiplication factor");
1128         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1129             SYSCTL_CHILDREN(rack_timely),
1130             OID_AUTO, "increase", CTLFLAG_RW,
1131             &rack_gp_increase_per, 2,
1132             "Rack timely increase perentage of our GP multiplication factor");
1133         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1134             SYSCTL_CHILDREN(rack_timely),
1135             OID_AUTO, "lowerbound", CTLFLAG_RW,
1136             &rack_per_lower_bound, 50,
1137             "Rack timely lowest percentage we allow GP multiplier to fall to");
1138         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1139             SYSCTL_CHILDREN(rack_timely),
1140             OID_AUTO, "upperboundss", CTLFLAG_RW,
1141             &rack_per_upper_bound_ss, 0,
1142             "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1143         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1144             SYSCTL_CHILDREN(rack_timely),
1145             OID_AUTO, "upperboundca", CTLFLAG_RW,
1146             &rack_per_upper_bound_ca, 0,
1147             "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1148         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1149             SYSCTL_CHILDREN(rack_timely),
1150             OID_AUTO, "dynamicgp", CTLFLAG_RW,
1151             &rack_do_dyn_mul, 0,
1152             "Rack timely do we enable dynmaic timely goodput by default");
1153         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1154             SYSCTL_CHILDREN(rack_timely),
1155             OID_AUTO, "no_rec_red", CTLFLAG_RW,
1156             &rack_gp_no_rec_chg, 1,
1157             "Rack timely do we prohibit the recovery multiplier from being lowered");
1158         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1159             SYSCTL_CHILDREN(rack_timely),
1160             OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1161             &rack_timely_dec_clear, 6,
1162             "Rack timely what threshold do we count to before another boost during b/w decent");
1163         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1164             SYSCTL_CHILDREN(rack_timely),
1165             OID_AUTO, "max_push_rise", CTLFLAG_RW,
1166             &rack_timely_max_push_rise, 3,
1167             "Rack timely how many times do we push up with b/w increase");
1168         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1169             SYSCTL_CHILDREN(rack_timely),
1170             OID_AUTO, "max_push_drop", CTLFLAG_RW,
1171             &rack_timely_max_push_drop, 3,
1172             "Rack timely how many times do we push back on b/w decent");
1173         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1174             SYSCTL_CHILDREN(rack_timely),
1175             OID_AUTO, "min_segs", CTLFLAG_RW,
1176             &rack_timely_min_segs, 4,
1177             "Rack timely when setting the cwnd what is the min num segments");
1178         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1179             SYSCTL_CHILDREN(rack_timely),
1180             OID_AUTO, "noback_max", CTLFLAG_RW,
1181             &rack_use_max_for_nobackoff, 0,
1182             "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1183         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1184             SYSCTL_CHILDREN(rack_timely),
1185             OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1186             &rack_timely_int_timely_only, 0,
1187             "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1188         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1189             SYSCTL_CHILDREN(rack_timely),
1190             OID_AUTO, "nonstop", CTLFLAG_RW,
1191             &rack_timely_no_stopping, 0,
1192             "Rack timely don't stop increase");
1193         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1194             SYSCTL_CHILDREN(rack_timely),
1195             OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1196             &rack_down_raise_thresh, 100,
1197             "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1198         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1199             SYSCTL_CHILDREN(rack_timely),
1200             OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1201             &rack_req_segs, 1,
1202             "Bottom dragging if not these many segments outstanding and room");
1203
1204         /* TLP and Rack related parameters */
1205         rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1206             SYSCTL_CHILDREN(rack_sysctl_root),
1207             OID_AUTO,
1208             "tlp",
1209             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1210             "TLP and Rack related Controls");
1211         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1212             SYSCTL_CHILDREN(rack_tlp),
1213             OID_AUTO, "use_rrr", CTLFLAG_RW,
1214             &use_rack_rr, 1,
1215             "Do we use Rack Rapid Recovery");
1216         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1217             SYSCTL_CHILDREN(rack_tlp),
1218             OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1219             &rack_max_abc_post_recovery, 2,
1220             "Since we do early recovery, do we override the l_abc to a value, if so what?");
1221         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1222             SYSCTL_CHILDREN(rack_tlp),
1223             OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1224             &rack_non_rxt_use_cr, 0,
1225             "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1226         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1227             SYSCTL_CHILDREN(rack_tlp),
1228             OID_AUTO, "tlpmethod", CTLFLAG_RW,
1229             &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1230             "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1231         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1232             SYSCTL_CHILDREN(rack_tlp),
1233             OID_AUTO, "limit", CTLFLAG_RW,
1234             &rack_tlp_limit, 2,
1235             "How many TLP's can be sent without sending new data");
1236         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1237             SYSCTL_CHILDREN(rack_tlp),
1238             OID_AUTO, "use_greater", CTLFLAG_RW,
1239             &rack_tlp_use_greater, 1,
1240             "Should we use the rack_rtt time if its greater than srtt");
1241         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1242             SYSCTL_CHILDREN(rack_tlp),
1243             OID_AUTO, "tlpminto", CTLFLAG_RW,
1244             &rack_tlp_min, 10000,
1245             "TLP minimum timeout per the specification (in microseconds)");
1246         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1247             SYSCTL_CHILDREN(rack_tlp),
1248             OID_AUTO, "send_oldest", CTLFLAG_RW,
1249             &rack_always_send_oldest, 0,
1250             "Should we always send the oldest TLP and RACK-TLP");
1251         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1252             SYSCTL_CHILDREN(rack_tlp),
1253             OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1254             &rack_limited_retran, 0,
1255             "How many times can a rack timeout drive out sends");
1256         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1257             SYSCTL_CHILDREN(rack_tlp),
1258             OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1259             &rack_lower_cwnd_at_tlp, 0,
1260             "When a TLP completes a retran should we enter recovery");
1261         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1262             SYSCTL_CHILDREN(rack_tlp),
1263             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1264             &rack_reorder_thresh, 2,
1265             "What factor for rack will be added when seeing reordering (shift right)");
1266         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1267             SYSCTL_CHILDREN(rack_tlp),
1268             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1269             &rack_tlp_thresh, 1,
1270             "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1271         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1272             SYSCTL_CHILDREN(rack_tlp),
1273             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1274             &rack_reorder_fade, 60000000,
1275             "Does reorder detection fade, if so how many microseconds (0 means never)");
1276         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1277             SYSCTL_CHILDREN(rack_tlp),
1278             OID_AUTO, "pktdelay", CTLFLAG_RW,
1279             &rack_pkt_delay, 1000,
1280             "Extra RACK time (in microseconds) besides reordering thresh");
1281
1282         /* Timer related controls */
1283         rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1284             SYSCTL_CHILDREN(rack_sysctl_root),
1285             OID_AUTO,
1286             "timers",
1287             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1288             "Timer related controls");
1289         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1290             SYSCTL_CHILDREN(rack_timers),
1291             OID_AUTO, "persmin", CTLFLAG_RW,
1292             &rack_persist_min, 250000,
1293             "What is the minimum time in microseconds between persists");
1294         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1295             SYSCTL_CHILDREN(rack_timers),
1296             OID_AUTO, "persmax", CTLFLAG_RW,
1297             &rack_persist_max, 2000000,
1298             "What is the largest delay in microseconds between persists");
1299         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1300             SYSCTL_CHILDREN(rack_timers),
1301             OID_AUTO, "delayed_ack", CTLFLAG_RW,
1302             &rack_delayed_ack_time, 40000,
1303             "Delayed ack time (40ms in microseconds)");
1304         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1305             SYSCTL_CHILDREN(rack_timers),
1306             OID_AUTO, "minrto", CTLFLAG_RW,
1307             &rack_rto_min, 30000,
1308             "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1309         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1310             SYSCTL_CHILDREN(rack_timers),
1311             OID_AUTO, "maxrto", CTLFLAG_RW,
1312             &rack_rto_max, 4000000,
1313             "Maximum RTO in microseconds -- should be at least as large as min_rto");
1314         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1315             SYSCTL_CHILDREN(rack_timers),
1316             OID_AUTO, "minto", CTLFLAG_RW,
1317             &rack_min_to, 1000,
1318             "Minimum rack timeout in microseconds");
1319         /* Measure controls */
1320         rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1321             SYSCTL_CHILDREN(rack_sysctl_root),
1322             OID_AUTO,
1323             "measure",
1324             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1325             "Measure related controls");
1326         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1327             SYSCTL_CHILDREN(rack_measure),
1328             OID_AUTO, "wma_divisor", CTLFLAG_RW,
1329             &rack_wma_divisor, 8,
1330             "When doing b/w calculation what is the  divisor for the WMA");
1331         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1332             SYSCTL_CHILDREN(rack_measure),
1333             OID_AUTO, "end_cwnd", CTLFLAG_RW,
1334             &rack_cwnd_block_ends_measure, 0,
1335             "Does a cwnd just-return end the measurement window (app limited)");
1336         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1337             SYSCTL_CHILDREN(rack_measure),
1338             OID_AUTO, "end_rwnd", CTLFLAG_RW,
1339             &rack_rwnd_block_ends_measure, 0,
1340             "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1341         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1342             SYSCTL_CHILDREN(rack_measure),
1343             OID_AUTO, "min_target", CTLFLAG_RW,
1344             &rack_def_data_window, 20,
1345             "What is the minimum target window (in mss) for a GP measurements");
1346         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1347             SYSCTL_CHILDREN(rack_measure),
1348             OID_AUTO, "goal_bdp", CTLFLAG_RW,
1349             &rack_goal_bdp, 2,
1350             "What is the goal BDP to measure");
1351         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1352             SYSCTL_CHILDREN(rack_measure),
1353             OID_AUTO, "min_srtts", CTLFLAG_RW,
1354             &rack_min_srtts, 1,
1355             "What is the goal BDP to measure");
1356         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1357             SYSCTL_CHILDREN(rack_measure),
1358             OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1359             &rack_min_measure_usec, 0,
1360             "What is the Minimum time time for a measurement if 0, this is off");
1361         /* Features */
1362         rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1363             SYSCTL_CHILDREN(rack_sysctl_root),
1364             OID_AUTO,
1365             "features",
1366             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1367             "Feature controls");
1368         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1369             SYSCTL_CHILDREN(rack_features),
1370             OID_AUTO, "cmpack", CTLFLAG_RW,
1371             &rack_use_cmp_acks, 1,
1372             "Should RACK have LRO send compressed acks");
1373         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1374             SYSCTL_CHILDREN(rack_features),
1375             OID_AUTO, "fsb", CTLFLAG_RW,
1376             &rack_use_fsb, 1,
1377             "Should RACK use the fast send block?");
1378         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1379             SYSCTL_CHILDREN(rack_features),
1380             OID_AUTO, "rfo", CTLFLAG_RW,
1381             &rack_use_rfo, 1,
1382             "Should RACK use rack_fast_output()?");
1383         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1384             SYSCTL_CHILDREN(rack_features),
1385             OID_AUTO, "rsmrfo", CTLFLAG_RW,
1386             &rack_use_rsm_rfo, 1,
1387             "Should RACK use rack_fast_rsm_output()?");
1388         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1389             SYSCTL_CHILDREN(rack_features),
1390             OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1391             &rack_enable_mqueue_for_nonpaced, 0,
1392             "Should RACK use mbuf queuing for non-paced connections");
1393         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1394             SYSCTL_CHILDREN(rack_features),
1395             OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1396             &rack_do_hystart, 0,
1397             "Should RACK enable HyStart++ on connections?");
1398         /* Misc rack controls */
1399         rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1400             SYSCTL_CHILDREN(rack_sysctl_root),
1401             OID_AUTO,
1402             "misc",
1403             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1404             "Misc related controls");
1405 #ifdef TCP_ACCOUNTING
1406         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1407             SYSCTL_CHILDREN(rack_misc),
1408             OID_AUTO, "tcp_acct", CTLFLAG_RW,
1409             &rack_tcp_accounting, 0,
1410             "Should we turn on TCP accounting for all rack sessions?");
1411 #endif
1412         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1413             SYSCTL_CHILDREN(rack_misc),
1414             OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1415             &rack_apply_rtt_with_reduced_conf, 0,
1416             "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1417         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1418             SYSCTL_CHILDREN(rack_misc),
1419             OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1420             &rack_dsack_std_based, 3,
1421             "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1422         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1423             SYSCTL_CHILDREN(rack_misc),
1424             OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1425             &rack_prr_addbackmax, 2,
1426             "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1427         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1428             SYSCTL_CHILDREN(rack_misc),
1429             OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1430             &rack_stats_gets_ms_rtt, 1,
1431             "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1432         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1433             SYSCTL_CHILDREN(rack_misc),
1434             OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1435             &rack_client_low_buf, 0,
1436             "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1437         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1438             SYSCTL_CHILDREN(rack_misc),
1439             OID_AUTO, "defprofile", CTLFLAG_RW,
1440             &rack_def_profile, 0,
1441             "Should RACK use a default profile (0=no, num == profile num)?");
1442         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1443             SYSCTL_CHILDREN(rack_misc),
1444             OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1445             &rack_enable_shared_cwnd, 1,
1446             "Should RACK try to use the shared cwnd on connections where allowed");
1447         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1448             SYSCTL_CHILDREN(rack_misc),
1449             OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1450             &rack_limits_scwnd, 1,
1451             "Should RACK place low end time limits on the shared cwnd feature");
1452         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1453             SYSCTL_CHILDREN(rack_misc),
1454             OID_AUTO, "iMac_dack", CTLFLAG_RW,
1455             &rack_use_imac_dack, 0,
1456             "Should RACK try to emulate iMac delayed ack");
1457         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1458             SYSCTL_CHILDREN(rack_misc),
1459             OID_AUTO, "no_prr", CTLFLAG_RW,
1460             &rack_disable_prr, 0,
1461             "Should RACK not use prr and only pace (must have pacing on)");
1462         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1463             SYSCTL_CHILDREN(rack_misc),
1464             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1465             &rack_verbose_logging, 0,
1466             "Should RACK black box logging be verbose");
1467         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1468             SYSCTL_CHILDREN(rack_misc),
1469             OID_AUTO, "data_after_close", CTLFLAG_RW,
1470             &rack_ignore_data_after_close, 1,
1471             "Do we hold off sending a RST until all pending data is ack'd");
1472         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1473             SYSCTL_CHILDREN(rack_misc),
1474             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1475             &rack_sack_not_required, 1,
1476             "Do we allow rack to run on connections not supporting SACK");
1477         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1478             SYSCTL_CHILDREN(rack_misc),
1479             OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1480             &rack_send_a_lot_in_prr, 1,
1481             "Send a lot in prr");
1482         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1483             SYSCTL_CHILDREN(rack_misc),
1484             OID_AUTO, "autoscale", CTLFLAG_RW,
1485             &rack_autosndbuf_inc, 20,
1486             "What percentage should rack scale up its snd buffer by?");
1487         /* Sack Attacker detection stuff */
1488         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1489             SYSCTL_CHILDREN(rack_attack),
1490             OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1491             &rack_highest_sack_thresh_seen, 0,
1492             "Highest sack to ack ratio seen");
1493         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1494             SYSCTL_CHILDREN(rack_attack),
1495             OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1496             &rack_highest_move_thresh_seen, 0,
1497             "Highest move to non-move ratio seen");
1498         rack_ack_total = counter_u64_alloc(M_WAITOK);
1499         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1500             SYSCTL_CHILDREN(rack_attack),
1501             OID_AUTO, "acktotal", CTLFLAG_RD,
1502             &rack_ack_total,
1503             "Total number of Ack's");
1504         rack_express_sack = counter_u64_alloc(M_WAITOK);
1505         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1506             SYSCTL_CHILDREN(rack_attack),
1507             OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1508             &rack_express_sack,
1509             "Total expresss number of Sack's");
1510         rack_sack_total = counter_u64_alloc(M_WAITOK);
1511         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1512             SYSCTL_CHILDREN(rack_attack),
1513             OID_AUTO, "sacktotal", CTLFLAG_RD,
1514             &rack_sack_total,
1515             "Total number of SACKs");
1516         rack_move_none = counter_u64_alloc(M_WAITOK);
1517         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1518             SYSCTL_CHILDREN(rack_attack),
1519             OID_AUTO, "move_none", CTLFLAG_RD,
1520             &rack_move_none,
1521             "Total number of SACK index reuse of positions under threshold");
1522         rack_move_some = counter_u64_alloc(M_WAITOK);
1523         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1524             SYSCTL_CHILDREN(rack_attack),
1525             OID_AUTO, "move_some", CTLFLAG_RD,
1526             &rack_move_some,
1527             "Total number of SACK index reuse of positions over threshold");
1528         rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1529         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1530             SYSCTL_CHILDREN(rack_attack),
1531             OID_AUTO, "attacks", CTLFLAG_RD,
1532             &rack_sack_attacks_detected,
1533             "Total number of SACK attackers that had sack disabled");
1534         rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1535         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1536             SYSCTL_CHILDREN(rack_attack),
1537             OID_AUTO, "reversed", CTLFLAG_RD,
1538             &rack_sack_attacks_reversed,
1539             "Total number of SACK attackers that were later determined false positive");
1540         rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1541         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1542             SYSCTL_CHILDREN(rack_attack),
1543             OID_AUTO, "nextmerge", CTLFLAG_RD,
1544             &rack_sack_used_next_merge,
1545             "Total number of times we used the next merge");
1546         rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1547         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1548             SYSCTL_CHILDREN(rack_attack),
1549             OID_AUTO, "prevmerge", CTLFLAG_RD,
1550             &rack_sack_used_prev_merge,
1551             "Total number of times we used the prev merge");
1552         /* Counters */
1553         rack_fto_send = counter_u64_alloc(M_WAITOK);
1554         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1555             SYSCTL_CHILDREN(rack_counters),
1556             OID_AUTO, "fto_send", CTLFLAG_RD,
1557             &rack_fto_send, "Total number of rack_fast_output sends");
1558         rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1559         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1560             SYSCTL_CHILDREN(rack_counters),
1561             OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1562             &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1563         rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1564         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1565             SYSCTL_CHILDREN(rack_counters),
1566             OID_AUTO, "nfto_resend", CTLFLAG_RD,
1567             &rack_nfto_resend, "Total number of rack_output retransmissions");
1568         rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1569         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1570             SYSCTL_CHILDREN(rack_counters),
1571             OID_AUTO, "nfto_send", CTLFLAG_RD,
1572             &rack_non_fto_send, "Total number of rack_output first sends");
1573         rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1574         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1575             SYSCTL_CHILDREN(rack_counters),
1576             OID_AUTO, "rfo_extended", CTLFLAG_RD,
1577             &rack_extended_rfo, "Total number of times we extended rfo");
1578
1579         rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1580         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1581             SYSCTL_CHILDREN(rack_counters),
1582             OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1583             &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1584         rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1585
1586         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1587             SYSCTL_CHILDREN(rack_counters),
1588             OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1589             &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1590         rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1591         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1592             SYSCTL_CHILDREN(rack_counters),
1593             OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1594             &rack_tlp_tot,
1595             "Total number of tail loss probe expirations");
1596         rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1597         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1598             SYSCTL_CHILDREN(rack_counters),
1599             OID_AUTO, "tlp_new", CTLFLAG_RD,
1600             &rack_tlp_newdata,
1601             "Total number of tail loss probe sending new data");
1602         rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1603         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1604             SYSCTL_CHILDREN(rack_counters),
1605             OID_AUTO, "tlp_retran", CTLFLAG_RD,
1606             &rack_tlp_retran,
1607             "Total number of tail loss probe sending retransmitted data");
1608         rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1609         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1610             SYSCTL_CHILDREN(rack_counters),
1611             OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1612             &rack_tlp_retran_bytes,
1613             "Total bytes of tail loss probe sending retransmitted data");
1614         rack_to_tot = counter_u64_alloc(M_WAITOK);
1615         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1616             SYSCTL_CHILDREN(rack_counters),
1617             OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1618             &rack_to_tot,
1619             "Total number of times the rack to expired");
1620         rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1621         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1622             SYSCTL_CHILDREN(rack_counters),
1623             OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1624             &rack_saw_enobuf,
1625             "Total number of times a sends returned enobuf for non-hdwr paced connections");
1626         rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1627         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1628             SYSCTL_CHILDREN(rack_counters),
1629             OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1630             &rack_saw_enobuf_hw,
1631             "Total number of times a send returned enobuf for hdwr paced connections");
1632         rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1633         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1634             SYSCTL_CHILDREN(rack_counters),
1635             OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1636             &rack_saw_enetunreach,
1637             "Total number of times a send received a enetunreachable");
1638         rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1639         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1640             SYSCTL_CHILDREN(rack_counters),
1641             OID_AUTO, "alloc_hot", CTLFLAG_RD,
1642             &rack_hot_alloc,
1643             "Total allocations from the top of our list");
1644         rack_to_alloc = counter_u64_alloc(M_WAITOK);
1645         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1646             SYSCTL_CHILDREN(rack_counters),
1647             OID_AUTO, "allocs", CTLFLAG_RD,
1648             &rack_to_alloc,
1649             "Total allocations of tracking structures");
1650         rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1651         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1652             SYSCTL_CHILDREN(rack_counters),
1653             OID_AUTO, "allochard", CTLFLAG_RD,
1654             &rack_to_alloc_hard,
1655             "Total allocations done with sleeping the hard way");
1656         rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1657         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1658             SYSCTL_CHILDREN(rack_counters),
1659             OID_AUTO, "allocemerg", CTLFLAG_RD,
1660             &rack_to_alloc_emerg,
1661             "Total allocations done from emergency cache");
1662         rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1663         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1664             SYSCTL_CHILDREN(rack_counters),
1665             OID_AUTO, "alloc_limited", CTLFLAG_RD,
1666             &rack_to_alloc_limited,
1667             "Total allocations dropped due to limit");
1668         rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1669         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1670             SYSCTL_CHILDREN(rack_counters),
1671             OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1672             &rack_alloc_limited_conns,
1673             "Connections with allocations dropped due to limit");
1674         rack_split_limited = counter_u64_alloc(M_WAITOK);
1675         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1676             SYSCTL_CHILDREN(rack_counters),
1677             OID_AUTO, "split_limited", CTLFLAG_RD,
1678             &rack_split_limited,
1679             "Split allocations dropped due to limit");
1680         rack_persists_sends = counter_u64_alloc(M_WAITOK);
1681         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1682             SYSCTL_CHILDREN(rack_counters),
1683             OID_AUTO, "persist_sends", CTLFLAG_RD,
1684             &rack_persists_sends,
1685             "Number of times we sent a persist probe");
1686         rack_persists_acks = counter_u64_alloc(M_WAITOK);
1687         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1688             SYSCTL_CHILDREN(rack_counters),
1689             OID_AUTO, "persist_acks", CTLFLAG_RD,
1690             &rack_persists_acks,
1691             "Number of times a persist probe was acked");
1692         rack_persists_loss = counter_u64_alloc(M_WAITOK);
1693         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1694             SYSCTL_CHILDREN(rack_counters),
1695             OID_AUTO, "persist_loss", CTLFLAG_RD,
1696             &rack_persists_loss,
1697             "Number of times we detected a lost persist probe (no ack)");
1698         rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1699         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1700             SYSCTL_CHILDREN(rack_counters),
1701             OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1702             &rack_persists_lost_ends,
1703             "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1704 #ifdef INVARIANTS
1705         rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1706         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1707             SYSCTL_CHILDREN(rack_counters),
1708             OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1709             &rack_adjust_map_bw,
1710             "Number of times we hit the case where the sb went up and down on a sendmap entry");
1711 #endif
1712         rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1713         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1714             SYSCTL_CHILDREN(rack_counters),
1715             OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1716             &rack_multi_single_eq,
1717             "Number of compressed acks total represented");
1718         rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1719         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1720             SYSCTL_CHILDREN(rack_counters),
1721             OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1722             &rack_proc_non_comp_ack,
1723             "Number of non compresseds acks that we processed");
1724
1725
1726         rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1727         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1728             SYSCTL_CHILDREN(rack_counters),
1729             OID_AUTO, "sack_long", CTLFLAG_RD,
1730             &rack_sack_proc_all,
1731             "Total times we had to walk whole list for sack processing");
1732         rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1733         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1734             SYSCTL_CHILDREN(rack_counters),
1735             OID_AUTO, "sack_restart", CTLFLAG_RD,
1736             &rack_sack_proc_restart,
1737             "Total times we had to walk whole list due to a restart");
1738         rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1739         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1740             SYSCTL_CHILDREN(rack_counters),
1741             OID_AUTO, "sack_short", CTLFLAG_RD,
1742             &rack_sack_proc_short,
1743             "Total times we took shortcut for sack processing");
1744         rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1745         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1746             SYSCTL_CHILDREN(rack_attack),
1747             OID_AUTO, "skipacked", CTLFLAG_RD,
1748             &rack_sack_skipped_acked,
1749             "Total number of times we skipped previously sacked");
1750         rack_sack_splits = counter_u64_alloc(M_WAITOK);
1751         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1752             SYSCTL_CHILDREN(rack_attack),
1753             OID_AUTO, "ofsplit", CTLFLAG_RD,
1754             &rack_sack_splits,
1755             "Total number of times we did the old fashion tree split");
1756         rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1757         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1758             SYSCTL_CHILDREN(rack_counters),
1759             OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1760             &rack_input_idle_reduces,
1761             "Total number of idle reductions on input");
1762         rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1763         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1764             SYSCTL_CHILDREN(rack_counters),
1765             OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1766             &rack_collapsed_win_seen,
1767             "Total number of collapsed window events seen (where our window shrinks)");
1768
1769         rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1770         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1771             SYSCTL_CHILDREN(rack_counters),
1772             OID_AUTO, "collapsed_win", CTLFLAG_RD,
1773             &rack_collapsed_win,
1774             "Total number of collapsed window events where we mark packets");
1775         rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1776         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1777             SYSCTL_CHILDREN(rack_counters),
1778             OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1779             &rack_collapsed_win_rxt,
1780             "Total number of packets that were retransmitted");
1781         rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1782         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1783             SYSCTL_CHILDREN(rack_counters),
1784             OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1785             &rack_collapsed_win_rxt_bytes,
1786             "Total number of bytes that were retransmitted");
1787         rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1788         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1789             SYSCTL_CHILDREN(rack_counters),
1790             OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1791             &rack_try_scwnd,
1792             "Total number of scwnd attempts");
1793         COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1794         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1795             OID_AUTO, "outsize", CTLFLAG_RD,
1796             rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1797         COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1798         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1799             OID_AUTO, "opts", CTLFLAG_RD,
1800             rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1801         SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1802             SYSCTL_CHILDREN(rack_sysctl_root),
1803             OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1804             &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1805 }
1806
1807 static __inline int
1808 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1809 {
1810         if (SEQ_GEQ(b->r_start, a->r_start) &&
1811             SEQ_LT(b->r_start, a->r_end)) {
1812                 /*
1813                  * The entry b is within the
1814                  * block a. i.e.:
1815                  * a --   |-------------|
1816                  * b --   |----|
1817                  * <or>
1818                  * b --       |------|
1819                  * <or>
1820                  * b --       |-----------|
1821                  */
1822                 return (0);
1823         } else if (SEQ_GEQ(b->r_start, a->r_end)) {
1824                 /*
1825                  * b falls as either the next
1826                  * sequence block after a so a
1827                  * is said to be smaller than b.
1828                  * i.e:
1829                  * a --   |------|
1830                  * b --          |--------|
1831                  * or
1832                  * b --              |-----|
1833                  */
1834                 return (1);
1835         }
1836         /*
1837          * Whats left is where a is
1838          * larger than b. i.e:
1839          * a --         |-------|
1840          * b --  |---|
1841          * or even possibly
1842          * b --   |--------------|
1843          */
1844         return (-1);
1845 }
1846
1847 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1848 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1849
1850 static uint32_t
1851 rc_init_window(struct tcp_rack *rack)
1852 {
1853         uint32_t win;
1854
1855         if (rack->rc_init_win == 0) {
1856                 /*
1857                  * Nothing set by the user, use the system stack
1858                  * default.
1859                  */
1860                 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1861         }
1862         win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1863         return (win);
1864 }
1865
1866 static uint64_t
1867 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1868 {
1869         if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1870                 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1871         else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1872                 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1873         else
1874                 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1875 }
1876
1877 static uint64_t
1878 rack_get_bw(struct tcp_rack *rack)
1879 {
1880         if (rack->use_fixed_rate) {
1881                 /* Return the fixed pacing rate */
1882                 return (rack_get_fixed_pacing_bw(rack));
1883         }
1884         if (rack->r_ctl.gp_bw == 0) {
1885                 /*
1886                  * We have yet no b/w measurement,
1887                  * if we have a user set initial bw
1888                  * return it. If we don't have that and
1889                  * we have an srtt, use the tcp IW (10) to
1890                  * calculate a fictional b/w over the SRTT
1891                  * which is more or less a guess. Note
1892                  * we don't use our IW from rack on purpose
1893                  * so if we have like IW=30, we are not
1894                  * calculating a "huge" b/w.
1895                  */
1896                 uint64_t bw, srtt;
1897                 if (rack->r_ctl.init_rate)
1898                         return (rack->r_ctl.init_rate);
1899
1900                 /* Has the user set a max peak rate? */
1901 #ifdef NETFLIX_PEAKRATE
1902                 if (rack->rc_tp->t_maxpeakrate)
1903                         return (rack->rc_tp->t_maxpeakrate);
1904 #endif
1905                 /* Ok lets come up with the IW guess, if we have a srtt */
1906                 if (rack->rc_tp->t_srtt == 0) {
1907                         /*
1908                          * Go with old pacing method
1909                          * i.e. burst mitigation only.
1910                          */
1911                         return (0);
1912                 }
1913                 /* Ok lets get the initial TCP win (not racks) */
1914                 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1915                 srtt = (uint64_t)rack->rc_tp->t_srtt;
1916                 bw *= (uint64_t)USECS_IN_SECOND;
1917                 bw /= srtt;
1918                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1919                         bw = rack->r_ctl.bw_rate_cap;
1920                 return (bw);
1921         } else {
1922                 uint64_t bw;
1923
1924                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
1925                         /* Averaging is done, we can return the value */
1926                         bw = rack->r_ctl.gp_bw;
1927                 } else {
1928                         /* Still doing initial average must calculate */
1929                         bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
1930                 }
1931 #ifdef NETFLIX_PEAKRATE
1932                 if ((rack->rc_tp->t_maxpeakrate) &&
1933                     (bw > rack->rc_tp->t_maxpeakrate)) {
1934                         /* The user has set a peak rate to pace at
1935                          * don't allow us to pace faster than that.
1936                          */
1937                         return (rack->rc_tp->t_maxpeakrate);
1938                 }
1939 #endif
1940                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1941                         bw = rack->r_ctl.bw_rate_cap;
1942                 return (bw);
1943         }
1944 }
1945
1946 static uint16_t
1947 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1948 {
1949         if (rack->use_fixed_rate) {
1950                 return (100);
1951         } else if (rack->in_probe_rtt && (rsm == NULL))
1952                 return (rack->r_ctl.rack_per_of_gp_probertt);
1953         else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
1954                   rack->r_ctl.rack_per_of_gp_rec)) {
1955                 if (rsm) {
1956                         /* a retransmission always use the recovery rate */
1957                         return (rack->r_ctl.rack_per_of_gp_rec);
1958                 } else if (rack->rack_rec_nonrxt_use_cr) {
1959                         /* Directed to use the configured rate */
1960                         goto configured_rate;
1961                 } else if (rack->rack_no_prr &&
1962                            (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1963                         /* No PRR, lets just use the b/w estimate only */
1964                         return (100);
1965                 } else {
1966                         /*
1967                          * Here we may have a non-retransmit but we
1968                          * have no overrides, so just use the recovery
1969                          * rate (prr is in effect).
1970                          */
1971                         return (rack->r_ctl.rack_per_of_gp_rec);
1972                 }
1973         }
1974 configured_rate:
1975         /* For the configured rate we look at our cwnd vs the ssthresh */
1976         if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1977                 return (rack->r_ctl.rack_per_of_gp_ss);
1978         else
1979                 return (rack->r_ctl.rack_per_of_gp_ca);
1980 }
1981
1982 static void
1983 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
1984 {
1985         /*
1986          * Types of logs (mod value)
1987          * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
1988          * 2 = a dsack round begins, persist is reset to 16.
1989          * 3 = a dsack round ends
1990          * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
1991          * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
1992          * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
1993          */
1994         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1995                 union tcp_log_stackspecific log;
1996                 struct timeval tv;
1997
1998                 memset(&log, 0, sizeof(log));
1999                 log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2000                 log.u_bbr.flex1 <<= 1;
2001                 log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2002                 log.u_bbr.flex1 <<= 1;
2003                 log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2004                 log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2005                 log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2006                 log.u_bbr.flex4 = flex4;
2007                 log.u_bbr.flex5 = flex5;
2008                 log.u_bbr.flex6 = flex6;
2009                 log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2010                 log.u_bbr.flex8 = mod;
2011                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2012                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2013                     &rack->rc_inp->inp_socket->so_rcv,
2014                     &rack->rc_inp->inp_socket->so_snd,
2015                     RACK_DSACK_HANDLING, 0,
2016                     0, &log, false, &tv);
2017         }
2018 }
2019
2020 static void
2021 rack_log_hdwr_pacing(struct tcp_rack *rack,
2022                      uint64_t rate, uint64_t hw_rate, int line,
2023                      int error, uint16_t mod)
2024 {
2025         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2026                 union tcp_log_stackspecific log;
2027                 struct timeval tv;
2028                 const struct ifnet *ifp;
2029
2030                 memset(&log, 0, sizeof(log));
2031                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2032                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2033                 if (rack->r_ctl.crte) {
2034                         ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2035                 } else if (rack->rc_inp->inp_route.ro_nh &&
2036                            rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2037                         ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2038                 } else
2039                         ifp = NULL;
2040                 if (ifp) {
2041                         log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2042                         log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2043                 }
2044                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2045                 log.u_bbr.bw_inuse = rate;
2046                 log.u_bbr.flex5 = line;
2047                 log.u_bbr.flex6 = error;
2048                 log.u_bbr.flex7 = mod;
2049                 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2050                 log.u_bbr.flex8 = rack->use_fixed_rate;
2051                 log.u_bbr.flex8 <<= 1;
2052                 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2053                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2054                 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2055                 if (rack->r_ctl.crte)
2056                         log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2057                 else
2058                         log.u_bbr.cur_del_rate = 0;
2059                 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2060                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2061                     &rack->rc_inp->inp_socket->so_rcv,
2062                     &rack->rc_inp->inp_socket->so_snd,
2063                     BBR_LOG_HDWR_PACE, 0,
2064                     0, &log, false, &tv);
2065         }
2066 }
2067
2068 static uint64_t
2069 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2070 {
2071         /*
2072          * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2073          */
2074         uint64_t bw_est, high_rate;
2075         uint64_t gain;
2076
2077         gain = (uint64_t)rack_get_output_gain(rack, rsm);
2078         bw_est = bw * gain;
2079         bw_est /= (uint64_t)100;
2080         /* Never fall below the minimum (def 64kbps) */
2081         if (bw_est < RACK_MIN_BW)
2082                 bw_est = RACK_MIN_BW;
2083         if (rack->r_rack_hw_rate_caps) {
2084                 /* Rate caps are in place */
2085                 if (rack->r_ctl.crte != NULL) {
2086                         /* We have a hdwr rate already */
2087                         high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2088                         if (bw_est >= high_rate) {
2089                                 /* We are capping bw at the highest rate table entry */
2090                                 rack_log_hdwr_pacing(rack,
2091                                                      bw_est, high_rate, __LINE__,
2092                                                      0, 3);
2093                                 bw_est = high_rate;
2094                                 if (capped)
2095                                         *capped = 1;
2096                         }
2097                 } else if ((rack->rack_hdrw_pacing == 0) &&
2098                            (rack->rack_hdw_pace_ena) &&
2099                            (rack->rack_attempt_hdwr_pace == 0) &&
2100                            (rack->rc_inp->inp_route.ro_nh != NULL) &&
2101                            (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2102                         /*
2103                          * Special case, we have not yet attempted hardware
2104                          * pacing, and yet we may, when we do, find out if we are
2105                          * above the highest rate. We need to know the maxbw for the interface
2106                          * in question (if it supports ratelimiting). We get back
2107                          * a 0, if the interface is not found in the RL lists.
2108                          */
2109                         high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2110                         if (high_rate) {
2111                                 /* Yep, we have a rate is it above this rate? */
2112                                 if (bw_est > high_rate) {
2113                                         bw_est = high_rate;
2114                                         if (capped)
2115                                                 *capped = 1;
2116                                 }
2117                         }
2118                 }
2119         }
2120         return (bw_est);
2121 }
2122
2123 static void
2124 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2125 {
2126         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2127                 union tcp_log_stackspecific log;
2128                 struct timeval tv;
2129
2130                 if ((mod != 1) && (rack_verbose_logging == 0)) {
2131                         /*
2132                          * We get 3 values currently for mod
2133                          * 1 - We are retransmitting and this tells the reason.
2134                          * 2 - We are clearing a dup-ack count.
2135                          * 3 - We are incrementing a dup-ack count.
2136                          *
2137                          * The clear/increment are only logged
2138                          * if you have BBverbose on.
2139                          */
2140                         return;
2141                 }
2142                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2143                 log.u_bbr.flex1 = tsused;
2144                 log.u_bbr.flex2 = thresh;
2145                 log.u_bbr.flex3 = rsm->r_flags;
2146                 log.u_bbr.flex4 = rsm->r_dupack;
2147                 log.u_bbr.flex5 = rsm->r_start;
2148                 log.u_bbr.flex6 = rsm->r_end;
2149                 log.u_bbr.flex8 = mod;
2150                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2151                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2152                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2153                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2154                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2155                 log.u_bbr.pacing_gain = rack->r_must_retran;
2156                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2157                     &rack->rc_inp->inp_socket->so_rcv,
2158                     &rack->rc_inp->inp_socket->so_snd,
2159                     BBR_LOG_SETTINGS_CHG, 0,
2160                     0, &log, false, &tv);
2161         }
2162 }
2163
2164 static void
2165 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2166 {
2167         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2168                 union tcp_log_stackspecific log;
2169                 struct timeval tv;
2170
2171                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2172                 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2173                 log.u_bbr.flex2 = to;
2174                 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2175                 log.u_bbr.flex4 = slot;
2176                 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2177                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2178                 log.u_bbr.flex7 = rack->rc_in_persist;
2179                 log.u_bbr.flex8 = which;
2180                 if (rack->rack_no_prr)
2181                         log.u_bbr.pkts_out = 0;
2182                 else
2183                         log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2184                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2185                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2186                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2187                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2188                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2189                 log.u_bbr.pacing_gain = rack->r_must_retran;
2190                 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2191                 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2192                 log.u_bbr.lost = rack_rto_min;
2193                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2194                     &rack->rc_inp->inp_socket->so_rcv,
2195                     &rack->rc_inp->inp_socket->so_snd,
2196                     BBR_LOG_TIMERSTAR, 0,
2197                     0, &log, false, &tv);
2198         }
2199 }
2200
2201 static void
2202 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2203 {
2204         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2205                 union tcp_log_stackspecific log;
2206                 struct timeval tv;
2207
2208                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2209                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2210                 log.u_bbr.flex8 = to_num;
2211                 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2212                 log.u_bbr.flex2 = rack->rc_rack_rtt;
2213                 if (rsm == NULL)
2214                         log.u_bbr.flex3 = 0;
2215                 else
2216                         log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2217                 if (rack->rack_no_prr)
2218                         log.u_bbr.flex5 = 0;
2219                 else
2220                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2221                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2222                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2223                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2224                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2225                 log.u_bbr.pacing_gain = rack->r_must_retran;
2226                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2227                     &rack->rc_inp->inp_socket->so_rcv,
2228                     &rack->rc_inp->inp_socket->so_snd,
2229                     BBR_LOG_RTO, 0,
2230                     0, &log, false, &tv);
2231         }
2232 }
2233
2234 static void
2235 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2236                  struct rack_sendmap *prev,
2237                  struct rack_sendmap *rsm,
2238                  struct rack_sendmap *next,
2239                  int flag, uint32_t th_ack, int line)
2240 {
2241         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2242                 union tcp_log_stackspecific log;
2243                 struct timeval tv;
2244
2245                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2246                 log.u_bbr.flex8 = flag;
2247                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2248                 log.u_bbr.cur_del_rate = (uint64_t)prev;
2249                 log.u_bbr.delRate = (uint64_t)rsm;
2250                 log.u_bbr.rttProp = (uint64_t)next;
2251                 log.u_bbr.flex7 = 0;
2252                 if (prev) {
2253                         log.u_bbr.flex1 = prev->r_start;
2254                         log.u_bbr.flex2 = prev->r_end;
2255                         log.u_bbr.flex7 |= 0x4;
2256                 }
2257                 if (rsm) {
2258                         log.u_bbr.flex3 = rsm->r_start;
2259                         log.u_bbr.flex4 = rsm->r_end;
2260                         log.u_bbr.flex7 |= 0x2;
2261                 }
2262                 if (next) {
2263                         log.u_bbr.flex5 = next->r_start;
2264                         log.u_bbr.flex6 = next->r_end;
2265                         log.u_bbr.flex7 |= 0x1;
2266                 }
2267                 log.u_bbr.applimited = line;
2268                 log.u_bbr.pkts_out = th_ack;
2269                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2270                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2271                 if (rack->rack_no_prr)
2272                         log.u_bbr.lost = 0;
2273                 else
2274                         log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2275                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2276                     &rack->rc_inp->inp_socket->so_rcv,
2277                     &rack->rc_inp->inp_socket->so_snd,
2278                     TCP_LOG_MAPCHG, 0,
2279                     0, &log, false, &tv);
2280         }
2281 }
2282
2283 static void
2284 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2285                  struct rack_sendmap *rsm, int conf)
2286 {
2287         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2288                 union tcp_log_stackspecific log;
2289                 struct timeval tv;
2290                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2291                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2292                 log.u_bbr.flex1 = t;
2293                 log.u_bbr.flex2 = len;
2294                 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2295                 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2296                 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2297                 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2298                 log.u_bbr.flex7 = conf;
2299                 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2300                 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2301                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2302                 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2303                 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2304                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2305                 if (rsm) {
2306                         log.u_bbr.pkt_epoch = rsm->r_start;
2307                         log.u_bbr.lost = rsm->r_end;
2308                         log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2309                         /* We loose any upper of the 24 bits */
2310                         log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2311                 } else {
2312                         /* Its a SYN */
2313                         log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2314                         log.u_bbr.lost = 0;
2315                         log.u_bbr.cwnd_gain = 0;
2316                         log.u_bbr.pacing_gain = 0;
2317                 }
2318                 /* Write out general bits of interest rrs here */
2319                 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2320                 log.u_bbr.use_lt_bw <<= 1;
2321                 log.u_bbr.use_lt_bw |= rack->forced_ack;
2322                 log.u_bbr.use_lt_bw <<= 1;
2323                 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2324                 log.u_bbr.use_lt_bw <<= 1;
2325                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2326                 log.u_bbr.use_lt_bw <<= 1;
2327                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2328                 log.u_bbr.use_lt_bw <<= 1;
2329                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2330                 log.u_bbr.use_lt_bw <<= 1;
2331                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2332                 log.u_bbr.use_lt_bw <<= 1;
2333                 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2334                 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2335                 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2336                 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2337                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2338                 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2339                 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2340                 log.u_bbr.bw_inuse <<= 32;
2341                 if (rsm)
2342                         log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2343                 TCP_LOG_EVENTP(tp, NULL,
2344                     &rack->rc_inp->inp_socket->so_rcv,
2345                     &rack->rc_inp->inp_socket->so_snd,
2346                     BBR_LOG_BBRRTT, 0,
2347                     0, &log, false, &tv);
2348
2349
2350         }
2351 }
2352
2353 static void
2354 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2355 {
2356         /*
2357          * Log the rtt sample we are
2358          * applying to the srtt algorithm in
2359          * useconds.
2360          */
2361         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2362                 union tcp_log_stackspecific log;
2363                 struct timeval tv;
2364
2365                 /* Convert our ms to a microsecond */
2366                 memset(&log, 0, sizeof(log));
2367                 log.u_bbr.flex1 = rtt;
2368                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2369                 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2370                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2371                 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2372                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2373                 log.u_bbr.flex7 = 1;
2374                 log.u_bbr.flex8 = rack->sack_attack_disable;
2375                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2376                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2377                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2378                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2379                 log.u_bbr.pacing_gain = rack->r_must_retran;
2380                 /*
2381                  * We capture in delRate the upper 32 bits as
2382                  * the confidence level we had declared, and the
2383                  * lower 32 bits as the actual RTT using the arrival
2384                  * timestamp.
2385                  */
2386                 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2387                 log.u_bbr.delRate <<= 32;
2388                 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2389                 /* Lets capture all the things that make up t_rtxcur */
2390                 log.u_bbr.applimited = rack_rto_min;
2391                 log.u_bbr.epoch = rack_rto_max;
2392                 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2393                 log.u_bbr.lost = rack_rto_min;
2394                 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2395                 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2396                 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2397                 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2398                 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2399                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2400                     &rack->rc_inp->inp_socket->so_rcv,
2401                     &rack->rc_inp->inp_socket->so_snd,
2402                     TCP_LOG_RTT, 0,
2403                     0, &log, false, &tv);
2404         }
2405 }
2406
2407 static void
2408 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2409 {
2410         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2411                 union tcp_log_stackspecific log;
2412                 struct timeval tv;
2413
2414                 /* Convert our ms to a microsecond */
2415                 memset(&log, 0, sizeof(log));
2416                 log.u_bbr.flex1 = rtt;
2417                 log.u_bbr.flex2 = send_time;
2418                 log.u_bbr.flex3 = ack_time;
2419                 log.u_bbr.flex4 = where;
2420                 log.u_bbr.flex7 = 2;
2421                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2422                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2423                     &rack->rc_inp->inp_socket->so_rcv,
2424                     &rack->rc_inp->inp_socket->so_snd,
2425                     TCP_LOG_RTT, 0,
2426                     0, &log, false, &tv);
2427         }
2428 }
2429
2430
2431
2432 static inline void
2433 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2434 {
2435         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2436                 union tcp_log_stackspecific log;
2437                 struct timeval tv;
2438
2439                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2440                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2441                 log.u_bbr.flex1 = line;
2442                 log.u_bbr.flex2 = tick;
2443                 log.u_bbr.flex3 = tp->t_maxunacktime;
2444                 log.u_bbr.flex4 = tp->t_acktime;
2445                 log.u_bbr.flex8 = event;
2446                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2447                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2448                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2449                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2450                 log.u_bbr.pacing_gain = rack->r_must_retran;
2451                 TCP_LOG_EVENTP(tp, NULL,
2452                     &rack->rc_inp->inp_socket->so_rcv,
2453                     &rack->rc_inp->inp_socket->so_snd,
2454                     BBR_LOG_PROGRESS, 0,
2455                     0, &log, false, &tv);
2456         }
2457 }
2458
2459 static void
2460 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2461 {
2462         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2463                 union tcp_log_stackspecific log;
2464
2465                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2466                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2467                 log.u_bbr.flex1 = slot;
2468                 if (rack->rack_no_prr)
2469                         log.u_bbr.flex2 = 0;
2470                 else
2471                         log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2472                 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2473                 log.u_bbr.flex8 = rack->rc_in_persist;
2474                 log.u_bbr.timeStamp = cts;
2475                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2476                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2477                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2478                 log.u_bbr.pacing_gain = rack->r_must_retran;
2479                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2480                     &rack->rc_inp->inp_socket->so_rcv,
2481                     &rack->rc_inp->inp_socket->so_snd,
2482                     BBR_LOG_BBRSND, 0,
2483                     0, &log, false, tv);
2484         }
2485 }
2486
2487 static void
2488 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2489 {
2490         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2491                 union tcp_log_stackspecific log;
2492                 struct timeval tv;
2493
2494                 memset(&log, 0, sizeof(log));
2495                 log.u_bbr.flex1 = did_out;
2496                 log.u_bbr.flex2 = nxt_pkt;
2497                 log.u_bbr.flex3 = way_out;
2498                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2499                 if (rack->rack_no_prr)
2500                         log.u_bbr.flex5 = 0;
2501                 else
2502                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2503                 log.u_bbr.flex6 = nsegs;
2504                 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2505                 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;        /* Do we have ack-can-send set */
2506                 log.u_bbr.flex7 <<= 1;
2507                 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2508                 log.u_bbr.flex7 <<= 1;
2509                 log.u_bbr.flex7 |= rack->r_wanted_output;       /* Do we want output */
2510                 log.u_bbr.flex8 = rack->rc_in_persist;
2511                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2512                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2513                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2514                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2515                 log.u_bbr.use_lt_bw <<= 1;
2516                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2517                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2518                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2519                 log.u_bbr.pacing_gain = rack->r_must_retran;
2520                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2521                     &rack->rc_inp->inp_socket->so_rcv,
2522                     &rack->rc_inp->inp_socket->so_snd,
2523                     BBR_LOG_DOSEG_DONE, 0,
2524                     0, &log, false, &tv);
2525         }
2526 }
2527
2528 static void
2529 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2530 {
2531         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2532                 union tcp_log_stackspecific log;
2533                 struct timeval tv;
2534
2535                 memset(&log, 0, sizeof(log));
2536                 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2537                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2538                 log.u_bbr.flex4 = arg1;
2539                 log.u_bbr.flex5 = arg2;
2540                 log.u_bbr.flex6 = arg3;
2541                 log.u_bbr.flex8 = frm;
2542                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2543                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2544                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2545                 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2546                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2547                 log.u_bbr.pacing_gain = rack->r_must_retran;
2548                 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
2549                     &tptosocket(tp)->so_snd,
2550                     TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
2551         }
2552 }
2553
2554 static void
2555 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2556                           uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2557 {
2558         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2559                 union tcp_log_stackspecific log;
2560                 struct timeval tv;
2561
2562                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2563                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2564                 log.u_bbr.flex1 = slot;
2565                 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2566                 log.u_bbr.flex4 = reason;
2567                 if (rack->rack_no_prr)
2568                         log.u_bbr.flex5 = 0;
2569                 else
2570                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2571                 log.u_bbr.flex7 = hpts_calling;
2572                 log.u_bbr.flex8 = rack->rc_in_persist;
2573                 log.u_bbr.lt_epoch = cwnd_to_use;
2574                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2575                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2576                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2577                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2578                 log.u_bbr.pacing_gain = rack->r_must_retran;
2579                 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2580                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2581                     &rack->rc_inp->inp_socket->so_rcv,
2582                     &rack->rc_inp->inp_socket->so_snd,
2583                     BBR_LOG_JUSTRET, 0,
2584                     tlen, &log, false, &tv);
2585         }
2586 }
2587
2588 static void
2589 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2590                    struct timeval *tv, uint32_t flags_on_entry)
2591 {
2592         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2593                 union tcp_log_stackspecific log;
2594
2595                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2596                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2597                 log.u_bbr.flex1 = line;
2598                 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2599                 log.u_bbr.flex3 = flags_on_entry;
2600                 log.u_bbr.flex4 = us_cts;
2601                 if (rack->rack_no_prr)
2602                         log.u_bbr.flex5 = 0;
2603                 else
2604                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2605                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2606                 log.u_bbr.flex7 = hpts_removed;
2607                 log.u_bbr.flex8 = 1;
2608                 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2609                 log.u_bbr.timeStamp = us_cts;
2610                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2611                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2612                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2613                 log.u_bbr.pacing_gain = rack->r_must_retran;
2614                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2615                     &rack->rc_inp->inp_socket->so_rcv,
2616                     &rack->rc_inp->inp_socket->so_snd,
2617                     BBR_LOG_TIMERCANC, 0,
2618                     0, &log, false, tv);
2619         }
2620 }
2621
2622 static void
2623 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2624                           uint32_t flex1, uint32_t flex2,
2625                           uint32_t flex3, uint32_t flex4,
2626                           uint32_t flex5, uint32_t flex6,
2627                           uint16_t flex7, uint8_t mod)
2628 {
2629         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2630                 union tcp_log_stackspecific log;
2631                 struct timeval tv;
2632
2633                 if (mod == 1) {
2634                         /* No you can't use 1, its for the real to cancel */
2635                         return;
2636                 }
2637                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2638                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2639                 log.u_bbr.flex1 = flex1;
2640                 log.u_bbr.flex2 = flex2;
2641                 log.u_bbr.flex3 = flex3;
2642                 log.u_bbr.flex4 = flex4;
2643                 log.u_bbr.flex5 = flex5;
2644                 log.u_bbr.flex6 = flex6;
2645                 log.u_bbr.flex7 = flex7;
2646                 log.u_bbr.flex8 = mod;
2647                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2648                     &rack->rc_inp->inp_socket->so_rcv,
2649                     &rack->rc_inp->inp_socket->so_snd,
2650                     BBR_LOG_TIMERCANC, 0,
2651                     0, &log, false, &tv);
2652         }
2653 }
2654
2655 static void
2656 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2657 {
2658         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2659                 union tcp_log_stackspecific log;
2660                 struct timeval tv;
2661
2662                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2663                 log.u_bbr.flex1 = timers;
2664                 log.u_bbr.flex2 = ret;
2665                 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2666                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2667                 log.u_bbr.flex5 = cts;
2668                 if (rack->rack_no_prr)
2669                         log.u_bbr.flex6 = 0;
2670                 else
2671                         log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2672                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2673                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2674                 log.u_bbr.pacing_gain = rack->r_must_retran;
2675                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2676                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2677                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2678                     &rack->rc_inp->inp_socket->so_rcv,
2679                     &rack->rc_inp->inp_socket->so_snd,
2680                     BBR_LOG_TO_PROCESS, 0,
2681                     0, &log, false, &tv);
2682         }
2683 }
2684
2685 static void
2686 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
2687 {
2688         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2689                 union tcp_log_stackspecific log;
2690                 struct timeval tv;
2691
2692                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2693                 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2694                 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2695                 if (rack->rack_no_prr)
2696                         log.u_bbr.flex3 = 0;
2697                 else
2698                         log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2699                 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2700                 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2701                 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2702                 log.u_bbr.flex7 = line;
2703                 log.u_bbr.flex8 = frm;
2704                 log.u_bbr.pkts_out = orig_cwnd;
2705                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2706                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2707                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2708                 log.u_bbr.use_lt_bw <<= 1;
2709                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2710                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2711                     &rack->rc_inp->inp_socket->so_rcv,
2712                     &rack->rc_inp->inp_socket->so_snd,
2713                     BBR_LOG_BBRUPD, 0,
2714                     0, &log, false, &tv);
2715         }
2716 }
2717
2718 #ifdef NETFLIX_EXP_DETECTION
2719 static void
2720 rack_log_sad(struct tcp_rack *rack, int event)
2721 {
2722         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2723                 union tcp_log_stackspecific log;
2724                 struct timeval tv;
2725
2726                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2727                 log.u_bbr.flex1 = rack->r_ctl.sack_count;
2728                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2729                 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2730                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2731                 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2732                 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2733                 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2734                 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2735                 log.u_bbr.lt_epoch |= rack->do_detection;
2736                 log.u_bbr.applimited = tcp_map_minimum;
2737                 log.u_bbr.flex7 = rack->sack_attack_disable;
2738                 log.u_bbr.flex8 = event;
2739                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2740                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2741                 log.u_bbr.delivered = tcp_sad_decay_val;
2742                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2743                     &rack->rc_inp->inp_socket->so_rcv,
2744                     &rack->rc_inp->inp_socket->so_snd,
2745                     TCP_SAD_DETECTION, 0,
2746                     0, &log, false, &tv);
2747         }
2748 }
2749 #endif
2750
2751 static void
2752 rack_counter_destroy(void)
2753 {
2754         counter_u64_free(rack_fto_send);
2755         counter_u64_free(rack_fto_rsm_send);
2756         counter_u64_free(rack_nfto_resend);
2757         counter_u64_free(rack_hw_pace_init_fail);
2758         counter_u64_free(rack_hw_pace_lost);
2759         counter_u64_free(rack_non_fto_send);
2760         counter_u64_free(rack_extended_rfo);
2761         counter_u64_free(rack_ack_total);
2762         counter_u64_free(rack_express_sack);
2763         counter_u64_free(rack_sack_total);
2764         counter_u64_free(rack_move_none);
2765         counter_u64_free(rack_move_some);
2766         counter_u64_free(rack_sack_attacks_detected);
2767         counter_u64_free(rack_sack_attacks_reversed);
2768         counter_u64_free(rack_sack_used_next_merge);
2769         counter_u64_free(rack_sack_used_prev_merge);
2770         counter_u64_free(rack_tlp_tot);
2771         counter_u64_free(rack_tlp_newdata);
2772         counter_u64_free(rack_tlp_retran);
2773         counter_u64_free(rack_tlp_retran_bytes);
2774         counter_u64_free(rack_to_tot);
2775         counter_u64_free(rack_saw_enobuf);
2776         counter_u64_free(rack_saw_enobuf_hw);
2777         counter_u64_free(rack_saw_enetunreach);
2778         counter_u64_free(rack_hot_alloc);
2779         counter_u64_free(rack_to_alloc);
2780         counter_u64_free(rack_to_alloc_hard);
2781         counter_u64_free(rack_to_alloc_emerg);
2782         counter_u64_free(rack_to_alloc_limited);
2783         counter_u64_free(rack_alloc_limited_conns);
2784         counter_u64_free(rack_split_limited);
2785         counter_u64_free(rack_multi_single_eq);
2786         counter_u64_free(rack_proc_non_comp_ack);
2787         counter_u64_free(rack_sack_proc_all);
2788         counter_u64_free(rack_sack_proc_restart);
2789         counter_u64_free(rack_sack_proc_short);
2790         counter_u64_free(rack_sack_skipped_acked);
2791         counter_u64_free(rack_sack_splits);
2792         counter_u64_free(rack_input_idle_reduces);
2793         counter_u64_free(rack_collapsed_win);
2794         counter_u64_free(rack_collapsed_win_rxt);
2795         counter_u64_free(rack_collapsed_win_rxt_bytes);
2796         counter_u64_free(rack_collapsed_win_seen);
2797         counter_u64_free(rack_try_scwnd);
2798         counter_u64_free(rack_persists_sends);
2799         counter_u64_free(rack_persists_acks);
2800         counter_u64_free(rack_persists_loss);
2801         counter_u64_free(rack_persists_lost_ends);
2802 #ifdef INVARIANTS
2803         counter_u64_free(rack_adjust_map_bw);
2804 #endif
2805         COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2806         COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2807 }
2808
2809 static struct rack_sendmap *
2810 rack_alloc(struct tcp_rack *rack)
2811 {
2812         struct rack_sendmap *rsm;
2813
2814         /*
2815          * First get the top of the list it in
2816          * theory is the "hottest" rsm we have,
2817          * possibly just freed by ack processing.
2818          */
2819         if (rack->rc_free_cnt > rack_free_cache) {
2820                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2821                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2822                 counter_u64_add(rack_hot_alloc, 1);
2823                 rack->rc_free_cnt--;
2824                 return (rsm);
2825         }
2826         /*
2827          * Once we get under our free cache we probably
2828          * no longer have a "hot" one available. Lets
2829          * get one from UMA.
2830          */
2831         rsm = uma_zalloc(rack_zone, M_NOWAIT);
2832         if (rsm) {
2833                 rack->r_ctl.rc_num_maps_alloced++;
2834                 counter_u64_add(rack_to_alloc, 1);
2835                 return (rsm);
2836         }
2837         /*
2838          * Dig in to our aux rsm's (the last two) since
2839          * UMA failed to get us one.
2840          */
2841         if (rack->rc_free_cnt) {
2842                 counter_u64_add(rack_to_alloc_emerg, 1);
2843                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2844                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2845                 rack->rc_free_cnt--;
2846                 return (rsm);
2847         }
2848         return (NULL);
2849 }
2850
2851 static struct rack_sendmap *
2852 rack_alloc_full_limit(struct tcp_rack *rack)
2853 {
2854         if ((V_tcp_map_entries_limit > 0) &&
2855             (rack->do_detection == 0) &&
2856             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2857                 counter_u64_add(rack_to_alloc_limited, 1);
2858                 if (!rack->alloc_limit_reported) {
2859                         rack->alloc_limit_reported = 1;
2860                         counter_u64_add(rack_alloc_limited_conns, 1);
2861                 }
2862                 return (NULL);
2863         }
2864         return (rack_alloc(rack));
2865 }
2866
2867 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2868 static struct rack_sendmap *
2869 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2870 {
2871         struct rack_sendmap *rsm;
2872
2873         if (limit_type) {
2874                 /* currently there is only one limit type */
2875                 if (V_tcp_map_split_limit > 0 &&
2876                     (rack->do_detection == 0) &&
2877                     rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2878                         counter_u64_add(rack_split_limited, 1);
2879                         if (!rack->alloc_limit_reported) {
2880                                 rack->alloc_limit_reported = 1;
2881                                 counter_u64_add(rack_alloc_limited_conns, 1);
2882                         }
2883                         return (NULL);
2884                 }
2885         }
2886
2887         /* allocate and mark in the limit type, if set */
2888         rsm = rack_alloc(rack);
2889         if (rsm != NULL && limit_type) {
2890                 rsm->r_limit_type = limit_type;
2891                 rack->r_ctl.rc_num_split_allocs++;
2892         }
2893         return (rsm);
2894 }
2895
2896 static void
2897 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2898 {
2899         if (rsm->r_flags & RACK_APP_LIMITED) {
2900                 if (rack->r_ctl.rc_app_limited_cnt > 0) {
2901                         rack->r_ctl.rc_app_limited_cnt--;
2902                 }
2903         }
2904         if (rsm->r_limit_type) {
2905                 /* currently there is only one limit type */
2906                 rack->r_ctl.rc_num_split_allocs--;
2907         }
2908         if (rsm == rack->r_ctl.rc_first_appl) {
2909                 if (rack->r_ctl.rc_app_limited_cnt == 0)
2910                         rack->r_ctl.rc_first_appl = NULL;
2911                 else {
2912                         /* Follow the next one out */
2913                         struct rack_sendmap fe;
2914
2915                         fe.r_start = rsm->r_nseq_appl;
2916                         rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2917                 }
2918         }
2919         if (rsm == rack->r_ctl.rc_resend)
2920                 rack->r_ctl.rc_resend = NULL;
2921         if (rsm == rack->r_ctl.rc_end_appl)
2922                 rack->r_ctl.rc_end_appl = NULL;
2923         if (rack->r_ctl.rc_tlpsend == rsm)
2924                 rack->r_ctl.rc_tlpsend = NULL;
2925         if (rack->r_ctl.rc_sacklast == rsm)
2926                 rack->r_ctl.rc_sacklast = NULL;
2927         memset(rsm, 0, sizeof(struct rack_sendmap));
2928         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
2929         rack->rc_free_cnt++;
2930 }
2931
2932 static void
2933 rack_free_trim(struct tcp_rack *rack)
2934 {
2935         struct rack_sendmap *rsm;
2936
2937         /*
2938          * Free up all the tail entries until
2939          * we get our list down to the limit.
2940          */
2941         while (rack->rc_free_cnt > rack_free_cache) {
2942                 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
2943                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2944                 rack->rc_free_cnt--;
2945                 uma_zfree(rack_zone, rsm);
2946         }
2947 }
2948
2949
2950 static uint32_t
2951 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2952 {
2953         uint64_t srtt, bw, len, tim;
2954         uint32_t segsiz, def_len, minl;
2955
2956         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2957         def_len = rack_def_data_window * segsiz;
2958         if (rack->rc_gp_filled == 0) {
2959                 /*
2960                  * We have no measurement (IW is in flight?) so
2961                  * we can only guess using our data_window sysctl
2962                  * value (usually 20MSS).
2963                  */
2964                 return (def_len);
2965         }
2966         /*
2967          * Now we have a number of factors to consider.
2968          *
2969          * 1) We have a desired BDP which is usually
2970          *    at least 2.
2971          * 2) We have a minimum number of rtt's usually 1 SRTT
2972          *    but we allow it too to be more.
2973          * 3) We want to make sure a measurement last N useconds (if
2974          *    we have set rack_min_measure_usec.
2975          *
2976          * We handle the first concern here by trying to create a data
2977          * window of max(rack_def_data_window, DesiredBDP). The
2978          * second concern we handle in not letting the measurement
2979          * window end normally until at least the required SRTT's
2980          * have gone by which is done further below in
2981          * rack_enough_for_measurement(). Finally the third concern
2982          * we also handle here by calculating how long that time
2983          * would take at the current BW and then return the
2984          * max of our first calculation and that length. Note
2985          * that if rack_min_measure_usec is 0, we don't deal
2986          * with concern 3. Also for both Concern 1 and 3 an
2987          * application limited period could end the measurement
2988          * earlier.
2989          *
2990          * So lets calculate the BDP with the "known" b/w using
2991          * the SRTT has our rtt and then multiply it by the
2992          * goal.
2993          */
2994         bw = rack_get_bw(rack);
2995         srtt = (uint64_t)tp->t_srtt;
2996         len = bw * srtt;
2997         len /= (uint64_t)HPTS_USEC_IN_SEC;
2998         len *= max(1, rack_goal_bdp);
2999         /* Now we need to round up to the nearest MSS */
3000         len = roundup(len, segsiz);
3001         if (rack_min_measure_usec) {
3002                 /* Now calculate our min length for this b/w */
3003                 tim = rack_min_measure_usec;
3004                 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3005                 if (minl == 0)
3006                         minl = 1;
3007                 minl = roundup(minl, segsiz);
3008                 if (len < minl)
3009                         len = minl;
3010         }
3011         /*
3012          * Now if we have a very small window we want
3013          * to attempt to get the window that is
3014          * as small as possible. This happens on
3015          * low b/w connections and we don't want to
3016          * span huge numbers of rtt's between measurements.
3017          *
3018          * We basically include 2 over our "MIN window" so
3019          * that the measurement can be shortened (possibly) by
3020          * an ack'ed packet.
3021          */
3022         if (len < def_len)
3023                 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3024         else
3025                 return (max((uint32_t)len, def_len));
3026
3027 }
3028
3029 static int
3030 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3031 {
3032         uint32_t tim, srtts, segsiz;
3033
3034         /*
3035          * Has enough time passed for the GP measurement to be valid?
3036          */
3037         if ((tp->snd_max == tp->snd_una) ||
3038             (th_ack == tp->snd_max)){
3039                 /* All is acked */
3040                 *quality = RACK_QUALITY_ALLACKED;
3041                 return (1);
3042         }
3043         if (SEQ_LT(th_ack, tp->gput_seq)) {
3044                 /* Not enough bytes yet */
3045                 return (0);
3046         }
3047         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3048         if (SEQ_LT(th_ack, tp->gput_ack) &&
3049             ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3050                 /* Not enough bytes yet */
3051                 return (0);
3052         }
3053         if (rack->r_ctl.rc_first_appl &&
3054             (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3055                 /*
3056                  * We are up to the app limited send point
3057                  * we have to measure irrespective of the time..
3058                  */
3059                 *quality = RACK_QUALITY_APPLIMITED;
3060                 return (1);
3061         }
3062         /* Now what about time? */
3063         srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3064         tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3065         if (tim >= srtts) {
3066                 *quality = RACK_QUALITY_HIGH;
3067                 return (1);
3068         }
3069         /* Nope not even a full SRTT has passed */
3070         return (0);
3071 }
3072
3073 static void
3074 rack_log_timely(struct tcp_rack *rack,
3075                 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3076                 uint64_t up_bnd, int line, uint8_t method)
3077 {
3078         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3079                 union tcp_log_stackspecific log;
3080                 struct timeval tv;
3081
3082                 memset(&log, 0, sizeof(log));
3083                 log.u_bbr.flex1 = logged;
3084                 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3085                 log.u_bbr.flex2 <<= 4;
3086                 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3087                 log.u_bbr.flex2 <<= 4;
3088                 log.u_bbr.flex2 |= rack->rc_gp_incr;
3089                 log.u_bbr.flex2 <<= 4;
3090                 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3091                 log.u_bbr.flex3 = rack->rc_gp_incr;
3092                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3093                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3094                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3095                 log.u_bbr.flex7 = rack->rc_gp_bwred;
3096                 log.u_bbr.flex8 = method;
3097                 log.u_bbr.cur_del_rate = cur_bw;
3098                 log.u_bbr.delRate = low_bnd;
3099                 log.u_bbr.bw_inuse = up_bnd;
3100                 log.u_bbr.rttProp = rack_get_bw(rack);
3101                 log.u_bbr.pkt_epoch = line;
3102                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3103                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3104                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3105                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3106                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3107                 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3108                 log.u_bbr.cwnd_gain <<= 1;
3109                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3110                 log.u_bbr.cwnd_gain <<= 1;
3111                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3112                 log.u_bbr.cwnd_gain <<= 1;
3113                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3114                 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3115                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3116                     &rack->rc_inp->inp_socket->so_rcv,
3117                     &rack->rc_inp->inp_socket->so_snd,
3118                     TCP_TIMELY_WORK, 0,
3119                     0, &log, false, &tv);
3120         }
3121 }
3122
3123 static int
3124 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3125 {
3126         /*
3127          * Before we increase we need to know if
3128          * the estimate just made was less than
3129          * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3130          *
3131          * If we already are pacing at a fast enough
3132          * rate to push us faster there is no sense of
3133          * increasing.
3134          *
3135          * We first caculate our actual pacing rate (ss or ca multiplier
3136          * times our cur_bw).
3137          *
3138          * Then we take the last measured rate and multipy by our
3139          * maximum pacing overage to give us a max allowable rate.
3140          *
3141          * If our act_rate is smaller than our max_allowable rate
3142          * then we should increase. Else we should hold steady.
3143          *
3144          */
3145         uint64_t act_rate, max_allow_rate;
3146
3147         if (rack_timely_no_stopping)
3148                 return (1);
3149
3150         if ((cur_bw == 0) || (last_bw_est == 0)) {
3151                 /*
3152                  * Initial startup case or
3153                  * everything is acked case.
3154                  */
3155                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3156                                 __LINE__, 9);
3157                 return (1);
3158         }
3159         if (mult <= 100) {
3160                 /*
3161                  * We can always pace at or slightly above our rate.
3162                  */
3163                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3164                                 __LINE__, 9);
3165                 return (1);
3166         }
3167         act_rate = cur_bw * (uint64_t)mult;
3168         act_rate /= 100;
3169         max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3170         max_allow_rate /= 100;
3171         if (act_rate < max_allow_rate) {
3172                 /*
3173                  * Here the rate we are actually pacing at
3174                  * is smaller than 10% above our last measurement.
3175                  * This means we are pacing below what we would
3176                  * like to try to achieve (plus some wiggle room).
3177                  */
3178                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3179                                 __LINE__, 9);
3180                 return (1);
3181         } else {
3182                 /*
3183                  * Here we are already pacing at least rack_max_per_above(10%)
3184                  * what we are getting back. This indicates most likely
3185                  * that we are being limited (cwnd/rwnd/app) and can't
3186                  * get any more b/w. There is no sense of trying to
3187                  * raise up the pacing rate its not speeding us up
3188                  * and we already are pacing faster than we are getting.
3189                  */
3190                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3191                                 __LINE__, 8);
3192                 return (0);
3193         }
3194 }
3195
3196 static void
3197 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3198 {
3199         /*
3200          * When we drag bottom, we want to assure
3201          * that no multiplier is below 1.0, if so
3202          * we want to restore it to at least that.
3203          */
3204         if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3205                 /* This is unlikely we usually do not touch recovery */
3206                 rack->r_ctl.rack_per_of_gp_rec = 100;
3207         }
3208         if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3209                 rack->r_ctl.rack_per_of_gp_ca = 100;
3210         }
3211         if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3212                 rack->r_ctl.rack_per_of_gp_ss = 100;
3213         }
3214 }
3215
3216 static void
3217 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3218 {
3219         if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3220                 rack->r_ctl.rack_per_of_gp_ca = 100;
3221         }
3222         if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3223                 rack->r_ctl.rack_per_of_gp_ss = 100;
3224         }
3225 }
3226
3227 static void
3228 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3229 {
3230         int32_t  calc, logged, plus;
3231
3232         logged = 0;
3233
3234         if (override) {
3235                 /*
3236                  * override is passed when we are
3237                  * loosing b/w and making one last
3238                  * gasp at trying to not loose out
3239                  * to a new-reno flow.
3240                  */
3241                 goto extra_boost;
3242         }
3243         /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3244         if (rack->rc_gp_incr &&
3245             ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3246                 /*
3247                  * Reset and get 5 strokes more before the boost. Note
3248                  * that the count is 0 based so we have to add one.
3249                  */
3250 extra_boost:
3251                 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3252                 rack->rc_gp_timely_inc_cnt = 0;
3253         } else
3254                 plus = (uint32_t)rack_gp_increase_per;
3255         /* Must be at least 1% increase for true timely increases */
3256         if ((plus < 1) &&
3257             ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3258                 plus = 1;
3259         if (rack->rc_gp_saw_rec &&
3260             (rack->rc_gp_no_rec_chg == 0) &&
3261             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3262                                   rack->r_ctl.rack_per_of_gp_rec)) {
3263                 /* We have been in recovery ding it too */
3264                 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3265                 if (calc > 0xffff)
3266                         calc = 0xffff;
3267                 logged |= 1;
3268                 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3269                 if (rack_per_upper_bound_ss &&
3270                     (rack->rc_dragged_bottom == 0) &&
3271                     (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3272                         rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3273         }
3274         if (rack->rc_gp_saw_ca &&
3275             (rack->rc_gp_saw_ss == 0) &&
3276             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3277                                   rack->r_ctl.rack_per_of_gp_ca)) {
3278                 /* In CA */
3279                 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3280                 if (calc > 0xffff)
3281                         calc = 0xffff;
3282                 logged |= 2;
3283                 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3284                 if (rack_per_upper_bound_ca &&
3285                     (rack->rc_dragged_bottom == 0) &&
3286                     (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3287                         rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3288         }
3289         if (rack->rc_gp_saw_ss &&
3290             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3291                                   rack->r_ctl.rack_per_of_gp_ss)) {
3292                 /* In SS */
3293                 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3294                 if (calc > 0xffff)
3295                         calc = 0xffff;
3296                 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3297                 if (rack_per_upper_bound_ss &&
3298                     (rack->rc_dragged_bottom == 0) &&
3299                     (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3300                         rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3301                 logged |= 4;
3302         }
3303         if (logged &&
3304             (rack->rc_gp_incr == 0)){
3305                 /* Go into increment mode */
3306                 rack->rc_gp_incr = 1;
3307                 rack->rc_gp_timely_inc_cnt = 0;
3308         }
3309         if (rack->rc_gp_incr &&
3310             logged &&
3311             (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3312                 rack->rc_gp_timely_inc_cnt++;
3313         }
3314         rack_log_timely(rack,  logged, plus, 0, 0,
3315                         __LINE__, 1);
3316 }
3317
3318 static uint32_t
3319 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3320 {
3321         /*
3322          * norm_grad = rtt_diff / minrtt;
3323          * new_per = curper * (1 - B * norm_grad)
3324          *
3325          * B = rack_gp_decrease_per (default 10%)
3326          * rtt_dif = input var current rtt-diff
3327          * curper = input var current percentage
3328          * minrtt = from rack filter
3329          *
3330          */
3331         uint64_t perf;
3332
3333         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3334                     ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3335                      (((uint64_t)rtt_diff * (uint64_t)1000000)/
3336                       (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3337                      (uint64_t)1000000)) /
3338                 (uint64_t)1000000);
3339         if (perf > curper) {
3340                 /* TSNH */
3341                 perf = curper - 1;
3342         }
3343         return ((uint32_t)perf);
3344 }
3345
3346 static uint32_t
3347 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3348 {
3349         /*
3350          *                                   highrttthresh
3351          * result = curper * (1 - (B * ( 1 -  ------          ))
3352          *                                     gp_srtt
3353          *
3354          * B = rack_gp_decrease_per (default 10%)
3355          * highrttthresh = filter_min * rack_gp_rtt_maxmul
3356          */
3357         uint64_t perf;
3358         uint32_t highrttthresh;
3359
3360         highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3361
3362         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3363                                      ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3364                                         ((uint64_t)highrttthresh * (uint64_t)1000000) /
3365                                                     (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3366         return (perf);
3367 }
3368
3369 static void
3370 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3371 {
3372         uint64_t logvar, logvar2, logvar3;
3373         uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3374
3375         if (rack->rc_gp_incr) {
3376                 /* Turn off increment counting */
3377                 rack->rc_gp_incr = 0;
3378                 rack->rc_gp_timely_inc_cnt = 0;
3379         }
3380         ss_red = ca_red = rec_red = 0;
3381         logged = 0;
3382         /* Calculate the reduction value */
3383         if (rtt_diff < 0) {
3384                 rtt_diff *= -1;
3385         }
3386         /* Must be at least 1% reduction */
3387         if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3388                 /* We have been in recovery ding it too */
3389                 if (timely_says == 2) {
3390                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3391                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3392                         if (alt < new_per)
3393                                 val = alt;
3394                         else
3395                                 val = new_per;
3396                 } else
3397                          val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3398                 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3399                         rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3400                         rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3401                 } else {
3402                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3403                         rec_red = 0;
3404                 }
3405                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3406                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3407                 logged |= 1;
3408         }
3409         if (rack->rc_gp_saw_ss) {
3410                 /* Sent in SS */
3411                 if (timely_says == 2) {
3412                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3413                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3414                         if (alt < new_per)
3415                                 val = alt;
3416                         else
3417                                 val = new_per;
3418                 } else
3419                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3420                 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3421                         ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3422                         rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3423                 } else {
3424                         ss_red = new_per;
3425                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3426                         logvar = new_per;
3427                         logvar <<= 32;
3428                         logvar |= alt;
3429                         logvar2 = (uint32_t)rtt;
3430                         logvar2 <<= 32;
3431                         logvar2 |= (uint32_t)rtt_diff;
3432                         logvar3 = rack_gp_rtt_maxmul;
3433                         logvar3 <<= 32;
3434                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3435                         rack_log_timely(rack, timely_says,
3436                                         logvar2, logvar3,
3437                                         logvar, __LINE__, 10);
3438                 }
3439                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3440                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3441                 logged |= 4;
3442         } else if (rack->rc_gp_saw_ca) {
3443                 /* Sent in CA */
3444                 if (timely_says == 2) {
3445                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3446                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3447                         if (alt < new_per)
3448                                 val = alt;
3449                         else
3450                                 val = new_per;
3451                 } else
3452                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3453                 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3454                         ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3455                         rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3456                 } else {
3457                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3458                         ca_red = 0;
3459                         logvar = new_per;
3460                         logvar <<= 32;
3461                         logvar |= alt;
3462                         logvar2 = (uint32_t)rtt;
3463                         logvar2 <<= 32;
3464                         logvar2 |= (uint32_t)rtt_diff;
3465                         logvar3 = rack_gp_rtt_maxmul;
3466                         logvar3 <<= 32;
3467                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3468                         rack_log_timely(rack, timely_says,
3469                                         logvar2, logvar3,
3470                                         logvar, __LINE__, 10);
3471                 }
3472                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3473                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3474                 logged |= 2;
3475         }
3476         if (rack->rc_gp_timely_dec_cnt < 0x7) {
3477                 rack->rc_gp_timely_dec_cnt++;
3478                 if (rack_timely_dec_clear &&
3479                     (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3480                         rack->rc_gp_timely_dec_cnt = 0;
3481         }
3482         logvar = ss_red;
3483         logvar <<= 32;
3484         logvar |= ca_red;
3485         rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3486                         __LINE__, 2);
3487 }
3488
3489 static void
3490 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3491                      uint32_t rtt, uint32_t line, uint8_t reas)
3492 {
3493         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3494                 union tcp_log_stackspecific log;
3495                 struct timeval tv;
3496
3497                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3498                 log.u_bbr.flex1 = line;
3499                 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3500                 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3501                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3502                 log.u_bbr.flex5 = rtt;
3503                 log.u_bbr.flex6 = rack->rc_highly_buffered;
3504                 log.u_bbr.flex6 <<= 1;
3505                 log.u_bbr.flex6 |= rack->forced_ack;
3506                 log.u_bbr.flex6 <<= 1;
3507                 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3508                 log.u_bbr.flex6 <<= 1;
3509                 log.u_bbr.flex6 |= rack->in_probe_rtt;
3510                 log.u_bbr.flex6 <<= 1;
3511                 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3512                 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3513                 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3514                 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3515                 log.u_bbr.flex8 = reas;
3516                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3517                 log.u_bbr.delRate = rack_get_bw(rack);
3518                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3519                 log.u_bbr.cur_del_rate <<= 32;
3520                 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3521                 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3522                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3523                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3524                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3525                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3526                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3527                 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3528                 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3529                 log.u_bbr.rttProp = us_cts;
3530                 log.u_bbr.rttProp <<= 32;
3531                 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3532                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3533                     &rack->rc_inp->inp_socket->so_rcv,
3534                     &rack->rc_inp->inp_socket->so_snd,
3535                     BBR_LOG_RTT_SHRINKS, 0,
3536                     0, &log, false, &rack->r_ctl.act_rcv_time);
3537         }
3538 }
3539
3540 static void
3541 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3542 {
3543         uint64_t bwdp;
3544
3545         bwdp = rack_get_bw(rack);
3546         bwdp *= (uint64_t)rtt;
3547         bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3548         rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3549         if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3550                 /*
3551                  * A window protocol must be able to have 4 packets
3552                  * outstanding as the floor in order to function
3553                  * (especially considering delayed ack :D).
3554                  */
3555                 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3556         }
3557 }
3558
3559 static void
3560 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3561 {
3562         /**
3563          * ProbeRTT is a bit different in rack_pacing than in
3564          * BBR. It is like BBR in that it uses the lowering of
3565          * the RTT as a signal that we saw something new and
3566          * counts from there for how long between. But it is
3567          * different in that its quite simple. It does not
3568          * play with the cwnd and wait until we get down
3569          * to N segments outstanding and hold that for
3570          * 200ms. Instead it just sets the pacing reduction
3571          * rate to a set percentage (70 by default) and hold
3572          * that for a number of recent GP Srtt's.
3573          */
3574         uint32_t segsiz;
3575
3576         if (rack->rc_gp_dyn_mul == 0)
3577                 return;
3578
3579         if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3580                 /* We are idle */
3581                 return;
3582         }
3583         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3584             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3585                 /*
3586                  * Stop the goodput now, the idea here is
3587                  * that future measurements with in_probe_rtt
3588                  * won't register if they are not greater so
3589                  * we want to get what info (if any) is available
3590                  * now.
3591                  */
3592                 rack_do_goodput_measurement(rack->rc_tp, rack,
3593                                             rack->rc_tp->snd_una, __LINE__,
3594                                             RACK_QUALITY_PROBERTT);
3595         }
3596         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3597         rack->r_ctl.rc_time_probertt_entered = us_cts;
3598         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3599                      rack->r_ctl.rc_pace_min_segs);
3600         rack->in_probe_rtt = 1;
3601         rack->measure_saw_probe_rtt = 1;
3602         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3603         rack->r_ctl.rc_time_probertt_starts = 0;
3604         rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3605         if (rack_probertt_use_min_rtt_entry)
3606                 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3607         else
3608                 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3609         rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3610                              __LINE__, RACK_RTTS_ENTERPROBE);
3611 }
3612
3613 static void
3614 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3615 {
3616         struct rack_sendmap *rsm;
3617         uint32_t segsiz;
3618
3619         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3620                      rack->r_ctl.rc_pace_min_segs);
3621         rack->in_probe_rtt = 0;
3622         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3623             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3624                 /*
3625                  * Stop the goodput now, the idea here is
3626                  * that future measurements with in_probe_rtt
3627                  * won't register if they are not greater so
3628                  * we want to get what info (if any) is available
3629                  * now.
3630                  */
3631                 rack_do_goodput_measurement(rack->rc_tp, rack,
3632                                             rack->rc_tp->snd_una, __LINE__,
3633                                             RACK_QUALITY_PROBERTT);
3634         } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3635                 /*
3636                  * We don't have enough data to make a measurement.
3637                  * So lets just stop and start here after exiting
3638                  * probe-rtt. We probably are not interested in
3639                  * the results anyway.
3640                  */
3641                 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3642         }
3643         /*
3644          * Measurements through the current snd_max are going
3645          * to be limited by the slower pacing rate.
3646          *
3647          * We need to mark these as app-limited so we
3648          * don't collapse the b/w.
3649          */
3650         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3651         if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3652                 if (rack->r_ctl.rc_app_limited_cnt == 0)
3653                         rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3654                 else {
3655                         /*
3656                          * Go out to the end app limited and mark
3657                          * this new one as next and move the end_appl up
3658                          * to this guy.
3659                          */
3660                         if (rack->r_ctl.rc_end_appl)
3661                                 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3662                         rack->r_ctl.rc_end_appl = rsm;
3663                 }
3664                 rsm->r_flags |= RACK_APP_LIMITED;
3665                 rack->r_ctl.rc_app_limited_cnt++;
3666         }
3667         /*
3668          * Now, we need to examine our pacing rate multipliers.
3669          * If its under 100%, we need to kick it back up to
3670          * 100%. We also don't let it be over our "max" above
3671          * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3672          * Note setting clamp_atexit_prtt to 0 has the effect
3673          * of setting CA/SS to 100% always at exit (which is
3674          * the default behavior).
3675          */
3676         if (rack_probertt_clear_is) {
3677                 rack->rc_gp_incr = 0;
3678                 rack->rc_gp_bwred = 0;
3679                 rack->rc_gp_timely_inc_cnt = 0;
3680                 rack->rc_gp_timely_dec_cnt = 0;
3681         }
3682         /* Do we do any clamping at exit? */
3683         if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3684                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3685                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3686         }
3687         if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3688                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3689                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3690         }
3691         /*
3692          * Lets set rtt_diff to 0, so that we will get a "boost"
3693          * after exiting.
3694          */
3695         rack->r_ctl.rc_rtt_diff = 0;
3696
3697         /* Clear all flags so we start fresh */
3698         rack->rc_tp->t_bytes_acked = 0;
3699         rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
3700         /*
3701          * If configured to, set the cwnd and ssthresh to
3702          * our targets.
3703          */
3704         if (rack_probe_rtt_sets_cwnd) {
3705                 uint64_t ebdp;
3706                 uint32_t setto;
3707
3708                 /* Set ssthresh so we get into CA once we hit our target */
3709                 if (rack_probertt_use_min_rtt_exit == 1) {
3710                         /* Set to min rtt */
3711                         rack_set_prtt_target(rack, segsiz,
3712                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3713                 } else if (rack_probertt_use_min_rtt_exit == 2) {
3714                         /* Set to current gp rtt */
3715                         rack_set_prtt_target(rack, segsiz,
3716                                              rack->r_ctl.rc_gp_srtt);
3717                 } else if (rack_probertt_use_min_rtt_exit == 3) {
3718                         /* Set to entry gp rtt */
3719                         rack_set_prtt_target(rack, segsiz,
3720                                              rack->r_ctl.rc_entry_gp_rtt);
3721                 } else {
3722                         uint64_t sum;
3723                         uint32_t setval;
3724
3725                         sum = rack->r_ctl.rc_entry_gp_rtt;
3726                         sum *= 10;
3727                         sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3728                         if (sum >= 20) {
3729                                 /*
3730                                  * A highly buffered path needs
3731                                  * cwnd space for timely to work.
3732                                  * Lets set things up as if
3733                                  * we are heading back here again.
3734                                  */
3735                                 setval = rack->r_ctl.rc_entry_gp_rtt;
3736                         } else if (sum >= 15) {
3737                                 /*
3738                                  * Lets take the smaller of the
3739                                  * two since we are just somewhat
3740                                  * buffered.
3741                                  */
3742                                 setval = rack->r_ctl.rc_gp_srtt;
3743                                 if (setval > rack->r_ctl.rc_entry_gp_rtt)
3744                                         setval = rack->r_ctl.rc_entry_gp_rtt;
3745                         } else {
3746                                 /*
3747                                  * Here we are not highly buffered
3748                                  * and should pick the min we can to
3749                                  * keep from causing loss.
3750                                  */
3751                                 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3752                         }
3753                         rack_set_prtt_target(rack, segsiz,
3754                                              setval);
3755                 }
3756                 if (rack_probe_rtt_sets_cwnd > 1) {
3757                         /* There is a percentage here to boost */
3758                         ebdp = rack->r_ctl.rc_target_probertt_flight;
3759                         ebdp *= rack_probe_rtt_sets_cwnd;
3760                         ebdp /= 100;
3761                         setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3762                 } else
3763                         setto = rack->r_ctl.rc_target_probertt_flight;
3764                 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3765                 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3766                         /* Enforce a min */
3767                         rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3768                 }
3769                 /* If we set in the cwnd also set the ssthresh point so we are in CA */
3770                 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3771         }
3772         rack_log_rtt_shrinks(rack,  us_cts,
3773                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3774                              __LINE__, RACK_RTTS_EXITPROBE);
3775         /* Clear times last so log has all the info */
3776         rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3777         rack->r_ctl.rc_time_probertt_entered = us_cts;
3778         rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3779         rack->r_ctl.rc_time_of_last_probertt = us_cts;
3780 }
3781
3782 static void
3783 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3784 {
3785         /* Check in on probe-rtt */
3786         if (rack->rc_gp_filled == 0) {
3787                 /* We do not do p-rtt unless we have gp measurements */
3788                 return;
3789         }
3790         if (rack->in_probe_rtt) {
3791                 uint64_t no_overflow;
3792                 uint32_t endtime, must_stay;
3793
3794                 if (rack->r_ctl.rc_went_idle_time &&
3795                     ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3796                         /*
3797                          * We went idle during prtt, just exit now.
3798                          */
3799                         rack_exit_probertt(rack, us_cts);
3800                 } else if (rack_probe_rtt_safety_val &&
3801                     TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3802                     ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3803                         /*
3804                          * Probe RTT safety value triggered!
3805                          */
3806                         rack_log_rtt_shrinks(rack,  us_cts,
3807                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3808                                              __LINE__, RACK_RTTS_SAFETY);
3809                         rack_exit_probertt(rack, us_cts);
3810                 }
3811                 /* Calculate the max we will wait */
3812                 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3813                 if (rack->rc_highly_buffered)
3814                         endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3815                 /* Calculate the min we must wait */
3816                 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3817                 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3818                     TSTMP_LT(us_cts, endtime)) {
3819                         uint32_t calc;
3820                         /* Do we lower more? */
3821 no_exit:
3822                         if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3823                                 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3824                         else
3825                                 calc = 0;
3826                         calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3827                         if (calc) {
3828                                 /* Maybe */
3829                                 calc *= rack_per_of_gp_probertt_reduce;
3830                                 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3831                                 /* Limit it too */
3832                                 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3833                                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3834                         }
3835                         /* We must reach target or the time set */
3836                         return;
3837                 }
3838                 if (rack->r_ctl.rc_time_probertt_starts == 0) {
3839                         if ((TSTMP_LT(us_cts, must_stay) &&
3840                              rack->rc_highly_buffered) ||
3841                              (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3842                               rack->r_ctl.rc_target_probertt_flight)) {
3843                                 /* We are not past the must_stay time */
3844                                 goto no_exit;
3845                         }
3846                         rack_log_rtt_shrinks(rack,  us_cts,
3847                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3848                                              __LINE__, RACK_RTTS_REACHTARGET);
3849                         rack->r_ctl.rc_time_probertt_starts = us_cts;
3850                         if (rack->r_ctl.rc_time_probertt_starts == 0)
3851                                 rack->r_ctl.rc_time_probertt_starts = 1;
3852                         /* Restore back to our rate we want to pace at in prtt */
3853                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3854                 }
3855                 /*
3856                  * Setup our end time, some number of gp_srtts plus 200ms.
3857                  */
3858                 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3859                                (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3860                 if (rack_probertt_gpsrtt_cnt_div)
3861                         endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3862                 else
3863                         endtime = 0;
3864                 endtime += rack_min_probertt_hold;
3865                 endtime += rack->r_ctl.rc_time_probertt_starts;
3866                 if (TSTMP_GEQ(us_cts,  endtime)) {
3867                         /* yes, exit probertt */
3868                         rack_exit_probertt(rack, us_cts);
3869                 }
3870
3871         } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3872                 /* Go into probertt, its been too long since we went lower */
3873                 rack_enter_probertt(rack, us_cts);
3874         }
3875 }
3876
3877 static void
3878 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3879                        uint32_t rtt, int32_t rtt_diff)
3880 {
3881         uint64_t cur_bw, up_bnd, low_bnd, subfr;
3882         uint32_t losses;
3883
3884         if ((rack->rc_gp_dyn_mul == 0) ||
3885             (rack->use_fixed_rate) ||
3886             (rack->in_probe_rtt) ||
3887             (rack->rc_always_pace == 0)) {
3888                 /* No dynamic GP multiplier in play */
3889                 return;
3890         }
3891         losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3892         cur_bw = rack_get_bw(rack);
3893         /* Calculate our up and down range */
3894         up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3895         up_bnd /= 100;
3896         up_bnd += rack->r_ctl.last_gp_comp_bw;
3897
3898         subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3899         subfr /= 100;
3900         low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3901         if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3902                 /*
3903                  * This is the case where our RTT is above
3904                  * the max target and we have been configured
3905                  * to just do timely no bonus up stuff in that case.
3906                  *
3907                  * There are two configurations, set to 1, and we
3908                  * just do timely if we are over our max. If its
3909                  * set above 1 then we slam the multipliers down
3910                  * to 100 and then decrement per timely.
3911                  */
3912                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3913                                 __LINE__, 3);
3914                 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3915                         rack_validate_multipliers_at_or_below_100(rack);
3916                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3917         } else if ((last_bw_est < low_bnd) && !losses) {
3918                 /*
3919                  * We are decreasing this is a bit complicated this
3920                  * means we are loosing ground. This could be
3921                  * because another flow entered and we are competing
3922                  * for b/w with it. This will push the RTT up which
3923                  * makes timely unusable unless we want to get shoved
3924                  * into a corner and just be backed off (the age
3925                  * old problem with delay based CC).
3926                  *
3927                  * On the other hand if it was a route change we
3928                  * would like to stay somewhat contained and not
3929                  * blow out the buffers.
3930                  */
3931                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3932                                 __LINE__, 3);
3933                 rack->r_ctl.last_gp_comp_bw = cur_bw;
3934                 if (rack->rc_gp_bwred == 0) {
3935                         /* Go into reduction counting */
3936                         rack->rc_gp_bwred = 1;
3937                         rack->rc_gp_timely_dec_cnt = 0;
3938                 }
3939                 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3940                     (timely_says == 0)) {
3941                         /*
3942                          * Push another time with a faster pacing
3943                          * to try to gain back (we include override to
3944                          * get a full raise factor).
3945                          */
3946                         if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3947                             (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3948                             (timely_says == 0) ||
3949                             (rack_down_raise_thresh == 0)) {
3950                                 /*
3951                                  * Do an override up in b/w if we were
3952                                  * below the threshold or if the threshold
3953                                  * is zero we always do the raise.
3954                                  */
3955                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3956                         } else {
3957                                 /* Log it stays the same */
3958                                 rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3959                                                 __LINE__, 11);
3960                         }
3961                         rack->rc_gp_timely_dec_cnt++;
3962                         /* We are not incrementing really no-count */
3963                         rack->rc_gp_incr = 0;
3964                         rack->rc_gp_timely_inc_cnt = 0;
3965                 } else {
3966                         /*
3967                          * Lets just use the RTT
3968                          * information and give up
3969                          * pushing.
3970                          */
3971                         goto use_timely;
3972                 }
3973         } else if ((timely_says != 2) &&
3974                     !losses &&
3975                     (last_bw_est > up_bnd)) {
3976                 /*
3977                  * We are increasing b/w lets keep going, updating
3978                  * our b/w and ignoring any timely input, unless
3979                  * of course we are at our max raise (if there is one).
3980                  */
3981
3982                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3983                                 __LINE__, 3);
3984                 rack->r_ctl.last_gp_comp_bw = cur_bw;
3985                 if (rack->rc_gp_saw_ss &&
3986                     rack_per_upper_bound_ss &&
3987                      (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3988                             /*
3989                              * In cases where we can't go higher
3990                              * we should just use timely.
3991                              */
3992                             goto use_timely;
3993                 }
3994                 if (rack->rc_gp_saw_ca &&
3995                     rack_per_upper_bound_ca &&
3996                     (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
3997                             /*
3998                              * In cases where we can't go higher
3999                              * we should just use timely.
4000                              */
4001                             goto use_timely;
4002                 }
4003                 rack->rc_gp_bwred = 0;
4004                 rack->rc_gp_timely_dec_cnt = 0;
4005                 /* You get a set number of pushes if timely is trying to reduce */
4006                 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4007                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4008                 } else {
4009                         /* Log it stays the same */
4010                         rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4011                             __LINE__, 12);
4012                 }
4013                 return;
4014         } else {
4015                 /*
4016                  * We are staying between the lower and upper range bounds
4017                  * so use timely to decide.
4018                  */
4019                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4020                                 __LINE__, 3);
4021 use_timely:
4022                 if (timely_says) {
4023                         rack->rc_gp_incr = 0;
4024                         rack->rc_gp_timely_inc_cnt = 0;
4025                         if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4026                             !losses &&
4027                             (last_bw_est < low_bnd)) {
4028                                 /* We are loosing ground */
4029                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4030                                 rack->rc_gp_timely_dec_cnt++;
4031                                 /* We are not incrementing really no-count */
4032                                 rack->rc_gp_incr = 0;
4033                                 rack->rc_gp_timely_inc_cnt = 0;
4034                         } else
4035                                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4036                 } else {
4037                         rack->rc_gp_bwred = 0;
4038                         rack->rc_gp_timely_dec_cnt = 0;
4039                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4040                 }
4041         }
4042 }
4043
4044 static int32_t
4045 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4046 {
4047         int32_t timely_says;
4048         uint64_t log_mult, log_rtt_a_diff;
4049
4050         log_rtt_a_diff = rtt;
4051         log_rtt_a_diff <<= 32;
4052         log_rtt_a_diff |= (uint32_t)rtt_diff;
4053         if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4054                     rack_gp_rtt_maxmul)) {
4055                 /* Reduce the b/w multiplier */
4056                 timely_says = 2;
4057                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4058                 log_mult <<= 32;
4059                 log_mult |= prev_rtt;
4060                 rack_log_timely(rack,  timely_says, log_mult,
4061                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4062                                 log_rtt_a_diff, __LINE__, 4);
4063         } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4064                            ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4065                             max(rack_gp_rtt_mindiv , 1)))) {
4066                 /* Increase the b/w multiplier */
4067                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4068                         ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4069                          max(rack_gp_rtt_mindiv , 1));
4070                 log_mult <<= 32;
4071                 log_mult |= prev_rtt;
4072                 timely_says = 0;
4073                 rack_log_timely(rack,  timely_says, log_mult ,
4074                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4075                                 log_rtt_a_diff, __LINE__, 5);
4076         } else {
4077                 /*
4078                  * Use a gradient to find it the timely gradient
4079                  * is:
4080                  * grad = rc_rtt_diff / min_rtt;
4081                  *
4082                  * anything below or equal to 0 will be
4083                  * a increase indication. Anything above
4084                  * zero is a decrease. Note we take care
4085                  * of the actual gradient calculation
4086                  * in the reduction (its not needed for
4087                  * increase).
4088                  */
4089                 log_mult = prev_rtt;
4090                 if (rtt_diff <= 0) {
4091                         /*
4092                          * Rttdiff is less than zero, increase the
4093                          * b/w multiplier (its 0 or negative)
4094                          */
4095                         timely_says = 0;
4096                         rack_log_timely(rack,  timely_says, log_mult,
4097                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4098                 } else {
4099                         /* Reduce the b/w multiplier */
4100                         timely_says = 1;
4101                         rack_log_timely(rack,  timely_says, log_mult,
4102                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4103                 }
4104         }
4105         return (timely_says);
4106 }
4107
4108 static void
4109 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4110                             tcp_seq th_ack, int line, uint8_t quality)
4111 {
4112         uint64_t tim, bytes_ps, ltim, stim, utim;
4113         uint32_t segsiz, bytes, reqbytes, us_cts;
4114         int32_t gput, new_rtt_diff, timely_says;
4115         uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4116         int did_add = 0;
4117
4118         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4119         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4120         if (TSTMP_GEQ(us_cts, tp->gput_ts))
4121                 tim = us_cts - tp->gput_ts;
4122         else
4123                 tim = 0;
4124         if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4125                 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4126         else
4127                 stim = 0;
4128         /*
4129          * Use the larger of the send time or ack time. This prevents us
4130          * from being influenced by ack artifacts to come up with too
4131          * high of measurement. Note that since we are spanning over many more
4132          * bytes in most of our measurements hopefully that is less likely to
4133          * occur.
4134          */
4135         if (tim > stim)
4136                 utim = max(tim, 1);
4137         else
4138                 utim = max(stim, 1);
4139         /* Lets get a msec time ltim too for the old stuff */
4140         ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4141         gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4142         reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4143         if ((tim == 0) && (stim == 0)) {
4144                 /*
4145                  * Invalid measurement time, maybe
4146                  * all on one ack/one send?
4147                  */
4148                 bytes = 0;
4149                 bytes_ps = 0;
4150                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4151                                            0, 0, 0, 10, __LINE__, NULL, quality);
4152                 goto skip_measurement;
4153         }
4154         if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4155                 /* We never made a us_rtt measurement? */
4156                 bytes = 0;
4157                 bytes_ps = 0;
4158                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4159                                            0, 0, 0, 10, __LINE__, NULL, quality);
4160                 goto skip_measurement;
4161         }
4162         /*
4163          * Calculate the maximum possible b/w this connection
4164          * could have. We base our calculation on the lowest
4165          * rtt we have seen during the measurement and the
4166          * largest rwnd the client has given us in that time. This
4167          * forms a BDP that is the maximum that we could ever
4168          * get to the client. Anything larger is not valid.
4169          *
4170          * I originally had code here that rejected measurements
4171          * where the time was less than 1/2 the latest us_rtt.
4172          * But after thinking on that I realized its wrong since
4173          * say you had a 150Mbps or even 1Gbps link, and you
4174          * were a long way away.. example I am in Europe (100ms rtt)
4175          * talking to my 1Gbps link in S.C. Now measuring say 150,000
4176          * bytes my time would be 1.2ms, and yet my rtt would say
4177          * the measurement was invalid the time was < 50ms. The
4178          * same thing is true for 150Mb (8ms of time).
4179          *
4180          * A better way I realized is to look at what the maximum
4181          * the connection could possibly do. This is gated on
4182          * the lowest RTT we have seen and the highest rwnd.
4183          * We should in theory never exceed that, if we are
4184          * then something on the path is storing up packets
4185          * and then feeding them all at once to our endpoint
4186          * messing up our measurement.
4187          */
4188         rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4189         rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4190         rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4191         if (SEQ_LT(th_ack, tp->gput_seq)) {
4192                 /* No measurement can be made */
4193                 bytes = 0;
4194                 bytes_ps = 0;
4195                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4196                                            0, 0, 0, 10, __LINE__, NULL, quality);
4197                 goto skip_measurement;
4198         } else
4199                 bytes = (th_ack - tp->gput_seq);
4200         bytes_ps = (uint64_t)bytes;
4201         /*
4202          * Don't measure a b/w for pacing unless we have gotten at least
4203          * an initial windows worth of data in this measurement interval.
4204          *
4205          * Small numbers of bytes get badly influenced by delayed ack and
4206          * other artifacts. Note we take the initial window or our
4207          * defined minimum GP (defaulting to 10 which hopefully is the
4208          * IW).
4209          */
4210         if (rack->rc_gp_filled == 0) {
4211                 /*
4212                  * The initial estimate is special. We
4213                  * have blasted out an IW worth of packets
4214                  * without a real valid ack ts results. We
4215                  * then setup the app_limited_needs_set flag,
4216                  * this should get the first ack in (probably 2
4217                  * MSS worth) to be recorded as the timestamp.
4218                  * We thus allow a smaller number of bytes i.e.
4219                  * IW - 2MSS.
4220                  */
4221                 reqbytes -= (2 * segsiz);
4222                 /* Also lets fill previous for our first measurement to be neutral */
4223                 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4224         }
4225         if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4226                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4227                                            rack->r_ctl.rc_app_limited_cnt,
4228                                            0, 0, 10, __LINE__, NULL, quality);
4229                 goto skip_measurement;
4230         }
4231         /*
4232          * We now need to calculate the Timely like status so
4233          * we can update (possibly) the b/w multipliers.
4234          */
4235         new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4236         if (rack->rc_gp_filled == 0) {
4237                 /* No previous reading */
4238                 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4239         } else {
4240                 if (rack->measure_saw_probe_rtt == 0) {
4241                         /*
4242                          * We don't want a probertt to be counted
4243                          * since it will be negative incorrectly. We
4244                          * expect to be reducing the RTT when we
4245                          * pace at a slower rate.
4246                          */
4247                         rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4248                         rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4249                 }
4250         }
4251         timely_says = rack_make_timely_judgement(rack,
4252                 rack->r_ctl.rc_gp_srtt,
4253                 rack->r_ctl.rc_rtt_diff,
4254                 rack->r_ctl.rc_prev_gp_srtt
4255                 );
4256         bytes_ps *= HPTS_USEC_IN_SEC;
4257         bytes_ps /= utim;
4258         if (bytes_ps > rack->r_ctl.last_max_bw) {
4259                 /*
4260                  * Something is on path playing
4261                  * since this b/w is not possible based
4262                  * on our BDP (highest rwnd and lowest rtt
4263                  * we saw in the measurement window).
4264                  *
4265                  * Another option here would be to
4266                  * instead skip the measurement.
4267                  */
4268                 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4269                                            bytes_ps, rack->r_ctl.last_max_bw, 0,
4270                                            11, __LINE__, NULL, quality);
4271                 bytes_ps = rack->r_ctl.last_max_bw;
4272         }
4273         /* We store gp for b/w in bytes per second */
4274         if (rack->rc_gp_filled == 0) {
4275                 /* Initial measurement */
4276                 if (bytes_ps) {
4277                         rack->r_ctl.gp_bw = bytes_ps;
4278                         rack->rc_gp_filled = 1;
4279                         rack->r_ctl.num_measurements = 1;
4280                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4281                 } else {
4282                         rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4283                                                    rack->r_ctl.rc_app_limited_cnt,
4284                                                    0, 0, 10, __LINE__, NULL, quality);
4285                 }
4286                 if (tcp_in_hpts(rack->rc_inp) &&
4287                     (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4288                         /*
4289                          * Ok we can't trust the pacer in this case
4290                          * where we transition from un-paced to paced.
4291                          * Or for that matter when the burst mitigation
4292                          * was making a wild guess and got it wrong.
4293                          * Stop the pacer and clear up all the aggregate
4294                          * delays etc.
4295                          */
4296                         tcp_hpts_remove(rack->rc_inp);
4297                         rack->r_ctl.rc_hpts_flags = 0;
4298                         rack->r_ctl.rc_last_output_to = 0;
4299                 }
4300                 did_add = 2;
4301         } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4302                 /* Still a small number run an average */
4303                 rack->r_ctl.gp_bw += bytes_ps;
4304                 addpart = rack->r_ctl.num_measurements;
4305                 rack->r_ctl.num_measurements++;
4306                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4307                         /* We have collected enough to move forward */
4308                         rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4309                 }
4310                 did_add = 3;
4311         } else {
4312                 /*
4313                  * We want to take 1/wma of the goodput and add in to 7/8th
4314                  * of the old value weighted by the srtt. So if your measurement
4315                  * period is say 2 SRTT's long you would get 1/4 as the
4316                  * value, if it was like 1/2 SRTT then you would get 1/16th.
4317                  *
4318                  * But we must be careful not to take too much i.e. if the
4319                  * srtt is say 20ms and the measurement is taken over
4320                  * 400ms our weight would be 400/20 i.e. 20. On the
4321                  * other hand if we get a measurement over 1ms with a
4322                  * 10ms rtt we only want to take a much smaller portion.
4323                  */
4324                 if (rack->r_ctl.num_measurements < 0xff) {
4325                         rack->r_ctl.num_measurements++;
4326                 }
4327                 srtt = (uint64_t)tp->t_srtt;
4328                 if (srtt == 0) {
4329                         /*
4330                          * Strange why did t_srtt go back to zero?
4331                          */
4332                         if (rack->r_ctl.rc_rack_min_rtt)
4333                                 srtt = rack->r_ctl.rc_rack_min_rtt;
4334                         else
4335                                 srtt = HPTS_USEC_IN_MSEC;
4336                 }
4337                 /*
4338                  * XXXrrs: Note for reviewers, in playing with
4339                  * dynamic pacing I discovered this GP calculation
4340                  * as done originally leads to some undesired results.
4341                  * Basically you can get longer measurements contributing
4342                  * too much to the WMA. Thus I changed it if you are doing
4343                  * dynamic adjustments to only do the aportioned adjustment
4344                  * if we have a very small (time wise) measurement. Longer
4345                  * measurements just get there weight (defaulting to 1/8)
4346                  * add to the WMA. We may want to think about changing
4347                  * this to always do that for both sides i.e. dynamic
4348                  * and non-dynamic... but considering lots of folks
4349                  * were playing with this I did not want to change the
4350                  * calculation per.se. without your thoughts.. Lawerence?
4351                  * Peter??
4352                  */
4353                 if (rack->rc_gp_dyn_mul == 0) {
4354                         subpart = rack->r_ctl.gp_bw * utim;
4355                         subpart /= (srtt * 8);
4356                         if (subpart < (rack->r_ctl.gp_bw / 2)) {
4357                                 /*
4358                                  * The b/w update takes no more
4359                                  * away then 1/2 our running total
4360                                  * so factor it in.
4361                                  */
4362                                 addpart = bytes_ps * utim;
4363                                 addpart /= (srtt * 8);
4364                         } else {
4365                                 /*
4366                                  * Don't allow a single measurement
4367                                  * to account for more than 1/2 of the
4368                                  * WMA. This could happen on a retransmission
4369                                  * where utim becomes huge compared to
4370                                  * srtt (multiple retransmissions when using
4371                                  * the sending rate which factors in all the
4372                                  * transmissions from the first one).
4373                                  */
4374                                 subpart = rack->r_ctl.gp_bw / 2;
4375                                 addpart = bytes_ps / 2;
4376                         }
4377                         resid_bw = rack->r_ctl.gp_bw - subpart;
4378                         rack->r_ctl.gp_bw = resid_bw + addpart;
4379                         did_add = 1;
4380                 } else {
4381                         if ((utim / srtt) <= 1) {
4382                                 /*
4383                                  * The b/w update was over a small period
4384                                  * of time. The idea here is to prevent a small
4385                                  * measurement time period from counting
4386                                  * too much. So we scale it based on the
4387                                  * time so it attributes less than 1/rack_wma_divisor
4388                                  * of its measurement.
4389                                  */
4390                                 subpart = rack->r_ctl.gp_bw * utim;
4391                                 subpart /= (srtt * rack_wma_divisor);
4392                                 addpart = bytes_ps * utim;
4393                                 addpart /= (srtt * rack_wma_divisor);
4394                         } else {
4395                                 /*
4396                                  * The scaled measurement was long
4397                                  * enough so lets just add in the
4398                                  * portion of the measurement i.e. 1/rack_wma_divisor
4399                                  */
4400                                 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4401                                 addpart = bytes_ps / rack_wma_divisor;
4402                         }
4403                         if ((rack->measure_saw_probe_rtt == 0) ||
4404                             (bytes_ps > rack->r_ctl.gp_bw)) {
4405                                 /*
4406                                  * For probe-rtt we only add it in
4407                                  * if its larger, all others we just
4408                                  * add in.
4409                                  */
4410                                 did_add = 1;
4411                                 resid_bw = rack->r_ctl.gp_bw - subpart;
4412                                 rack->r_ctl.gp_bw = resid_bw + addpart;
4413                         }
4414                 }
4415         }
4416         if ((rack->gp_ready == 0) &&
4417             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4418                 /* We have enough measurements now */
4419                 rack->gp_ready = 1;
4420                 rack_set_cc_pacing(rack);
4421                 if (rack->defer_options)
4422                         rack_apply_deferred_options(rack);
4423         }
4424         rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4425                                    rack_get_bw(rack), 22, did_add, NULL, quality);
4426         /* We do not update any multipliers if we are in or have seen a probe-rtt */
4427         if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4428                 rack_update_multiplier(rack, timely_says, bytes_ps,
4429                                        rack->r_ctl.rc_gp_srtt,
4430                                        rack->r_ctl.rc_rtt_diff);
4431         rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4432                                    rack_get_bw(rack), 3, line, NULL, quality);
4433         /* reset the gp srtt and setup the new prev */
4434         rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4435         /* Record the lost count for the next measurement */
4436         rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4437         /*
4438          * We restart our diffs based on the gpsrtt in the
4439          * measurement window.
4440          */
4441         rack->rc_gp_rtt_set = 0;
4442         rack->rc_gp_saw_rec = 0;
4443         rack->rc_gp_saw_ca = 0;
4444         rack->rc_gp_saw_ss = 0;
4445         rack->rc_dragged_bottom = 0;
4446 skip_measurement:
4447
4448 #ifdef STATS
4449         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4450                                  gput);
4451         /*
4452          * XXXLAS: This is a temporary hack, and should be
4453          * chained off VOI_TCP_GPUT when stats(9) grows an
4454          * API to deal with chained VOIs.
4455          */
4456         if (tp->t_stats_gput_prev > 0)
4457                 stats_voi_update_abs_s32(tp->t_stats,
4458                                          VOI_TCP_GPUT_ND,
4459                                          ((gput - tp->t_stats_gput_prev) * 100) /
4460                                          tp->t_stats_gput_prev);
4461 #endif
4462         tp->t_flags &= ~TF_GPUTINPROG;
4463         tp->t_stats_gput_prev = gput;
4464         /*
4465          * Now are we app limited now and there is space from where we
4466          * were to where we want to go?
4467          *
4468          * We don't do the other case i.e. non-applimited here since
4469          * the next send will trigger us picking up the missing data.
4470          */
4471         if (rack->r_ctl.rc_first_appl &&
4472             TCPS_HAVEESTABLISHED(tp->t_state) &&
4473             rack->r_ctl.rc_app_limited_cnt &&
4474             (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4475             ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4476              max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4477                 /*
4478                  * Yep there is enough outstanding to make a measurement here.
4479                  */
4480                 struct rack_sendmap *rsm, fe;
4481
4482                 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4483                 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4484                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4485                 rack->app_limited_needs_set = 0;
4486                 tp->gput_seq = th_ack;
4487                 if (rack->in_probe_rtt)
4488                         rack->measure_saw_probe_rtt = 1;
4489                 else if ((rack->measure_saw_probe_rtt) &&
4490                          (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4491                         rack->measure_saw_probe_rtt = 0;
4492                 if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4493                         /* There is a full window to gain info from */
4494                         tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4495                 } else {
4496                         /* We can only measure up to the applimited point */
4497                         tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4498                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4499                                 /*
4500                                  * We don't have enough to make a measurement.
4501                                  */
4502                                 tp->t_flags &= ~TF_GPUTINPROG;
4503                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4504                                                            0, 0, 0, 6, __LINE__, NULL, quality);
4505                                 return;
4506                         }
4507                 }
4508                 if (tp->t_state >= TCPS_FIN_WAIT_1) {
4509                         /*
4510                          * We will get no more data into the SB
4511                          * this means we need to have the data available
4512                          * before we start a measurement.
4513                          */
4514                         if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4515                                 /* Nope not enough data. */
4516                                 return;
4517                         }
4518                 }
4519                 tp->t_flags |= TF_GPUTINPROG;
4520                 /*
4521                  * Now we need to find the timestamp of the send at tp->gput_seq
4522                  * for the send based measurement.
4523                  */
4524                 fe.r_start = tp->gput_seq;
4525                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4526                 if (rsm) {
4527                         /* Ok send-based limit is set */
4528                         if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4529                                 /*
4530                                  * Move back to include the earlier part
4531                                  * so our ack time lines up right (this may
4532                                  * make an overlapping measurement but thats
4533                                  * ok).
4534                                  */
4535                                 tp->gput_seq = rsm->r_start;
4536                         }
4537                         if (rsm->r_flags & RACK_ACKED)
4538                                 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4539                         else
4540                                 rack->app_limited_needs_set = 1;
4541                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4542                 } else {
4543                         /*
4544                          * If we don't find the rsm due to some
4545                          * send-limit set the current time, which
4546                          * basically disables the send-limit.
4547                          */
4548                         struct timeval tv;
4549
4550                         microuptime(&tv);
4551                         rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4552                 }
4553                 rack_log_pacing_delay_calc(rack,
4554                                            tp->gput_seq,
4555                                            tp->gput_ack,
4556                                            (uint64_t)rsm,
4557                                            tp->gput_ts,
4558                                            rack->r_ctl.rc_app_limited_cnt,
4559                                            9,
4560                                            __LINE__, NULL, quality);
4561         }
4562 }
4563
4564 /*
4565  * CC wrapper hook functions
4566  */
4567 static void
4568 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4569     uint16_t type, int32_t recovery)
4570 {
4571         uint32_t prior_cwnd, acked;
4572         struct tcp_log_buffer *lgb = NULL;
4573         uint8_t labc_to_use, quality;
4574
4575         INP_WLOCK_ASSERT(tptoinpcb(tp));
4576         tp->t_ccv.nsegs = nsegs;
4577         acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
4578         if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4579                 uint32_t max;
4580
4581                 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4582                 if (tp->t_ccv.bytes_this_ack > max) {
4583                         tp->t_ccv.bytes_this_ack = max;
4584                 }
4585         }
4586 #ifdef STATS
4587         stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4588             ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4589 #endif
4590         quality = RACK_QUALITY_NONE;
4591         if ((tp->t_flags & TF_GPUTINPROG) &&
4592             rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4593                 /* Measure the Goodput */
4594                 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4595 #ifdef NETFLIX_PEAKRATE
4596                 if ((type == CC_ACK) &&
4597                     (tp->t_maxpeakrate)) {
4598                         /*
4599                          * We update t_peakrate_thr. This gives us roughly
4600                          * one update per round trip time. Note
4601                          * it will only be used if pace_always is off i.e
4602                          * we don't do this for paced flows.
4603                          */
4604                         rack_update_peakrate_thr(tp);
4605                 }
4606 #endif
4607         }
4608         /* Which way our we limited, if not cwnd limited no advance in CA */
4609         if (tp->snd_cwnd <= tp->snd_wnd)
4610                 tp->t_ccv.flags |= CCF_CWND_LIMITED;
4611         else
4612                 tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
4613         if (tp->snd_cwnd > tp->snd_ssthresh) {
4614                 tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
4615                          nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4616                 /* For the setting of a window past use the actual scwnd we are using */
4617                 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4618                         tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4619                         tp->t_ccv.flags |= CCF_ABC_SENTAWND;
4620                 }
4621         } else {
4622                 tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4623                 tp->t_bytes_acked = 0;
4624         }
4625         prior_cwnd = tp->snd_cwnd;
4626         if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4627             (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4628                 labc_to_use = rack->rc_labc;
4629         else
4630                 labc_to_use = rack_max_abc_post_recovery;
4631         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4632                 union tcp_log_stackspecific log;
4633                 struct timeval tv;
4634
4635                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4636                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4637                 log.u_bbr.flex1 = th_ack;
4638                 log.u_bbr.flex2 = tp->t_ccv.flags;
4639                 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
4640                 log.u_bbr.flex4 = tp->t_ccv.nsegs;
4641                 log.u_bbr.flex5 = labc_to_use;
4642                 log.u_bbr.flex6 = prior_cwnd;
4643                 log.u_bbr.flex7 = V_tcp_do_newsack;
4644                 log.u_bbr.flex8 = 1;
4645                 lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4646                                      0, &log, false, NULL, NULL, 0, &tv);
4647         }
4648         if (CC_ALGO(tp)->ack_received != NULL) {
4649                 /* XXXLAS: Find a way to live without this */
4650                 tp->t_ccv.curack = th_ack;
4651                 tp->t_ccv.labc = labc_to_use;
4652                 tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
4653                 CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
4654         }
4655         if (lgb) {
4656                 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4657         }
4658         if (rack->r_must_retran) {
4659                 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4660                         /*
4661                          * We now are beyond the rxt point so lets disable
4662                          * the flag.
4663                          */
4664                         rack->r_ctl.rc_out_at_rto = 0;
4665                         rack->r_must_retran = 0;
4666                 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4667                         /*
4668                          * Only decrement the rc_out_at_rto if the cwnd advances
4669                          * at least a whole segment. Otherwise next time the peer
4670                          * acks, we won't be able to send this generaly happens
4671                          * when we are in Congestion Avoidance.
4672                          */
4673                         if (acked <= rack->r_ctl.rc_out_at_rto){
4674                                 rack->r_ctl.rc_out_at_rto -= acked;
4675                         } else {
4676                                 rack->r_ctl.rc_out_at_rto = 0;
4677                         }
4678                 }
4679         }
4680 #ifdef STATS
4681         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4682 #endif
4683         if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4684                 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4685         }
4686 #ifdef NETFLIX_PEAKRATE
4687         /* we enforce max peak rate if it is set and we are not pacing */
4688         if ((rack->rc_always_pace == 0) &&
4689             tp->t_peakrate_thr &&
4690             (tp->snd_cwnd > tp->t_peakrate_thr)) {
4691                 tp->snd_cwnd = tp->t_peakrate_thr;
4692         }
4693 #endif
4694 }
4695
4696 static void
4697 tcp_rack_partialack(struct tcpcb *tp)
4698 {
4699         struct tcp_rack *rack;
4700
4701         rack = (struct tcp_rack *)tp->t_fb_ptr;
4702         INP_WLOCK_ASSERT(tptoinpcb(tp));
4703         /*
4704          * If we are doing PRR and have enough
4705          * room to send <or> we are pacing and prr
4706          * is disabled we will want to see if we
4707          * can send data (by setting r_wanted_output to
4708          * true).
4709          */
4710         if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4711             rack->rack_no_prr)
4712                 rack->r_wanted_output = 1;
4713 }
4714
4715 static void
4716 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4717 {
4718         struct tcp_rack *rack;
4719         uint32_t orig_cwnd;
4720
4721         orig_cwnd = tp->snd_cwnd;
4722         INP_WLOCK_ASSERT(tptoinpcb(tp));
4723         rack = (struct tcp_rack *)tp->t_fb_ptr;
4724         /* only alert CC if we alerted when we entered */
4725         if (CC_ALGO(tp)->post_recovery != NULL) {
4726                 tp->t_ccv.curack = th_ack;
4727                 CC_ALGO(tp)->post_recovery(&tp->t_ccv);
4728                 if (tp->snd_cwnd < tp->snd_ssthresh) {
4729                         /*
4730                          * Rack has burst control and pacing
4731                          * so lets not set this any lower than
4732                          * snd_ssthresh per RFC-6582 (option 2).
4733                          */
4734                         tp->snd_cwnd = tp->snd_ssthresh;
4735                 }
4736         }
4737         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4738                 union tcp_log_stackspecific log;
4739                 struct timeval tv;
4740
4741                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4742                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4743                 log.u_bbr.flex1 = th_ack;
4744                 log.u_bbr.flex2 = tp->t_ccv.flags;
4745                 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
4746                 log.u_bbr.flex4 = tp->t_ccv.nsegs;
4747                 log.u_bbr.flex5 = V_tcp_abc_l_var;
4748                 log.u_bbr.flex6 = orig_cwnd;
4749                 log.u_bbr.flex7 = V_tcp_do_newsack;
4750                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4751                 log.u_bbr.flex8 = 2;
4752                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4753                                0, &log, false, NULL, NULL, 0, &tv);
4754         }
4755         if ((rack->rack_no_prr == 0) &&
4756             (rack->no_prr_addback == 0) &&
4757             (rack->r_ctl.rc_prr_sndcnt > 0)) {
4758                 /*
4759                  * Suck the next prr cnt back into cwnd, but
4760                  * only do that if we are not application limited.
4761                  */
4762                 if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
4763                         /*
4764                          * We are allowed to add back to the cwnd the amount we did
4765                          * not get out if:
4766                          * a) no_prr_addback is off.
4767                          * b) we are not app limited
4768                          * c) we are doing prr
4769                          * <and>
4770                          * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4771                          */
4772                         tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4773                                             rack->r_ctl.rc_prr_sndcnt);
4774                 }
4775                 rack->r_ctl.rc_prr_sndcnt = 0;
4776                 rack_log_to_prr(rack, 1, 0, __LINE__);
4777         }
4778         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
4779         tp->snd_recover = tp->snd_una;
4780         if (rack->r_ctl.dsack_persist) {
4781                 rack->r_ctl.dsack_persist--;
4782                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4783                         rack->r_ctl.num_dsack = 0;
4784                 }
4785                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4786         }
4787         EXIT_RECOVERY(tp->t_flags);
4788 }
4789
4790 static void
4791 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
4792 {
4793         struct tcp_rack *rack;
4794         uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4795
4796         INP_WLOCK_ASSERT(tptoinpcb(tp));
4797 #ifdef STATS
4798         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4799 #endif
4800         if (IN_RECOVERY(tp->t_flags) == 0) {
4801                 in_rec_at_entry = 0;
4802                 ssthresh_enter = tp->snd_ssthresh;
4803                 cwnd_enter = tp->snd_cwnd;
4804         } else
4805                 in_rec_at_entry = 1;
4806         rack = (struct tcp_rack *)tp->t_fb_ptr;
4807         switch (type) {
4808         case CC_NDUPACK:
4809                 tp->t_flags &= ~TF_WASFRECOVERY;
4810                 tp->t_flags &= ~TF_WASCRECOVERY;
4811                 if (!IN_FASTRECOVERY(tp->t_flags)) {
4812                         rack->r_ctl.rc_prr_delivered = 0;
4813                         rack->r_ctl.rc_prr_out = 0;
4814                         if (rack->rack_no_prr == 0) {
4815                                 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4816                                 rack_log_to_prr(rack, 2, in_rec_at_entry, line);
4817                         }
4818                         rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4819                         tp->snd_recover = tp->snd_max;
4820                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4821                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4822                 }
4823                 break;
4824         case CC_ECN:
4825                 if (!IN_CONGRECOVERY(tp->t_flags) ||
4826                     /*
4827                      * Allow ECN reaction on ACK to CWR, if
4828                      * that data segment was also CE marked.
4829                      */
4830                     SEQ_GEQ(ack, tp->snd_recover)) {
4831                         EXIT_CONGRECOVERY(tp->t_flags);
4832                         KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4833                         tp->snd_recover = tp->snd_max + 1;
4834                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4835                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4836                 }
4837                 break;
4838         case CC_RTO:
4839                 tp->t_dupacks = 0;
4840                 tp->t_bytes_acked = 0;
4841                 EXIT_RECOVERY(tp->t_flags);
4842                 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4843                     ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4844                 orig_cwnd = tp->snd_cwnd;
4845                 tp->snd_cwnd = ctf_fixed_maxseg(tp);
4846                 rack_log_to_prr(rack, 16, orig_cwnd, line);
4847                 if (tp->t_flags2 & TF2_ECN_PERMIT)
4848                         tp->t_flags2 |= TF2_ECN_SND_CWR;
4849                 break;
4850         case CC_RTO_ERR:
4851                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4852                 /* RTO was unnecessary, so reset everything. */
4853                 tp->snd_cwnd = tp->snd_cwnd_prev;
4854                 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4855                 tp->snd_recover = tp->snd_recover_prev;
4856                 if (tp->t_flags & TF_WASFRECOVERY) {
4857                         ENTER_FASTRECOVERY(tp->t_flags);
4858                         tp->t_flags &= ~TF_WASFRECOVERY;
4859                 }
4860                 if (tp->t_flags & TF_WASCRECOVERY) {
4861                         ENTER_CONGRECOVERY(tp->t_flags);
4862                         tp->t_flags &= ~TF_WASCRECOVERY;
4863                 }
4864                 tp->snd_nxt = tp->snd_max;
4865                 tp->t_badrxtwin = 0;
4866                 break;
4867         }
4868         if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4869             (type != CC_RTO)){
4870                 tp->t_ccv.curack = ack;
4871                 CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
4872         }
4873         if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4874                 rack_log_to_prr(rack, 15, cwnd_enter, line);
4875                 rack->r_ctl.dsack_byte_cnt = 0;
4876                 rack->r_ctl.retran_during_recovery = 0;
4877                 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4878                 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4879                 rack->r_ent_rec_ns = 1;
4880         }
4881 }
4882
4883 static inline void
4884 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4885 {
4886         uint32_t i_cwnd;
4887
4888         INP_WLOCK_ASSERT(tptoinpcb(tp));
4889
4890 #ifdef NETFLIX_STATS
4891         KMOD_TCPSTAT_INC(tcps_idle_restarts);
4892         if (tp->t_state == TCPS_ESTABLISHED)
4893                 KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4894 #endif
4895         if (CC_ALGO(tp)->after_idle != NULL)
4896                 CC_ALGO(tp)->after_idle(&tp->t_ccv);
4897
4898         if (tp->snd_cwnd == 1)
4899                 i_cwnd = tp->t_maxseg;          /* SYN(-ACK) lost */
4900         else
4901                 i_cwnd = rc_init_window(rack);
4902
4903         /*
4904          * Being idle is no different than the initial window. If the cc
4905          * clamps it down below the initial window raise it to the initial
4906          * window.
4907          */
4908         if (tp->snd_cwnd < i_cwnd) {
4909                 tp->snd_cwnd = i_cwnd;
4910         }
4911 }
4912
4913 /*
4914  * Indicate whether this ack should be delayed.  We can delay the ack if
4915  * following conditions are met:
4916  *      - There is no delayed ack timer in progress.
4917  *      - Our last ack wasn't a 0-sized window. We never want to delay
4918  *        the ack that opens up a 0-sized window.
4919  *      - LRO wasn't used for this segment. We make sure by checking that the
4920  *        segment size is not larger than the MSS.
4921  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4922  *        connection.
4923  */
4924 #define DELAY_ACK(tp, tlen)                      \
4925         (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4926         ((tp->t_flags & TF_DELACK) == 0) &&      \
4927         (tlen <= tp->t_maxseg) &&                \
4928         (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4929
4930 static struct rack_sendmap *
4931 rack_find_lowest_rsm(struct tcp_rack *rack)
4932 {
4933         struct rack_sendmap *rsm;
4934
4935         /*
4936          * Walk the time-order transmitted list looking for an rsm that is
4937          * not acked. This will be the one that was sent the longest time
4938          * ago that is still outstanding.
4939          */
4940         TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4941                 if (rsm->r_flags & RACK_ACKED) {
4942                         continue;
4943                 }
4944                 goto finish;
4945         }
4946 finish:
4947         return (rsm);
4948 }
4949
4950 static struct rack_sendmap *
4951 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4952 {
4953         struct rack_sendmap *prsm;
4954
4955         /*
4956          * Walk the sequence order list backward until we hit and arrive at
4957          * the highest seq not acked. In theory when this is called it
4958          * should be the last segment (which it was not).
4959          */
4960         prsm = rsm;
4961         RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4962                 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4963                         continue;
4964                 }
4965                 return (prsm);
4966         }
4967         return (NULL);
4968 }
4969
4970 static uint32_t
4971 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4972 {
4973         int32_t lro;
4974         uint32_t thresh;
4975
4976         /*
4977          * lro is the flag we use to determine if we have seen reordering.
4978          * If it gets set we have seen reordering. The reorder logic either
4979          * works in one of two ways:
4980          *
4981          * If reorder-fade is configured, then we track the last time we saw
4982          * re-ordering occur. If we reach the point where enough time as
4983          * passed we no longer consider reordering has occuring.
4984          *
4985          * Or if reorder-face is 0, then once we see reordering we consider
4986          * the connection to alway be subject to reordering and just set lro
4987          * to 1.
4988          *
4989          * In the end if lro is non-zero we add the extra time for
4990          * reordering in.
4991          */
4992         if (srtt == 0)
4993                 srtt = 1;
4994         if (rack->r_ctl.rc_reorder_ts) {
4995                 if (rack->r_ctl.rc_reorder_fade) {
4996                         if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
4997                                 lro = cts - rack->r_ctl.rc_reorder_ts;
4998                                 if (lro == 0) {
4999                                         /*
5000                                          * No time as passed since the last
5001                                          * reorder, mark it as reordering.
5002                                          */
5003                                         lro = 1;
5004                                 }
5005                         } else {
5006                                 /* Negative time? */
5007                                 lro = 0;
5008                         }
5009                         if (lro > rack->r_ctl.rc_reorder_fade) {
5010                                 /* Turn off reordering seen too */
5011                                 rack->r_ctl.rc_reorder_ts = 0;
5012                                 lro = 0;
5013                         }
5014                 } else {
5015                         /* Reodering does not fade */
5016                         lro = 1;
5017                 }
5018         } else {
5019                 lro = 0;
5020         }
5021         if (rack->rc_rack_tmr_std_based == 0) {
5022                 thresh = srtt + rack->r_ctl.rc_pkt_delay;
5023         } else {
5024                 /* Standards based pkt-delay is 1/4 srtt */
5025                 thresh = srtt +  (srtt >> 2);
5026         }
5027         if (lro && (rack->rc_rack_tmr_std_based == 0)) {
5028                 /* It must be set, if not you get 1/4 rtt */
5029                 if (rack->r_ctl.rc_reorder_shift)
5030                         thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5031                 else
5032                         thresh += (srtt >> 2);
5033         }
5034         if (rack->rc_rack_use_dsack &&
5035             lro &&
5036             (rack->r_ctl.num_dsack > 0)) {
5037                 /*
5038                  * We only increase the reordering window if we
5039                  * have seen reordering <and> we have a DSACK count.
5040                  */
5041                 thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5042                 rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
5043         }
5044         /* SRTT * 2 is the ceiling */
5045         if (thresh > (srtt * 2)) {
5046                 thresh = srtt * 2;
5047         }
5048         /* And we don't want it above the RTO max either */
5049         if (thresh > rack_rto_max) {
5050                 thresh = rack_rto_max;
5051         }
5052         rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
5053         return (thresh);
5054 }
5055
5056 static uint32_t
5057 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5058                      struct rack_sendmap *rsm, uint32_t srtt)
5059 {
5060         struct rack_sendmap *prsm;
5061         uint32_t thresh, len;
5062         int segsiz;
5063
5064         if (srtt == 0)
5065                 srtt = 1;
5066         if (rack->r_ctl.rc_tlp_threshold)
5067                 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5068         else
5069                 thresh = (srtt * 2);
5070
5071         /* Get the previous sent packet, if any */
5072         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5073         len = rsm->r_end - rsm->r_start;
5074         if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5075                 /* Exactly like the ID */
5076                 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5077                         uint32_t alt_thresh;
5078                         /*
5079                          * Compensate for delayed-ack with the d-ack time.
5080                          */
5081                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5082                         if (alt_thresh > thresh)
5083                                 thresh = alt_thresh;
5084                 }
5085         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5086                 /* 2.1 behavior */
5087                 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5088                 if (prsm && (len <= segsiz)) {
5089                         /*
5090                          * Two packets outstanding, thresh should be (2*srtt) +
5091                          * possible inter-packet delay (if any).
5092                          */
5093                         uint32_t inter_gap = 0;
5094                         int idx, nidx;
5095
5096                         idx = rsm->r_rtr_cnt - 1;
5097                         nidx = prsm->r_rtr_cnt - 1;
5098                         if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5099                                 /* Yes it was sent later (or at the same time) */
5100                                 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5101                         }
5102                         thresh += inter_gap;
5103                 } else if (len <= segsiz) {
5104                         /*
5105                          * Possibly compensate for delayed-ack.
5106                          */
5107                         uint32_t alt_thresh;
5108
5109                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5110                         if (alt_thresh > thresh)
5111                                 thresh = alt_thresh;
5112                 }
5113         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5114                 /* 2.2 behavior */
5115                 if (len <= segsiz) {
5116                         uint32_t alt_thresh;
5117                         /*
5118                          * Compensate for delayed-ack with the d-ack time.
5119                          */
5120                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5121                         if (alt_thresh > thresh)
5122                                 thresh = alt_thresh;
5123                 }
5124         }
5125         /* Not above an RTO */
5126         if (thresh > tp->t_rxtcur) {
5127                 thresh = tp->t_rxtcur;
5128         }
5129         /* Not above a RTO max */
5130         if (thresh > rack_rto_max) {
5131                 thresh = rack_rto_max;
5132         }
5133         /* Apply user supplied min TLP */
5134         if (thresh < rack_tlp_min) {
5135                 thresh = rack_tlp_min;
5136         }
5137         return (thresh);
5138 }
5139
5140 static uint32_t
5141 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5142 {
5143         /*
5144          * We want the rack_rtt which is the
5145          * last rtt we measured. However if that
5146          * does not exist we fallback to the srtt (which
5147          * we probably will never do) and then as a last
5148          * resort we use RACK_INITIAL_RTO if no srtt is
5149          * yet set.
5150          */
5151         if (rack->rc_rack_rtt)
5152                 return (rack->rc_rack_rtt);
5153         else if (tp->t_srtt == 0)
5154                 return (RACK_INITIAL_RTO);
5155         return (tp->t_srtt);
5156 }
5157
5158 static struct rack_sendmap *
5159 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5160 {
5161         /*
5162          * Check to see that we don't need to fall into recovery. We will
5163          * need to do so if our oldest transmit is past the time we should
5164          * have had an ack.
5165          */
5166         struct tcp_rack *rack;
5167         struct rack_sendmap *rsm;
5168         int32_t idx;
5169         uint32_t srtt, thresh;
5170
5171         rack = (struct tcp_rack *)tp->t_fb_ptr;
5172         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5173                 return (NULL);
5174         }
5175         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5176         if (rsm == NULL)
5177                 return (NULL);
5178
5179
5180         if (rsm->r_flags & RACK_ACKED) {
5181                 rsm = rack_find_lowest_rsm(rack);
5182                 if (rsm == NULL)
5183                         return (NULL);
5184         }
5185         idx = rsm->r_rtr_cnt - 1;
5186         srtt = rack_grab_rtt(tp, rack);
5187         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5188         if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5189                 return (NULL);
5190         }
5191         if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5192                 return (NULL);
5193         }
5194         /* Ok if we reach here we are over-due and this guy can be sent */
5195         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
5196         return (rsm);
5197 }
5198
5199 static uint32_t
5200 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5201 {
5202         int32_t t;
5203         int32_t tt;
5204         uint32_t ret_val;
5205
5206         t = (tp->t_srtt + (tp->t_rttvar << 2));
5207         RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5208             rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5209         rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5210         ret_val = (uint32_t)tt;
5211         return (ret_val);
5212 }
5213
5214 static uint32_t
5215 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5216 {
5217         /*
5218          * Start the FR timer, we do this based on getting the first one in
5219          * the rc_tmap. Note that if its NULL we must stop the timer. in all
5220          * events we need to stop the running timer (if its running) before
5221          * starting the new one.
5222          */
5223         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5224         uint32_t srtt_cur;
5225         int32_t idx;
5226         int32_t is_tlp_timer = 0;
5227         struct rack_sendmap *rsm;
5228
5229         if (rack->t_timers_stopped) {
5230                 /* All timers have been stopped none are to run */
5231                 return (0);
5232         }
5233         if (rack->rc_in_persist) {
5234                 /* We can't start any timer in persists */
5235                 return (rack_get_persists_timer_val(tp, rack));
5236         }
5237         rack->rc_on_min_to = 0;
5238         if ((tp->t_state < TCPS_ESTABLISHED) ||
5239             ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5240                 goto activate_rxt;
5241         }
5242         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5243         if ((rsm == NULL) || sup_rack) {
5244                 /* Nothing on the send map or no rack */
5245 activate_rxt:
5246                 time_since_sent = 0;
5247                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5248                 if (rsm) {
5249                         /*
5250                          * Should we discount the RTX timer any?
5251                          *
5252                          * We want to discount it the smallest amount.
5253                          * If a timer (Rack/TLP or RXT) has gone off more
5254                          * recently thats the discount we want to use (now - timer time).
5255                          * If the retransmit of the oldest packet was more recent then
5256                          * we want to use that (now - oldest-packet-last_transmit_time).
5257                          *
5258                          */
5259                         idx = rsm->r_rtr_cnt - 1;
5260                         if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5261                                 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5262                         else
5263                                 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5264                         if (TSTMP_GT(cts, tstmp_touse))
5265                             time_since_sent = cts - tstmp_touse;
5266                 }
5267                 if (SEQ_LT(tp->snd_una, tp->snd_max) ||
5268                     sbavail(&tptosocket(tp)->so_snd)) {
5269                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5270                         to = tp->t_rxtcur;
5271                         if (to > time_since_sent)
5272                                 to -= time_since_sent;
5273                         else
5274                                 to = rack->r_ctl.rc_min_to;
5275                         if (to == 0)
5276                                 to = 1;
5277                         /* Special case for KEEPINIT */
5278                         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5279                             (TP_KEEPINIT(tp) != 0) &&
5280                             rsm) {
5281                                 /*
5282                                  * We have to put a ceiling on the rxt timer
5283                                  * of the keep-init timeout.
5284                                  */
5285                                 uint32_t max_time, red;
5286
5287                                 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5288                                 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5289                                         red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5290                                         if (red < max_time)
5291                                                 max_time -= red;
5292                                         else
5293                                                 max_time = 1;
5294                                 }
5295                                 /* Reduce timeout to the keep value if needed */
5296                                 if (max_time < to)
5297                                         to = max_time;
5298                         }
5299                         return (to);
5300                 }
5301                 return (0);
5302         }
5303         if (rsm->r_flags & RACK_ACKED) {
5304                 rsm = rack_find_lowest_rsm(rack);
5305                 if (rsm == NULL) {
5306                         /* No lowest? */
5307                         goto activate_rxt;
5308                 }
5309         }
5310         if (rack->sack_attack_disable) {
5311                 /*
5312                  * We don't want to do
5313                  * any TLP's if you are an attacker.
5314                  * Though if you are doing what
5315                  * is expected you may still have
5316                  * SACK-PASSED marks.
5317                  */
5318                 goto activate_rxt;
5319         }
5320         /* Convert from ms to usecs */
5321         if ((rsm->r_flags & RACK_SACK_PASSED) ||
5322             (rsm->r_flags & RACK_RWND_COLLAPSED) ||
5323             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5324                 if ((tp->t_flags & TF_SENTFIN) &&
5325                     ((tp->snd_max - tp->snd_una) == 1) &&
5326                     (rsm->r_flags & RACK_HAS_FIN)) {
5327                         /*
5328                          * We don't start a rack timer if all we have is a
5329                          * FIN outstanding.
5330                          */
5331                         goto activate_rxt;
5332                 }
5333                 if ((rack->use_rack_rr == 0) &&
5334                     (IN_FASTRECOVERY(tp->t_flags)) &&
5335                     (rack->rack_no_prr == 0) &&
5336                      (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5337                         /*
5338                          * We are not cheating, in recovery  and
5339                          * not enough ack's to yet get our next
5340                          * retransmission out.
5341                          *
5342                          * Note that classified attackers do not
5343                          * get to use the rack-cheat.
5344                          */
5345                         goto activate_tlp;
5346                 }
5347                 srtt = rack_grab_rtt(tp, rack);
5348                 thresh = rack_calc_thresh_rack(rack, srtt, cts);
5349                 idx = rsm->r_rtr_cnt - 1;
5350                 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5351                 if (SEQ_GEQ(exp, cts)) {
5352                         to = exp - cts;
5353                         if (to < rack->r_ctl.rc_min_to) {
5354                                 to = rack->r_ctl.rc_min_to;
5355                                 if (rack->r_rr_config == 3)
5356                                         rack->rc_on_min_to = 1;
5357                         }
5358                 } else {
5359                         to = rack->r_ctl.rc_min_to;
5360                         if (rack->r_rr_config == 3)
5361                                 rack->rc_on_min_to = 1;
5362                 }
5363         } else {
5364                 /* Ok we need to do a TLP not RACK */
5365 activate_tlp:
5366                 if ((rack->rc_tlp_in_progress != 0) &&
5367                     (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5368                         /*
5369                          * The previous send was a TLP and we have sent
5370                          * N TLP's without sending new data.
5371                          */
5372                         goto activate_rxt;
5373                 }
5374                 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5375                 if (rsm == NULL) {
5376                         /* We found no rsm to TLP with. */
5377                         goto activate_rxt;
5378                 }
5379                 if (rsm->r_flags & RACK_HAS_FIN) {
5380                         /* If its a FIN we dont do TLP */
5381                         rsm = NULL;
5382                         goto activate_rxt;
5383                 }
5384                 idx = rsm->r_rtr_cnt - 1;
5385                 time_since_sent = 0;
5386                 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5387                         tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5388                 else
5389                         tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5390                 if (TSTMP_GT(cts, tstmp_touse))
5391                     time_since_sent = cts - tstmp_touse;
5392                 is_tlp_timer = 1;
5393                 if (tp->t_srtt) {
5394                         if ((rack->rc_srtt_measure_made == 0) &&
5395                             (tp->t_srtt == 1)) {
5396                                 /*
5397                                  * If another stack as run and set srtt to 1,
5398                                  * then the srtt was 0, so lets use the initial.
5399                                  */
5400                                 srtt = RACK_INITIAL_RTO;
5401                         } else {
5402                                 srtt_cur = tp->t_srtt;
5403                                 srtt = srtt_cur;
5404                         }
5405                 } else
5406                         srtt = RACK_INITIAL_RTO;
5407                 /*
5408                  * If the SRTT is not keeping up and the
5409                  * rack RTT has spiked we want to use
5410                  * the last RTT not the smoothed one.
5411                  */
5412                 if (rack_tlp_use_greater &&
5413                     tp->t_srtt &&
5414                     (srtt < rack_grab_rtt(tp, rack))) {
5415                         srtt = rack_grab_rtt(tp, rack);
5416                 }
5417                 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5418                 if (thresh > time_since_sent) {
5419                         to = thresh - time_since_sent;
5420                 } else {
5421                         to = rack->r_ctl.rc_min_to;
5422                         rack_log_alt_to_to_cancel(rack,
5423                                                   thresh,               /* flex1 */
5424                                                   time_since_sent,      /* flex2 */
5425                                                   tstmp_touse,          /* flex3 */
5426                                                   rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5427                                                   (uint32_t)rsm->r_tim_lastsent[idx],
5428                                                   srtt,
5429                                                   idx, 99);
5430                 }
5431                 if (to < rack_tlp_min) {
5432                         to = rack_tlp_min;
5433                 }
5434                 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5435                         /*
5436                          * If the TLP time works out to larger than the max
5437                          * RTO lets not do TLP.. just RTO.
5438                          */
5439                         goto activate_rxt;
5440                 }
5441         }
5442         if (is_tlp_timer == 0) {
5443                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5444         } else {
5445                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5446         }
5447         if (to == 0)
5448                 to = 1;
5449         return (to);
5450 }
5451
5452 static void
5453 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5454 {
5455         if (rack->rc_in_persist == 0) {
5456                 if (tp->t_flags & TF_GPUTINPROG) {
5457                         /*
5458                          * Stop the goodput now, the calling of the
5459                          * measurement function clears the flag.
5460                          */
5461                         rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5462                                                     RACK_QUALITY_PERSIST);
5463                 }
5464 #ifdef NETFLIX_SHARED_CWND
5465                 if (rack->r_ctl.rc_scw) {
5466                         tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5467                         rack->rack_scwnd_is_idle = 1;
5468                 }
5469 #endif
5470                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5471                 if (rack->r_ctl.rc_went_idle_time == 0)
5472                         rack->r_ctl.rc_went_idle_time = 1;
5473                 rack_timer_cancel(tp, rack, cts, __LINE__);
5474                 rack->r_ctl.persist_lost_ends = 0;
5475                 rack->probe_not_answered = 0;
5476                 rack->forced_ack = 0;
5477                 tp->t_rxtshift = 0;
5478                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5479                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5480                 rack->rc_in_persist = 1;
5481         }
5482 }
5483
5484 static void
5485 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5486 {
5487         if (tcp_in_hpts(rack->rc_inp)) {
5488                 tcp_hpts_remove(rack->rc_inp);
5489                 rack->r_ctl.rc_hpts_flags = 0;
5490         }
5491 #ifdef NETFLIX_SHARED_CWND
5492         if (rack->r_ctl.rc_scw) {
5493                 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5494                 rack->rack_scwnd_is_idle = 0;
5495         }
5496 #endif
5497         if (rack->rc_gp_dyn_mul &&
5498             (rack->use_fixed_rate == 0) &&
5499             (rack->rc_always_pace)) {
5500                 /*
5501                  * Do we count this as if a probe-rtt just
5502                  * finished?
5503                  */
5504                 uint32_t time_idle, idle_min;
5505
5506                 time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5507                 idle_min = rack_min_probertt_hold;
5508                 if (rack_probertt_gpsrtt_cnt_div) {
5509                         uint64_t extra;
5510                         extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5511                                 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
5512                         extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5513                         idle_min += (uint32_t)extra;
5514                 }
5515                 if (time_idle >= idle_min) {
5516                         /* Yes, we count it as a probe-rtt. */
5517                         uint32_t us_cts;
5518
5519                         us_cts = tcp_get_usecs(NULL);
5520                         if (rack->in_probe_rtt == 0) {
5521                                 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5522                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5523                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5524                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5525                         } else {
5526                                 rack_exit_probertt(rack, us_cts);
5527                         }
5528                 }
5529         }
5530         rack->rc_in_persist = 0;
5531         rack->r_ctl.rc_went_idle_time = 0;
5532         tp->t_rxtshift = 0;
5533         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5534            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5535         rack->r_ctl.rc_agg_delayed = 0;
5536         rack->r_early = 0;
5537         rack->r_late = 0;
5538         rack->r_ctl.rc_agg_early = 0;
5539 }
5540
5541 static void
5542 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5543                    struct hpts_diag *diag, struct timeval *tv)
5544 {
5545         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5546                 union tcp_log_stackspecific log;
5547
5548                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5549                 log.u_bbr.flex1 = diag->p_nxt_slot;
5550                 log.u_bbr.flex2 = diag->p_cur_slot;
5551                 log.u_bbr.flex3 = diag->slot_req;
5552                 log.u_bbr.flex4 = diag->inp_hptsslot;
5553                 log.u_bbr.flex5 = diag->slot_remaining;
5554                 log.u_bbr.flex6 = diag->need_new_to;
5555                 log.u_bbr.flex7 = diag->p_hpts_active;
5556                 log.u_bbr.flex8 = diag->p_on_min_sleep;
5557                 /* Hijack other fields as needed */
5558                 log.u_bbr.epoch = diag->have_slept;
5559                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
5560                 log.u_bbr.pkts_out = diag->co_ret;
5561                 log.u_bbr.applimited = diag->hpts_sleep_time;
5562                 log.u_bbr.delivered = diag->p_prev_slot;
5563                 log.u_bbr.inflight = diag->p_runningslot;
5564                 log.u_bbr.bw_inuse = diag->wheel_slot;
5565                 log.u_bbr.rttProp = diag->wheel_cts;
5566                 log.u_bbr.timeStamp = cts;
5567                 log.u_bbr.delRate = diag->maxslots;
5568                 log.u_bbr.cur_del_rate = diag->p_curtick;
5569                 log.u_bbr.cur_del_rate <<= 32;
5570                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
5571                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5572                     &rack->rc_inp->inp_socket->so_rcv,
5573                     &rack->rc_inp->inp_socket->so_snd,
5574                     BBR_LOG_HPTSDIAG, 0,
5575                     0, &log, false, tv);
5576         }
5577
5578 }
5579
5580 static void
5581 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5582 {
5583         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5584                 union tcp_log_stackspecific log;
5585                 struct timeval tv;
5586
5587                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5588                 log.u_bbr.flex1 = sb->sb_flags;
5589                 log.u_bbr.flex2 = len;
5590                 log.u_bbr.flex3 = sb->sb_state;
5591                 log.u_bbr.flex8 = type;
5592                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5593                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5594                     &rack->rc_inp->inp_socket->so_rcv,
5595                     &rack->rc_inp->inp_socket->so_snd,
5596                     TCP_LOG_SB_WAKE, 0,
5597                     len, &log, false, &tv);
5598         }
5599 }
5600
5601 static void
5602 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5603       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5604 {
5605         struct hpts_diag diag;
5606         struct inpcb *inp = tptoinpcb(tp);
5607         struct timeval tv;
5608         uint32_t delayed_ack = 0;
5609         uint32_t hpts_timeout;
5610         uint32_t entry_slot = slot;
5611         uint8_t stopped;
5612         uint32_t left = 0;
5613         uint32_t us_cts;
5614
5615         if ((tp->t_state == TCPS_CLOSED) ||
5616             (tp->t_state == TCPS_LISTEN)) {
5617                 return;
5618         }
5619         if (tcp_in_hpts(inp)) {
5620                 /* Already on the pacer */
5621                 return;
5622         }
5623         stopped = rack->rc_tmr_stopped;
5624         if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5625                 left = rack->r_ctl.rc_timer_exp - cts;
5626         }
5627         rack->r_ctl.rc_timer_exp = 0;
5628         rack->r_ctl.rc_hpts_flags = 0;
5629         us_cts = tcp_get_usecs(&tv);
5630         /* Now early/late accounting */
5631         rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5632         if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5633                 /*
5634                  * We have a early carry over set,
5635                  * we can always add more time so we
5636                  * can always make this compensation.
5637                  *
5638                  * Note if ack's are allowed to wake us do not
5639                  * penalize the next timer for being awoke
5640                  * by an ack aka the rc_agg_early (non-paced mode).
5641                  */
5642                 slot += rack->r_ctl.rc_agg_early;
5643                 rack->r_early = 0;
5644                 rack->r_ctl.rc_agg_early = 0;
5645         }
5646         if (rack->r_late) {
5647                 /*
5648                  * This is harder, we can
5649                  * compensate some but it
5650                  * really depends on what
5651                  * the current pacing time is.
5652                  */
5653                 if (rack->r_ctl.rc_agg_delayed >= slot) {
5654                         /*
5655                          * We can't compensate for it all.
5656                          * And we have to have some time
5657                          * on the clock. We always have a min
5658                          * 10 slots (10 x 10 i.e. 100 usecs).
5659                          */
5660                         if (slot <= HPTS_TICKS_PER_SLOT) {
5661                                 /* We gain delay */
5662                                 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5663                                 slot = HPTS_TICKS_PER_SLOT;
5664                         } else {
5665                                 /* We take off some */
5666                                 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5667                                 slot = HPTS_TICKS_PER_SLOT;
5668                         }
5669                 } else {
5670                         slot -= rack->r_ctl.rc_agg_delayed;
5671                         rack->r_ctl.rc_agg_delayed = 0;
5672                         /* Make sure we have 100 useconds at minimum */
5673                         if (slot < HPTS_TICKS_PER_SLOT) {
5674                                 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5675                                 slot = HPTS_TICKS_PER_SLOT;
5676                         }
5677                         if (rack->r_ctl.rc_agg_delayed == 0)
5678                                 rack->r_late = 0;
5679                 }
5680         }
5681         if (slot) {
5682                 /* We are pacing too */
5683                 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5684         }
5685         hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5686 #ifdef NETFLIX_EXP_DETECTION
5687         if (rack->sack_attack_disable &&
5688             (slot < tcp_sad_pacing_interval)) {
5689                 /*
5690                  * We have a potential attacker on
5691                  * the line. We have possibly some
5692                  * (or now) pacing time set. We want to
5693                  * slow down the processing of sacks by some
5694                  * amount (if it is an attacker). Set the default
5695                  * slot for attackers in place (unless the orginal
5696                  * interval is longer). Its stored in
5697                  * micro-seconds, so lets convert to msecs.
5698                  */
5699                 slot = tcp_sad_pacing_interval;
5700         }
5701 #endif
5702         if (tp->t_flags & TF_DELACK) {
5703                 delayed_ack = TICKS_2_USEC(tcp_delacktime);
5704                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5705         }
5706         if (delayed_ack && ((hpts_timeout == 0) ||
5707                             (delayed_ack < hpts_timeout)))
5708                 hpts_timeout = delayed_ack;
5709         else
5710                 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5711         /*
5712          * If no timers are going to run and we will fall off the hptsi
5713          * wheel, we resort to a keep-alive timer if its configured.
5714          */
5715         if ((hpts_timeout == 0) &&
5716             (slot == 0)) {
5717                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5718                     (tp->t_state <= TCPS_CLOSING)) {
5719                         /*
5720                          * Ok we have no timer (persists, rack, tlp, rxt  or
5721                          * del-ack), we don't have segments being paced. So
5722                          * all that is left is the keepalive timer.
5723                          */
5724                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5725                                 /* Get the established keep-alive time */
5726                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5727                         } else {
5728                                 /*
5729                                  * Get the initial setup keep-alive time,
5730                                  * note that this is probably not going to
5731                                  * happen, since rack will be running a rxt timer
5732                                  * if a SYN of some sort is outstanding. It is
5733                                  * actually handled in rack_timeout_rxt().
5734                                  */
5735                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5736                         }
5737                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5738                         if (rack->in_probe_rtt) {
5739                                 /*
5740                                  * We want to instead not wake up a long time from
5741                                  * now but to wake up about the time we would
5742                                  * exit probe-rtt and initiate a keep-alive ack.
5743                                  * This will get us out of probe-rtt and update
5744                                  * our min-rtt.
5745                                  */
5746                                 hpts_timeout = rack_min_probertt_hold;
5747                         }
5748                 }
5749         }
5750         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5751             (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5752                 /*
5753                  * RACK, TLP, persists and RXT timers all are restartable
5754                  * based on actions input .. i.e we received a packet (ack
5755                  * or sack) and that changes things (rw, or snd_una etc).
5756                  * Thus we can restart them with a new value. For
5757                  * keep-alive, delayed_ack we keep track of what was left
5758                  * and restart the timer with a smaller value.
5759                  */
5760                 if (left < hpts_timeout)
5761                         hpts_timeout = left;
5762         }
5763         if (hpts_timeout) {
5764                 /*
5765                  * Hack alert for now we can't time-out over 2,147,483
5766                  * seconds (a bit more than 596 hours), which is probably ok
5767                  * :).
5768                  */
5769                 if (hpts_timeout > 0x7ffffffe)
5770                         hpts_timeout = 0x7ffffffe;
5771                 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5772         }
5773         rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5774         if ((rack->gp_ready == 0) &&
5775             (rack->use_fixed_rate == 0) &&
5776             (hpts_timeout < slot) &&
5777             (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5778                 /*
5779                  * We have no good estimate yet for the
5780                  * old clunky burst mitigation or the
5781                  * real pacing. And the tlp or rxt is smaller
5782                  * than the pacing calculation. Lets not
5783                  * pace that long since we know the calculation
5784                  * so far is not accurate.
5785                  */
5786                 slot = hpts_timeout;
5787         }
5788         /**
5789          * Turn off all the flags for queuing by default. The
5790          * flags have important meanings to what happens when
5791          * LRO interacts with the transport. Most likely (by default now)
5792          * mbuf_queueing and ack compression are on. So the transport
5793          * has a couple of flags that control what happens (if those
5794          * are not on then these flags won't have any effect since it
5795          * won't go through the queuing LRO path).
5796          *
5797          * INP_MBUF_QUEUE_READY - This flags says that I am busy
5798          *                        pacing output, so don't disturb. But
5799          *                        it also means LRO can wake me if there
5800          *                        is a SACK arrival.
5801          *
5802          * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5803          *                       with the above flag (QUEUE_READY) and
5804          *                       when present it says don't even wake me
5805          *                       if a SACK arrives.
5806          *
5807          * The idea behind these flags is that if we are pacing we
5808          * set the MBUF_QUEUE_READY and only get woken up if
5809          * a SACK arrives (which could change things) or if
5810          * our pacing timer expires. If, however, we have a rack
5811          * timer running, then we don't even want a sack to wake
5812          * us since the rack timer has to expire before we can send.
5813          *
5814          * Other cases should usually have none of the flags set
5815          * so LRO can call into us.
5816          */
5817         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5818         if (slot) {
5819                 rack->r_ctl.rc_last_output_to = us_cts + slot;
5820                 /*
5821                  * A pacing timer (slot) is being set, in
5822                  * such a case we cannot send (we are blocked by
5823                  * the timer). So lets tell LRO that it should not
5824                  * wake us unless there is a SACK. Note this only
5825                  * will be effective if mbuf queueing is on or
5826                  * compressed acks are being processed.
5827                  */
5828                 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5829                 /*
5830                  * But wait if we have a Rack timer running
5831                  * even a SACK should not disturb us (with
5832                  * the exception of r_rr_config 3).
5833                  */
5834                 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5835                     (rack->r_rr_config != 3))
5836                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5837                 if (rack->rc_ack_can_sendout_data) {
5838                         /*
5839                          * Ahh but wait, this is that special case
5840                          * where the pacing timer can be disturbed
5841                          * backout the changes (used for non-paced
5842                          * burst limiting).
5843                          */
5844                         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5845                 }
5846                 if ((rack->use_rack_rr) &&
5847                     (rack->r_rr_config < 2) &&
5848                     ((hpts_timeout) && (hpts_timeout < slot))) {
5849                         /*
5850                          * Arrange for the hpts to kick back in after the
5851                          * t-o if the t-o does not cause a send.
5852                          */
5853                         (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5854                                                    __LINE__, &diag);
5855                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5856                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5857                 } else {
5858                         (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(slot),
5859                                                    __LINE__, &diag);
5860                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5861                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5862                 }
5863         } else if (hpts_timeout) {
5864                 /*
5865                  * With respect to inp_flags2 here, lets let any new acks wake
5866                  * us up here. Since we are not pacing (no pacing timer), output
5867                  * can happen so we should let it. If its a Rack timer, then any inbound
5868                  * packet probably won't change the sending (we will be blocked)
5869                  * but it may change the prr stats so letting it in (the set defaults
5870                  * at the start of this block) are good enough.
5871                  */
5872                 (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5873                                            __LINE__, &diag);
5874                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5875                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5876         } else {
5877                 /* No timer starting */
5878 #ifdef INVARIANTS
5879                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5880                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5881                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5882                 }
5883 #endif
5884         }
5885         rack->rc_tmr_stopped = 0;
5886         if (slot)
5887                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5888 }
5889
5890 /*
5891  * RACK Timer, here we simply do logging and house keeping.
5892  * the normal rack_output() function will call the
5893  * appropriate thing to check if we need to do a RACK retransmit.
5894  * We return 1, saying don't proceed with rack_output only
5895  * when all timers have been stopped (destroyed PCB?).
5896  */
5897 static int
5898 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5899 {
5900         /*
5901          * This timer simply provides an internal trigger to send out data.
5902          * The check_recovery_mode call will see if there are needed
5903          * retransmissions, if so we will enter fast-recovery. The output
5904          * call may or may not do the same thing depending on sysctl
5905          * settings.
5906          */
5907         struct rack_sendmap *rsm;
5908
5909         counter_u64_add(rack_to_tot, 1);
5910         if (rack->r_state && (rack->r_state != tp->t_state))
5911                 rack_set_state(tp, rack);
5912         rack->rc_on_min_to = 0;
5913         rsm = rack_check_recovery_mode(tp, cts);
5914         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5915         if (rsm) {
5916                 rack->r_ctl.rc_resend = rsm;
5917                 rack->r_timer_override = 1;
5918                 if (rack->use_rack_rr) {
5919                         /*
5920                          * Don't accumulate extra pacing delay
5921                          * we are allowing the rack timer to
5922                          * over-ride pacing i.e. rrr takes precedence
5923                          * if the pacing interval is longer than the rrr
5924                          * time (in other words we get the min pacing
5925                          * time versus rrr pacing time).
5926                          */
5927                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5928                 }
5929         }
5930         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5931         if (rsm == NULL) {
5932                 /* restart a timer and return 1 */
5933                 rack_start_hpts_timer(rack, tp, cts,
5934                                       0, 0, 0);
5935                 return (1);
5936         }
5937         return (0);
5938 }
5939
5940 static void
5941 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5942 {
5943         if (rsm->m->m_len > rsm->orig_m_len) {
5944                 /*
5945                  * Mbuf grew, caused by sbcompress, our offset does
5946                  * not change.
5947                  */
5948                 rsm->orig_m_len = rsm->m->m_len;
5949         } else if (rsm->m->m_len < rsm->orig_m_len) {
5950                 /*
5951                  * Mbuf shrank, trimmed off the top by an ack, our
5952                  * offset changes.
5953                  */
5954                 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
5955                 rsm->orig_m_len = rsm->m->m_len;
5956         }
5957 }
5958
5959 static void
5960 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
5961 {
5962         struct mbuf *m;
5963         uint32_t soff;
5964
5965         if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
5966                 /* Fix up the orig_m_len and possibly the mbuf offset */
5967                 rack_adjust_orig_mlen(src_rsm);
5968         }
5969         m = src_rsm->m;
5970         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
5971         while (soff >= m->m_len) {
5972                 /* Move out past this mbuf */
5973                 soff -= m->m_len;
5974                 m = m->m_next;
5975                 KASSERT((m != NULL),
5976                         ("rsm:%p nrsm:%p hit at soff:%u null m",
5977                          src_rsm, rsm, soff));
5978         }
5979         rsm->m = m;
5980         rsm->soff = soff;
5981         rsm->orig_m_len = m->m_len;
5982 }
5983
5984 static __inline void
5985 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5986                struct rack_sendmap *rsm, uint32_t start)
5987 {
5988         int idx;
5989
5990         nrsm->r_start = start;
5991         nrsm->r_end = rsm->r_end;
5992         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
5993         nrsm->r_flags = rsm->r_flags;
5994         nrsm->r_dupack = rsm->r_dupack;
5995         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
5996         nrsm->r_rtr_bytes = 0;
5997         nrsm->r_fas = rsm->r_fas;
5998         rsm->r_end = nrsm->r_start;
5999         nrsm->r_just_ret = rsm->r_just_ret;
6000         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6001                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6002         }
6003         /* Now if we have SYN flag we keep it on the left edge */
6004         if (nrsm->r_flags & RACK_HAS_SYN)
6005                 nrsm->r_flags &= ~RACK_HAS_SYN;
6006         /* Now if we have a FIN flag we keep it on the right edge */
6007         if (rsm->r_flags & RACK_HAS_FIN)
6008                 rsm->r_flags &= ~RACK_HAS_FIN;
6009         /* Push bit must go to the right edge as well */
6010         if (rsm->r_flags & RACK_HAD_PUSH)
6011                 rsm->r_flags &= ~RACK_HAD_PUSH;
6012         /* Clone over the state of the hw_tls flag */
6013         nrsm->r_hw_tls = rsm->r_hw_tls;
6014         /*
6015          * Now we need to find nrsm's new location in the mbuf chain
6016          * we basically calculate a new offset, which is soff +
6017          * how much is left in original rsm. Then we walk out the mbuf
6018          * chain to find the righ position, it may be the same mbuf
6019          * or maybe not.
6020          */
6021         KASSERT(((rsm->m != NULL) ||
6022                  (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
6023                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6024         if (rsm->m)
6025                 rack_setup_offset_for_rsm(rsm, nrsm);
6026 }
6027
6028 static struct rack_sendmap *
6029 rack_merge_rsm(struct tcp_rack *rack,
6030                struct rack_sendmap *l_rsm,
6031                struct rack_sendmap *r_rsm)
6032 {
6033         /*
6034          * We are merging two ack'd RSM's,
6035          * the l_rsm is on the left (lower seq
6036          * values) and the r_rsm is on the right
6037          * (higher seq value). The simplest way
6038          * to merge these is to move the right
6039          * one into the left. I don't think there
6040          * is any reason we need to try to find
6041          * the oldest (or last oldest retransmitted).
6042          */
6043 #ifdef INVARIANTS
6044         struct rack_sendmap *rm;
6045 #endif
6046         rack_log_map_chg(rack->rc_tp, rack, NULL,
6047                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6048         l_rsm->r_end = r_rsm->r_end;
6049         if (l_rsm->r_dupack < r_rsm->r_dupack)
6050                 l_rsm->r_dupack = r_rsm->r_dupack;
6051         if (r_rsm->r_rtr_bytes)
6052                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6053         if (r_rsm->r_in_tmap) {
6054                 /* This really should not happen */
6055                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6056                 r_rsm->r_in_tmap = 0;
6057         }
6058
6059         /* Now the flags */
6060         if (r_rsm->r_flags & RACK_HAS_FIN)
6061                 l_rsm->r_flags |= RACK_HAS_FIN;
6062         if (r_rsm->r_flags & RACK_TLP)
6063                 l_rsm->r_flags |= RACK_TLP;
6064         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6065                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6066         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6067             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6068                 /*
6069                  * If both are app-limited then let the
6070                  * free lower the count. If right is app
6071                  * limited and left is not, transfer.
6072                  */
6073                 l_rsm->r_flags |= RACK_APP_LIMITED;
6074                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
6075                 if (r_rsm == rack->r_ctl.rc_first_appl)
6076                         rack->r_ctl.rc_first_appl = l_rsm;
6077         }
6078 #ifndef INVARIANTS
6079         (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6080 #else
6081         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6082         if (rm != r_rsm) {
6083                 panic("removing head in rack:%p rsm:%p rm:%p",
6084                       rack, r_rsm, rm);
6085         }
6086 #endif
6087         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6088                 /* Transfer the split limit to the map we free */
6089                 r_rsm->r_limit_type = l_rsm->r_limit_type;
6090                 l_rsm->r_limit_type = 0;
6091         }
6092         rack_free(rack, r_rsm);
6093         return (l_rsm);
6094 }
6095
6096 /*
6097  * TLP Timer, here we simply setup what segment we want to
6098  * have the TLP expire on, the normal rack_output() will then
6099  * send it out.
6100  *
6101  * We return 1, saying don't proceed with rack_output only
6102  * when all timers have been stopped (destroyed PCB?).
6103  */
6104 static int
6105 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6106 {
6107         /*
6108          * Tail Loss Probe.
6109          */
6110         struct rack_sendmap *rsm = NULL;
6111 #ifdef INVARIANTS
6112         struct rack_sendmap *insret;
6113 #endif
6114         struct socket *so = tptosocket(tp);
6115         uint32_t amm;
6116         uint32_t out, avail;
6117         int collapsed_win = 0;
6118
6119         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6120                 /* Its not time yet */
6121                 return (0);
6122         }
6123         if (ctf_progress_timeout_check(tp, true)) {
6124                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6125                 return (-ETIMEDOUT);    /* tcp_drop() */
6126         }
6127         /*
6128          * A TLP timer has expired. We have been idle for 2 rtts. So we now
6129          * need to figure out how to force a full MSS segment out.
6130          */
6131         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6132         rack->r_ctl.retran_during_recovery = 0;
6133         rack->r_ctl.dsack_byte_cnt = 0;
6134         counter_u64_add(rack_tlp_tot, 1);
6135         if (rack->r_state && (rack->r_state != tp->t_state))
6136                 rack_set_state(tp, rack);
6137         avail = sbavail(&so->so_snd);
6138         out = tp->snd_max - tp->snd_una;
6139         if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
6140                 /* special case, we need a retransmission */
6141                 collapsed_win = 1;
6142                 goto need_retran;
6143         }
6144         if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6145                 rack->r_ctl.dsack_persist--;
6146                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6147                         rack->r_ctl.num_dsack = 0;
6148                 }
6149                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6150         }
6151         if ((tp->t_flags & TF_GPUTINPROG) &&
6152             (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6153                 /*
6154                  * If this is the second in a row
6155                  * TLP and we are doing a measurement
6156                  * its time to abandon the measurement.
6157                  * Something is likely broken on
6158                  * the clients network and measuring a
6159                  * broken network does us no good.
6160                  */
6161                 tp->t_flags &= ~TF_GPUTINPROG;
6162                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6163                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
6164                                            tp->gput_seq,
6165                                            0, 0, 18, __LINE__, NULL, 0);
6166         }
6167         /*
6168          * Check our send oldest always settings, and if
6169          * there is an oldest to send jump to the need_retran.
6170          */
6171         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6172                 goto need_retran;
6173
6174         if (avail > out) {
6175                 /* New data is available */
6176                 amm = avail - out;
6177                 if (amm > ctf_fixed_maxseg(tp)) {
6178                         amm = ctf_fixed_maxseg(tp);
6179                         if ((amm + out) > tp->snd_wnd) {
6180                                 /* We are rwnd limited */
6181                                 goto need_retran;
6182                         }
6183                 } else if (amm < ctf_fixed_maxseg(tp)) {
6184                         /* not enough to fill a MTU */
6185                         goto need_retran;
6186                 }
6187                 if (IN_FASTRECOVERY(tp->t_flags)) {
6188                         /* Unlikely */
6189                         if (rack->rack_no_prr == 0) {
6190                                 if (out + amm <= tp->snd_wnd) {
6191                                         rack->r_ctl.rc_prr_sndcnt = amm;
6192                                         rack->r_ctl.rc_tlp_new_data = amm;
6193                                         rack_log_to_prr(rack, 4, 0, __LINE__);
6194                                 }
6195                         } else
6196                                 goto need_retran;
6197                 } else {
6198                         /* Set the send-new override */
6199                         if (out + amm <= tp->snd_wnd)
6200                                 rack->r_ctl.rc_tlp_new_data = amm;
6201                         else
6202                                 goto need_retran;
6203                 }
6204                 rack->r_ctl.rc_tlpsend = NULL;
6205                 counter_u64_add(rack_tlp_newdata, 1);
6206                 goto send;
6207         }
6208 need_retran:
6209         /*
6210          * Ok we need to arrange the last un-acked segment to be re-sent, or
6211          * optionally the first un-acked segment.
6212          */
6213         if (collapsed_win == 0) {
6214                 if (rack_always_send_oldest)
6215                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6216                 else {
6217                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6218                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6219                                 rsm = rack_find_high_nonack(rack, rsm);
6220                         }
6221                 }
6222                 if (rsm == NULL) {
6223 #ifdef TCP_BLACKBOX
6224                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6225 #endif
6226                         goto out;
6227                 }
6228         } else {
6229                 /*
6230                  * We must find the last segment
6231                  * that was acceptable by the client.
6232                  */
6233                 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6234                         if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6235                                 /* Found one */
6236                                 break;
6237                         }
6238                 }
6239                 if (rsm == NULL) {
6240                         /* None? if so send the first */
6241                         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6242                         if (rsm == NULL) {
6243 #ifdef TCP_BLACKBOX
6244                                 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6245 #endif
6246                                 goto out;
6247                         }
6248                 }
6249         }
6250         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6251                 /*
6252                  * We need to split this the last segment in two.
6253                  */
6254                 struct rack_sendmap *nrsm;
6255
6256                 nrsm = rack_alloc_full_limit(rack);
6257                 if (nrsm == NULL) {
6258                         /*
6259                          * No memory to split, we will just exit and punt
6260                          * off to the RXT timer.
6261                          */
6262                         goto out;
6263                 }
6264                 rack_clone_rsm(rack, nrsm, rsm,
6265                                (rsm->r_end - ctf_fixed_maxseg(tp)));
6266                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6267 #ifndef INVARIANTS
6268                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6269 #else
6270                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6271                 if (insret != NULL) {
6272                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6273                               nrsm, insret, rack, rsm);
6274                 }
6275 #endif
6276                 if (rsm->r_in_tmap) {
6277                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6278                         nrsm->r_in_tmap = 1;
6279                 }
6280                 rsm = nrsm;
6281         }
6282         rack->r_ctl.rc_tlpsend = rsm;
6283 send:
6284         /* Make sure output path knows we are doing a TLP */
6285         *doing_tlp = 1;
6286         rack->r_timer_override = 1;
6287         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6288         return (0);
6289 out:
6290         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6291         return (0);
6292 }
6293
6294 /*
6295  * Delayed ack Timer, here we simply need to setup the
6296  * ACK_NOW flag and remove the DELACK flag. From there
6297  * the output routine will send the ack out.
6298  *
6299  * We only return 1, saying don't proceed, if all timers
6300  * are stopped (destroyed PCB?).
6301  */
6302 static int
6303 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6304 {
6305
6306         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6307         tp->t_flags &= ~TF_DELACK;
6308         tp->t_flags |= TF_ACKNOW;
6309         KMOD_TCPSTAT_INC(tcps_delack);
6310         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6311         return (0);
6312 }
6313
6314 /*
6315  * Persists timer, here we simply send the
6316  * same thing as a keepalive will.
6317  * the one byte send.
6318  *
6319  * We only return 1, saying don't proceed, if all timers
6320  * are stopped (destroyed PCB?).
6321  */
6322 static int
6323 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6324 {
6325         struct tcptemp *t_template;
6326         int32_t retval = 1;
6327
6328         if (rack->rc_in_persist == 0)
6329                 return (0);
6330         if (ctf_progress_timeout_check(tp, false)) {
6331                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6332                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6333                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6334                 return (-ETIMEDOUT);    /* tcp_drop() */
6335         }
6336         /*
6337          * Persistence timer into zero window. Force a byte to be output, if
6338          * possible.
6339          */
6340         KMOD_TCPSTAT_INC(tcps_persisttimeo);
6341         /*
6342          * Hack: if the peer is dead/unreachable, we do not time out if the
6343          * window is closed.  After a full backoff, drop the connection if
6344          * the idle time (no responses to probes) reaches the maximum
6345          * backoff that we would use if retransmitting.
6346          */
6347         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6348             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6349              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6350                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6351                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6352                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6353                 retval = -ETIMEDOUT;    /* tcp_drop() */
6354                 goto out;
6355         }
6356         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6357             tp->snd_una == tp->snd_max)
6358                 rack_exit_persist(tp, rack, cts);
6359         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6360         /*
6361          * If the user has closed the socket then drop a persisting
6362          * connection after a much reduced timeout.
6363          */
6364         if (tp->t_state > TCPS_CLOSE_WAIT &&
6365             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6366                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6367                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6368                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6369                 retval = -ETIMEDOUT;    /* tcp_drop() */
6370                 goto out;
6371         }
6372         t_template = tcpip_maketemplate(rack->rc_inp);
6373         if (t_template) {
6374                 /* only set it if we were answered */
6375                 if (rack->forced_ack == 0) {
6376                         rack->forced_ack = 1;
6377                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6378                 } else {
6379                         rack->probe_not_answered = 1;
6380                         counter_u64_add(rack_persists_loss, 1);
6381                         rack->r_ctl.persist_lost_ends++;
6382                 }
6383                 counter_u64_add(rack_persists_sends, 1);
6384                 tcp_respond(tp, t_template->tt_ipgen,
6385                             &t_template->tt_t, (struct mbuf *)NULL,
6386                             tp->rcv_nxt, tp->snd_una - 1, 0);
6387                 /* This sends an ack */
6388                 if (tp->t_flags & TF_DELACK)
6389                         tp->t_flags &= ~TF_DELACK;
6390                 free(t_template, M_TEMP);
6391         }
6392         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6393                 tp->t_rxtshift++;
6394 out:
6395         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6396         rack_start_hpts_timer(rack, tp, cts,
6397                               0, 0, 0);
6398         return (retval);
6399 }
6400
6401 /*
6402  * If a keepalive goes off, we had no other timers
6403  * happening. We always return 1 here since this
6404  * routine either drops the connection or sends
6405  * out a segment with respond.
6406  */
6407 static int
6408 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6409 {
6410         struct tcptemp *t_template;
6411         struct inpcb *inp = tptoinpcb(tp);
6412
6413         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6414         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6415         /*
6416          * Keep-alive timer went off; send something or drop connection if
6417          * idle for too long.
6418          */
6419         KMOD_TCPSTAT_INC(tcps_keeptimeo);
6420         if (tp->t_state < TCPS_ESTABLISHED)
6421                 goto dropit;
6422         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6423             tp->t_state <= TCPS_CLOSING) {
6424                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6425                         goto dropit;
6426                 /*
6427                  * Send a packet designed to force a response if the peer is
6428                  * up and reachable: either an ACK if the connection is
6429                  * still alive, or an RST if the peer has closed the
6430                  * connection due to timeout or reboot. Using sequence
6431                  * number tp->snd_una-1 causes the transmitted zero-length
6432                  * segment to lie outside the receive window; by the
6433                  * protocol spec, this requires the correspondent TCP to
6434                  * respond.
6435                  */
6436                 KMOD_TCPSTAT_INC(tcps_keepprobe);
6437                 t_template = tcpip_maketemplate(inp);
6438                 if (t_template) {
6439                         if (rack->forced_ack == 0) {
6440                                 rack->forced_ack = 1;
6441                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6442                         } else {
6443                                 rack->probe_not_answered = 1;
6444                         }
6445                         tcp_respond(tp, t_template->tt_ipgen,
6446                             &t_template->tt_t, (struct mbuf *)NULL,
6447                             tp->rcv_nxt, tp->snd_una - 1, 0);
6448                         free(t_template, M_TEMP);
6449                 }
6450         }
6451         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6452         return (1);
6453 dropit:
6454         KMOD_TCPSTAT_INC(tcps_keepdrops);
6455         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6456         return (-ETIMEDOUT);    /* tcp_drop() */
6457 }
6458
6459 /*
6460  * Retransmit helper function, clear up all the ack
6461  * flags and take care of important book keeping.
6462  */
6463 static void
6464 rack_remxt_tmr(struct tcpcb *tp)
6465 {
6466         /*
6467          * The retransmit timer went off, all sack'd blocks must be
6468          * un-acked.
6469          */
6470         struct rack_sendmap *rsm, *trsm = NULL;
6471         struct tcp_rack *rack;
6472
6473         rack = (struct tcp_rack *)tp->t_fb_ptr;
6474         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6475         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6476         if (rack->r_state && (rack->r_state != tp->t_state))
6477                 rack_set_state(tp, rack);
6478         /*
6479          * Ideally we would like to be able to
6480          * mark SACK-PASS on anything not acked here.
6481          *
6482          * However, if we do that we would burst out
6483          * all that data 1ms apart. This would be unwise,
6484          * so for now we will just let the normal rxt timer
6485          * and tlp timer take care of it.
6486          *
6487          * Also we really need to stick them back in sequence
6488          * order. This way we send in the proper order and any
6489          * sacks that come floating in will "re-ack" the data.
6490          * To do this we zap the tmap with an INIT and then
6491          * walk through and place every rsm in the RB tree
6492          * back in its seq ordered place.
6493          */
6494         TAILQ_INIT(&rack->r_ctl.rc_tmap);
6495         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6496                 rsm->r_dupack = 0;
6497                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6498                 /* We must re-add it back to the tlist */
6499                 if (trsm == NULL) {
6500                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6501                 } else {
6502                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6503                 }
6504                 rsm->r_in_tmap = 1;
6505                 trsm = rsm;
6506                 if (rsm->r_flags & RACK_ACKED)
6507                         rsm->r_flags |= RACK_WAS_ACKED;
6508                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
6509                 rsm->r_flags |= RACK_MUST_RXT;
6510         }
6511         /* Clear the count (we just un-acked them) */
6512         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6513         rack->r_ctl.rc_sacked = 0;
6514         rack->r_ctl.rc_sacklast = NULL;
6515         rack->r_ctl.rc_agg_delayed = 0;
6516         rack->r_early = 0;
6517         rack->r_ctl.rc_agg_early = 0;
6518         rack->r_late = 0;
6519         /* Clear the tlp rtx mark */
6520         rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6521         if (rack->r_ctl.rc_resend != NULL)
6522                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6523         rack->r_ctl.rc_prr_sndcnt = 0;
6524         rack_log_to_prr(rack, 6, 0, __LINE__);
6525         rack->r_timer_override = 1;
6526         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6527 #ifdef NETFLIX_EXP_DETECTION
6528             || (rack->sack_attack_disable != 0)
6529 #endif
6530                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6531                 /*
6532                  * For non-sack customers new data
6533                  * needs to go out as retransmits until
6534                  * we retransmit up to snd_max.
6535                  */
6536                 rack->r_must_retran = 1;
6537                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6538                                                 rack->r_ctl.rc_sacked);
6539         }
6540         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6541 }
6542
6543 static void
6544 rack_convert_rtts(struct tcpcb *tp)
6545 {
6546         if (tp->t_srtt > 1) {
6547                 uint32_t val, frac;
6548
6549                 val = tp->t_srtt >> TCP_RTT_SHIFT;
6550                 frac = tp->t_srtt & 0x1f;
6551                 tp->t_srtt = TICKS_2_USEC(val);
6552                 /*
6553                  * frac is the fractional part of the srtt (if any)
6554                  * but its in ticks and every bit represents
6555                  * 1/32nd of a hz.
6556                  */
6557                 if (frac) {
6558                         if (hz == 1000) {
6559                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6560                         } else {
6561                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6562                         }
6563                         tp->t_srtt += frac;
6564                 }
6565         }
6566         if (tp->t_rttvar) {
6567                 uint32_t val, frac;
6568
6569                 val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6570                 frac = tp->t_rttvar & 0x1f;
6571                 tp->t_rttvar = TICKS_2_USEC(val);
6572                 /*
6573                  * frac is the fractional part of the srtt (if any)
6574                  * but its in ticks and every bit represents
6575                  * 1/32nd of a hz.
6576                  */
6577                 if (frac) {
6578                         if (hz == 1000) {
6579                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6580                         } else {
6581                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6582                         }
6583                         tp->t_rttvar += frac;
6584                 }
6585         }
6586         tp->t_rxtcur = RACK_REXMTVAL(tp);
6587         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6588                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6589         }
6590         if (tp->t_rxtcur > rack_rto_max) {
6591                 tp->t_rxtcur = rack_rto_max;
6592         }
6593 }
6594
6595 static void
6596 rack_cc_conn_init(struct tcpcb *tp)
6597 {
6598         struct tcp_rack *rack;
6599         uint32_t srtt;
6600
6601         rack = (struct tcp_rack *)tp->t_fb_ptr;
6602         srtt = tp->t_srtt;
6603         cc_conn_init(tp);
6604         /*
6605          * Now convert to rack's internal format,
6606          * if required.
6607          */
6608         if ((srtt == 0) && (tp->t_srtt != 0))
6609                 rack_convert_rtts(tp);
6610         /*
6611          * We want a chance to stay in slowstart as
6612          * we create a connection. TCP spec says that
6613          * initially ssthresh is infinite. For our
6614          * purposes that is the snd_wnd.
6615          */
6616         if (tp->snd_ssthresh < tp->snd_wnd) {
6617                 tp->snd_ssthresh = tp->snd_wnd;
6618         }
6619         /*
6620          * We also want to assure a IW worth of
6621          * data can get inflight.
6622          */
6623         if (rc_init_window(rack) < tp->snd_cwnd)
6624                 tp->snd_cwnd = rc_init_window(rack);
6625 }
6626
6627 /*
6628  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6629  * we will setup to retransmit the lowest seq number outstanding.
6630  */
6631 static int
6632 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6633 {
6634         struct inpcb *inp = tptoinpcb(tp);
6635         int32_t rexmt;
6636         int32_t retval = 0;
6637         bool isipv6;
6638
6639         if ((tp->t_flags & TF_GPUTINPROG) &&
6640             (tp->t_rxtshift)) {
6641                 /*
6642                  * We have had a second timeout
6643                  * measurements on successive rxt's are not profitable.
6644                  * It is unlikely to be of any use (the network is
6645                  * broken or the client went away).
6646                  */
6647                 tp->t_flags &= ~TF_GPUTINPROG;
6648                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6649                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
6650                                            tp->gput_seq,
6651                                            0, 0, 18, __LINE__, NULL, 0);
6652         }
6653         if (ctf_progress_timeout_check(tp, false)) {
6654                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6655                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6656                 return (-ETIMEDOUT);    /* tcp_drop() */
6657         }
6658         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6659         rack->r_ctl.retran_during_recovery = 0;
6660         rack->rc_ack_required = 1;
6661         rack->r_ctl.dsack_byte_cnt = 0;
6662         if (IN_FASTRECOVERY(tp->t_flags))
6663                 tp->t_flags |= TF_WASFRECOVERY;
6664         else
6665                 tp->t_flags &= ~TF_WASFRECOVERY;
6666         if (IN_CONGRECOVERY(tp->t_flags))
6667                 tp->t_flags |= TF_WASCRECOVERY;
6668         else
6669                 tp->t_flags &= ~TF_WASCRECOVERY;
6670         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6671             (tp->snd_una == tp->snd_max)) {
6672                 /* Nothing outstanding .. nothing to do */
6673                 return (0);
6674         }
6675         if (rack->r_ctl.dsack_persist) {
6676                 rack->r_ctl.dsack_persist--;
6677                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6678                         rack->r_ctl.num_dsack = 0;
6679                 }
6680                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6681         }
6682         /*
6683          * Rack can only run one timer  at a time, so we cannot
6684          * run a KEEPINIT (gating SYN sending) and a retransmit
6685          * timer for the SYN. So if we are in a front state and
6686          * have a KEEPINIT timer we need to check the first transmit
6687          * against now to see if we have exceeded the KEEPINIT time
6688          * (if one is set).
6689          */
6690         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6691             (TP_KEEPINIT(tp) != 0)) {
6692                 struct rack_sendmap *rsm;
6693
6694                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6695                 if (rsm) {
6696                         /* Ok we have something outstanding to test keepinit with */
6697                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6698                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6699                                 /* We have exceeded the KEEPINIT time */
6700                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6701                                 goto drop_it;
6702                         }
6703                 }
6704         }
6705         /*
6706          * Retransmission timer went off.  Message has not been acked within
6707          * retransmit interval.  Back off to a longer retransmit interval
6708          * and retransmit one segment.
6709          */
6710         rack_remxt_tmr(tp);
6711         if ((rack->r_ctl.rc_resend == NULL) ||
6712             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6713                 /*
6714                  * If the rwnd collapsed on
6715                  * the one we are retransmitting
6716                  * it does not count against the
6717                  * rxt count.
6718                  */
6719                 tp->t_rxtshift++;
6720         }
6721         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6722                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6723 drop_it:
6724                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6725                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6726                 /* XXXGL: previously t_softerror was casted to uint16_t */
6727                 MPASS(tp->t_softerror >= 0);
6728                 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
6729                 goto out;       /* tcp_drop() */
6730         }
6731         if (tp->t_state == TCPS_SYN_SENT) {
6732                 /*
6733                  * If the SYN was retransmitted, indicate CWND to be limited
6734                  * to 1 segment in cc_conn_init().
6735                  */
6736                 tp->snd_cwnd = 1;
6737         } else if (tp->t_rxtshift == 1) {
6738                 /*
6739                  * first retransmit; record ssthresh and cwnd so they can be
6740                  * recovered if this turns out to be a "bad" retransmit. A
6741                  * retransmit is considered "bad" if an ACK for this segment
6742                  * is received within RTT/2 interval; the assumption here is
6743                  * that the ACK was already in flight.  See "On Estimating
6744                  * End-to-End Network Path Properties" by Allman and Paxson
6745                  * for more details.
6746                  */
6747                 tp->snd_cwnd_prev = tp->snd_cwnd;
6748                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6749                 tp->snd_recover_prev = tp->snd_recover;
6750                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6751                 tp->t_flags |= TF_PREVVALID;
6752         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6753                 tp->t_flags &= ~TF_PREVVALID;
6754         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6755         if ((tp->t_state == TCPS_SYN_SENT) ||
6756             (tp->t_state == TCPS_SYN_RECEIVED))
6757                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6758         else
6759                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6760
6761         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6762            max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6763         /*
6764          * We enter the path for PLMTUD if connection is established or, if
6765          * connection is FIN_WAIT_1 status, reason for the last is that if
6766          * amount of data we send is very small, we could send it in couple
6767          * of packets and process straight to FIN. In that case we won't
6768          * catch ESTABLISHED state.
6769          */
6770 #ifdef INET6
6771         isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
6772 #else
6773         isipv6 = false;
6774 #endif
6775         if (((V_tcp_pmtud_blackhole_detect == 1) ||
6776             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6777             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6778             ((tp->t_state == TCPS_ESTABLISHED) ||
6779             (tp->t_state == TCPS_FIN_WAIT_1))) {
6780                 /*
6781                  * Idea here is that at each stage of mtu probe (usually,
6782                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
6783                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6784                  * should take care of that.
6785                  */
6786                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6787                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6788                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6789                     tp->t_rxtshift % 2 == 0)) {
6790                         /*
6791                          * Enter Path MTU Black-hole Detection mechanism: -
6792                          * Disable Path MTU Discovery (IP "DF" bit). -
6793                          * Reduce MTU to lower value than what we negotiated
6794                          * with peer.
6795                          */
6796                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6797                                 /* Record that we may have found a black hole. */
6798                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6799                                 /* Keep track of previous MSS. */
6800                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6801                         }
6802
6803                         /*
6804                          * Reduce the MSS to blackhole value or to the
6805                          * default in an attempt to retransmit.
6806                          */
6807 #ifdef INET6
6808                         if (isipv6 &&
6809                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6810                                 /* Use the sysctl tuneable blackhole MSS. */
6811                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6812                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6813                         } else if (isipv6) {
6814                                 /* Use the default MSS. */
6815                                 tp->t_maxseg = V_tcp_v6mssdflt;
6816                                 /*
6817                                  * Disable Path MTU Discovery when we switch
6818                                  * to minmss.
6819                                  */
6820                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6821                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6822                         }
6823 #endif
6824 #if defined(INET6) && defined(INET)
6825                         else
6826 #endif
6827 #ifdef INET
6828                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6829                                 /* Use the sysctl tuneable blackhole MSS. */
6830                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6831                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6832                         } else {
6833                                 /* Use the default MSS. */
6834                                 tp->t_maxseg = V_tcp_mssdflt;
6835                                 /*
6836                                  * Disable Path MTU Discovery when we switch
6837                                  * to minmss.
6838                                  */
6839                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6840                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6841                         }
6842 #endif
6843                 } else {
6844                         /*
6845                          * If further retransmissions are still unsuccessful
6846                          * with a lowered MTU, maybe this isn't a blackhole
6847                          * and we restore the previous MSS and blackhole
6848                          * detection flags. The limit '6' is determined by
6849                          * giving each probe stage (1448, 1188, 524) 2
6850                          * chances to recover.
6851                          */
6852                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6853                             (tp->t_rxtshift >= 6)) {
6854                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6855                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6856                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6857                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6858                         }
6859                 }
6860         }
6861         /*
6862          * Disable RFC1323 and SACK if we haven't got any response to
6863          * our third SYN to work-around some broken terminal servers
6864          * (most of which have hopefully been retired) that have bad VJ
6865          * header compression code which trashes TCP segments containing
6866          * unknown-to-them TCP options.
6867          */
6868         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6869             (tp->t_rxtshift == 3))
6870                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6871         /*
6872          * If we backed off this far, our srtt estimate is probably bogus.
6873          * Clobber it so we'll take the next rtt measurement as our srtt;
6874          * move the current srtt into rttvar to keep the current retransmit
6875          * times until then.
6876          */
6877         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6878 #ifdef INET6
6879                 if ((inp->inp_vflag & INP_IPV6) != 0)
6880                         in6_losing(inp);
6881                 else
6882 #endif
6883                         in_losing(inp);
6884                 tp->t_rttvar += tp->t_srtt;
6885                 tp->t_srtt = 0;
6886         }
6887         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6888         tp->snd_recover = tp->snd_max;
6889         tp->t_flags |= TF_ACKNOW;
6890         tp->t_rtttime = 0;
6891         rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
6892 out:
6893         return (retval);
6894 }
6895
6896 static int
6897 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
6898 {
6899         int32_t ret = 0;
6900         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6901
6902         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
6903             (tp->t_flags & TF_GPUTINPROG)) {
6904                 /*
6905                  * We have a goodput in progress
6906                  * and we have entered a late state.
6907                  * Do we have enough data in the sb
6908                  * to handle the GPUT request?
6909                  */
6910                 uint32_t bytes;
6911
6912                 bytes = tp->gput_ack - tp->gput_seq;
6913                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
6914                         bytes += tp->gput_seq - tp->snd_una;
6915                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
6916                         /*
6917                          * There are not enough bytes in the socket
6918                          * buffer that have been sent to cover this
6919                          * measurement. Cancel it.
6920                          */
6921                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6922                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
6923                                                    tp->gput_seq,
6924                                                    0, 0, 18, __LINE__, NULL, 0);
6925                         tp->t_flags &= ~TF_GPUTINPROG;
6926                 }
6927         }
6928         if (timers == 0) {
6929                 return (0);
6930         }
6931         if (tp->t_state == TCPS_LISTEN) {
6932                 /* no timers on listen sockets */
6933                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6934                         return (0);
6935                 return (1);
6936         }
6937         if ((timers & PACE_TMR_RACK) &&
6938             rack->rc_on_min_to) {
6939                 /*
6940                  * For the rack timer when we
6941                  * are on a min-timeout (which means rrr_conf = 3)
6942                  * we don't want to check the timer. It may
6943                  * be going off for a pace and thats ok we
6944                  * want to send the retransmit (if its ready).
6945                  *
6946                  * If its on a normal rack timer (non-min) then
6947                  * we will check if its expired.
6948                  */
6949                 goto skip_time_check;
6950         }
6951         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6952                 uint32_t left;
6953
6954                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6955                         ret = -1;
6956                         rack_log_to_processing(rack, cts, ret, 0);
6957                         return (0);
6958                 }
6959                 if (hpts_calling == 0) {
6960                         /*
6961                          * A user send or queued mbuf (sack) has called us? We
6962                          * return 0 and let the pacing guards
6963                          * deal with it if they should or
6964                          * should not cause a send.
6965                          */
6966                         ret = -2;
6967                         rack_log_to_processing(rack, cts, ret, 0);
6968                         return (0);
6969                 }
6970                 /*
6971                  * Ok our timer went off early and we are not paced false
6972                  * alarm, go back to sleep.
6973                  */
6974                 ret = -3;
6975                 left = rack->r_ctl.rc_timer_exp - cts;
6976                 tcp_hpts_insert(tptoinpcb(tp), HPTS_MS_TO_SLOTS(left));
6977                 rack_log_to_processing(rack, cts, ret, left);
6978                 return (1);
6979         }
6980 skip_time_check:
6981         rack->rc_tmr_stopped = 0;
6982         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6983         if (timers & PACE_TMR_DELACK) {
6984                 ret = rack_timeout_delack(tp, rack, cts);
6985         } else if (timers & PACE_TMR_RACK) {
6986                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6987                 rack->r_fast_output = 0;
6988                 ret = rack_timeout_rack(tp, rack, cts);
6989         } else if (timers & PACE_TMR_TLP) {
6990                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6991                 ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
6992         } else if (timers & PACE_TMR_RXT) {
6993                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6994                 rack->r_fast_output = 0;
6995                 ret = rack_timeout_rxt(tp, rack, cts);
6996         } else if (timers & PACE_TMR_PERSIT) {
6997                 ret = rack_timeout_persist(tp, rack, cts);
6998         } else if (timers & PACE_TMR_KEEP) {
6999                 ret = rack_timeout_keepalive(tp, rack, cts);
7000         }
7001         rack_log_to_processing(rack, cts, ret, timers);
7002         return (ret);
7003 }
7004
7005 static void
7006 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
7007 {
7008         struct timeval tv;
7009         uint32_t us_cts, flags_on_entry;
7010         uint8_t hpts_removed = 0;
7011
7012         flags_on_entry = rack->r_ctl.rc_hpts_flags;
7013         us_cts = tcp_get_usecs(&tv);
7014         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
7015             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
7016              ((tp->snd_max - tp->snd_una) == 0))) {
7017                 tcp_hpts_remove(rack->rc_inp);
7018                 hpts_removed = 1;
7019                 /* If we were not delayed cancel out the flag. */
7020                 if ((tp->snd_max - tp->snd_una) == 0)
7021                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7022                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7023         }
7024         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7025                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7026                 if (tcp_in_hpts(rack->rc_inp) &&
7027                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7028                         /*
7029                          * Canceling timer's when we have no output being
7030                          * paced. We also must remove ourselves from the
7031                          * hpts.
7032                          */
7033                         tcp_hpts_remove(rack->rc_inp);
7034                         hpts_removed = 1;
7035                 }
7036                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7037         }
7038         if (hpts_removed == 0)
7039                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7040 }
7041
7042 static int
7043 rack_stopall(struct tcpcb *tp)
7044 {
7045         struct tcp_rack *rack;
7046         rack = (struct tcp_rack *)tp->t_fb_ptr;
7047         rack->t_timers_stopped = 1;
7048         return (0);
7049 }
7050
7051 static void
7052 rack_stop_all_timers(struct tcpcb *tp)
7053 {
7054         struct tcp_rack *rack;
7055
7056         /*
7057          * Assure no timers are running.
7058          */
7059         if (tcp_timer_active(tp, TT_PERSIST)) {
7060                 /* We enter in persists, set the flag appropriately */
7061                 rack = (struct tcp_rack *)tp->t_fb_ptr;
7062                 rack->rc_in_persist = 1;
7063         }
7064 }
7065
7066 static void
7067 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7068     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7069 {
7070         int32_t idx;
7071
7072         rsm->r_rtr_cnt++;
7073         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7074         rsm->r_dupack = 0;
7075         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7076                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7077                 rsm->r_flags |= RACK_OVERMAX;
7078         }
7079         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7080                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7081                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7082         }
7083         idx = rsm->r_rtr_cnt - 1;
7084         rsm->r_tim_lastsent[idx] = ts;
7085         /*
7086          * Here we don't add in the len of send, since its already
7087          * in snduna <->snd_max.
7088          */
7089         rsm->r_fas = ctf_flight_size(rack->rc_tp,
7090                                      rack->r_ctl.rc_sacked);
7091         if (rsm->r_flags & RACK_ACKED) {
7092                 /* Problably MTU discovery messing with us */
7093                 rsm->r_flags &= ~RACK_ACKED;
7094                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7095         }
7096         if (rsm->r_in_tmap) {
7097                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7098                 rsm->r_in_tmap = 0;
7099         }
7100         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7101         rsm->r_in_tmap = 1;
7102         /* Take off the must retransmit flag, if its on */
7103         if (rsm->r_flags & RACK_MUST_RXT) {
7104                 if (rack->r_must_retran)
7105                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
7106                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
7107                         /*
7108                          * We have retransmitted all we need. Clear
7109                          * any must retransmit flags.
7110                          */
7111                         rack->r_must_retran = 0;
7112                         rack->r_ctl.rc_out_at_rto = 0;
7113                 }
7114                 rsm->r_flags &= ~RACK_MUST_RXT;
7115         }
7116         if (rsm->r_flags & RACK_SACK_PASSED) {
7117                 /* We have retransmitted due to the SACK pass */
7118                 rsm->r_flags &= ~RACK_SACK_PASSED;
7119                 rsm->r_flags |= RACK_WAS_SACKPASS;
7120         }
7121 }
7122
7123 static uint32_t
7124 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7125     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7126 {
7127         /*
7128          * We (re-)transmitted starting at rsm->r_start for some length
7129          * (possibly less than r_end.
7130          */
7131         struct rack_sendmap *nrsm;
7132 #ifdef INVARIANTS
7133         struct rack_sendmap *insret;
7134 #endif
7135         uint32_t c_end;
7136         int32_t len;
7137
7138         len = *lenp;
7139         c_end = rsm->r_start + len;
7140         if (SEQ_GEQ(c_end, rsm->r_end)) {
7141                 /*
7142                  * We retransmitted the whole piece or more than the whole
7143                  * slopping into the next rsm.
7144                  */
7145                 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7146                 if (c_end == rsm->r_end) {
7147                         *lenp = 0;
7148                         return (0);
7149                 } else {
7150                         int32_t act_len;
7151
7152                         /* Hangs over the end return whats left */
7153                         act_len = rsm->r_end - rsm->r_start;
7154                         *lenp = (len - act_len);
7155                         return (rsm->r_end);
7156                 }
7157                 /* We don't get out of this block. */
7158         }
7159         /*
7160          * Here we retransmitted less than the whole thing which means we
7161          * have to split this into what was transmitted and what was not.
7162          */
7163         nrsm = rack_alloc_full_limit(rack);
7164         if (nrsm == NULL) {
7165                 /*
7166                  * We can't get memory, so lets not proceed.
7167                  */
7168                 *lenp = 0;
7169                 return (0);
7170         }
7171         /*
7172          * So here we are going to take the original rsm and make it what we
7173          * retransmitted. nrsm will be the tail portion we did not
7174          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7175          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7176          * 1, 6 and the new piece will be 6, 11.
7177          */
7178         rack_clone_rsm(rack, nrsm, rsm, c_end);
7179         nrsm->r_dupack = 0;
7180         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7181 #ifndef INVARIANTS
7182         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7183 #else
7184         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7185         if (insret != NULL) {
7186                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7187                       nrsm, insret, rack, rsm);
7188         }
7189 #endif
7190         if (rsm->r_in_tmap) {
7191                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7192                 nrsm->r_in_tmap = 1;
7193         }
7194         rsm->r_flags &= (~RACK_HAS_FIN);
7195         rack_update_rsm(tp, rack, rsm, ts, add_flag);
7196         /* Log a split of rsm into rsm and nrsm */
7197         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7198         *lenp = 0;
7199         return (0);
7200 }
7201
7202 static void
7203 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7204                 uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
7205                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7206 {
7207         struct tcp_rack *rack;
7208         struct rack_sendmap *rsm, *nrsm, fe;
7209 #ifdef INVARIANTS
7210         struct rack_sendmap *insret;
7211 #endif
7212         register uint32_t snd_max, snd_una;
7213
7214         /*
7215          * Add to the RACK log of packets in flight or retransmitted. If
7216          * there is a TS option we will use the TS echoed, if not we will
7217          * grab a TS.
7218          *
7219          * Retransmissions will increment the count and move the ts to its
7220          * proper place. Note that if options do not include TS's then we
7221          * won't be able to effectively use the ACK for an RTT on a retran.
7222          *
7223          * Notes about r_start and r_end. Lets consider a send starting at
7224          * sequence 1 for 10 bytes. In such an example the r_start would be
7225          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7226          * This means that r_end is actually the first sequence for the next
7227          * slot (11).
7228          *
7229          */
7230         /*
7231          * If err is set what do we do XXXrrs? should we not add the thing?
7232          * -- i.e. return if err != 0 or should we pretend we sent it? --
7233          * i.e. proceed with add ** do this for now.
7234          */
7235         INP_WLOCK_ASSERT(tptoinpcb(tp));
7236         if (err)
7237                 /*
7238                  * We don't log errors -- we could but snd_max does not
7239                  * advance in this case either.
7240                  */
7241                 return;
7242
7243         if (th_flags & TH_RST) {
7244                 /*
7245                  * We don't log resets and we return immediately from
7246                  * sending
7247                  */
7248                 return;
7249         }
7250         rack = (struct tcp_rack *)tp->t_fb_ptr;
7251         snd_una = tp->snd_una;
7252         snd_max = tp->snd_max;
7253         if (th_flags & (TH_SYN | TH_FIN)) {
7254                 /*
7255                  * The call to rack_log_output is made before bumping
7256                  * snd_max. This means we can record one extra byte on a SYN
7257                  * or FIN if seq_out is adding more on and a FIN is present
7258                  * (and we are not resending).
7259                  */
7260                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7261                         len++;
7262                 if (th_flags & TH_FIN)
7263                         len++;
7264                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7265                         /*
7266                          * The add/update as not been done for the FIN/SYN
7267                          * yet.
7268                          */
7269                         snd_max = tp->snd_nxt;
7270                 }
7271         }
7272         if (SEQ_LEQ((seq_out + len), snd_una)) {
7273                 /* Are sending an old segment to induce an ack (keep-alive)? */
7274                 return;
7275         }
7276         if (SEQ_LT(seq_out, snd_una)) {
7277                 /* huh? should we panic? */
7278                 uint32_t end;
7279
7280                 end = seq_out + len;
7281                 seq_out = snd_una;
7282                 if (SEQ_GEQ(end, seq_out))
7283                         len = end - seq_out;
7284                 else
7285                         len = 0;
7286         }
7287         if (len == 0) {
7288                 /* We don't log zero window probes */
7289                 return;
7290         }
7291         if (IN_FASTRECOVERY(tp->t_flags)) {
7292                 rack->r_ctl.rc_prr_out += len;
7293         }
7294         /* First question is it a retransmission or new? */
7295         if (seq_out == snd_max) {
7296                 /* Its new */
7297 again:
7298                 rsm = rack_alloc(rack);
7299                 if (rsm == NULL) {
7300                         /*
7301                          * Hmm out of memory and the tcb got destroyed while
7302                          * we tried to wait.
7303                          */
7304                         return;
7305                 }
7306                 if (th_flags & TH_FIN) {
7307                         rsm->r_flags = RACK_HAS_FIN|add_flag;
7308                 } else {
7309                         rsm->r_flags = add_flag;
7310                 }
7311                 if (hw_tls)
7312                         rsm->r_hw_tls = 1;
7313                 rsm->r_tim_lastsent[0] = cts;
7314                 rsm->r_rtr_cnt = 1;
7315                 rsm->r_rtr_bytes = 0;
7316                 if (th_flags & TH_SYN) {
7317                         /* The data space is one beyond snd_una */
7318                         rsm->r_flags |= RACK_HAS_SYN;
7319                 }
7320                 rsm->r_start = seq_out;
7321                 rsm->r_end = rsm->r_start + len;
7322                 rsm->r_dupack = 0;
7323                 /*
7324                  * save off the mbuf location that
7325                  * sndmbuf_noadv returned (which is
7326                  * where we started copying from)..
7327                  */
7328                 rsm->m = s_mb;
7329                 rsm->soff = s_moff;
7330                 /*
7331                  * Here we do add in the len of send, since its not yet
7332                  * reflected in in snduna <->snd_max
7333                  */
7334                 rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7335                                               rack->r_ctl.rc_sacked) +
7336                               (rsm->r_end - rsm->r_start));
7337                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7338                 if (rsm->m) {
7339                         if (rsm->m->m_len <= rsm->soff) {
7340                                 /*
7341                                  * XXXrrs Question, will this happen?
7342                                  *
7343                                  * If sbsndptr is set at the correct place
7344                                  * then s_moff should always be somewhere
7345                                  * within rsm->m. But if the sbsndptr was
7346                                  * off then that won't be true. If it occurs
7347                                  * we need to walkout to the correct location.
7348                                  */
7349                                 struct mbuf *lm;
7350
7351                                 lm = rsm->m;
7352                                 while (lm->m_len <= rsm->soff) {
7353                                         rsm->soff -= lm->m_len;
7354                                         lm = lm->m_next;
7355                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7356                                                              __func__, rack, s_moff, s_mb, rsm->soff));
7357                                 }
7358                                 rsm->m = lm;
7359                         }
7360                         rsm->orig_m_len = rsm->m->m_len;
7361                 } else
7362                         rsm->orig_m_len = 0;
7363                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7364                 /* Log a new rsm */
7365                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7366 #ifndef INVARIANTS
7367                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7368 #else
7369                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7370                 if (insret != NULL) {
7371                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7372                               nrsm, insret, rack, rsm);
7373                 }
7374 #endif
7375                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7376                 rsm->r_in_tmap = 1;
7377                 /*
7378                  * Special case detection, is there just a single
7379                  * packet outstanding when we are not in recovery?
7380                  *
7381                  * If this is true mark it so.
7382                  */
7383                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7384                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7385                         struct rack_sendmap *prsm;
7386
7387                         prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7388                         if (prsm)
7389                                 prsm->r_one_out_nr = 1;
7390                 }
7391                 return;
7392         }
7393         /*
7394          * If we reach here its a retransmission and we need to find it.
7395          */
7396         memset(&fe, 0, sizeof(fe));
7397 more:
7398         if (hintrsm && (hintrsm->r_start == seq_out)) {
7399                 rsm = hintrsm;
7400                 hintrsm = NULL;
7401         } else {
7402                 /* No hints sorry */
7403                 rsm = NULL;
7404         }
7405         if ((rsm) && (rsm->r_start == seq_out)) {
7406                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7407                 if (len == 0) {
7408                         return;
7409                 } else {
7410                         goto more;
7411                 }
7412         }
7413         /* Ok it was not the last pointer go through it the hard way. */
7414 refind:
7415         fe.r_start = seq_out;
7416         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7417         if (rsm) {
7418                 if (rsm->r_start == seq_out) {
7419                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7420                         if (len == 0) {
7421                                 return;
7422                         } else {
7423                                 goto refind;
7424                         }
7425                 }
7426                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7427                         /* Transmitted within this piece */
7428                         /*
7429                          * Ok we must split off the front and then let the
7430                          * update do the rest
7431                          */
7432                         nrsm = rack_alloc_full_limit(rack);
7433                         if (nrsm == NULL) {
7434                                 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7435                                 return;
7436                         }
7437                         /*
7438                          * copy rsm to nrsm and then trim the front of rsm
7439                          * to not include this part.
7440                          */
7441                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
7442                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7443 #ifndef INVARIANTS
7444                         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7445 #else
7446                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7447                         if (insret != NULL) {
7448                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7449                                       nrsm, insret, rack, rsm);
7450                         }
7451 #endif
7452                         if (rsm->r_in_tmap) {
7453                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7454                                 nrsm->r_in_tmap = 1;
7455                         }
7456                         rsm->r_flags &= (~RACK_HAS_FIN);
7457                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7458                         if (len == 0) {
7459                                 return;
7460                         } else if (len > 0)
7461                                 goto refind;
7462                 }
7463         }
7464         /*
7465          * Hmm not found in map did they retransmit both old and on into the
7466          * new?
7467          */
7468         if (seq_out == tp->snd_max) {
7469                 goto again;
7470         } else if (SEQ_LT(seq_out, tp->snd_max)) {
7471 #ifdef INVARIANTS
7472                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7473                        seq_out, len, tp->snd_una, tp->snd_max);
7474                 printf("Starting Dump of all rack entries\n");
7475                 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7476                         printf("rsm:%p start:%u end:%u\n",
7477                                rsm, rsm->r_start, rsm->r_end);
7478                 }
7479                 printf("Dump complete\n");
7480                 panic("seq_out not found rack:%p tp:%p",
7481                       rack, tp);
7482 #endif
7483         } else {
7484 #ifdef INVARIANTS
7485                 /*
7486                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
7487                  * flag)
7488                  */
7489                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7490                       seq_out, len, tp->snd_max, tp);
7491 #endif
7492         }
7493 }
7494
7495 /*
7496  * Record one of the RTT updates from an ack into
7497  * our sample structure.
7498  */
7499
7500 static void
7501 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7502                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7503 {
7504         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7505             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7506                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7507         }
7508         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7509             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7510                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7511         }
7512         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7513             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7514                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7515             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7516                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7517         }
7518         if ((confidence == 1) &&
7519             ((rsm == NULL) ||
7520              (rsm->r_just_ret) ||
7521              (rsm->r_one_out_nr &&
7522               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7523                 /*
7524                  * If the rsm had a just return
7525                  * hit it then we can't trust the
7526                  * rtt measurement for buffer deterimination
7527                  * Note that a confidence of 2, indicates
7528                  * SACK'd which overrides the r_just_ret or
7529                  * the r_one_out_nr. If it was a CUM-ACK and
7530                  * we had only two outstanding, but get an
7531                  * ack for only 1. Then that also lowers our
7532                  * confidence.
7533                  */
7534                 confidence = 0;
7535         }
7536         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7537             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7538                 if (rack->r_ctl.rack_rs.confidence == 0) {
7539                         /*
7540                          * We take anything with no current confidence
7541                          * saved.
7542                          */
7543                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7544                         rack->r_ctl.rack_rs.confidence = confidence;
7545                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7546                 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7547                         /*
7548                          * Once we have a confident number,
7549                          * we can update it with a smaller
7550                          * value since this confident number
7551                          * may include the DSACK time until
7552                          * the next segment (the second one) arrived.
7553                          */
7554                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7555                         rack->r_ctl.rack_rs.confidence = confidence;
7556                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7557                 }
7558         }
7559         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7560         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7561         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7562         rack->r_ctl.rack_rs.rs_rtt_cnt++;
7563 }
7564
7565 /*
7566  * Collect new round-trip time estimate
7567  * and update averages and current timeout.
7568  */
7569 static void
7570 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7571 {
7572         int32_t delta;
7573         int32_t rtt;
7574
7575         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7576                 /* No valid sample */
7577                 return;
7578         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7579                 /* We are to use the lowest RTT seen in a single ack */
7580                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7581         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7582                 /* We are to use the highest RTT seen in a single ack */
7583                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7584         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7585                 /* We are to use the average RTT seen in a single ack */
7586                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7587                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7588         } else {
7589 #ifdef INVARIANTS
7590                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7591 #endif
7592                 return;
7593         }
7594         if (rtt == 0)
7595                 rtt = 1;
7596         if (rack->rc_gp_rtt_set == 0) {
7597                 /*
7598                  * With no RTT we have to accept
7599                  * even one we are not confident of.
7600                  */
7601                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7602                 rack->rc_gp_rtt_set = 1;
7603         } else if (rack->r_ctl.rack_rs.confidence) {
7604                 /* update the running gp srtt */
7605                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7606                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7607         }
7608         if (rack->r_ctl.rack_rs.confidence) {
7609                 /*
7610                  * record the low and high for highly buffered path computation,
7611                  * we only do this if we are confident (not a retransmission).
7612                  */
7613                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7614                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7615                 }
7616                 if (rack->rc_highly_buffered == 0) {
7617                         /*
7618                          * Currently once we declare a path has
7619                          * highly buffered there is no going
7620                          * back, which may be a problem...
7621                          */
7622                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7623                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7624                                                      rack->r_ctl.rc_highest_us_rtt,
7625                                                      rack->r_ctl.rc_lowest_us_rtt,
7626                                                      RACK_RTTS_SEEHBP);
7627                                 rack->rc_highly_buffered = 1;
7628                         }
7629                 }
7630         }
7631         if ((rack->r_ctl.rack_rs.confidence) ||
7632             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7633                 /*
7634                  * If we are highly confident of it <or> it was
7635                  * never retransmitted we accept it as the last us_rtt.
7636                  */
7637                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7638                 /* The lowest rtt can be set if its was not retransmited */
7639                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7640                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7641                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
7642                                 rack->r_ctl.rc_lowest_us_rtt = 1;
7643                 }
7644         }
7645         rack = (struct tcp_rack *)tp->t_fb_ptr;
7646         if (tp->t_srtt != 0) {
7647                 /*
7648                  * We keep a simple srtt in microseconds, like our rtt
7649                  * measurement. We don't need to do any tricks with shifting
7650                  * etc. Instead we just add in 1/8th of the new measurement
7651                  * and subtract out 1/8 of the old srtt. We do the same with
7652                  * the variance after finding the absolute value of the
7653                  * difference between this sample and the current srtt.
7654                  */
7655                 delta = tp->t_srtt - rtt;
7656                 /* Take off 1/8th of the current sRTT */
7657                 tp->t_srtt -= (tp->t_srtt >> 3);
7658                 /* Add in 1/8th of the new RTT just measured */
7659                 tp->t_srtt += (rtt >> 3);
7660                 if (tp->t_srtt <= 0)
7661                         tp->t_srtt = 1;
7662                 /* Now lets make the absolute value of the variance */
7663                 if (delta < 0)
7664                         delta = -delta;
7665                 /* Subtract out 1/8th */
7666                 tp->t_rttvar -= (tp->t_rttvar >> 3);
7667                 /* Add in 1/8th of the new variance we just saw */
7668                 tp->t_rttvar += (delta >> 3);
7669                 if (tp->t_rttvar <= 0)
7670                         tp->t_rttvar = 1;
7671         } else {
7672                 /*
7673                  * No rtt measurement yet - use the unsmoothed rtt. Set the
7674                  * variance to half the rtt (so our first retransmit happens
7675                  * at 3*rtt).
7676                  */
7677                 tp->t_srtt = rtt;
7678                 tp->t_rttvar = rtt >> 1;
7679         }
7680         rack->rc_srtt_measure_made = 1;
7681         KMOD_TCPSTAT_INC(tcps_rttupdated);
7682         tp->t_rttupdated++;
7683 #ifdef STATS
7684         if (rack_stats_gets_ms_rtt == 0) {
7685                 /* Send in the microsecond rtt used for rxt timeout purposes */
7686                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7687         } else if (rack_stats_gets_ms_rtt == 1) {
7688                 /* Send in the millisecond rtt used for rxt timeout purposes */
7689                 int32_t ms_rtt;
7690
7691                 /* Round up */
7692                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7693                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7694         } else if (rack_stats_gets_ms_rtt == 2) {
7695                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
7696                 int32_t ms_rtt;
7697
7698                 /* Round up */
7699                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7700                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7701         }  else {
7702                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
7703                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7704         }
7705
7706 #endif
7707         /*
7708          * the retransmit should happen at rtt + 4 * rttvar. Because of the
7709          * way we do the smoothing, srtt and rttvar will each average +1/2
7710          * tick of bias.  When we compute the retransmit timer, we want 1/2
7711          * tick of rounding and 1 extra tick because of +-1/2 tick
7712          * uncertainty in the firing of the timer.  The bias will give us
7713          * exactly the 1.5 tick we need.  But, because the bias is
7714          * statistical, we have to test that we don't drop below the minimum
7715          * feasible timer (which is 2 ticks).
7716          */
7717         tp->t_rxtshift = 0;
7718         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7719                       max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7720         rack_log_rtt_sample(rack, rtt);
7721         tp->t_softerror = 0;
7722 }
7723
7724
7725 static void
7726 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7727 {
7728         /*
7729          * Apply to filter the inbound us-rtt at us_cts.
7730          */
7731         uint32_t old_rtt;
7732
7733         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7734         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7735                                us_rtt, us_cts);
7736         if (old_rtt > us_rtt) {
7737                 /* We just hit a new lower rtt time */
7738                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7739                                      __LINE__, RACK_RTTS_NEWRTT);
7740                 /*
7741                  * Only count it if its lower than what we saw within our
7742                  * calculated range.
7743                  */
7744                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7745                         if (rack_probertt_lower_within &&
7746                             rack->rc_gp_dyn_mul &&
7747                             (rack->use_fixed_rate == 0) &&
7748                             (rack->rc_always_pace)) {
7749                                 /*
7750                                  * We are seeing a new lower rtt very close
7751                                  * to the time that we would have entered probe-rtt.
7752                                  * This is probably due to the fact that a peer flow
7753                                  * has entered probe-rtt. Lets go in now too.
7754                                  */
7755                                 uint32_t val;
7756
7757                                 val = rack_probertt_lower_within * rack_time_between_probertt;
7758                                 val /= 100;
7759                                 if ((rack->in_probe_rtt == 0)  &&
7760                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7761                                         rack_enter_probertt(rack, us_cts);
7762                                 }
7763                         }
7764                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7765                 }
7766         }
7767 }
7768
7769 static int
7770 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7771     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7772 {
7773         uint32_t us_rtt;
7774         int32_t i, all;
7775         uint32_t t, len_acked;
7776
7777         if ((rsm->r_flags & RACK_ACKED) ||
7778             (rsm->r_flags & RACK_WAS_ACKED))
7779                 /* Already done */
7780                 return (0);
7781         if (rsm->r_no_rtt_allowed) {
7782                 /* Not allowed */
7783                 return (0);
7784         }
7785         if (ack_type == CUM_ACKED) {
7786                 if (SEQ_GT(th_ack, rsm->r_end)) {
7787                         len_acked = rsm->r_end - rsm->r_start;
7788                         all = 1;
7789                 } else {
7790                         len_acked = th_ack - rsm->r_start;
7791                         all = 0;
7792                 }
7793         } else {
7794                 len_acked = rsm->r_end - rsm->r_start;
7795                 all = 0;
7796         }
7797         if (rsm->r_rtr_cnt == 1) {
7798
7799                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7800                 if ((int)t <= 0)
7801                         t = 1;
7802                 if (!tp->t_rttlow || tp->t_rttlow > t)
7803                         tp->t_rttlow = t;
7804                 if (!rack->r_ctl.rc_rack_min_rtt ||
7805                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7806                         rack->r_ctl.rc_rack_min_rtt = t;
7807                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7808                                 rack->r_ctl.rc_rack_min_rtt = 1;
7809                         }
7810                 }
7811                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7812                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7813                 else
7814                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7815                 if (us_rtt == 0)
7816                         us_rtt = 1;
7817                 if (CC_ALGO(tp)->rttsample != NULL) {
7818                         /* Kick the RTT to the CC */
7819                         CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
7820                 }
7821                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7822                 if (ack_type == SACKED) {
7823                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7824                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7825                 } else {
7826                         /*
7827                          * We need to setup what our confidence
7828                          * is in this ack.
7829                          *
7830                          * If the rsm was app limited and it is
7831                          * less than a mss in length (the end
7832                          * of the send) then we have a gap. If we
7833                          * were app limited but say we were sending
7834                          * multiple MSS's then we are more confident
7835                          * int it.
7836                          *
7837                          * When we are not app-limited then we see if
7838                          * the rsm is being included in the current
7839                          * measurement, we tell this by the app_limited_needs_set
7840                          * flag.
7841                          *
7842                          * Note that being cwnd blocked is not applimited
7843                          * as well as the pacing delay between packets which
7844                          * are sending only 1 or 2 MSS's also will show up
7845                          * in the RTT. We probably need to examine this algorithm
7846                          * a bit more and enhance it to account for the delay
7847                          * between rsm's. We could do that by saving off the
7848                          * pacing delay of each rsm (in an rsm) and then
7849                          * factoring that in somehow though for now I am
7850                          * not sure how :)
7851                          */
7852                         int calc_conf = 0;
7853
7854                         if (rsm->r_flags & RACK_APP_LIMITED) {
7855                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7856                                         calc_conf = 0;
7857                                 else
7858                                         calc_conf = 1;
7859                         } else if (rack->app_limited_needs_set == 0) {
7860                                 calc_conf = 1;
7861                         } else {
7862                                 calc_conf = 0;
7863                         }
7864                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7865                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7866                                             calc_conf, rsm, rsm->r_rtr_cnt);
7867                 }
7868                 if ((rsm->r_flags & RACK_TLP) &&
7869                     (!IN_FASTRECOVERY(tp->t_flags))) {
7870                         /* Segment was a TLP and our retrans matched */
7871                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7872                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
7873                         }
7874                 }
7875                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7876                         /* New more recent rack_tmit_time */
7877                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7878                         rack->rc_rack_rtt = t;
7879                 }
7880                 return (1);
7881         }
7882         /*
7883          * We clear the soft/rxtshift since we got an ack.
7884          * There is no assurance we will call the commit() function
7885          * so we need to clear these to avoid incorrect handling.
7886          */
7887         tp->t_rxtshift = 0;
7888         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7889                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7890         tp->t_softerror = 0;
7891         if (to && (to->to_flags & TOF_TS) &&
7892             (ack_type == CUM_ACKED) &&
7893             (to->to_tsecr) &&
7894             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7895                 /*
7896                  * Now which timestamp does it match? In this block the ACK
7897                  * must be coming from a previous transmission.
7898                  */
7899                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7900                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7901                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7902                                 if ((int)t <= 0)
7903                                         t = 1;
7904                                 if (CC_ALGO(tp)->rttsample != NULL) {
7905                                         /*
7906                                          * Kick the RTT to the CC, here
7907                                          * we lie a bit in that we know the
7908                                          * retransmission is correct even though
7909                                          * we retransmitted. This is because
7910                                          * we match the timestamps.
7911                                          */
7912                                         if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
7913                                                 us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
7914                                         else
7915                                                 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
7916                                         CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
7917                                 }
7918                                 if ((i + 1) < rsm->r_rtr_cnt) {
7919                                         /*
7920                                          * The peer ack'd from our previous
7921                                          * transmission. We have a spurious
7922                                          * retransmission and thus we dont
7923                                          * want to update our rack_rtt.
7924                                          *
7925                                          * Hmm should there be a CC revert here?
7926                                          *
7927                                          */
7928                                         return (0);
7929                                 }
7930                                 if (!tp->t_rttlow || tp->t_rttlow > t)
7931                                         tp->t_rttlow = t;
7932                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7933                                         rack->r_ctl.rc_rack_min_rtt = t;
7934                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7935                                                 rack->r_ctl.rc_rack_min_rtt = 1;
7936                                         }
7937                                 }
7938                                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7939                                            (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7940                                         /* New more recent rack_tmit_time */
7941                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7942                                         rack->rc_rack_rtt = t;
7943                                 }
7944                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7945                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7946                                                     rsm->r_rtr_cnt);
7947                                 return (1);
7948                         }
7949                 }
7950                 goto ts_not_found;
7951         } else {
7952                 /*
7953                  * Ok its a SACK block that we retransmitted. or a windows
7954                  * machine without timestamps. We can tell nothing from the
7955                  * time-stamp since its not there or the time the peer last
7956                  * recieved a segment that moved forward its cum-ack point.
7957                  */
7958 ts_not_found:
7959                 i = rsm->r_rtr_cnt - 1;
7960                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7961                 if ((int)t <= 0)
7962                         t = 1;
7963                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7964                         /*
7965                          * We retransmitted and the ack came back in less
7966                          * than the smallest rtt we have observed. We most
7967                          * likely did an improper retransmit as outlined in
7968                          * 6.2 Step 2 point 2 in the rack-draft so we
7969                          * don't want to update our rack_rtt. We in
7970                          * theory (in future) might want to think about reverting our
7971                          * cwnd state but we won't for now.
7972                          */
7973                         return (0);
7974                 } else if (rack->r_ctl.rc_rack_min_rtt) {
7975                         /*
7976                          * We retransmitted it and the retransmit did the
7977                          * job.
7978                          */
7979                         if (!rack->r_ctl.rc_rack_min_rtt ||
7980                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7981                                 rack->r_ctl.rc_rack_min_rtt = t;
7982                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7983                                         rack->r_ctl.rc_rack_min_rtt = 1;
7984                                 }
7985                         }
7986                         if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7987                                 /* New more recent rack_tmit_time */
7988                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7989                                 rack->rc_rack_rtt = t;
7990                         }
7991                         return (1);
7992                 }
7993         }
7994         return (0);
7995 }
7996
7997 /*
7998  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7999  */
8000 static void
8001 rack_log_sack_passed(struct tcpcb *tp,
8002     struct tcp_rack *rack, struct rack_sendmap *rsm)
8003 {
8004         struct rack_sendmap *nrsm;
8005
8006         nrsm = rsm;
8007         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
8008             rack_head, r_tnext) {
8009                 if (nrsm == rsm) {
8010                         /* Skip orginal segment he is acked */
8011                         continue;
8012                 }
8013                 if (nrsm->r_flags & RACK_ACKED) {
8014                         /*
8015                          * Skip ack'd segments, though we
8016                          * should not see these, since tmap
8017                          * should not have ack'd segments.
8018                          */
8019                         continue;
8020                 }
8021                 if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
8022                         /*
8023                          * If the peer dropped the rwnd on
8024                          * these then we don't worry about them.
8025                          */
8026                         continue;
8027                 }
8028                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8029                         /*
8030                          * We found one that is already marked
8031                          * passed, we have been here before and
8032                          * so all others below this are marked.
8033                          */
8034                         break;
8035                 }
8036                 nrsm->r_flags |= RACK_SACK_PASSED;
8037                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8038         }
8039 }
8040
8041 static void
8042 rack_need_set_test(struct tcpcb *tp,
8043                    struct tcp_rack *rack,
8044                    struct rack_sendmap *rsm,
8045                    tcp_seq th_ack,
8046                    int line,
8047                    int use_which)
8048 {
8049
8050         if ((tp->t_flags & TF_GPUTINPROG) &&
8051             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8052                 /*
8053                  * We were app limited, and this ack
8054                  * butts up or goes beyond the point where we want
8055                  * to start our next measurement. We need
8056                  * to record the new gput_ts as here and
8057                  * possibly update the start sequence.
8058                  */
8059                 uint32_t seq, ts;
8060
8061                 if (rsm->r_rtr_cnt > 1) {
8062                         /*
8063                          * This is a retransmit, can we
8064                          * really make any assessment at this
8065                          * point?  We are not really sure of
8066                          * the timestamp, is it this or the
8067                          * previous transmission?
8068                          *
8069                          * Lets wait for something better that
8070                          * is not retransmitted.
8071                          */
8072                         return;
8073                 }
8074                 seq = tp->gput_seq;
8075                 ts = tp->gput_ts;
8076                 rack->app_limited_needs_set = 0;
8077                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8078                 /* Do we start at a new end? */
8079                 if ((use_which == RACK_USE_BEG) &&
8080                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8081                         /*
8082                          * When we get an ACK that just eats
8083                          * up some of the rsm, we set RACK_USE_BEG
8084                          * since whats at r_start (i.e. th_ack)
8085                          * is left unacked and thats where the
8086                          * measurement not starts.
8087                          */
8088                         tp->gput_seq = rsm->r_start;
8089                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8090                 }
8091                 if ((use_which == RACK_USE_END) &&
8092                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8093                             /*
8094                              * We use the end when the cumack
8095                              * is moving forward and completely
8096                              * deleting the rsm passed so basically
8097                              * r_end holds th_ack.
8098                              *
8099                              * For SACK's we also want to use the end
8100                              * since this piece just got sacked and
8101                              * we want to target anything after that
8102                              * in our measurement.
8103                              */
8104                             tp->gput_seq = rsm->r_end;
8105                             rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8106                 }
8107                 if (use_which == RACK_USE_END_OR_THACK) {
8108                         /*
8109                          * special case for ack moving forward,
8110                          * not a sack, we need to move all the
8111                          * way up to where this ack cum-ack moves
8112                          * to.
8113                          */
8114                         if (SEQ_GT(th_ack, rsm->r_end))
8115                                 tp->gput_seq = th_ack;
8116                         else
8117                                 tp->gput_seq = rsm->r_end;
8118                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8119                 }
8120                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8121                         /*
8122                          * We moved beyond this guy's range, re-calculate
8123                          * the new end point.
8124                          */
8125                         if (rack->rc_gp_filled == 0) {
8126                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8127                         } else {
8128                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8129                         }
8130                 }
8131                 /*
8132                  * We are moving the goal post, we may be able to clear the
8133                  * measure_saw_probe_rtt flag.
8134                  */
8135                 if ((rack->in_probe_rtt == 0) &&
8136                     (rack->measure_saw_probe_rtt) &&
8137                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8138                         rack->measure_saw_probe_rtt = 0;
8139                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8140                                            seq, tp->gput_seq, 0, 5, line, NULL, 0);
8141                 if (rack->rc_gp_filled &&
8142                     ((tp->gput_ack - tp->gput_seq) <
8143                      max(rc_init_window(rack), (MIN_GP_WIN *
8144                                                 ctf_fixed_maxseg(tp))))) {
8145                         uint32_t ideal_amount;
8146
8147                         ideal_amount = rack_get_measure_window(tp, rack);
8148                         if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
8149                                 /*
8150                                  * There is no sense of continuing this measurement
8151                                  * because its too small to gain us anything we
8152                                  * trust. Skip it and that way we can start a new
8153                                  * measurement quicker.
8154                                  */
8155                                 tp->t_flags &= ~TF_GPUTINPROG;
8156                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8157                                                            0, 0, 0, 6, __LINE__, NULL, 0);
8158                         } else {
8159                                 /*
8160                                  * Reset the window further out.
8161                                  */
8162                                 tp->gput_ack = tp->gput_seq + ideal_amount;
8163                         }
8164                 }
8165         }
8166 }
8167
8168 static inline int
8169 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8170 {
8171         if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8172                 /* Behind our TLP definition or right at */
8173                 return (0);
8174         }
8175         if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8176                 /* The start is beyond or right at our end of TLP definition */
8177                 return (0);
8178         }
8179         /* It has to be a sub-part of the original TLP recorded */
8180         return (1);
8181 }
8182
8183
8184 static uint32_t
8185 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8186                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8187 {
8188         uint32_t start, end, changed = 0;
8189         struct rack_sendmap stack_map;
8190         struct rack_sendmap *rsm, *nrsm, fe, *prev, *next;
8191 #ifdef INVARIANTS
8192         struct rack_sendmap *insret;
8193 #endif
8194         int32_t used_ref = 1;
8195         int moved = 0;
8196
8197         start = sack->start;
8198         end = sack->end;
8199         rsm = *prsm;
8200         memset(&fe, 0, sizeof(fe));
8201 do_rest_ofb:
8202         if ((rsm == NULL) ||
8203             (SEQ_LT(end, rsm->r_start)) ||
8204             (SEQ_GEQ(start, rsm->r_end)) ||
8205             (SEQ_LT(start, rsm->r_start))) {
8206                 /*
8207                  * We are not in the right spot,
8208                  * find the correct spot in the tree.
8209                  */
8210                 used_ref = 0;
8211                 fe.r_start = start;
8212                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8213                 moved++;
8214         }
8215         if (rsm == NULL) {
8216                 /* TSNH */
8217                 goto out;
8218         }
8219         /* Ok we have an ACK for some piece of this rsm */
8220         if (rsm->r_start != start) {
8221                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8222                         /*
8223                          * Before any splitting or hookery is
8224                          * done is it a TLP of interest i.e. rxt?
8225                          */
8226                         if ((rsm->r_flags & RACK_TLP) &&
8227                             (rsm->r_rtr_cnt > 1)) {
8228                                 /*
8229                                  * We are splitting a rxt TLP, check
8230                                  * if we need to save off the start/end
8231                                  */
8232                                 if (rack->rc_last_tlp_acked_set &&
8233                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8234                                         /*
8235                                          * We already turned this on since we are inside
8236                                          * the previous one was a partially sack now we
8237                                          * are getting another one (maybe all of it).
8238                                          *
8239                                          */
8240                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8241                                         /*
8242                                          * Lets make sure we have all of it though.
8243                                          */
8244                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8245                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8246                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8247                                                                      rack->r_ctl.last_tlp_acked_end);
8248                                         }
8249                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8250                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8251                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8252                                                                      rack->r_ctl.last_tlp_acked_end);
8253                                         }
8254                                 } else {
8255                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8256                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8257                                         rack->rc_last_tlp_past_cumack = 0;
8258                                         rack->rc_last_tlp_acked_set = 1;
8259                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8260                                 }
8261                         }
8262                         /**
8263                          * Need to split this in two pieces the before and after,
8264                          * the before remains in the map, the after must be
8265                          * added. In other words we have:
8266                          * rsm        |--------------|
8267                          * sackblk        |------->
8268                          * rsm will become
8269                          *     rsm    |---|
8270                          * and nrsm will be  the sacked piece
8271                          *     nrsm       |----------|
8272                          *
8273                          * But before we start down that path lets
8274                          * see if the sack spans over on top of
8275                          * the next guy and it is already sacked.
8276                          *
8277                          */
8278                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8279                         if (next && (next->r_flags & RACK_ACKED) &&
8280                             SEQ_GEQ(end, next->r_start)) {
8281                                 /**
8282                                  * So the next one is already acked, and
8283                                  * we can thus by hookery use our stack_map
8284                                  * to reflect the piece being sacked and
8285                                  * then adjust the two tree entries moving
8286                                  * the start and ends around. So we start like:
8287                                  *  rsm     |------------|             (not-acked)
8288                                  *  next                 |-----------| (acked)
8289                                  *  sackblk        |-------->
8290                                  *  We want to end like so:
8291                                  *  rsm     |------|                   (not-acked)
8292                                  *  next           |-----------------| (acked)
8293                                  *  nrsm           |-----|
8294                                  * Where nrsm is a temporary stack piece we
8295                                  * use to update all the gizmos.
8296                                  */
8297                                 /* Copy up our fudge block */
8298                                 nrsm = &stack_map;
8299                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8300                                 /* Now adjust our tree blocks */
8301                                 rsm->r_end = start;
8302                                 next->r_start = start;
8303                                 /* Now we must adjust back where next->m is */
8304                                 rack_setup_offset_for_rsm(rsm, next);
8305
8306                                 /* We don't need to adjust rsm, it did not change */
8307                                 /* Clear out the dup ack count of the remainder */
8308                                 rsm->r_dupack = 0;
8309                                 rsm->r_just_ret = 0;
8310                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8311                                 /* Now lets make sure our fudge block is right */
8312                                 nrsm->r_start = start;
8313                                 /* Now lets update all the stats and such */
8314                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8315                                 if (rack->app_limited_needs_set)
8316                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8317                                 changed += (nrsm->r_end - nrsm->r_start);
8318                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8319                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8320                                         rack->r_ctl.rc_reorder_ts = cts;
8321                                 }
8322                                 /*
8323                                  * Now we want to go up from rsm (the
8324                                  * one left un-acked) to the next one
8325                                  * in the tmap. We do this so when
8326                                  * we walk backwards we include marking
8327                                  * sack-passed on rsm (The one passed in
8328                                  * is skipped since it is generally called
8329                                  * on something sacked before removing it
8330                                  * from the tmap).
8331                                  */
8332                                 if (rsm->r_in_tmap) {
8333                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
8334                                         /*
8335                                          * Now that we have the next
8336                                          * one walk backwards from there.
8337                                          */
8338                                         if (nrsm && nrsm->r_in_tmap)
8339                                                 rack_log_sack_passed(tp, rack, nrsm);
8340                                 }
8341                                 /* Now are we done? */
8342                                 if (SEQ_LT(end, next->r_end) ||
8343                                     (end == next->r_end)) {
8344                                         /* Done with block */
8345                                         goto out;
8346                                 }
8347                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8348                                 counter_u64_add(rack_sack_used_next_merge, 1);
8349                                 /* Postion for the next block */
8350                                 start = next->r_end;
8351                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8352                                 if (rsm == NULL)
8353                                         goto out;
8354                         } else {
8355                                 /**
8356                                  * We can't use any hookery here, so we
8357                                  * need to split the map. We enter like
8358                                  * so:
8359                                  *  rsm      |--------|
8360                                  *  sackblk       |----->
8361                                  * We will add the new block nrsm and
8362                                  * that will be the new portion, and then
8363                                  * fall through after reseting rsm. So we
8364                                  * split and look like this:
8365                                  *  rsm      |----|
8366                                  *  sackblk       |----->
8367                                  *  nrsm          |---|
8368                                  * We then fall through reseting
8369                                  * rsm to nrsm, so the next block
8370                                  * picks it up.
8371                                  */
8372                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8373                                 if (nrsm == NULL) {
8374                                         /*
8375                                          * failed XXXrrs what can we do but loose the sack
8376                                          * info?
8377                                          */
8378                                         goto out;
8379                                 }
8380                                 counter_u64_add(rack_sack_splits, 1);
8381                                 rack_clone_rsm(rack, nrsm, rsm, start);
8382                                 rsm->r_just_ret = 0;
8383 #ifndef INVARIANTS
8384                                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8385 #else
8386                                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8387                                 if (insret != NULL) {
8388                                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8389                                               nrsm, insret, rack, rsm);
8390                                 }
8391 #endif
8392                                 if (rsm->r_in_tmap) {
8393                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8394                                         nrsm->r_in_tmap = 1;
8395                                 }
8396                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8397                                 rsm->r_flags &= (~RACK_HAS_FIN);
8398                                 /* Position us to point to the new nrsm that starts the sack blk */
8399                                 rsm = nrsm;
8400                         }
8401                 } else {
8402                         /* Already sacked this piece */
8403                         counter_u64_add(rack_sack_skipped_acked, 1);
8404                         moved++;
8405                         if (end == rsm->r_end) {
8406                                 /* Done with block */
8407                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8408                                 goto out;
8409                         } else if (SEQ_LT(end, rsm->r_end)) {
8410                                 /* A partial sack to a already sacked block */
8411                                 moved++;
8412                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8413                                 goto out;
8414                         } else {
8415                                 /*
8416                                  * The end goes beyond this guy
8417                                  * reposition the start to the
8418                                  * next block.
8419                                  */
8420                                 start = rsm->r_end;
8421                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8422                                 if (rsm == NULL)
8423                                         goto out;
8424                         }
8425                 }
8426         }
8427         if (SEQ_GEQ(end, rsm->r_end)) {
8428                 /**
8429                  * The end of this block is either beyond this guy or right
8430                  * at this guy. I.e.:
8431                  *  rsm ---                 |-----|
8432                  *  end                     |-----|
8433                  *  <or>
8434                  *  end                     |---------|
8435                  */
8436                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8437                         /*
8438                          * Is it a TLP of interest?
8439                          */
8440                         if ((rsm->r_flags & RACK_TLP) &&
8441                             (rsm->r_rtr_cnt > 1)) {
8442                                 /*
8443                                  * We are splitting a rxt TLP, check
8444                                  * if we need to save off the start/end
8445                                  */
8446                                 if (rack->rc_last_tlp_acked_set &&
8447                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8448                                         /*
8449                                          * We already turned this on since we are inside
8450                                          * the previous one was a partially sack now we
8451                                          * are getting another one (maybe all of it).
8452                                          */
8453                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8454                                         /*
8455                                          * Lets make sure we have all of it though.
8456                                          */
8457                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8458                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8459                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8460                                                                      rack->r_ctl.last_tlp_acked_end);
8461                                         }
8462                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8463                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8464                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8465                                                                      rack->r_ctl.last_tlp_acked_end);
8466                                         }
8467                                 } else {
8468                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8469                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8470                                         rack->rc_last_tlp_past_cumack = 0;
8471                                         rack->rc_last_tlp_acked_set = 1;
8472                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8473                                 }
8474                         }
8475                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8476                         changed += (rsm->r_end - rsm->r_start);
8477                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8478                         if (rsm->r_in_tmap) /* should be true */
8479                                 rack_log_sack_passed(tp, rack, rsm);
8480                         /* Is Reordering occuring? */
8481                         if (rsm->r_flags & RACK_SACK_PASSED) {
8482                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8483                                 rack->r_ctl.rc_reorder_ts = cts;
8484                         }
8485                         if (rack->app_limited_needs_set)
8486                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8487                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8488                         rsm->r_flags |= RACK_ACKED;
8489                         if (rsm->r_in_tmap) {
8490                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8491                                 rsm->r_in_tmap = 0;
8492                         }
8493                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8494                 } else {
8495                         counter_u64_add(rack_sack_skipped_acked, 1);
8496                         moved++;
8497                 }
8498                 if (end == rsm->r_end) {
8499                         /* This block only - done, setup for next */
8500                         goto out;
8501                 }
8502                 /*
8503                  * There is more not coverend by this rsm move on
8504                  * to the next block in the RB tree.
8505                  */
8506                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8507                 start = rsm->r_end;
8508                 rsm = nrsm;
8509                 if (rsm == NULL)
8510                         goto out;
8511                 goto do_rest_ofb;
8512         }
8513         /**
8514          * The end of this sack block is smaller than
8515          * our rsm i.e.:
8516          *  rsm ---                 |-----|
8517          *  end                     |--|
8518          */
8519         if ((rsm->r_flags & RACK_ACKED) == 0) {
8520                 /*
8521                  * Is it a TLP of interest?
8522                  */
8523                 if ((rsm->r_flags & RACK_TLP) &&
8524                     (rsm->r_rtr_cnt > 1)) {
8525                         /*
8526                          * We are splitting a rxt TLP, check
8527                          * if we need to save off the start/end
8528                          */
8529                         if (rack->rc_last_tlp_acked_set &&
8530                             (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8531                                 /*
8532                                  * We already turned this on since we are inside
8533                                  * the previous one was a partially sack now we
8534                                  * are getting another one (maybe all of it).
8535                                  */
8536                                 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8537                                 /*
8538                                  * Lets make sure we have all of it though.
8539                                  */
8540                                 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8541                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8542                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8543                                                              rack->r_ctl.last_tlp_acked_end);
8544                                 }
8545                                 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8546                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8547                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8548                                                              rack->r_ctl.last_tlp_acked_end);
8549                                 }
8550                         } else {
8551                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8552                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8553                                 rack->rc_last_tlp_past_cumack = 0;
8554                                 rack->rc_last_tlp_acked_set = 1;
8555                                 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8556                         }
8557                 }
8558                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8559                 if (prev &&
8560                     (prev->r_flags & RACK_ACKED)) {
8561                         /**
8562                          * Goal, we want the right remainder of rsm to shrink
8563                          * in place and span from (rsm->r_start = end) to rsm->r_end.
8564                          * We want to expand prev to go all the way
8565                          * to prev->r_end <- end.
8566                          * so in the tree we have before:
8567                          *   prev     |--------|         (acked)
8568                          *   rsm               |-------| (non-acked)
8569                          *   sackblk           |-|
8570                          * We churn it so we end up with
8571                          *   prev     |----------|       (acked)
8572                          *   rsm                 |-----| (non-acked)
8573                          *   nrsm              |-| (temporary)
8574                          *
8575                          * Note if either prev/rsm is a TLP we don't
8576                          * do this.
8577                          */
8578                         nrsm = &stack_map;
8579                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8580                         prev->r_end = end;
8581                         rsm->r_start = end;
8582                         /* Now adjust nrsm (stack copy) to be
8583                          * the one that is the small
8584                          * piece that was "sacked".
8585                          */
8586                         nrsm->r_end = end;
8587                         rsm->r_dupack = 0;
8588                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8589                         /*
8590                          * Now that the rsm has had its start moved forward
8591                          * lets go ahead and get its new place in the world.
8592                          */
8593                         rack_setup_offset_for_rsm(prev, rsm);
8594                         /*
8595                          * Now nrsm is our new little piece
8596                          * that is acked (which was merged
8597                          * to prev). Update the rtt and changed
8598                          * based on that. Also check for reordering.
8599                          */
8600                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8601                         if (rack->app_limited_needs_set)
8602                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8603                         changed += (nrsm->r_end - nrsm->r_start);
8604                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8605                         if (nrsm->r_flags & RACK_SACK_PASSED) {
8606                                 rack->r_ctl.rc_reorder_ts = cts;
8607                         }
8608                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8609                         rsm = prev;
8610                         counter_u64_add(rack_sack_used_prev_merge, 1);
8611                 } else {
8612                         /**
8613                          * This is the case where our previous
8614                          * block is not acked either, so we must
8615                          * split the block in two.
8616                          */
8617                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8618                         if (nrsm == NULL) {
8619                                 /* failed rrs what can we do but loose the sack info? */
8620                                 goto out;
8621                         }
8622                         if ((rsm->r_flags & RACK_TLP) &&
8623                             (rsm->r_rtr_cnt > 1)) {
8624                                 /*
8625                                  * We are splitting a rxt TLP, check
8626                                  * if we need to save off the start/end
8627                                  */
8628                                 if (rack->rc_last_tlp_acked_set &&
8629                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8630                                             /*
8631                                              * We already turned this on since this block is inside
8632                                              * the previous one was a partially sack now we
8633                                              * are getting another one (maybe all of it).
8634                                              */
8635                                             rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8636                                             /*
8637                                              * Lets make sure we have all of it though.
8638                                              */
8639                                             if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8640                                                     rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8641                                                     rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8642                                                                          rack->r_ctl.last_tlp_acked_end);
8643                                             }
8644                                             if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8645                                                     rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8646                                                     rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8647                                                                          rack->r_ctl.last_tlp_acked_end);
8648                                             }
8649                                     } else {
8650                                             rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8651                                             rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8652                                             rack->rc_last_tlp_acked_set = 1;
8653                                             rack->rc_last_tlp_past_cumack = 0;
8654                                             rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8655                                     }
8656                         }
8657                         /**
8658                          * In this case nrsm becomes
8659                          * nrsm->r_start = end;
8660                          * nrsm->r_end = rsm->r_end;
8661                          * which is un-acked.
8662                          * <and>
8663                          * rsm->r_end = nrsm->r_start;
8664                          * i.e. the remaining un-acked
8665                          * piece is left on the left
8666                          * hand side.
8667                          *
8668                          * So we start like this
8669                          * rsm      |----------| (not acked)
8670                          * sackblk  |---|
8671                          * build it so we have
8672                          * rsm      |---|         (acked)
8673                          * nrsm         |------|  (not acked)
8674                          */
8675                         counter_u64_add(rack_sack_splits, 1);
8676                         rack_clone_rsm(rack, nrsm, rsm, end);
8677                         rsm->r_flags &= (~RACK_HAS_FIN);
8678                         rsm->r_just_ret = 0;
8679 #ifndef INVARIANTS
8680                         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8681 #else
8682                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8683                         if (insret != NULL) {
8684                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8685                                       nrsm, insret, rack, rsm);
8686                         }
8687 #endif
8688                         if (rsm->r_in_tmap) {
8689                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8690                                 nrsm->r_in_tmap = 1;
8691                         }
8692                         nrsm->r_dupack = 0;
8693                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8694                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8695                         changed += (rsm->r_end - rsm->r_start);
8696                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8697                         if (rsm->r_in_tmap) /* should be true */
8698                                 rack_log_sack_passed(tp, rack, rsm);
8699                         /* Is Reordering occuring? */
8700                         if (rsm->r_flags & RACK_SACK_PASSED) {
8701                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8702                                 rack->r_ctl.rc_reorder_ts = cts;
8703                         }
8704                         if (rack->app_limited_needs_set)
8705                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8706                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8707                         rsm->r_flags |= RACK_ACKED;
8708                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8709                         if (rsm->r_in_tmap) {
8710                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8711                                 rsm->r_in_tmap = 0;
8712                         }
8713                 }
8714         } else if (start != end){
8715                 /*
8716                  * The block was already acked.
8717                  */
8718                 counter_u64_add(rack_sack_skipped_acked, 1);
8719                 moved++;
8720         }
8721 out:
8722         if (rsm &&
8723             ((rsm->r_flags & RACK_TLP) == 0) &&
8724             (rsm->r_flags & RACK_ACKED)) {
8725                 /*
8726                  * Now can we merge where we worked
8727                  * with either the previous or
8728                  * next block?
8729                  */
8730                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8731                 while (next) {
8732                         if (next->r_flags & RACK_TLP)
8733                                 break;
8734                         if (next->r_flags & RACK_ACKED) {
8735                         /* yep this and next can be merged */
8736                                 rsm = rack_merge_rsm(rack, rsm, next);
8737                                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8738                         } else
8739                                 break;
8740                 }
8741                 /* Now what about the previous? */
8742                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8743                 while (prev) {
8744                         if (prev->r_flags & RACK_TLP)
8745                                 break;
8746                         if (prev->r_flags & RACK_ACKED) {
8747                                 /* yep the previous and this can be merged */
8748                                 rsm = rack_merge_rsm(rack, prev, rsm);
8749                                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8750                         } else
8751                                 break;
8752                 }
8753         }
8754         if (used_ref == 0) {
8755                 counter_u64_add(rack_sack_proc_all, 1);
8756         } else {
8757                 counter_u64_add(rack_sack_proc_short, 1);
8758         }
8759         /* Save off the next one for quick reference. */
8760         if (rsm)
8761                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8762         else
8763                 nrsm = NULL;
8764         *prsm = rack->r_ctl.rc_sacklast = nrsm;
8765         /* Pass back the moved. */
8766         *moved_two = moved;
8767         return (changed);
8768 }
8769
8770 static void inline
8771 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8772 {
8773         struct rack_sendmap *tmap;
8774
8775         tmap = NULL;
8776         while (rsm && (rsm->r_flags & RACK_ACKED)) {
8777                 /* Its no longer sacked, mark it so */
8778                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8779 #ifdef INVARIANTS
8780                 if (rsm->r_in_tmap) {
8781                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
8782                               rack, rsm, rsm->r_flags);
8783                 }
8784 #endif
8785                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8786                 /* Rebuild it into our tmap */
8787                 if (tmap == NULL) {
8788                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8789                         tmap = rsm;
8790                 } else {
8791                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8792                         tmap = rsm;
8793                 }
8794                 tmap->r_in_tmap = 1;
8795                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8796         }
8797         /*
8798          * Now lets possibly clear the sack filter so we start
8799          * recognizing sacks that cover this area.
8800          */
8801         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8802
8803 }
8804
8805 static void
8806 rack_do_decay(struct tcp_rack *rack)
8807 {
8808         struct timeval res;
8809
8810 #define timersub(tvp, uvp, vvp)                                         \
8811         do {                                                            \
8812                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
8813                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
8814                 if ((vvp)->tv_usec < 0) {                               \
8815                         (vvp)->tv_sec--;                                \
8816                         (vvp)->tv_usec += 1000000;                      \
8817                 }                                                       \
8818         } while (0)
8819
8820         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8821 #undef timersub
8822
8823         rack->r_ctl.input_pkt++;
8824         if ((rack->rc_in_persist) ||
8825             (res.tv_sec >= 1) ||
8826             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8827                 /*
8828                  * Check for decay of non-SAD,
8829                  * we want all SAD detection metrics to
8830                  * decay 1/4 per second (or more) passed.
8831                  */
8832 #ifdef NETFLIX_EXP_DETECTION
8833                 uint32_t pkt_delta;
8834
8835                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8836 #endif
8837                 /* Update our saved tracking values */
8838                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8839                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8840                 /* Now do we escape without decay? */
8841 #ifdef NETFLIX_EXP_DETECTION
8842                 if (rack->rc_in_persist ||
8843                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8844                     (pkt_delta < tcp_sad_low_pps)){
8845                         /*
8846                          * We don't decay idle connections
8847                          * or ones that have a low input pps.
8848                          */
8849                         return;
8850                 }
8851                 /* Decay the counters */
8852                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8853                                                         tcp_sad_decay_val);
8854                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8855                                                          tcp_sad_decay_val);
8856                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8857                                                                tcp_sad_decay_val);
8858                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8859                                                                 tcp_sad_decay_val);
8860 #endif
8861         }
8862 }
8863
8864 static void
8865 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8866 {
8867         struct rack_sendmap *rsm;
8868 #ifdef INVARIANTS
8869         struct rack_sendmap *rm;
8870 #endif
8871
8872         /*
8873          * The ACK point is advancing to th_ack, we must drop off
8874          * the packets in the rack log and calculate any eligble
8875          * RTT's.
8876          */
8877         rack->r_wanted_output = 1;
8878
8879         /* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
8880         if ((rack->rc_last_tlp_acked_set == 1)&&
8881             (rack->rc_last_tlp_past_cumack == 1) &&
8882             (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
8883                 /*
8884                  * We have reached the point where our last rack
8885                  * tlp retransmit sequence is ahead of the cum-ack.
8886                  * This can only happen when the cum-ack moves all
8887                  * the way around (its been a full 2^^31+1 bytes
8888                  * or more since we sent a retransmitted TLP). Lets
8889                  * turn off the valid flag since its not really valid.
8890                  *
8891                  * Note since sack's also turn on this event we have
8892                  * a complication, we have to wait to age it out until
8893                  * the cum-ack is by the TLP before checking which is
8894                  * what the next else clause does.
8895                  */
8896                 rack_log_dsack_event(rack, 9, __LINE__,
8897                                      rack->r_ctl.last_tlp_acked_start,
8898                                      rack->r_ctl.last_tlp_acked_end);
8899                 rack->rc_last_tlp_acked_set = 0;
8900                 rack->rc_last_tlp_past_cumack = 0;
8901         } else if ((rack->rc_last_tlp_acked_set == 1) &&
8902                    (rack->rc_last_tlp_past_cumack == 0) &&
8903                    (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
8904                 /*
8905                  * It is safe to start aging TLP's out.
8906                  */
8907                 rack->rc_last_tlp_past_cumack = 1;
8908         }
8909         /* We do the same for the tlp send seq as well */
8910         if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8911             (rack->rc_last_sent_tlp_past_cumack == 1) &&
8912             (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
8913                 rack_log_dsack_event(rack, 9, __LINE__,
8914                                      rack->r_ctl.last_sent_tlp_seq,
8915                                      (rack->r_ctl.last_sent_tlp_seq +
8916                                       rack->r_ctl.last_sent_tlp_len));
8917                 rack->rc_last_sent_tlp_seq_valid = 0;
8918                 rack->rc_last_sent_tlp_past_cumack = 0;
8919         } else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8920                    (rack->rc_last_sent_tlp_past_cumack == 0) &&
8921                    (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
8922                 /*
8923                  * It is safe to start aging TLP's send.
8924                  */
8925                 rack->rc_last_sent_tlp_past_cumack = 1;
8926         }
8927 more:
8928         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8929         if (rsm == NULL) {
8930                 if ((th_ack - 1) == tp->iss) {
8931                         /*
8932                          * For the SYN incoming case we will not
8933                          * have called tcp_output for the sending of
8934                          * the SYN, so there will be no map. All
8935                          * other cases should probably be a panic.
8936                          */
8937                         return;
8938                 }
8939                 if (tp->t_flags & TF_SENTFIN) {
8940                         /* if we sent a FIN we often will not have map */
8941                         return;
8942                 }
8943 #ifdef INVARIANTS
8944                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8945                       tp,
8946                       tp->t_state, th_ack, rack,
8947                       tp->snd_una, tp->snd_max, tp->snd_nxt);
8948 #endif
8949                 return;
8950         }
8951         if (SEQ_LT(th_ack, rsm->r_start)) {
8952                 /* Huh map is missing this */
8953 #ifdef INVARIANTS
8954                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8955                        rsm->r_start,
8956                        th_ack, tp->t_state, rack->r_state);
8957 #endif
8958                 return;
8959         }
8960         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8961
8962         /* Now was it a retransmitted TLP? */
8963         if ((rsm->r_flags & RACK_TLP) &&
8964             (rsm->r_rtr_cnt > 1)) {
8965                 /*
8966                  * Yes, this rsm was a TLP and retransmitted, remember that
8967                  * since if a DSACK comes back on this we don't want
8968                  * to think of it as a reordered segment. This may
8969                  * get updated again with possibly even other TLPs
8970                  * in flight, but thats ok. Only when we don't send
8971                  * a retransmitted TLP for 1/2 the sequences space
8972                  * will it get turned off (above).
8973                  */
8974                 if (rack->rc_last_tlp_acked_set &&
8975                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8976                         /*
8977                          * We already turned this on since the end matches,
8978                          * the previous one was a partially ack now we
8979                          * are getting another one (maybe all of it).
8980                          */
8981                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8982                         /*
8983                          * Lets make sure we have all of it though.
8984                          */
8985                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8986                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8987                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8988                                                      rack->r_ctl.last_tlp_acked_end);
8989                         }
8990                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8991                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8992                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8993                                                      rack->r_ctl.last_tlp_acked_end);
8994                         }
8995                 } else {
8996                         rack->rc_last_tlp_past_cumack = 1;
8997                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8998                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8999                         rack->rc_last_tlp_acked_set = 1;
9000                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9001                 }
9002         }
9003         /* Now do we consume the whole thing? */
9004         if (SEQ_GEQ(th_ack, rsm->r_end)) {
9005                 /* Its all consumed. */
9006                 uint32_t left;
9007                 uint8_t newly_acked;
9008
9009                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
9010                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
9011                 rsm->r_rtr_bytes = 0;
9012                 /* Record the time of highest cumack sent */
9013                 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9014 #ifndef INVARIANTS
9015                 (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
9016 #else
9017                 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
9018                 if (rm != rsm) {
9019                         panic("removing head in rack:%p rsm:%p rm:%p",
9020                               rack, rsm, rm);
9021                 }
9022 #endif
9023                 if (rsm->r_in_tmap) {
9024                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9025                         rsm->r_in_tmap = 0;
9026                 }
9027                 newly_acked = 1;
9028                 if (rsm->r_flags & RACK_ACKED) {
9029                         /*
9030                          * It was acked on the scoreboard -- remove
9031                          * it from total
9032                          */
9033                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
9034                         newly_acked = 0;
9035                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
9036                         /*
9037                          * There are segments ACKED on the
9038                          * scoreboard further up. We are seeing
9039                          * reordering.
9040                          */
9041                         rsm->r_flags &= ~RACK_SACK_PASSED;
9042                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9043                         rsm->r_flags |= RACK_ACKED;
9044                         rack->r_ctl.rc_reorder_ts = cts;
9045                         if (rack->r_ent_rec_ns) {
9046                                 /*
9047                                  * We have sent no more, and we saw an sack
9048                                  * then ack arrive.
9049                                  */
9050                                 rack->r_might_revert = 1;
9051                         }
9052                 }
9053                 if ((rsm->r_flags & RACK_TO_REXT) &&
9054                     (tp->t_flags & TF_RCVD_TSTMP) &&
9055                     (to->to_flags & TOF_TS) &&
9056                     (to->to_tsecr != 0) &&
9057                     (tp->t_flags & TF_PREVVALID)) {
9058                         /*
9059                          * We can use the timestamp to see
9060                          * if this retransmission was from the
9061                          * first transmit. If so we made a mistake.
9062                          */
9063                         tp->t_flags &= ~TF_PREVVALID;
9064                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
9065                                 /* The first transmit is what this ack is for */
9066                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
9067                         }
9068                 }
9069                 left = th_ack - rsm->r_end;
9070                 if (rack->app_limited_needs_set && newly_acked)
9071                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
9072                 /* Free back to zone */
9073                 rack_free(rack, rsm);
9074                 if (left) {
9075                         goto more;
9076                 }
9077                 /* Check for reneging */
9078                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9079                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9080                         /*
9081                          * The peer has moved snd_una up to
9082                          * the edge of this send, i.e. one
9083                          * that it had previously acked. The only
9084                          * way that can be true if the peer threw
9085                          * away data (space issues) that it had
9086                          * previously sacked (else it would have
9087                          * given us snd_una up to (rsm->r_end).
9088                          * We need to undo the acked markings here.
9089                          *
9090                          * Note we have to look to make sure th_ack is
9091                          * our rsm->r_start in case we get an old ack
9092                          * where th_ack is behind snd_una.
9093                          */
9094                         rack_peer_reneges(rack, rsm, th_ack);
9095                 }
9096                 return;
9097         }
9098         if (rsm->r_flags & RACK_ACKED) {
9099                 /*
9100                  * It was acked on the scoreboard -- remove it from
9101                  * total for the part being cum-acked.
9102                  */
9103                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9104         }
9105         /*
9106          * Clear the dup ack count for
9107          * the piece that remains.
9108          */
9109         rsm->r_dupack = 0;
9110         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9111         if (rsm->r_rtr_bytes) {
9112                 /*
9113                  * It was retransmitted adjust the
9114                  * sack holes for what was acked.
9115                  */
9116                 int ack_am;
9117
9118                 ack_am = (th_ack - rsm->r_start);
9119                 if (ack_am >= rsm->r_rtr_bytes) {
9120                         rack->r_ctl.rc_holes_rxt -= ack_am;
9121                         rsm->r_rtr_bytes -= ack_am;
9122                 }
9123         }
9124         /*
9125          * Update where the piece starts and record
9126          * the time of send of highest cumack sent.
9127          */
9128         rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9129         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9130         /* Now we need to move our offset forward too */
9131         if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9132                 /* Fix up the orig_m_len and possibly the mbuf offset */
9133                 rack_adjust_orig_mlen(rsm);
9134         }
9135         rsm->soff += (th_ack - rsm->r_start);
9136         rsm->r_start = th_ack;
9137         /* Now do we need to move the mbuf fwd too? */
9138         if (rsm->m) {
9139                 while (rsm->soff >= rsm->m->m_len) {
9140                         rsm->soff -= rsm->m->m_len;
9141                         rsm->m = rsm->m->m_next;
9142                         KASSERT((rsm->m != NULL),
9143                                 (" nrsm:%p hit at soff:%u null m",
9144                                  rsm, rsm->soff));
9145                 }
9146                 rsm->orig_m_len = rsm->m->m_len;
9147         }
9148         if (rack->app_limited_needs_set)
9149                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9150 }
9151
9152 static void
9153 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9154 {
9155         struct rack_sendmap *rsm;
9156         int sack_pass_fnd = 0;
9157
9158         if (rack->r_might_revert) {
9159                 /*
9160                  * Ok we have reordering, have not sent anything, we
9161                  * might want to revert the congestion state if nothing
9162                  * further has SACK_PASSED on it. Lets check.
9163                  *
9164                  * We also get here when we have DSACKs come in for
9165                  * all the data that we FR'd. Note that a rxt or tlp
9166                  * timer clears this from happening.
9167                  */
9168
9169                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9170                         if (rsm->r_flags & RACK_SACK_PASSED) {
9171                                 sack_pass_fnd = 1;
9172                                 break;
9173                         }
9174                 }
9175                 if (sack_pass_fnd == 0) {
9176                         /*
9177                          * We went into recovery
9178                          * incorrectly due to reordering!
9179                          */
9180                         int orig_cwnd;
9181
9182                         rack->r_ent_rec_ns = 0;
9183                         orig_cwnd = tp->snd_cwnd;
9184                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9185                         tp->snd_recover = tp->snd_una;
9186                         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
9187                         EXIT_RECOVERY(tp->t_flags);
9188                 }
9189                 rack->r_might_revert = 0;
9190         }
9191 }
9192
9193 #ifdef NETFLIX_EXP_DETECTION
9194 static void
9195 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9196 {
9197         if ((rack->do_detection || tcp_force_detection) &&
9198             tcp_sack_to_ack_thresh &&
9199             tcp_sack_to_move_thresh &&
9200             ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9201                 /*
9202                  * We have thresholds set to find
9203                  * possible attackers and disable sack.
9204                  * Check them.
9205                  */
9206                 uint64_t ackratio, moveratio, movetotal;
9207
9208                 /* Log detecting */
9209                 rack_log_sad(rack, 1);
9210                 ackratio = (uint64_t)(rack->r_ctl.sack_count);
9211                 ackratio *= (uint64_t)(1000);
9212                 if (rack->r_ctl.ack_count)
9213                         ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9214                 else {
9215                         /* We really should not hit here */
9216                         ackratio = 1000;
9217                 }
9218                 if ((rack->sack_attack_disable == 0) &&
9219                     (ackratio > rack_highest_sack_thresh_seen))
9220                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9221                 movetotal = rack->r_ctl.sack_moved_extra;
9222                 movetotal += rack->r_ctl.sack_noextra_move;
9223                 moveratio = rack->r_ctl.sack_moved_extra;
9224                 moveratio *= (uint64_t)1000;
9225                 if (movetotal)
9226                         moveratio /= movetotal;
9227                 else {
9228                         /* No moves, thats pretty good */
9229                         moveratio = 0;
9230                 }
9231                 if ((rack->sack_attack_disable == 0) &&
9232                     (moveratio > rack_highest_move_thresh_seen))
9233                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
9234                 if (rack->sack_attack_disable == 0) {
9235                         if ((ackratio > tcp_sack_to_ack_thresh) &&
9236                             (moveratio > tcp_sack_to_move_thresh)) {
9237                                 /* Disable sack processing */
9238                                 rack->sack_attack_disable = 1;
9239                                 if (rack->r_rep_attack == 0) {
9240                                         rack->r_rep_attack = 1;
9241                                         counter_u64_add(rack_sack_attacks_detected, 1);
9242                                 }
9243                                 if (tcp_attack_on_turns_on_logging) {
9244                                         /*
9245                                          * Turn on logging, used for debugging
9246                                          * false positives.
9247                                          */
9248                                         rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9249                                 }
9250                                 /* Clamp the cwnd at flight size */
9251                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9252                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9253                                 rack_log_sad(rack, 2);
9254                         }
9255                 } else {
9256                         /* We are sack-disabled check for false positives */
9257                         if ((ackratio <= tcp_restoral_thresh) ||
9258                             (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9259                                 rack->sack_attack_disable = 0;
9260                                 rack_log_sad(rack, 3);
9261                                 /* Restart counting */
9262                                 rack->r_ctl.sack_count = 0;
9263                                 rack->r_ctl.sack_moved_extra = 0;
9264                                 rack->r_ctl.sack_noextra_move = 1;
9265                                 rack->r_ctl.ack_count = max(1,
9266                                       (bytes_this_ack / segsiz));
9267
9268                                 if (rack->r_rep_reverse == 0) {
9269                                         rack->r_rep_reverse = 1;
9270                                         counter_u64_add(rack_sack_attacks_reversed, 1);
9271                                 }
9272                                 /* Restore the cwnd */
9273                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9274                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9275                         }
9276                 }
9277         }
9278 }
9279 #endif
9280
9281 static int
9282 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9283 {
9284
9285         uint32_t am, l_end;
9286         int was_tlp = 0;
9287
9288         if (SEQ_GT(end, start))
9289                 am = end - start;
9290         else
9291                 am = 0;
9292         if ((rack->rc_last_tlp_acked_set ) &&
9293             (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9294             (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9295                 /*
9296                  * The DSACK is because of a TLP which we don't
9297                  * do anything with the reordering window over since
9298                  * it was not reordering that caused the DSACK but
9299                  * our previous retransmit TLP.
9300                  */
9301                 rack_log_dsack_event(rack, 7, __LINE__, start, end);
9302                 was_tlp = 1;
9303                 goto skip_dsack_round;
9304         }
9305         if (rack->rc_last_sent_tlp_seq_valid) {
9306                 l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9307                 if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9308                     (SEQ_LEQ(end, l_end))) {
9309                         /*
9310                          * This dsack is from the last sent TLP, ignore it
9311                          * for reordering purposes.
9312                          */
9313                         rack_log_dsack_event(rack, 7, __LINE__, start, end);
9314                         was_tlp = 1;
9315                         goto skip_dsack_round;
9316                 }
9317         }
9318         if (rack->rc_dsack_round_seen == 0) {
9319                 rack->rc_dsack_round_seen = 1;
9320                 rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9321                 rack->r_ctl.num_dsack++;
9322                 rack->r_ctl.dsack_persist = 16; /* 16 is from the standard */
9323                 rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9324         }
9325 skip_dsack_round:
9326         /*
9327          * We keep track of how many DSACK blocks we get
9328          * after a recovery incident.
9329          */
9330         rack->r_ctl.dsack_byte_cnt += am;
9331         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9332             rack->r_ctl.retran_during_recovery &&
9333             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9334                 /*
9335                  * False recovery most likely culprit is reordering. If
9336                  * nothing else is missing we need to revert.
9337                  */
9338                 rack->r_might_revert = 1;
9339                 rack_handle_might_revert(rack->rc_tp, rack);
9340                 rack->r_might_revert = 0;
9341                 rack->r_ctl.retran_during_recovery = 0;
9342                 rack->r_ctl.dsack_byte_cnt = 0;
9343         }
9344         return (was_tlp);
9345 }
9346
9347 static uint32_t
9348 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
9349 {
9350         return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
9351 }
9352
9353 static int32_t
9354 rack_compute_pipe(struct tcpcb *tp)
9355 {
9356         return ((int32_t)do_rack_compute_pipe(tp,
9357                                               (struct tcp_rack *)tp->t_fb_ptr,
9358                                               tp->snd_una));
9359 }
9360
9361 static void
9362 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9363 {
9364         /* Deal with changed and PRR here (in recovery only) */
9365         uint32_t pipe, snd_una;
9366
9367         rack->r_ctl.rc_prr_delivered += changed;
9368
9369         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9370                 /*
9371                  * It is all outstanding, we are application limited
9372                  * and thus we don't need more room to send anything.
9373                  * Note we use tp->snd_una here and not th_ack because
9374                  * the data as yet not been cut from the sb.
9375                  */
9376                 rack->r_ctl.rc_prr_sndcnt = 0;
9377                 return;
9378         }
9379         /* Compute prr_sndcnt */
9380         if (SEQ_GT(tp->snd_una, th_ack)) {
9381                 snd_una = tp->snd_una;
9382         } else {
9383                 snd_una = th_ack;
9384         }
9385         pipe = do_rack_compute_pipe(tp, rack, snd_una);
9386         if (pipe > tp->snd_ssthresh) {
9387                 long sndcnt;
9388
9389                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9390                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9391                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9392                 else {
9393                         rack->r_ctl.rc_prr_sndcnt = 0;
9394                         rack_log_to_prr(rack, 9, 0, __LINE__);
9395                         sndcnt = 0;
9396                 }
9397                 sndcnt++;
9398                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9399                         sndcnt -= rack->r_ctl.rc_prr_out;
9400                 else
9401                         sndcnt = 0;
9402                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9403                 rack_log_to_prr(rack, 10, 0, __LINE__);
9404         } else {
9405                 uint32_t limit;
9406
9407                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9408                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9409                 else
9410                         limit = 0;
9411                 if (changed > limit)
9412                         limit = changed;
9413                 limit += ctf_fixed_maxseg(tp);
9414                 if (tp->snd_ssthresh > pipe) {
9415                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9416                         rack_log_to_prr(rack, 11, 0, __LINE__);
9417                 } else {
9418                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9419                         rack_log_to_prr(rack, 12, 0, __LINE__);
9420                 }
9421         }
9422 }
9423
9424 static void
9425 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9426 {
9427         uint32_t changed;
9428         struct tcp_rack *rack;
9429         struct rack_sendmap *rsm;
9430         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9431         register uint32_t th_ack;
9432         int32_t i, j, k, num_sack_blks = 0;
9433         uint32_t cts, acked, ack_point;
9434         int loop_start = 0, moved_two = 0;
9435         uint32_t tsused;
9436
9437
9438         INP_WLOCK_ASSERT(tptoinpcb(tp));
9439         if (tcp_get_flags(th) & TH_RST) {
9440                 /* We don't log resets */
9441                 return;
9442         }
9443         rack = (struct tcp_rack *)tp->t_fb_ptr;
9444         cts = tcp_get_usecs(NULL);
9445         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9446         changed = 0;
9447         th_ack = th->th_ack;
9448         if (rack->sack_attack_disable == 0)
9449                 rack_do_decay(rack);
9450         if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9451                 /*
9452                  * You only get credit for
9453                  * MSS and greater (and you get extra
9454                  * credit for larger cum-ack moves).
9455                  */
9456                 int ac;
9457
9458                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9459                 rack->r_ctl.ack_count += ac;
9460                 counter_u64_add(rack_ack_total, ac);
9461         }
9462         if (rack->r_ctl.ack_count > 0xfff00000) {
9463                 /*
9464                  * reduce the number to keep us under
9465                  * a uint32_t.
9466                  */
9467                 rack->r_ctl.ack_count /= 2;
9468                 rack->r_ctl.sack_count /= 2;
9469         }
9470         if (SEQ_GT(th_ack, tp->snd_una)) {
9471                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9472                 tp->t_acktime = ticks;
9473         }
9474         if (rsm && SEQ_GT(th_ack, rsm->r_start))
9475                 changed = th_ack - rsm->r_start;
9476         if (changed) {
9477                 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9478         }
9479         if ((to->to_flags & TOF_SACK) == 0) {
9480                 /* We are done nothing left and no sack. */
9481                 rack_handle_might_revert(tp, rack);
9482                 /*
9483                  * For cases where we struck a dup-ack
9484                  * with no SACK, add to the changes so
9485                  * PRR will work right.
9486                  */
9487                 if (dup_ack_struck && (changed == 0)) {
9488                         changed += ctf_fixed_maxseg(rack->rc_tp);
9489                 }
9490                 goto out;
9491         }
9492         /* Sack block processing */
9493         if (SEQ_GT(th_ack, tp->snd_una))
9494                 ack_point = th_ack;
9495         else
9496                 ack_point = tp->snd_una;
9497         for (i = 0; i < to->to_nsacks; i++) {
9498                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9499                       &sack, sizeof(sack));
9500                 sack.start = ntohl(sack.start);
9501                 sack.end = ntohl(sack.end);
9502                 if (SEQ_GT(sack.end, sack.start) &&
9503                     SEQ_GT(sack.start, ack_point) &&
9504                     SEQ_LT(sack.start, tp->snd_max) &&
9505                     SEQ_GT(sack.end, ack_point) &&
9506                     SEQ_LEQ(sack.end, tp->snd_max)) {
9507                         sack_blocks[num_sack_blks] = sack;
9508                         num_sack_blks++;
9509                 } else if (SEQ_LEQ(sack.start, th_ack) &&
9510                            SEQ_LEQ(sack.end, th_ack)) {
9511                         int was_tlp;
9512
9513                         was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9514                         /*
9515                          * Its a D-SACK block.
9516                          */
9517                         tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9518                 }
9519         }
9520         if (rack->rc_dsack_round_seen) {
9521                 /* Is the dsack roound over? */
9522                 if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9523                         /* Yes it is */
9524                         rack->rc_dsack_round_seen = 0;
9525                         rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9526                 }
9527         }
9528         /*
9529          * Sort the SACK blocks so we can update the rack scoreboard with
9530          * just one pass.
9531          */
9532         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9533                                          num_sack_blks, th->th_ack);
9534         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9535         if (num_sack_blks == 0) {
9536                 /* Nothing to sack (DSACKs?) */
9537                 goto out_with_totals;
9538         }
9539         if (num_sack_blks < 2) {
9540                 /* Only one, we don't need to sort */
9541                 goto do_sack_work;
9542         }
9543         /* Sort the sacks */
9544         for (i = 0; i < num_sack_blks; i++) {
9545                 for (j = i + 1; j < num_sack_blks; j++) {
9546                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9547                                 sack = sack_blocks[i];
9548                                 sack_blocks[i] = sack_blocks[j];
9549                                 sack_blocks[j] = sack;
9550                         }
9551                 }
9552         }
9553         /*
9554          * Now are any of the sack block ends the same (yes some
9555          * implementations send these)?
9556          */
9557 again:
9558         if (num_sack_blks == 0)
9559                 goto out_with_totals;
9560         if (num_sack_blks > 1) {
9561                 for (i = 0; i < num_sack_blks; i++) {
9562                         for (j = i + 1; j < num_sack_blks; j++) {
9563                                 if (sack_blocks[i].end == sack_blocks[j].end) {
9564                                         /*
9565                                          * Ok these two have the same end we
9566                                          * want the smallest end and then
9567                                          * throw away the larger and start
9568                                          * again.
9569                                          */
9570                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9571                                                 /*
9572                                                  * The second block covers
9573                                                  * more area use that
9574                                                  */
9575                                                 sack_blocks[i].start = sack_blocks[j].start;
9576                                         }
9577                                         /*
9578                                          * Now collapse out the dup-sack and
9579                                          * lower the count
9580                                          */
9581                                         for (k = (j + 1); k < num_sack_blks; k++) {
9582                                                 sack_blocks[j].start = sack_blocks[k].start;
9583                                                 sack_blocks[j].end = sack_blocks[k].end;
9584                                                 j++;
9585                                         }
9586                                         num_sack_blks--;
9587                                         goto again;
9588                                 }
9589                         }
9590                 }
9591         }
9592 do_sack_work:
9593         /*
9594          * First lets look to see if
9595          * we have retransmitted and
9596          * can use the transmit next?
9597          */
9598         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9599         if (rsm &&
9600             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9601             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9602                 /*
9603                  * We probably did the FR and the next
9604                  * SACK in continues as we would expect.
9605                  */
9606                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9607                 if (acked) {
9608                         rack->r_wanted_output = 1;
9609                         changed += acked;
9610                 }
9611                 if (num_sack_blks == 1) {
9612                         /*
9613                          * This is what we would expect from
9614                          * a normal implementation to happen
9615                          * after we have retransmitted the FR,
9616                          * i.e the sack-filter pushes down
9617                          * to 1 block and the next to be retransmitted
9618                          * is the sequence in the sack block (has more
9619                          * are acked). Count this as ACK'd data to boost
9620                          * up the chances of recovering any false positives.
9621                          */
9622                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9623                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9624                         counter_u64_add(rack_express_sack, 1);
9625                         if (rack->r_ctl.ack_count > 0xfff00000) {
9626                                 /*
9627                                  * reduce the number to keep us under
9628                                  * a uint32_t.
9629                                  */
9630                                 rack->r_ctl.ack_count /= 2;
9631                                 rack->r_ctl.sack_count /= 2;
9632                         }
9633                         goto out_with_totals;
9634                 } else {
9635                         /*
9636                          * Start the loop through the
9637                          * rest of blocks, past the first block.
9638                          */
9639                         moved_two = 0;
9640                         loop_start = 1;
9641                 }
9642         }
9643         /* Its a sack of some sort */
9644         rack->r_ctl.sack_count++;
9645         if (rack->r_ctl.sack_count > 0xfff00000) {
9646                 /*
9647                  * reduce the number to keep us under
9648                  * a uint32_t.
9649                  */
9650                 rack->r_ctl.ack_count /= 2;
9651                 rack->r_ctl.sack_count /= 2;
9652         }
9653         counter_u64_add(rack_sack_total, 1);
9654         if (rack->sack_attack_disable) {
9655                 /* An attacker disablement is in place */
9656                 if (num_sack_blks > 1) {
9657                         rack->r_ctl.sack_count += (num_sack_blks - 1);
9658                         rack->r_ctl.sack_moved_extra++;
9659                         counter_u64_add(rack_move_some, 1);
9660                         if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9661                                 rack->r_ctl.sack_moved_extra /= 2;
9662                                 rack->r_ctl.sack_noextra_move /= 2;
9663                         }
9664                 }
9665                 goto out;
9666         }
9667         rsm = rack->r_ctl.rc_sacklast;
9668         for (i = loop_start; i < num_sack_blks; i++) {
9669                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9670                 if (acked) {
9671                         rack->r_wanted_output = 1;
9672                         changed += acked;
9673                 }
9674                 if (moved_two) {
9675                         /*
9676                          * If we did not get a SACK for at least a MSS and
9677                          * had to move at all, or if we moved more than our
9678                          * threshold, it counts against the "extra" move.
9679                          */
9680                         rack->r_ctl.sack_moved_extra += moved_two;
9681                         counter_u64_add(rack_move_some, 1);
9682                 } else {
9683                         /*
9684                          * else we did not have to move
9685                          * any more than we would expect.
9686                          */
9687                         rack->r_ctl.sack_noextra_move++;
9688                         counter_u64_add(rack_move_none, 1);
9689                 }
9690                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9691                         /*
9692                          * If the SACK was not a full MSS then
9693                          * we add to sack_count the number of
9694                          * MSS's (or possibly more than
9695                          * a MSS if its a TSO send) we had to skip by.
9696                          */
9697                         rack->r_ctl.sack_count += moved_two;
9698                         counter_u64_add(rack_sack_total, moved_two);
9699                 }
9700                 /*
9701                  * Now we need to setup for the next
9702                  * round. First we make sure we won't
9703                  * exceed the size of our uint32_t on
9704                  * the various counts, and then clear out
9705                  * moved_two.
9706                  */
9707                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9708                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9709                         rack->r_ctl.sack_moved_extra /= 2;
9710                         rack->r_ctl.sack_noextra_move /= 2;
9711                 }
9712                 if (rack->r_ctl.sack_count > 0xfff00000) {
9713                         rack->r_ctl.ack_count /= 2;
9714                         rack->r_ctl.sack_count /= 2;
9715                 }
9716                 moved_two = 0;
9717         }
9718 out_with_totals:
9719         if (num_sack_blks > 1) {
9720                 /*
9721                  * You get an extra stroke if
9722                  * you have more than one sack-blk, this
9723                  * could be where we are skipping forward
9724                  * and the sack-filter is still working, or
9725                  * it could be an attacker constantly
9726                  * moving us.
9727                  */
9728                 rack->r_ctl.sack_moved_extra++;
9729                 counter_u64_add(rack_move_some, 1);
9730         }
9731 out:
9732 #ifdef NETFLIX_EXP_DETECTION
9733         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9734 #endif
9735         if (changed) {
9736                 /* Something changed cancel the rack timer */
9737                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9738         }
9739         tsused = tcp_get_usecs(NULL);
9740         rsm = tcp_rack_output(tp, rack, tsused);
9741         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9742             rsm &&
9743             ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
9744                 /* Enter recovery */
9745                 entered_recovery = 1;
9746                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9747                 /*
9748                  * When we enter recovery we need to assure we send
9749                  * one packet.
9750                  */
9751                 if (rack->rack_no_prr == 0) {
9752                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9753                         rack_log_to_prr(rack, 8, 0, __LINE__);
9754                 }
9755                 rack->r_timer_override = 1;
9756                 rack->r_early = 0;
9757                 rack->r_ctl.rc_agg_early = 0;
9758         } else if (IN_FASTRECOVERY(tp->t_flags) &&
9759                    rsm &&
9760                    (rack->r_rr_config == 3)) {
9761                 /*
9762                  * Assure we can output and we get no
9763                  * remembered pace time except the retransmit.
9764                  */
9765                 rack->r_timer_override = 1;
9766                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9767                 rack->r_ctl.rc_resend = rsm;
9768         }
9769         if (IN_FASTRECOVERY(tp->t_flags) &&
9770             (rack->rack_no_prr == 0) &&
9771             (entered_recovery == 0)) {
9772                 rack_update_prr(tp, rack, changed, th_ack);
9773                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9774                      ((tcp_in_hpts(rack->rc_inp) == 0) &&
9775                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9776                         /*
9777                          * If you are pacing output you don't want
9778                          * to override.
9779                          */
9780                         rack->r_early = 0;
9781                         rack->r_ctl.rc_agg_early = 0;
9782                         rack->r_timer_override = 1;
9783                 }
9784         }
9785 }
9786
9787 static void
9788 rack_strike_dupack(struct tcp_rack *rack)
9789 {
9790         struct rack_sendmap *rsm;
9791
9792         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9793         while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9794                 rsm = TAILQ_NEXT(rsm, r_tnext);
9795                 if (rsm->r_flags & RACK_MUST_RXT) {
9796                         /* Sendmap entries that are marked to
9797                          * be retransmitted do not need dupack's
9798                          * struck. We get these marks for a number
9799                          * of reasons (rxt timeout with no sack,
9800                          * mtu change, or rwnd collapses). When
9801                          * these events occur, we know we must retransmit
9802                          * them and mark the sendmap entries. Dupack counting
9803                          * is not needed since we are already set to retransmit
9804                          * it as soon as we can.
9805                          */
9806                         continue;
9807                 }
9808         }
9809         if (rsm && (rsm->r_dupack < 0xff)) {
9810                 rsm->r_dupack++;
9811                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9812                         struct timeval tv;
9813                         uint32_t cts;
9814                         /*
9815                          * Here we see if we need to retransmit. For
9816                          * a SACK type connection if enough time has passed
9817                          * we will get a return of the rsm. For a non-sack
9818                          * connection we will get the rsm returned if the
9819                          * dupack value is 3 or more.
9820                          */
9821                         cts = tcp_get_usecs(&tv);
9822                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9823                         if (rack->r_ctl.rc_resend != NULL) {
9824                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9825                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9826                                                          rack->rc_tp->snd_una, __LINE__);
9827                                 }
9828                                 rack->r_wanted_output = 1;
9829                                 rack->r_timer_override = 1;
9830                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9831                         }
9832                 } else {
9833                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9834                 }
9835         }
9836 }
9837
9838 static void
9839 rack_check_bottom_drag(struct tcpcb *tp,
9840                        struct tcp_rack *rack,
9841                        struct socket *so, int32_t acked)
9842 {
9843         uint32_t segsiz, minseg;
9844
9845         segsiz = ctf_fixed_maxseg(tp);
9846         minseg = segsiz;
9847
9848         if (tp->snd_max == tp->snd_una) {
9849                 /*
9850                  * We are doing dynamic pacing and we are way
9851                  * under. Basically everything got acked while
9852                  * we were still waiting on the pacer to expire.
9853                  *
9854                  * This means we need to boost the b/w in
9855                  * addition to any earlier boosting of
9856                  * the multiplier.
9857                  */
9858                 rack->rc_dragged_bottom = 1;
9859                 rack_validate_multipliers_at_or_above100(rack);
9860                 /*
9861                  * Lets use the segment bytes acked plus
9862                  * the lowest RTT seen as the basis to
9863                  * form a b/w estimate. This will be off
9864                  * due to the fact that the true estimate
9865                  * should be around 1/2 the time of the RTT
9866                  * but we can settle for that.
9867                  */
9868                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9869                     acked) {
9870                         uint64_t bw, calc_bw, rtt;
9871
9872                         rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9873                         if (rtt == 0) {
9874                                 /* no us sample is there a ms one? */
9875                                 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9876                                         rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9877                                 } else {
9878                                         goto no_measurement;
9879                                 }
9880                         }
9881                         bw = acked;
9882                         calc_bw = bw * 1000000;
9883                         calc_bw /= rtt;
9884                         if (rack->r_ctl.last_max_bw &&
9885                             (rack->r_ctl.last_max_bw < calc_bw)) {
9886                                 /*
9887                                  * If we have a last calculated max bw
9888                                  * enforce it.
9889                                  */
9890                                 calc_bw = rack->r_ctl.last_max_bw;
9891                         }
9892                         /* now plop it in */
9893                         if (rack->rc_gp_filled == 0) {
9894                                 if (calc_bw > ONE_POINT_TWO_MEG) {
9895                                         /*
9896                                          * If we have no measurement
9897                                          * don't let us set in more than
9898                                          * 1.2Mbps. If we are still too
9899                                          * low after pacing with this we
9900                                          * will hopefully have a max b/w
9901                                          * available to sanity check things.
9902                                          */
9903                                         calc_bw = ONE_POINT_TWO_MEG;
9904                                 }
9905                                 rack->r_ctl.rc_rtt_diff = 0;
9906                                 rack->r_ctl.gp_bw = calc_bw;
9907                                 rack->rc_gp_filled = 1;
9908                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9909                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9910                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9911                         } else if (calc_bw > rack->r_ctl.gp_bw) {
9912                                 rack->r_ctl.rc_rtt_diff = 0;
9913                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9914                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9915                                 rack->r_ctl.gp_bw = calc_bw;
9916                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9917                         } else
9918                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9919                         if ((rack->gp_ready == 0) &&
9920                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9921                                 /* We have enough measurements now */
9922                                 rack->gp_ready = 1;
9923                                 rack_set_cc_pacing(rack);
9924                                 if (rack->defer_options)
9925                                         rack_apply_deferred_options(rack);
9926                         }
9927                         /*
9928                          * For acks over 1mss we do a extra boost to simulate
9929                          * where we would get 2 acks (we want 110 for the mul).
9930                          */
9931                         if (acked > segsiz)
9932                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9933                 } else {
9934                         /*
9935                          * zero rtt possibly?, settle for just an old increase.
9936                          */
9937 no_measurement:
9938                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
9939                 }
9940         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9941                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9942                                                minseg)) &&
9943                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9944                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9945                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9946                     (segsiz * rack_req_segs))) {
9947                 /*
9948                  * We are doing dynamic GP pacing and
9949                  * we have everything except 1MSS or less
9950                  * bytes left out. We are still pacing away.
9951                  * And there is data that could be sent, This
9952                  * means we are inserting delayed ack time in
9953                  * our measurements because we are pacing too slow.
9954                  */
9955                 rack_validate_multipliers_at_or_above100(rack);
9956                 rack->rc_dragged_bottom = 1;
9957                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9958         }
9959 }
9960
9961
9962
9963 static void
9964 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9965 {
9966         /*
9967          * The fast output path is enabled and we
9968          * have moved the cumack forward. Lets see if
9969          * we can expand forward the fast path length by
9970          * that amount. What we would ideally like to
9971          * do is increase the number of bytes in the
9972          * fast path block (left_to_send) by the
9973          * acked amount. However we have to gate that
9974          * by two factors:
9975          * 1) The amount outstanding and the rwnd of the peer
9976          *    (i.e. we don't want to exceed the rwnd of the peer).
9977          *    <and>
9978          * 2) The amount of data left in the socket buffer (i.e.
9979          *    we can't send beyond what is in the buffer).
9980          *
9981          * Note that this does not take into account any increase
9982          * in the cwnd. We will only extend the fast path by
9983          * what was acked.
9984          */
9985         uint32_t new_total, gating_val;
9986
9987         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9988         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9989                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9990         if (new_total <= gating_val) {
9991                 /* We can increase left_to_send by the acked amount */
9992                 counter_u64_add(rack_extended_rfo, 1);
9993                 rack->r_ctl.fsb.left_to_send = new_total;
9994                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9995                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
9996                          rack, rack->r_ctl.fsb.left_to_send,
9997                          sbavail(&rack->rc_inp->inp_socket->so_snd),
9998                          (tp->snd_max - tp->snd_una)));
9999
10000         }
10001 }
10002
10003 static void
10004 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
10005 {
10006         /*
10007          * Here any sendmap entry that points to the
10008          * beginning mbuf must be adjusted to the correct
10009          * offset. This must be called with:
10010          * 1) The socket buffer locked
10011          * 2) snd_una adjusted to its new postion.
10012          *
10013          * Note that (2) implies rack_ack_received has also
10014          * been called.
10015          *
10016          * We grab the first mbuf in the socket buffer and
10017          * then go through the front of the sendmap, recalculating
10018          * the stored offset for any sendmap entry that has
10019          * that mbuf. We must use the sb functions to do this
10020          * since its possible an add was done has well as
10021          * the subtraction we may have just completed. This should
10022          * not be a penalty though, since we just referenced the sb
10023          * to go in and trim off the mbufs that we freed (of course
10024          * there will be a penalty for the sendmap references though).
10025          */
10026         struct mbuf *m;
10027         struct rack_sendmap *rsm;
10028
10029         SOCKBUF_LOCK_ASSERT(sb);
10030         m = sb->sb_mb;
10031         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
10032         if ((rsm == NULL) || (m == NULL)) {
10033                 /* Nothing outstanding */
10034                 return;
10035         }
10036         while (rsm->m && (rsm->m == m)) {
10037                 /* one to adjust */
10038 #ifdef INVARIANTS
10039                 struct mbuf *tm;
10040                 uint32_t soff;
10041
10042                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
10043                 if (rsm->orig_m_len != m->m_len) {
10044                         rack_adjust_orig_mlen(rsm);
10045                 }
10046                 if (rsm->soff != soff) {
10047                         /*
10048                          * This is not a fatal error, we anticipate it
10049                          * might happen (the else code), so we count it here
10050                          * so that under invariant we can see that it really
10051                          * does happen.
10052                          */
10053                         counter_u64_add(rack_adjust_map_bw, 1);
10054                 }
10055                 rsm->m = tm;
10056                 rsm->soff = soff;
10057                 if (tm)
10058                         rsm->orig_m_len = rsm->m->m_len;
10059                 else
10060                         rsm->orig_m_len = 0;
10061 #else
10062                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
10063                 if (rsm->m)
10064                         rsm->orig_m_len = rsm->m->m_len;
10065                 else
10066                         rsm->orig_m_len = 0;
10067 #endif
10068                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
10069                               rsm);
10070                 if (rsm == NULL)
10071                         break;
10072         }
10073 }
10074
10075 /*
10076  * Return value of 1, we do not need to call rack_process_data().
10077  * return value of 0, rack_process_data can be called.
10078  * For ret_val if its 0 the TCP is locked, if its non-zero
10079  * its unlocked and probably unsafe to touch the TCB.
10080  */
10081 static int
10082 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10083     struct tcpcb *tp, struct tcpopt *to,
10084     uint32_t tiwin, int32_t tlen,
10085     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10086 {
10087         int32_t ourfinisacked = 0;
10088         int32_t nsegs, acked_amount;
10089         int32_t acked;
10090         struct mbuf *mfree;
10091         struct tcp_rack *rack;
10092         int32_t under_pacing = 0;
10093         int32_t recovery = 0;
10094
10095         INP_WLOCK_ASSERT(tptoinpcb(tp));
10096
10097         rack = (struct tcp_rack *)tp->t_fb_ptr;
10098         if (SEQ_GT(th->th_ack, tp->snd_max)) {
10099                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10100                                       &rack->r_ctl.challenge_ack_ts,
10101                                       &rack->r_ctl.challenge_ack_cnt);
10102                 rack->r_wanted_output = 1;
10103                 return (1);
10104         }
10105         if (rack->gp_ready &&
10106             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10107                 under_pacing = 1;
10108         }
10109         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10110                 int in_rec, dup_ack_struck = 0;
10111
10112                 in_rec = IN_FASTRECOVERY(tp->t_flags);
10113                 if (rack->rc_in_persist) {
10114                         tp->t_rxtshift = 0;
10115                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10116                                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10117                 }
10118                 if ((th->th_ack == tp->snd_una) &&
10119                     (tiwin == tp->snd_wnd) &&
10120                     ((to->to_flags & TOF_SACK) == 0)) {
10121                         rack_strike_dupack(rack);
10122                         dup_ack_struck = 1;
10123                 }
10124                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10125         }
10126         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10127                 /*
10128                  * Old ack, behind (or duplicate to) the last one rcv'd
10129                  * Note: We mark reordering is occuring if its
10130                  * less than and we have not closed our window.
10131                  */
10132                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10133                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10134                 }
10135                 return (0);
10136         }
10137         /*
10138          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10139          * something we sent.
10140          */
10141         if (tp->t_flags & TF_NEEDSYN) {
10142                 /*
10143                  * T/TCP: Connection was half-synchronized, and our SYN has
10144                  * been ACK'd (so connection is now fully synchronized).  Go
10145                  * to non-starred state, increment snd_una for ACK of SYN,
10146                  * and check if we can do window scaling.
10147                  */
10148                 tp->t_flags &= ~TF_NEEDSYN;
10149                 tp->snd_una++;
10150                 /* Do window scaling? */
10151                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10152                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10153                         tp->rcv_scale = tp->request_r_scale;
10154                         /* Send window already scaled. */
10155                 }
10156         }
10157         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10158
10159         acked = BYTES_THIS_ACK(tp, th);
10160         if (acked) {
10161                 /*
10162                  * Any time we move the cum-ack forward clear
10163                  * keep-alive tied probe-not-answered. The
10164                  * persists clears its own on entry.
10165                  */
10166                 rack->probe_not_answered = 0;
10167         }
10168         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10169         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10170         /*
10171          * If we just performed our first retransmit, and the ACK arrives
10172          * within our recovery window, then it was a mistake to do the
10173          * retransmit in the first place.  Recover our original cwnd and
10174          * ssthresh, and proceed to transmit where we left off.
10175          */
10176         if ((tp->t_flags & TF_PREVVALID) &&
10177             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10178                 tp->t_flags &= ~TF_PREVVALID;
10179                 if (tp->t_rxtshift == 1 &&
10180                     (int)(ticks - tp->t_badrxtwin) < 0)
10181                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10182         }
10183         if (acked) {
10184                 /* assure we are not backed off */
10185                 tp->t_rxtshift = 0;
10186                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10187                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10188                 rack->rc_tlp_in_progress = 0;
10189                 rack->r_ctl.rc_tlp_cnt_out = 0;
10190                 /*
10191                  * If it is the RXT timer we want to
10192                  * stop it, so we can restart a TLP.
10193                  */
10194                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10195                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10196 #ifdef NETFLIX_HTTP_LOGGING
10197                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10198 #endif
10199         }
10200         /*
10201          * If we have a timestamp reply, update smoothed round trip time. If
10202          * no timestamp is present but transmit timer is running and timed
10203          * sequence number was acked, update smoothed round trip time. Since
10204          * we now have an rtt measurement, cancel the timer backoff (cf.,
10205          * Phil Karn's retransmit alg.). Recompute the initial retransmit
10206          * timer.
10207          *
10208          * Some boxes send broken timestamp replies during the SYN+ACK
10209          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10210          * and blow up the retransmit timer.
10211          */
10212         /*
10213          * If all outstanding data is acked, stop retransmit timer and
10214          * remember to restart (more output or persist). If there is more
10215          * data to be acked, restart retransmit timer, using current
10216          * (possibly backed-off) value.
10217          */
10218         if (acked == 0) {
10219                 if (ofia)
10220                         *ofia = ourfinisacked;
10221                 return (0);
10222         }
10223         if (IN_RECOVERY(tp->t_flags)) {
10224                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10225                     (SEQ_LT(th->th_ack, tp->snd_max))) {
10226                         tcp_rack_partialack(tp);
10227                 } else {
10228                         rack_post_recovery(tp, th->th_ack);
10229                         recovery = 1;
10230                 }
10231         }
10232         /*
10233          * Let the congestion control algorithm update congestion control
10234          * related information. This typically means increasing the
10235          * congestion window.
10236          */
10237         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10238         SOCKBUF_LOCK(&so->so_snd);
10239         acked_amount = min(acked, (int)sbavail(&so->so_snd));
10240         tp->snd_wnd -= acked_amount;
10241         mfree = sbcut_locked(&so->so_snd, acked_amount);
10242         if ((sbused(&so->so_snd) == 0) &&
10243             (acked > acked_amount) &&
10244             (tp->t_state >= TCPS_FIN_WAIT_1) &&
10245             (tp->t_flags & TF_SENTFIN)) {
10246                 /*
10247                  * We must be sure our fin
10248                  * was sent and acked (we can be
10249                  * in FIN_WAIT_1 without having
10250                  * sent the fin).
10251                  */
10252                 ourfinisacked = 1;
10253         }
10254         tp->snd_una = th->th_ack;
10255         if (acked_amount && sbavail(&so->so_snd))
10256                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10257         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10258         /* NB: sowwakeup_locked() does an implicit unlock. */
10259         sowwakeup_locked(so);
10260         m_freem(mfree);
10261         if (SEQ_GT(tp->snd_una, tp->snd_recover))
10262                 tp->snd_recover = tp->snd_una;
10263
10264         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10265                 tp->snd_nxt = tp->snd_una;
10266         }
10267         if (under_pacing &&
10268             (rack->use_fixed_rate == 0) &&
10269             (rack->in_probe_rtt == 0) &&
10270             rack->rc_gp_dyn_mul &&
10271             rack->rc_always_pace) {
10272                 /* Check if we are dragging bottom */
10273                 rack_check_bottom_drag(tp, rack, so, acked);
10274         }
10275         if (tp->snd_una == tp->snd_max) {
10276                 /* Nothing left outstanding */
10277                 tp->t_flags &= ~TF_PREVVALID;
10278                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10279                 rack->r_ctl.retran_during_recovery = 0;
10280                 rack->r_ctl.dsack_byte_cnt = 0;
10281                 if (rack->r_ctl.rc_went_idle_time == 0)
10282                         rack->r_ctl.rc_went_idle_time = 1;
10283                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10284                 if (sbavail(&tptosocket(tp)->so_snd) == 0)
10285                         tp->t_acktime = 0;
10286                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10287                 /* Set need output so persist might get set */
10288                 rack->r_wanted_output = 1;
10289                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10290                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10291                     (sbavail(&so->so_snd) == 0) &&
10292                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10293                         /*
10294                          * The socket was gone and the
10295                          * peer sent data (now or in the past), time to
10296                          * reset him.
10297                          */
10298                         *ret_val = 1;
10299                         /* tcp_close will kill the inp pre-log the Reset */
10300                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10301                         tp = tcp_close(tp);
10302                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10303                         return (1);
10304                 }
10305         }
10306         if (ofia)
10307                 *ofia = ourfinisacked;
10308         return (0);
10309 }
10310
10311
10312 static void
10313 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
10314                   int dir, uint32_t flags, struct rack_sendmap *rsm)
10315 {
10316         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
10317                 union tcp_log_stackspecific log;
10318                 struct timeval tv;
10319
10320                 memset(&log, 0, sizeof(log));
10321                 log.u_bbr.flex1 = cnt;
10322                 log.u_bbr.flex2 = split;
10323                 log.u_bbr.flex3 = out;
10324                 log.u_bbr.flex4 = line;
10325                 log.u_bbr.flex5 = rack->r_must_retran;
10326                 log.u_bbr.flex6 = flags;
10327                 log.u_bbr.flex7 = rack->rc_has_collapsed;
10328                 log.u_bbr.flex8 = dir;  /*
10329                                          * 1 is collapsed, 0 is uncollapsed,
10330                                          * 2 is log of a rsm being marked, 3 is a split.
10331                                          */
10332                 if (rsm == NULL)
10333                         log.u_bbr.rttProp = 0;
10334                 else
10335                         log.u_bbr.rttProp = (uint64_t)rsm;
10336                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
10337                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10338                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
10339                     &rack->rc_inp->inp_socket->so_rcv,
10340                     &rack->rc_inp->inp_socket->so_snd,
10341                     TCP_RACK_LOG_COLLAPSE, 0,
10342                     0, &log, false, &tv);
10343         }
10344 }
10345
10346 static void
10347 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, int line)
10348 {
10349         /*
10350          * Here all we do is mark the collapsed point and set the flag.
10351          * This may happen again and again, but there is no
10352          * sense splitting our map until we know where the
10353          * peer finally lands in the collapse.
10354          */
10355         rack_trace_point(rack, RACK_TP_COLLAPSED_WND);
10356         if ((rack->rc_has_collapsed == 0) ||
10357             (rack->r_ctl.last_collapse_point != (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)))
10358                 counter_u64_add(rack_collapsed_win_seen, 1);
10359         rack->r_ctl.last_collapse_point = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10360         rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
10361         rack->rc_has_collapsed = 1;
10362         rack->r_collapse_point_valid = 1;
10363         rack_log_collapse(rack, 0, 0, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
10364 }
10365
10366 static void
10367 rack_un_collapse_window(struct tcp_rack *rack, int line)
10368 {
10369         struct rack_sendmap *nrsm, *rsm, fe;
10370         int cnt = 0, split = 0;
10371 #ifdef INVARIANTS
10372         struct rack_sendmap *insret;
10373 #endif
10374
10375         memset(&fe, 0, sizeof(fe));
10376         rack->rc_has_collapsed = 0;
10377         fe.r_start = rack->r_ctl.last_collapse_point;
10378         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10379         if (rsm == NULL) {
10380                 /* Nothing to do maybe the peer ack'ed it all */
10381                 rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10382                 return;
10383         }
10384         /* Now do we need to split this one? */
10385         if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
10386                 rack_log_collapse(rack, rsm->r_start, rsm->r_end,
10387                                   rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
10388                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10389                 if (nrsm == NULL) {
10390                         /* We can't get a rsm, mark all? */
10391                         nrsm = rsm;
10392                         goto no_split;
10393                 }
10394                 /* Clone it */
10395                 split = 1;
10396                 rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
10397 #ifndef INVARIANTS
10398                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10399 #else
10400                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10401                 if (insret != NULL) {
10402                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10403                               nrsm, insret, rack, rsm);
10404                 }
10405 #endif
10406                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
10407                                  rack->r_ctl.last_collapse_point, __LINE__);
10408                 if (rsm->r_in_tmap) {
10409                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10410                         nrsm->r_in_tmap = 1;
10411                 }
10412                 /*
10413                  * Set in the new RSM as the
10414                  * collapsed starting point
10415                  */
10416                 rsm = nrsm;
10417         }
10418 no_split:
10419         RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10420                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
10421                 rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
10422                 cnt++;
10423         }
10424         if (cnt) {
10425                 counter_u64_add(rack_collapsed_win, 1);
10426         }
10427         rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10428 }
10429
10430 static void
10431 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10432                         int32_t tlen, int32_t tfo_syn)
10433 {
10434         if (DELAY_ACK(tp, tlen) || tfo_syn) {
10435                 if (rack->rc_dack_mode &&
10436                     (tlen > 500) &&
10437                     (rack->rc_dack_toggle == 1)) {
10438                         goto no_delayed_ack;
10439                 }
10440                 rack_timer_cancel(tp, rack,
10441                                   rack->r_ctl.rc_rcvtime, __LINE__);
10442                 tp->t_flags |= TF_DELACK;
10443         } else {
10444 no_delayed_ack:
10445                 rack->r_wanted_output = 1;
10446                 tp->t_flags |= TF_ACKNOW;
10447                 if (rack->rc_dack_mode) {
10448                         if (tp->t_flags & TF_DELACK)
10449                                 rack->rc_dack_toggle = 1;
10450                         else
10451                                 rack->rc_dack_toggle = 0;
10452                 }
10453         }
10454 }
10455
10456 static void
10457 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10458 {
10459         /*
10460          * If fast output is in progress, lets validate that
10461          * the new window did not shrink on us and make it
10462          * so fast output should end.
10463          */
10464         if (rack->r_fast_output) {
10465                 uint32_t out;
10466
10467                 /*
10468                  * Calculate what we will send if left as is
10469                  * and compare that to our send window.
10470                  */
10471                 out = ctf_outstanding(tp);
10472                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10473                         /* ok we have an issue */
10474                         if (out >= tp->snd_wnd) {
10475                                 /* Turn off fast output the window is met or collapsed */
10476                                 rack->r_fast_output = 0;
10477                         } else {
10478                                 /* we have some room left */
10479                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10480                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10481                                         /* If not at least 1 full segment never mind */
10482                                         rack->r_fast_output = 0;
10483                                 }
10484                         }
10485                 }
10486         }
10487 }
10488
10489
10490 /*
10491  * Return value of 1, the TCB is unlocked and most
10492  * likely gone, return value of 0, the TCP is still
10493  * locked.
10494  */
10495 static int
10496 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10497     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10498     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10499 {
10500         /*
10501          * Update window information. Don't look at window if no ACK: TAC's
10502          * send garbage on first SYN.
10503          */
10504         int32_t nsegs;
10505         int32_t tfo_syn;
10506         struct tcp_rack *rack;
10507
10508         INP_WLOCK_ASSERT(tptoinpcb(tp));
10509
10510         rack = (struct tcp_rack *)tp->t_fb_ptr;
10511         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10512         if ((thflags & TH_ACK) &&
10513             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10514             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10515             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10516                 /* keep track of pure window updates */
10517                 if (tlen == 0 &&
10518                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10519                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10520                 tp->snd_wnd = tiwin;
10521                 rack_validate_fo_sendwin_up(tp, rack);
10522                 tp->snd_wl1 = th->th_seq;
10523                 tp->snd_wl2 = th->th_ack;
10524                 if (tp->snd_wnd > tp->max_sndwnd)
10525                         tp->max_sndwnd = tp->snd_wnd;
10526                 rack->r_wanted_output = 1;
10527         } else if (thflags & TH_ACK) {
10528                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10529                         tp->snd_wnd = tiwin;
10530                         rack_validate_fo_sendwin_up(tp, rack);
10531                         tp->snd_wl1 = th->th_seq;
10532                         tp->snd_wl2 = th->th_ack;
10533                 }
10534         }
10535         if (tp->snd_wnd < ctf_outstanding(tp))
10536                 /* The peer collapsed the window */
10537                 rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
10538         else if (rack->rc_has_collapsed)
10539                 rack_un_collapse_window(rack, __LINE__);
10540         if ((rack->r_collapse_point_valid) &&
10541             (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
10542                 rack->r_collapse_point_valid = 0;
10543         /* Was persist timer active and now we have window space? */
10544         if ((rack->rc_in_persist != 0) &&
10545             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10546                                 rack->r_ctl.rc_pace_min_segs))) {
10547                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10548                 tp->snd_nxt = tp->snd_max;
10549                 /* Make sure we output to start the timer */
10550                 rack->r_wanted_output = 1;
10551         }
10552         /* Do we enter persists? */
10553         if ((rack->rc_in_persist == 0) &&
10554             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10555             TCPS_HAVEESTABLISHED(tp->t_state) &&
10556             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10557             sbavail(&tptosocket(tp)->so_snd) &&
10558             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
10559                 /*
10560                  * Here the rwnd is less than
10561                  * the pacing size, we are established,
10562                  * nothing is outstanding, and there is
10563                  * data to send. Enter persists.
10564                  */
10565                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10566         }
10567         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10568                 m_freem(m);
10569                 return (0);
10570         }
10571         /*
10572          * don't process the URG bit, ignore them drag
10573          * along the up.
10574          */
10575         tp->rcv_up = tp->rcv_nxt;
10576
10577         /*
10578          * Process the segment text, merging it into the TCP sequencing
10579          * queue, and arranging for acknowledgment of receipt if necessary.
10580          * This process logically involves adjusting tp->rcv_wnd as data is
10581          * presented to the user (this happens in tcp_usrreq.c, case
10582          * PRU_RCVD).  If a FIN has already been received on this connection
10583          * then we just ignore the text.
10584          */
10585         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10586                    IS_FASTOPEN(tp->t_flags));
10587         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10588             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10589                 tcp_seq save_start = th->th_seq;
10590                 tcp_seq save_rnxt  = tp->rcv_nxt;
10591                 int     save_tlen  = tlen;
10592
10593                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10594                 /*
10595                  * Insert segment which includes th into TCP reassembly
10596                  * queue with control block tp.  Set thflags to whether
10597                  * reassembly now includes a segment with FIN.  This handles
10598                  * the common case inline (segment is the next to be
10599                  * received on an established connection, and the queue is
10600                  * empty), avoiding linkage into and removal from the queue
10601                  * and repetition of various conversions. Set DELACK for
10602                  * segments received in order, but ack immediately when
10603                  * segments are out of order (so fast retransmit can work).
10604                  */
10605                 if (th->th_seq == tp->rcv_nxt &&
10606                     SEGQ_EMPTY(tp) &&
10607                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
10608                     tfo_syn)) {
10609 #ifdef NETFLIX_SB_LIMITS
10610                         u_int mcnt, appended;
10611
10612                         if (so->so_rcv.sb_shlim) {
10613                                 mcnt = m_memcnt(m);
10614                                 appended = 0;
10615                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10616                                     CFO_NOSLEEP, NULL) == false) {
10617                                         counter_u64_add(tcp_sb_shlim_fails, 1);
10618                                         m_freem(m);
10619                                         return (0);
10620                                 }
10621                         }
10622 #endif
10623                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10624                         tp->rcv_nxt += tlen;
10625                         if (tlen &&
10626                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10627                             (tp->t_fbyte_in == 0)) {
10628                                 tp->t_fbyte_in = ticks;
10629                                 if (tp->t_fbyte_in == 0)
10630                                         tp->t_fbyte_in = 1;
10631                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10632                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10633                         }
10634                         thflags = tcp_get_flags(th) & TH_FIN;
10635                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10636                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10637                         SOCKBUF_LOCK(&so->so_rcv);
10638                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10639                                 m_freem(m);
10640                         } else
10641 #ifdef NETFLIX_SB_LIMITS
10642                                 appended =
10643 #endif
10644                                         sbappendstream_locked(&so->so_rcv, m, 0);
10645
10646                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10647                         /* NB: sorwakeup_locked() does an implicit unlock. */
10648                         sorwakeup_locked(so);
10649 #ifdef NETFLIX_SB_LIMITS
10650                         if (so->so_rcv.sb_shlim && appended != mcnt)
10651                                 counter_fo_release(so->so_rcv.sb_shlim,
10652                                     mcnt - appended);
10653 #endif
10654                 } else {
10655                         /*
10656                          * XXX: Due to the header drop above "th" is
10657                          * theoretically invalid by now.  Fortunately
10658                          * m_adj() doesn't actually frees any mbufs when
10659                          * trimming from the head.
10660                          */
10661                         tcp_seq temp = save_start;
10662
10663                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
10664                         tp->t_flags |= TF_ACKNOW;
10665                         if (tp->t_flags & TF_WAKESOR) {
10666                                 tp->t_flags &= ~TF_WAKESOR;
10667                                 /* NB: sorwakeup_locked() does an implicit unlock. */
10668                                 sorwakeup_locked(so);
10669                         }
10670                 }
10671                 if ((tp->t_flags & TF_SACK_PERMIT) &&
10672                     (save_tlen > 0) &&
10673                     TCPS_HAVEESTABLISHED(tp->t_state)) {
10674                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10675                                 /*
10676                                  * DSACK actually handled in the fastpath
10677                                  * above.
10678                                  */
10679                                 RACK_OPTS_INC(tcp_sack_path_1);
10680                                 tcp_update_sack_list(tp, save_start,
10681                                     save_start + save_tlen);
10682                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10683                                 if ((tp->rcv_numsacks >= 1) &&
10684                                     (tp->sackblks[0].end == save_start)) {
10685                                         /*
10686                                          * Partial overlap, recorded at todrop
10687                                          * above.
10688                                          */
10689                                         RACK_OPTS_INC(tcp_sack_path_2a);
10690                                         tcp_update_sack_list(tp,
10691                                             tp->sackblks[0].start,
10692                                             tp->sackblks[0].end);
10693                                 } else {
10694                                         RACK_OPTS_INC(tcp_sack_path_2b);
10695                                         tcp_update_dsack_list(tp, save_start,
10696                                             save_start + save_tlen);
10697                                 }
10698                         } else if (tlen >= save_tlen) {
10699                                 /* Update of sackblks. */
10700                                 RACK_OPTS_INC(tcp_sack_path_3);
10701                                 tcp_update_dsack_list(tp, save_start,
10702                                     save_start + save_tlen);
10703                         } else if (tlen > 0) {
10704                                 RACK_OPTS_INC(tcp_sack_path_4);
10705                                 tcp_update_dsack_list(tp, save_start,
10706                                     save_start + tlen);
10707                         }
10708                 }
10709         } else {
10710                 m_freem(m);
10711                 thflags &= ~TH_FIN;
10712         }
10713
10714         /*
10715          * If FIN is received ACK the FIN and let the user know that the
10716          * connection is closing.
10717          */
10718         if (thflags & TH_FIN) {
10719                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10720                         /* The socket upcall is handled by socantrcvmore. */
10721                         socantrcvmore(so);
10722                         /*
10723                          * If connection is half-synchronized (ie NEEDSYN
10724                          * flag on) then delay ACK, so it may be piggybacked
10725                          * when SYN is sent. Otherwise, since we received a
10726                          * FIN then no more input can be expected, send ACK
10727                          * now.
10728                          */
10729                         if (tp->t_flags & TF_NEEDSYN) {
10730                                 rack_timer_cancel(tp, rack,
10731                                     rack->r_ctl.rc_rcvtime, __LINE__);
10732                                 tp->t_flags |= TF_DELACK;
10733                         } else {
10734                                 tp->t_flags |= TF_ACKNOW;
10735                         }
10736                         tp->rcv_nxt++;
10737                 }
10738                 switch (tp->t_state) {
10739                         /*
10740                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
10741                          * CLOSE_WAIT state.
10742                          */
10743                 case TCPS_SYN_RECEIVED:
10744                         tp->t_starttime = ticks;
10745                         /* FALLTHROUGH */
10746                 case TCPS_ESTABLISHED:
10747                         rack_timer_cancel(tp, rack,
10748                             rack->r_ctl.rc_rcvtime, __LINE__);
10749                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
10750                         break;
10751
10752                         /*
10753                          * If still in FIN_WAIT_1 STATE FIN has not been
10754                          * acked so enter the CLOSING state.
10755                          */
10756                 case TCPS_FIN_WAIT_1:
10757                         rack_timer_cancel(tp, rack,
10758                             rack->r_ctl.rc_rcvtime, __LINE__);
10759                         tcp_state_change(tp, TCPS_CLOSING);
10760                         break;
10761
10762                         /*
10763                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
10764                          * starting the time-wait timer, turning off the
10765                          * other standard timers.
10766                          */
10767                 case TCPS_FIN_WAIT_2:
10768                         rack_timer_cancel(tp, rack,
10769                             rack->r_ctl.rc_rcvtime, __LINE__);
10770                         tcp_twstart(tp);
10771                         return (1);
10772                 }
10773         }
10774         /*
10775          * Return any desired output.
10776          */
10777         if ((tp->t_flags & TF_ACKNOW) ||
10778             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10779                 rack->r_wanted_output = 1;
10780         }
10781         return (0);
10782 }
10783
10784 /*
10785  * Here nothing is really faster, its just that we
10786  * have broken out the fast-data path also just like
10787  * the fast-ack.
10788  */
10789 static int
10790 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10791     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10792     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10793 {
10794         int32_t nsegs;
10795         int32_t newsize = 0;    /* automatic sockbuf scaling */
10796         struct tcp_rack *rack;
10797 #ifdef NETFLIX_SB_LIMITS
10798         u_int mcnt, appended;
10799 #endif
10800 #ifdef TCPDEBUG
10801         /*
10802          * The size of tcp_saveipgen must be the size of the max ip header,
10803          * now IPv6.
10804          */
10805         u_char tcp_saveipgen[IP6_HDR_LEN];
10806         struct tcphdr tcp_savetcp;
10807         short ostate = 0;
10808
10809 #endif
10810         /*
10811          * If last ACK falls within this segment's sequence numbers, record
10812          * the timestamp. NOTE that the test is modified according to the
10813          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10814          */
10815         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10816                 return (0);
10817         }
10818         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10819                 return (0);
10820         }
10821         if (tiwin && tiwin != tp->snd_wnd) {
10822                 return (0);
10823         }
10824         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10825                 return (0);
10826         }
10827         if (__predict_false((to->to_flags & TOF_TS) &&
10828             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10829                 return (0);
10830         }
10831         if (__predict_false((th->th_ack != tp->snd_una))) {
10832                 return (0);
10833         }
10834         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10835                 return (0);
10836         }
10837         if ((to->to_flags & TOF_TS) != 0 &&
10838             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10839                 tp->ts_recent_age = tcp_ts_getticks();
10840                 tp->ts_recent = to->to_tsval;
10841         }
10842         rack = (struct tcp_rack *)tp->t_fb_ptr;
10843         /*
10844          * This is a pure, in-sequence data packet with nothing on the
10845          * reassembly queue and we have enough buffer space to take it.
10846          */
10847         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10848
10849 #ifdef NETFLIX_SB_LIMITS
10850         if (so->so_rcv.sb_shlim) {
10851                 mcnt = m_memcnt(m);
10852                 appended = 0;
10853                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10854                     CFO_NOSLEEP, NULL) == false) {
10855                         counter_u64_add(tcp_sb_shlim_fails, 1);
10856                         m_freem(m);
10857                         return (1);
10858                 }
10859         }
10860 #endif
10861         /* Clean receiver SACK report if present */
10862         if (tp->rcv_numsacks)
10863                 tcp_clean_sackreport(tp);
10864         KMOD_TCPSTAT_INC(tcps_preddat);
10865         tp->rcv_nxt += tlen;
10866         if (tlen &&
10867             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10868             (tp->t_fbyte_in == 0)) {
10869                 tp->t_fbyte_in = ticks;
10870                 if (tp->t_fbyte_in == 0)
10871                         tp->t_fbyte_in = 1;
10872                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10873                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10874         }
10875         /*
10876          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10877          */
10878         tp->snd_wl1 = th->th_seq;
10879         /*
10880          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10881          */
10882         tp->rcv_up = tp->rcv_nxt;
10883         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10884         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10885 #ifdef TCPDEBUG
10886         if (so->so_options & SO_DEBUG)
10887                 tcp_trace(TA_INPUT, ostate, tp,
10888                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
10889 #endif
10890         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10891
10892         /* Add data to socket buffer. */
10893         SOCKBUF_LOCK(&so->so_rcv);
10894         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10895                 m_freem(m);
10896         } else {
10897                 /*
10898                  * Set new socket buffer size. Give up when limit is
10899                  * reached.
10900                  */
10901                 if (newsize)
10902                         if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
10903                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10904                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10905 #ifdef NETFLIX_SB_LIMITS
10906                 appended =
10907 #endif
10908                         sbappendstream_locked(&so->so_rcv, m, 0);
10909                 ctf_calc_rwin(so, tp);
10910         }
10911         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10912         /* NB: sorwakeup_locked() does an implicit unlock. */
10913         sorwakeup_locked(so);
10914 #ifdef NETFLIX_SB_LIMITS
10915         if (so->so_rcv.sb_shlim && mcnt != appended)
10916                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10917 #endif
10918         rack_handle_delayed_ack(tp, rack, tlen, 0);
10919         if (tp->snd_una == tp->snd_max)
10920                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10921         return (1);
10922 }
10923
10924 /*
10925  * This subfunction is used to try to highly optimize the
10926  * fast path. We again allow window updates that are
10927  * in sequence to remain in the fast-path. We also add
10928  * in the __predict's to attempt to help the compiler.
10929  * Note that if we return a 0, then we can *not* process
10930  * it and the caller should push the packet into the
10931  * slow-path.
10932  */
10933 static int
10934 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10935     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10936     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10937 {
10938         int32_t acked;
10939         int32_t nsegs;
10940 #ifdef TCPDEBUG
10941         /*
10942          * The size of tcp_saveipgen must be the size of the max ip header,
10943          * now IPv6.
10944          */
10945         u_char tcp_saveipgen[IP6_HDR_LEN];
10946         struct tcphdr tcp_savetcp;
10947         short ostate = 0;
10948 #endif
10949         int32_t under_pacing = 0;
10950         struct tcp_rack *rack;
10951
10952         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10953                 /* Old ack, behind (or duplicate to) the last one rcv'd */
10954                 return (0);
10955         }
10956         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10957                 /* Above what we have sent? */
10958                 return (0);
10959         }
10960         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10961                 /* We are retransmitting */
10962                 return (0);
10963         }
10964         if (__predict_false(tiwin == 0)) {
10965                 /* zero window */
10966                 return (0);
10967         }
10968         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10969                 /* We need a SYN or a FIN, unlikely.. */
10970                 return (0);
10971         }
10972         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10973                 /* Timestamp is behind .. old ack with seq wrap? */
10974                 return (0);
10975         }
10976         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10977                 /* Still recovering */
10978                 return (0);
10979         }
10980         rack = (struct tcp_rack *)tp->t_fb_ptr;
10981         if (rack->r_ctl.rc_sacked) {
10982                 /* We have sack holes on our scoreboard */
10983                 return (0);
10984         }
10985         /* Ok if we reach here, we can process a fast-ack */
10986         if (rack->gp_ready &&
10987             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10988                 under_pacing = 1;
10989         }
10990         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10991         rack_log_ack(tp, to, th, 0, 0);
10992         /* Did the window get updated? */
10993         if (tiwin != tp->snd_wnd) {
10994                 tp->snd_wnd = tiwin;
10995                 rack_validate_fo_sendwin_up(tp, rack);
10996                 tp->snd_wl1 = th->th_seq;
10997                 if (tp->snd_wnd > tp->max_sndwnd)
10998                         tp->max_sndwnd = tp->snd_wnd;
10999         }
11000         /* Do we exit persists? */
11001         if ((rack->rc_in_persist != 0) &&
11002             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
11003                                rack->r_ctl.rc_pace_min_segs))) {
11004                 rack_exit_persist(tp, rack, cts);
11005         }
11006         /* Do we enter persists? */
11007         if ((rack->rc_in_persist == 0) &&
11008             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
11009             TCPS_HAVEESTABLISHED(tp->t_state) &&
11010             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
11011             sbavail(&tptosocket(tp)->so_snd) &&
11012             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
11013                 /*
11014                  * Here the rwnd is less than
11015                  * the pacing size, we are established,
11016                  * nothing is outstanding, and there is
11017                  * data to send. Enter persists.
11018                  */
11019                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
11020         }
11021         /*
11022          * If last ACK falls within this segment's sequence numbers, record
11023          * the timestamp. NOTE that the test is modified according to the
11024          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
11025          */
11026         if ((to->to_flags & TOF_TS) != 0 &&
11027             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
11028                 tp->ts_recent_age = tcp_ts_getticks();
11029                 tp->ts_recent = to->to_tsval;
11030         }
11031         /*
11032          * This is a pure ack for outstanding data.
11033          */
11034         KMOD_TCPSTAT_INC(tcps_predack);
11035
11036         /*
11037          * "bad retransmit" recovery.
11038          */
11039         if ((tp->t_flags & TF_PREVVALID) &&
11040             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
11041                 tp->t_flags &= ~TF_PREVVALID;
11042                 if (tp->t_rxtshift == 1 &&
11043                     (int)(ticks - tp->t_badrxtwin) < 0)
11044                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
11045         }
11046         /*
11047          * Recalculate the transmit timer / rtt.
11048          *
11049          * Some boxes send broken timestamp replies during the SYN+ACK
11050          * phase, ignore timestamps of 0 or we could calculate a huge RTT
11051          * and blow up the retransmit timer.
11052          */
11053         acked = BYTES_THIS_ACK(tp, th);
11054
11055 #ifdef TCP_HHOOK
11056         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
11057         hhook_run_tcp_est_in(tp, th, to);
11058 #endif
11059         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11060         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11061         if (acked) {
11062                 struct mbuf *mfree;
11063
11064                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
11065                 SOCKBUF_LOCK(&so->so_snd);
11066                 mfree = sbcut_locked(&so->so_snd, acked);
11067                 tp->snd_una = th->th_ack;
11068                 /* Note we want to hold the sb lock through the sendmap adjust */
11069                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
11070                 /* Wake up the socket if we have room to write more */
11071                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11072                 sowwakeup_locked(so);
11073                 m_freem(mfree);
11074                 tp->t_rxtshift = 0;
11075                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11076                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11077                 rack->rc_tlp_in_progress = 0;
11078                 rack->r_ctl.rc_tlp_cnt_out = 0;
11079                 /*
11080                  * If it is the RXT timer we want to
11081                  * stop it, so we can restart a TLP.
11082                  */
11083                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11084                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11085 #ifdef NETFLIX_HTTP_LOGGING
11086                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
11087 #endif
11088         }
11089         /*
11090          * Let the congestion control algorithm update congestion control
11091          * related information. This typically means increasing the
11092          * congestion window.
11093          */
11094         if (tp->snd_wnd < ctf_outstanding(tp)) {
11095                 /* The peer collapsed the window */
11096                 rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
11097         } else if (rack->rc_has_collapsed)
11098                 rack_un_collapse_window(rack, __LINE__);
11099         if ((rack->r_collapse_point_valid) &&
11100             (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
11101                 rack->r_collapse_point_valid = 0;
11102         /*
11103          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
11104          */
11105         tp->snd_wl2 = th->th_ack;
11106         tp->t_dupacks = 0;
11107         m_freem(m);
11108         /* ND6_HINT(tp);         *//* Some progress has been made. */
11109
11110         /*
11111          * If all outstanding data are acked, stop retransmit timer,
11112          * otherwise restart timer using current (possibly backed-off)
11113          * value. If process is waiting for space, wakeup/selwakeup/signal.
11114          * If data are ready to send, let tcp_output decide between more
11115          * output or persist.
11116          */
11117 #ifdef TCPDEBUG
11118         if (so->so_options & SO_DEBUG)
11119                 tcp_trace(TA_INPUT, ostate, tp,
11120                     (void *)tcp_saveipgen,
11121                     &tcp_savetcp, 0);
11122 #endif
11123         if (under_pacing &&
11124             (rack->use_fixed_rate == 0) &&
11125             (rack->in_probe_rtt == 0) &&
11126             rack->rc_gp_dyn_mul &&
11127             rack->rc_always_pace) {
11128                 /* Check if we are dragging bottom */
11129                 rack_check_bottom_drag(tp, rack, so, acked);
11130         }
11131         if (tp->snd_una == tp->snd_max) {
11132                 tp->t_flags &= ~TF_PREVVALID;
11133                 rack->r_ctl.retran_during_recovery = 0;
11134                 rack->r_ctl.dsack_byte_cnt = 0;
11135                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11136                 if (rack->r_ctl.rc_went_idle_time == 0)
11137                         rack->r_ctl.rc_went_idle_time = 1;
11138                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11139                 if (sbavail(&tptosocket(tp)->so_snd) == 0)
11140                         tp->t_acktime = 0;
11141                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11142         }
11143         if (acked && rack->r_fast_output)
11144                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11145         if (sbavail(&so->so_snd)) {
11146                 rack->r_wanted_output = 1;
11147         }
11148         return (1);
11149 }
11150
11151 /*
11152  * Return value of 1, the TCB is unlocked and most
11153  * likely gone, return value of 0, the TCP is still
11154  * locked.
11155  */
11156 static int
11157 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11158     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11159     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11160 {
11161         int32_t ret_val = 0;
11162         int32_t todrop;
11163         int32_t ourfinisacked = 0;
11164         struct tcp_rack *rack;
11165
11166         INP_WLOCK_ASSERT(tptoinpcb(tp));
11167
11168         ctf_calc_rwin(so, tp);
11169         /*
11170          * If the state is SYN_SENT: if seg contains an ACK, but not for our
11171          * SYN, drop the input. if seg contains a RST, then drop the
11172          * connection. if seg does not contain SYN, then drop it. Otherwise
11173          * this is an acceptable SYN segment initialize tp->rcv_nxt and
11174          * tp->irs if seg contains ack then advance tp->snd_una if seg
11175          * contains an ECE and ECN support is enabled, the stream is ECN
11176          * capable. if SYN has been acked change to ESTABLISHED else
11177          * SYN_RCVD state arrange for segment to be acked (eventually)
11178          * continue processing rest of data/controls.
11179          */
11180         if ((thflags & TH_ACK) &&
11181             (SEQ_LEQ(th->th_ack, tp->iss) ||
11182             SEQ_GT(th->th_ack, tp->snd_max))) {
11183                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11184                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11185                 return (1);
11186         }
11187         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11188                 TCP_PROBE5(connect__refused, NULL, tp,
11189                     mtod(m, const char *), tp, th);
11190                 tp = tcp_drop(tp, ECONNREFUSED);
11191                 ctf_do_drop(m, tp);
11192                 return (1);
11193         }
11194         if (thflags & TH_RST) {
11195                 ctf_do_drop(m, tp);
11196                 return (1);
11197         }
11198         if (!(thflags & TH_SYN)) {
11199                 ctf_do_drop(m, tp);
11200                 return (1);
11201         }
11202         tp->irs = th->th_seq;
11203         tcp_rcvseqinit(tp);
11204         rack = (struct tcp_rack *)tp->t_fb_ptr;
11205         if (thflags & TH_ACK) {
11206                 int tfo_partial = 0;
11207
11208                 KMOD_TCPSTAT_INC(tcps_connects);
11209                 soisconnected(so);
11210 #ifdef MAC
11211                 mac_socketpeer_set_from_mbuf(m, so);
11212 #endif
11213                 /* Do window scaling on this connection? */
11214                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11215                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11216                         tp->rcv_scale = tp->request_r_scale;
11217                 }
11218                 tp->rcv_adv += min(tp->rcv_wnd,
11219                     TCP_MAXWIN << tp->rcv_scale);
11220                 /*
11221                  * If not all the data that was sent in the TFO SYN
11222                  * has been acked, resend the remainder right away.
11223                  */
11224                 if (IS_FASTOPEN(tp->t_flags) &&
11225                     (tp->snd_una != tp->snd_max)) {
11226                         tp->snd_nxt = th->th_ack;
11227                         tfo_partial = 1;
11228                 }
11229                 /*
11230                  * If there's data, delay ACK; if there's also a FIN ACKNOW
11231                  * will be turned on later.
11232                  */
11233                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11234                         rack_timer_cancel(tp, rack,
11235                                           rack->r_ctl.rc_rcvtime, __LINE__);
11236                         tp->t_flags |= TF_DELACK;
11237                 } else {
11238                         rack->r_wanted_output = 1;
11239                         tp->t_flags |= TF_ACKNOW;
11240                         rack->rc_dack_toggle = 0;
11241                 }
11242
11243                 tcp_ecn_input_syn_sent(tp, thflags, iptos);
11244
11245                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
11246                         /*
11247                          * We advance snd_una for the
11248                          * fast open case. If th_ack is
11249                          * acknowledging data beyond
11250                          * snd_una we can't just call
11251                          * ack-processing since the
11252                          * data stream in our send-map
11253                          * will start at snd_una + 1 (one
11254                          * beyond the SYN). If its just
11255                          * equal we don't need to do that
11256                          * and there is no send_map.
11257                          */
11258                         tp->snd_una++;
11259                 }
11260                 /*
11261                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11262                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11263                  */
11264                 tp->t_starttime = ticks;
11265                 if (tp->t_flags & TF_NEEDFIN) {
11266                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
11267                         tp->t_flags &= ~TF_NEEDFIN;
11268                         thflags &= ~TH_SYN;
11269                 } else {
11270                         tcp_state_change(tp, TCPS_ESTABLISHED);
11271                         TCP_PROBE5(connect__established, NULL, tp,
11272                             mtod(m, const char *), tp, th);
11273                         rack_cc_conn_init(tp);
11274                 }
11275         } else {
11276                 /*
11277                  * Received initial SYN in SYN-SENT[*] state => simultaneous
11278                  * open.  If segment contains CC option and there is a
11279                  * cached CC, apply TAO test. If it succeeds, connection is *
11280                  * half-synchronized. Otherwise, do 3-way handshake:
11281                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11282                  * there was no CC option, clear cached CC value.
11283                  */
11284                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
11285                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
11286         }
11287         /*
11288          * Advance th->th_seq to correspond to first data byte. If data,
11289          * trim to stay within window, dropping FIN if necessary.
11290          */
11291         th->th_seq++;
11292         if (tlen > tp->rcv_wnd) {
11293                 todrop = tlen - tp->rcv_wnd;
11294                 m_adj(m, -todrop);
11295                 tlen = tp->rcv_wnd;
11296                 thflags &= ~TH_FIN;
11297                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11298                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11299         }
11300         tp->snd_wl1 = th->th_seq - 1;
11301         tp->rcv_up = th->th_seq;
11302         /*
11303          * Client side of transaction: already sent SYN and data. If the
11304          * remote host used T/TCP to validate the SYN, our data will be
11305          * ACK'd; if so, enter normal data segment processing in the middle
11306          * of step 5, ack processing. Otherwise, goto step 6.
11307          */
11308         if (thflags & TH_ACK) {
11309                 /* For syn-sent we need to possibly update the rtt */
11310                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11311                         uint32_t t, mcts;
11312
11313                         mcts = tcp_ts_getticks();
11314                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11315                         if (!tp->t_rttlow || tp->t_rttlow > t)
11316                                 tp->t_rttlow = t;
11317                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11318                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11319                         tcp_rack_xmit_timer_commit(rack, tp);
11320                 }
11321                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11322                         return (ret_val);
11323                 /* We may have changed to FIN_WAIT_1 above */
11324                 if (tp->t_state == TCPS_FIN_WAIT_1) {
11325                         /*
11326                          * In FIN_WAIT_1 STATE in addition to the processing
11327                          * for the ESTABLISHED state if our FIN is now
11328                          * acknowledged then enter FIN_WAIT_2.
11329                          */
11330                         if (ourfinisacked) {
11331                                 /*
11332                                  * If we can't receive any more data, then
11333                                  * closing user can proceed. Starting the
11334                                  * timer is contrary to the specification,
11335                                  * but if we don't get a FIN we'll hang
11336                                  * forever.
11337                                  *
11338                                  * XXXjl: we should release the tp also, and
11339                                  * use a compressed state.
11340                                  */
11341                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11342                                         soisdisconnected(so);
11343                                         tcp_timer_activate(tp, TT_2MSL,
11344                                             (tcp_fast_finwait2_recycle ?
11345                                             tcp_finwait2_timeout :
11346                                             TP_MAXIDLE(tp)));
11347                                 }
11348                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11349                         }
11350                 }
11351         }
11352         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11353            tiwin, thflags, nxt_pkt));
11354 }
11355
11356 /*
11357  * Return value of 1, the TCB is unlocked and most
11358  * likely gone, return value of 0, the TCP is still
11359  * locked.
11360  */
11361 static int
11362 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11363     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11364     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11365 {
11366         struct tcp_rack *rack;
11367         int32_t ret_val = 0;
11368         int32_t ourfinisacked = 0;
11369
11370         ctf_calc_rwin(so, tp);
11371         if ((thflags & TH_ACK) &&
11372             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11373             SEQ_GT(th->th_ack, tp->snd_max))) {
11374                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11375                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11376                 return (1);
11377         }
11378         rack = (struct tcp_rack *)tp->t_fb_ptr;
11379         if (IS_FASTOPEN(tp->t_flags)) {
11380                 /*
11381                  * When a TFO connection is in SYN_RECEIVED, the
11382                  * only valid packets are the initial SYN, a
11383                  * retransmit/copy of the initial SYN (possibly with
11384                  * a subset of the original data), a valid ACK, a
11385                  * FIN, or a RST.
11386                  */
11387                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11388                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11389                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11390                         return (1);
11391                 } else if (thflags & TH_SYN) {
11392                         /* non-initial SYN is ignored */
11393                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11394                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11395                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11396                                 ctf_do_drop(m, NULL);
11397                                 return (0);
11398                         }
11399                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11400                         ctf_do_drop(m, NULL);
11401                         return (0);
11402                 }
11403         }
11404
11405         if ((thflags & TH_RST) ||
11406             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11407                 return (__ctf_process_rst(m, th, so, tp,
11408                                           &rack->r_ctl.challenge_ack_ts,
11409                                           &rack->r_ctl.challenge_ack_cnt));
11410         /*
11411          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11412          * it's less than ts_recent, drop it.
11413          */
11414         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11415             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11416                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11417                         return (ret_val);
11418         }
11419         /*
11420          * In the SYN-RECEIVED state, validate that the packet belongs to
11421          * this connection before trimming the data to fit the receive
11422          * window.  Check the sequence number versus IRS since we know the
11423          * sequence numbers haven't wrapped.  This is a partial fix for the
11424          * "LAND" DoS attack.
11425          */
11426         if (SEQ_LT(th->th_seq, tp->irs)) {
11427                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11428                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11429                 return (1);
11430         }
11431         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11432                               &rack->r_ctl.challenge_ack_ts,
11433                               &rack->r_ctl.challenge_ack_cnt)) {
11434                 return (ret_val);
11435         }
11436         /*
11437          * If last ACK falls within this segment's sequence numbers, record
11438          * its timestamp. NOTE: 1) That the test incorporates suggestions
11439          * from the latest proposal of the tcplw@cray.com list (Braden
11440          * 1993/04/26). 2) That updating only on newer timestamps interferes
11441          * with our earlier PAWS tests, so this check should be solely
11442          * predicated on the sequence space of this segment. 3) That we
11443          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11444          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11445          * SEG.Len, This modified check allows us to overcome RFC1323's
11446          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11447          * p.869. In such cases, we can still calculate the RTT correctly
11448          * when RCV.NXT == Last.ACK.Sent.
11449          */
11450         if ((to->to_flags & TOF_TS) != 0 &&
11451             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11452             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11453             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11454                 tp->ts_recent_age = tcp_ts_getticks();
11455                 tp->ts_recent = to->to_tsval;
11456         }
11457         tp->snd_wnd = tiwin;
11458         rack_validate_fo_sendwin_up(tp, rack);
11459         /*
11460          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11461          * is on (half-synchronized state), then queue data for later
11462          * processing; else drop segment and return.
11463          */
11464         if ((thflags & TH_ACK) == 0) {
11465                 if (IS_FASTOPEN(tp->t_flags)) {
11466                         rack_cc_conn_init(tp);
11467                 }
11468                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11469                     tiwin, thflags, nxt_pkt));
11470         }
11471         KMOD_TCPSTAT_INC(tcps_connects);
11472         if (tp->t_flags & TF_SONOTCONN) {
11473                 tp->t_flags &= ~TF_SONOTCONN;
11474                 soisconnected(so);
11475         }
11476         /* Do window scaling? */
11477         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11478             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11479                 tp->rcv_scale = tp->request_r_scale;
11480         }
11481         /*
11482          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11483          * FIN-WAIT-1
11484          */
11485         tp->t_starttime = ticks;
11486         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11487                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11488                 tp->t_tfo_pending = NULL;
11489         }
11490         if (tp->t_flags & TF_NEEDFIN) {
11491                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11492                 tp->t_flags &= ~TF_NEEDFIN;
11493         } else {
11494                 tcp_state_change(tp, TCPS_ESTABLISHED);
11495                 TCP_PROBE5(accept__established, NULL, tp,
11496                     mtod(m, const char *), tp, th);
11497                 /*
11498                  * TFO connections call cc_conn_init() during SYN
11499                  * processing.  Calling it again here for such connections
11500                  * is not harmless as it would undo the snd_cwnd reduction
11501                  * that occurs when a TFO SYN|ACK is retransmitted.
11502                  */
11503                 if (!IS_FASTOPEN(tp->t_flags))
11504                         rack_cc_conn_init(tp);
11505         }
11506         /*
11507          * Account for the ACK of our SYN prior to
11508          * regular ACK processing below, except for
11509          * simultaneous SYN, which is handled later.
11510          */
11511         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11512                 tp->snd_una++;
11513         /*
11514          * If segment contains data or ACK, will call tcp_reass() later; if
11515          * not, do so now to pass queued data to user.
11516          */
11517         if (tlen == 0 && (thflags & TH_FIN) == 0) {
11518                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11519                     (struct mbuf *)0);
11520                 if (tp->t_flags & TF_WAKESOR) {
11521                         tp->t_flags &= ~TF_WAKESOR;
11522                         /* NB: sorwakeup_locked() does an implicit unlock. */
11523                         sorwakeup_locked(so);
11524                 }
11525         }
11526         tp->snd_wl1 = th->th_seq - 1;
11527         /* For syn-recv we need to possibly update the rtt */
11528         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11529                 uint32_t t, mcts;
11530
11531                 mcts = tcp_ts_getticks();
11532                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11533                 if (!tp->t_rttlow || tp->t_rttlow > t)
11534                         tp->t_rttlow = t;
11535                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11536                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11537                 tcp_rack_xmit_timer_commit(rack, tp);
11538         }
11539         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11540                 return (ret_val);
11541         }
11542         if (tp->t_state == TCPS_FIN_WAIT_1) {
11543                 /* We could have went to FIN_WAIT_1 (or EST) above */
11544                 /*
11545                  * In FIN_WAIT_1 STATE in addition to the processing for the
11546                  * ESTABLISHED state if our FIN is now acknowledged then
11547                  * enter FIN_WAIT_2.
11548                  */
11549                 if (ourfinisacked) {
11550                         /*
11551                          * If we can't receive any more data, then closing
11552                          * user can proceed. Starting the timer is contrary
11553                          * to the specification, but if we don't get a FIN
11554                          * we'll hang forever.
11555                          *
11556                          * XXXjl: we should release the tp also, and use a
11557                          * compressed state.
11558                          */
11559                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11560                                 soisdisconnected(so);
11561                                 tcp_timer_activate(tp, TT_2MSL,
11562                                     (tcp_fast_finwait2_recycle ?
11563                                     tcp_finwait2_timeout :
11564                                     TP_MAXIDLE(tp)));
11565                         }
11566                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
11567                 }
11568         }
11569         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11570             tiwin, thflags, nxt_pkt));
11571 }
11572
11573 /*
11574  * Return value of 1, the TCB is unlocked and most
11575  * likely gone, return value of 0, the TCP is still
11576  * locked.
11577  */
11578 static int
11579 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11580     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11581     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11582 {
11583         int32_t ret_val = 0;
11584         struct tcp_rack *rack;
11585
11586         /*
11587          * Header prediction: check for the two common cases of a
11588          * uni-directional data xfer.  If the packet has no control flags,
11589          * is in-sequence, the window didn't change and we're not
11590          * retransmitting, it's a candidate.  If the length is zero and the
11591          * ack moved forward, we're the sender side of the xfer.  Just free
11592          * the data acked & wake any higher level process that was blocked
11593          * waiting for space.  If the length is non-zero and the ack didn't
11594          * move, we're the receiver side.  If we're getting packets in-order
11595          * (the reassembly queue is empty), add the data toc The socket
11596          * buffer and note that we need a delayed ack. Make sure that the
11597          * hidden state-flags are also off. Since we check for
11598          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11599          */
11600         rack = (struct tcp_rack *)tp->t_fb_ptr;
11601         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11602             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11603             __predict_true(SEGQ_EMPTY(tp)) &&
11604             __predict_true(th->th_seq == tp->rcv_nxt)) {
11605                 if (tlen == 0) {
11606                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11607                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11608                                 return (0);
11609                         }
11610                 } else {
11611                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11612                             tiwin, nxt_pkt, iptos)) {
11613                                 return (0);
11614                         }
11615                 }
11616         }
11617         ctf_calc_rwin(so, tp);
11618
11619         if ((thflags & TH_RST) ||
11620             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11621                 return (__ctf_process_rst(m, th, so, tp,
11622                                           &rack->r_ctl.challenge_ack_ts,
11623                                           &rack->r_ctl.challenge_ack_cnt));
11624
11625         /*
11626          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11627          * synchronized state.
11628          */
11629         if (thflags & TH_SYN) {
11630                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11631                 return (ret_val);
11632         }
11633         /*
11634          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11635          * it's less than ts_recent, drop it.
11636          */
11637         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11638             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11639                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11640                         return (ret_val);
11641         }
11642         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11643                               &rack->r_ctl.challenge_ack_ts,
11644                               &rack->r_ctl.challenge_ack_cnt)) {
11645                 return (ret_val);
11646         }
11647         /*
11648          * If last ACK falls within this segment's sequence numbers, record
11649          * its timestamp. NOTE: 1) That the test incorporates suggestions
11650          * from the latest proposal of the tcplw@cray.com list (Braden
11651          * 1993/04/26). 2) That updating only on newer timestamps interferes
11652          * with our earlier PAWS tests, so this check should be solely
11653          * predicated on the sequence space of this segment. 3) That we
11654          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11655          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11656          * SEG.Len, This modified check allows us to overcome RFC1323's
11657          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11658          * p.869. In such cases, we can still calculate the RTT correctly
11659          * when RCV.NXT == Last.ACK.Sent.
11660          */
11661         if ((to->to_flags & TOF_TS) != 0 &&
11662             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11663             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11664             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11665                 tp->ts_recent_age = tcp_ts_getticks();
11666                 tp->ts_recent = to->to_tsval;
11667         }
11668         /*
11669          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11670          * is on (half-synchronized state), then queue data for later
11671          * processing; else drop segment and return.
11672          */
11673         if ((thflags & TH_ACK) == 0) {
11674                 if (tp->t_flags & TF_NEEDSYN) {
11675                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11676                             tiwin, thflags, nxt_pkt));
11677
11678                 } else if (tp->t_flags & TF_ACKNOW) {
11679                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11680                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11681                         return (ret_val);
11682                 } else {
11683                         ctf_do_drop(m, NULL);
11684                         return (0);
11685                 }
11686         }
11687         /*
11688          * Ack processing.
11689          */
11690         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11691                 return (ret_val);
11692         }
11693         if (sbavail(&so->so_snd)) {
11694                 if (ctf_progress_timeout_check(tp, true)) {
11695                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11696                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11697                         return (1);
11698                 }
11699         }
11700         /* State changes only happen in rack_process_data() */
11701         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11702             tiwin, thflags, nxt_pkt));
11703 }
11704
11705 /*
11706  * Return value of 1, the TCB is unlocked and most
11707  * likely gone, return value of 0, the TCP is still
11708  * locked.
11709  */
11710 static int
11711 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11712     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11713     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11714 {
11715         int32_t ret_val = 0;
11716         struct tcp_rack *rack;
11717
11718         rack = (struct tcp_rack *)tp->t_fb_ptr;
11719         ctf_calc_rwin(so, tp);
11720         if ((thflags & TH_RST) ||
11721             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11722                 return (__ctf_process_rst(m, th, so, tp,
11723                                           &rack->r_ctl.challenge_ack_ts,
11724                                           &rack->r_ctl.challenge_ack_cnt));
11725         /*
11726          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11727          * synchronized state.
11728          */
11729         if (thflags & TH_SYN) {
11730                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11731                 return (ret_val);
11732         }
11733         /*
11734          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11735          * it's less than ts_recent, drop it.
11736          */
11737         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11738             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11739                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11740                         return (ret_val);
11741         }
11742         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11743                               &rack->r_ctl.challenge_ack_ts,
11744                               &rack->r_ctl.challenge_ack_cnt)) {
11745                 return (ret_val);
11746         }
11747         /*
11748          * If last ACK falls within this segment's sequence numbers, record
11749          * its timestamp. NOTE: 1) That the test incorporates suggestions
11750          * from the latest proposal of the tcplw@cray.com list (Braden
11751          * 1993/04/26). 2) That updating only on newer timestamps interferes
11752          * with our earlier PAWS tests, so this check should be solely
11753          * predicated on the sequence space of this segment. 3) That we
11754          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11755          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11756          * SEG.Len, This modified check allows us to overcome RFC1323's
11757          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11758          * p.869. In such cases, we can still calculate the RTT correctly
11759          * when RCV.NXT == Last.ACK.Sent.
11760          */
11761         if ((to->to_flags & TOF_TS) != 0 &&
11762             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11763             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11764             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11765                 tp->ts_recent_age = tcp_ts_getticks();
11766                 tp->ts_recent = to->to_tsval;
11767         }
11768         /*
11769          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11770          * is on (half-synchronized state), then queue data for later
11771          * processing; else drop segment and return.
11772          */
11773         if ((thflags & TH_ACK) == 0) {
11774                 if (tp->t_flags & TF_NEEDSYN) {
11775                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11776                             tiwin, thflags, nxt_pkt));
11777
11778                 } else if (tp->t_flags & TF_ACKNOW) {
11779                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11780                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11781                         return (ret_val);
11782                 } else {
11783                         ctf_do_drop(m, NULL);
11784                         return (0);
11785                 }
11786         }
11787         /*
11788          * Ack processing.
11789          */
11790         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11791                 return (ret_val);
11792         }
11793         if (sbavail(&so->so_snd)) {
11794                 if (ctf_progress_timeout_check(tp, true)) {
11795                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11796                                                 tp, tick, PROGRESS_DROP, __LINE__);
11797                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11798                         return (1);
11799                 }
11800         }
11801         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11802             tiwin, thflags, nxt_pkt));
11803 }
11804
11805 static int
11806 rack_check_data_after_close(struct mbuf *m,
11807     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11808 {
11809         struct tcp_rack *rack;
11810
11811         rack = (struct tcp_rack *)tp->t_fb_ptr;
11812         if (rack->rc_allow_data_af_clo == 0) {
11813         close_now:
11814                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11815                 /* tcp_close will kill the inp pre-log the Reset */
11816                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11817                 tp = tcp_close(tp);
11818                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11819                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11820                 return (1);
11821         }
11822         if (sbavail(&so->so_snd) == 0)
11823                 goto close_now;
11824         /* Ok we allow data that is ignored and a followup reset */
11825         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11826         tp->rcv_nxt = th->th_seq + *tlen;
11827         tp->t_flags2 |= TF2_DROP_AF_DATA;
11828         rack->r_wanted_output = 1;
11829         *tlen = 0;
11830         return (0);
11831 }
11832
11833 /*
11834  * Return value of 1, the TCB is unlocked and most
11835  * likely gone, return value of 0, the TCP is still
11836  * locked.
11837  */
11838 static int
11839 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11840     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11841     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11842 {
11843         int32_t ret_val = 0;
11844         int32_t ourfinisacked = 0;
11845         struct tcp_rack *rack;
11846
11847         rack = (struct tcp_rack *)tp->t_fb_ptr;
11848         ctf_calc_rwin(so, tp);
11849
11850         if ((thflags & TH_RST) ||
11851             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11852                 return (__ctf_process_rst(m, th, so, tp,
11853                                           &rack->r_ctl.challenge_ack_ts,
11854                                           &rack->r_ctl.challenge_ack_cnt));
11855         /*
11856          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11857          * synchronized state.
11858          */
11859         if (thflags & TH_SYN) {
11860                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11861                 return (ret_val);
11862         }
11863         /*
11864          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11865          * it's less than ts_recent, drop it.
11866          */
11867         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11868             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11869                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11870                         return (ret_val);
11871         }
11872         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11873                               &rack->r_ctl.challenge_ack_ts,
11874                               &rack->r_ctl.challenge_ack_cnt)) {
11875                 return (ret_val);
11876         }
11877         /*
11878          * If new data are received on a connection after the user processes
11879          * are gone, then RST the other end.
11880          */
11881         if ((tp->t_flags & TF_CLOSED) && tlen &&
11882             rack_check_data_after_close(m, tp, &tlen, th, so))
11883                 return (1);
11884         /*
11885          * If last ACK falls within this segment's sequence numbers, record
11886          * its timestamp. NOTE: 1) That the test incorporates suggestions
11887          * from the latest proposal of the tcplw@cray.com list (Braden
11888          * 1993/04/26). 2) That updating only on newer timestamps interferes
11889          * with our earlier PAWS tests, so this check should be solely
11890          * predicated on the sequence space of this segment. 3) That we
11891          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11892          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11893          * SEG.Len, This modified check allows us to overcome RFC1323's
11894          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11895          * p.869. In such cases, we can still calculate the RTT correctly
11896          * when RCV.NXT == Last.ACK.Sent.
11897          */
11898         if ((to->to_flags & TOF_TS) != 0 &&
11899             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11900             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11901             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11902                 tp->ts_recent_age = tcp_ts_getticks();
11903                 tp->ts_recent = to->to_tsval;
11904         }
11905         /*
11906          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11907          * is on (half-synchronized state), then queue data for later
11908          * processing; else drop segment and return.
11909          */
11910         if ((thflags & TH_ACK) == 0) {
11911                 if (tp->t_flags & TF_NEEDSYN) {
11912                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11913                             tiwin, thflags, nxt_pkt));
11914                 } else if (tp->t_flags & TF_ACKNOW) {
11915                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11916                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11917                         return (ret_val);
11918                 } else {
11919                         ctf_do_drop(m, NULL);
11920                         return (0);
11921                 }
11922         }
11923         /*
11924          * Ack processing.
11925          */
11926         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11927                 return (ret_val);
11928         }
11929         if (ourfinisacked) {
11930                 /*
11931                  * If we can't receive any more data, then closing user can
11932                  * proceed. Starting the timer is contrary to the
11933                  * specification, but if we don't get a FIN we'll hang
11934                  * forever.
11935                  *
11936                  * XXXjl: we should release the tp also, and use a
11937                  * compressed state.
11938                  */
11939                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11940                         soisdisconnected(so);
11941                         tcp_timer_activate(tp, TT_2MSL,
11942                             (tcp_fast_finwait2_recycle ?
11943                             tcp_finwait2_timeout :
11944                             TP_MAXIDLE(tp)));
11945                 }
11946                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11947         }
11948         if (sbavail(&so->so_snd)) {
11949                 if (ctf_progress_timeout_check(tp, true)) {
11950                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11951                                                 tp, tick, PROGRESS_DROP, __LINE__);
11952                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11953                         return (1);
11954                 }
11955         }
11956         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11957             tiwin, thflags, nxt_pkt));
11958 }
11959
11960 /*
11961  * Return value of 1, the TCB is unlocked and most
11962  * likely gone, return value of 0, the TCP is still
11963  * locked.
11964  */
11965 static int
11966 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11967     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11968     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11969 {
11970         int32_t ret_val = 0;
11971         int32_t ourfinisacked = 0;
11972         struct tcp_rack *rack;
11973
11974         rack = (struct tcp_rack *)tp->t_fb_ptr;
11975         ctf_calc_rwin(so, tp);
11976
11977         if ((thflags & TH_RST) ||
11978             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11979                 return (__ctf_process_rst(m, th, so, tp,
11980                                           &rack->r_ctl.challenge_ack_ts,
11981                                           &rack->r_ctl.challenge_ack_cnt));
11982         /*
11983          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11984          * synchronized state.
11985          */
11986         if (thflags & TH_SYN) {
11987                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11988                 return (ret_val);
11989         }
11990         /*
11991          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11992          * it's less than ts_recent, drop it.
11993          */
11994         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11995             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11996                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11997                         return (ret_val);
11998         }
11999         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12000                               &rack->r_ctl.challenge_ack_ts,
12001                               &rack->r_ctl.challenge_ack_cnt)) {
12002                 return (ret_val);
12003         }
12004         /*
12005          * If new data are received on a connection after the user processes
12006          * are gone, then RST the other end.
12007          */
12008         if ((tp->t_flags & TF_CLOSED) && tlen &&
12009             rack_check_data_after_close(m, tp, &tlen, th, so))
12010                 return (1);
12011         /*
12012          * If last ACK falls within this segment's sequence numbers, record
12013          * its timestamp. NOTE: 1) That the test incorporates suggestions
12014          * from the latest proposal of the tcplw@cray.com list (Braden
12015          * 1993/04/26). 2) That updating only on newer timestamps interferes
12016          * with our earlier PAWS tests, so this check should be solely
12017          * predicated on the sequence space of this segment. 3) That we
12018          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12019          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12020          * SEG.Len, This modified check allows us to overcome RFC1323's
12021          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12022          * p.869. In such cases, we can still calculate the RTT correctly
12023          * when RCV.NXT == Last.ACK.Sent.
12024          */
12025         if ((to->to_flags & TOF_TS) != 0 &&
12026             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12027             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12028             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12029                 tp->ts_recent_age = tcp_ts_getticks();
12030                 tp->ts_recent = to->to_tsval;
12031         }
12032         /*
12033          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12034          * is on (half-synchronized state), then queue data for later
12035          * processing; else drop segment and return.
12036          */
12037         if ((thflags & TH_ACK) == 0) {
12038                 if (tp->t_flags & TF_NEEDSYN) {
12039                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12040                             tiwin, thflags, nxt_pkt));
12041                 } else if (tp->t_flags & TF_ACKNOW) {
12042                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12043                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12044                         return (ret_val);
12045                 } else {
12046                         ctf_do_drop(m, NULL);
12047                         return (0);
12048                 }
12049         }
12050         /*
12051          * Ack processing.
12052          */
12053         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12054                 return (ret_val);
12055         }
12056         if (ourfinisacked) {
12057                 tcp_twstart(tp);
12058                 m_freem(m);
12059                 return (1);
12060         }
12061         if (sbavail(&so->so_snd)) {
12062                 if (ctf_progress_timeout_check(tp, true)) {
12063                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12064                                                 tp, tick, PROGRESS_DROP, __LINE__);
12065                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12066                         return (1);
12067                 }
12068         }
12069         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12070             tiwin, thflags, nxt_pkt));
12071 }
12072
12073 /*
12074  * Return value of 1, the TCB is unlocked and most
12075  * likely gone, return value of 0, the TCP is still
12076  * locked.
12077  */
12078 static int
12079 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12080     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12081     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12082 {
12083         int32_t ret_val = 0;
12084         int32_t ourfinisacked = 0;
12085         struct tcp_rack *rack;
12086
12087         rack = (struct tcp_rack *)tp->t_fb_ptr;
12088         ctf_calc_rwin(so, tp);
12089
12090         if ((thflags & TH_RST) ||
12091             (tp->t_fin_is_rst && (thflags & TH_FIN)))
12092                 return (__ctf_process_rst(m, th, so, tp,
12093                                           &rack->r_ctl.challenge_ack_ts,
12094                                           &rack->r_ctl.challenge_ack_cnt));
12095         /*
12096          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12097          * synchronized state.
12098          */
12099         if (thflags & TH_SYN) {
12100                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12101                 return (ret_val);
12102         }
12103         /*
12104          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12105          * it's less than ts_recent, drop it.
12106          */
12107         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12108             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12109                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12110                         return (ret_val);
12111         }
12112         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12113                               &rack->r_ctl.challenge_ack_ts,
12114                               &rack->r_ctl.challenge_ack_cnt)) {
12115                 return (ret_val);
12116         }
12117         /*
12118          * If new data are received on a connection after the user processes
12119          * are gone, then RST the other end.
12120          */
12121         if ((tp->t_flags & TF_CLOSED) && tlen &&
12122             rack_check_data_after_close(m, tp, &tlen, th, so))
12123                 return (1);
12124         /*
12125          * If last ACK falls within this segment's sequence numbers, record
12126          * its timestamp. NOTE: 1) That the test incorporates suggestions
12127          * from the latest proposal of the tcplw@cray.com list (Braden
12128          * 1993/04/26). 2) That updating only on newer timestamps interferes
12129          * with our earlier PAWS tests, so this check should be solely
12130          * predicated on the sequence space of this segment. 3) That we
12131          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12132          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12133          * SEG.Len, This modified check allows us to overcome RFC1323's
12134          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12135          * p.869. In such cases, we can still calculate the RTT correctly
12136          * when RCV.NXT == Last.ACK.Sent.
12137          */
12138         if ((to->to_flags & TOF_TS) != 0 &&
12139             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12140             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12141             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12142                 tp->ts_recent_age = tcp_ts_getticks();
12143                 tp->ts_recent = to->to_tsval;
12144         }
12145         /*
12146          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12147          * is on (half-synchronized state), then queue data for later
12148          * processing; else drop segment and return.
12149          */
12150         if ((thflags & TH_ACK) == 0) {
12151                 if (tp->t_flags & TF_NEEDSYN) {
12152                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12153                             tiwin, thflags, nxt_pkt));
12154                 } else if (tp->t_flags & TF_ACKNOW) {
12155                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12156                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12157                         return (ret_val);
12158                 } else {
12159                         ctf_do_drop(m, NULL);
12160                         return (0);
12161                 }
12162         }
12163         /*
12164          * case TCPS_LAST_ACK: Ack processing.
12165          */
12166         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12167                 return (ret_val);
12168         }
12169         if (ourfinisacked) {
12170                 tp = tcp_close(tp);
12171                 ctf_do_drop(m, tp);
12172                 return (1);
12173         }
12174         if (sbavail(&so->so_snd)) {
12175                 if (ctf_progress_timeout_check(tp, true)) {
12176                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12177                                                 tp, tick, PROGRESS_DROP, __LINE__);
12178                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12179                         return (1);
12180                 }
12181         }
12182         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12183             tiwin, thflags, nxt_pkt));
12184 }
12185
12186 /*
12187  * Return value of 1, the TCB is unlocked and most
12188  * likely gone, return value of 0, the TCP is still
12189  * locked.
12190  */
12191 static int
12192 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12193     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12194     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12195 {
12196         int32_t ret_val = 0;
12197         int32_t ourfinisacked = 0;
12198         struct tcp_rack *rack;
12199
12200         rack = (struct tcp_rack *)tp->t_fb_ptr;
12201         ctf_calc_rwin(so, tp);
12202
12203         /* Reset receive buffer auto scaling when not in bulk receive mode. */
12204         if ((thflags & TH_RST) ||
12205             (tp->t_fin_is_rst && (thflags & TH_FIN)))
12206                 return (__ctf_process_rst(m, th, so, tp,
12207                                           &rack->r_ctl.challenge_ack_ts,
12208                                           &rack->r_ctl.challenge_ack_cnt));
12209         /*
12210          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12211          * synchronized state.
12212          */
12213         if (thflags & TH_SYN) {
12214                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12215                 return (ret_val);
12216         }
12217         /*
12218          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12219          * it's less than ts_recent, drop it.
12220          */
12221         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12222             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12223                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12224                         return (ret_val);
12225         }
12226         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12227                               &rack->r_ctl.challenge_ack_ts,
12228                               &rack->r_ctl.challenge_ack_cnt)) {
12229                 return (ret_val);
12230         }
12231         /*
12232          * If new data are received on a connection after the user processes
12233          * are gone, then RST the other end.
12234          */
12235         if ((tp->t_flags & TF_CLOSED) && tlen &&
12236             rack_check_data_after_close(m, tp, &tlen, th, so))
12237                 return (1);
12238         /*
12239          * If last ACK falls within this segment's sequence numbers, record
12240          * its timestamp. NOTE: 1) That the test incorporates suggestions
12241          * from the latest proposal of the tcplw@cray.com list (Braden
12242          * 1993/04/26). 2) That updating only on newer timestamps interferes
12243          * with our earlier PAWS tests, so this check should be solely
12244          * predicated on the sequence space of this segment. 3) That we
12245          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12246          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12247          * SEG.Len, This modified check allows us to overcome RFC1323's
12248          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12249          * p.869. In such cases, we can still calculate the RTT correctly
12250          * when RCV.NXT == Last.ACK.Sent.
12251          */
12252         if ((to->to_flags & TOF_TS) != 0 &&
12253             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12254             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12255             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12256                 tp->ts_recent_age = tcp_ts_getticks();
12257                 tp->ts_recent = to->to_tsval;
12258         }
12259         /*
12260          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12261          * is on (half-synchronized state), then queue data for later
12262          * processing; else drop segment and return.
12263          */
12264         if ((thflags & TH_ACK) == 0) {
12265                 if (tp->t_flags & TF_NEEDSYN) {
12266                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12267                             tiwin, thflags, nxt_pkt));
12268                 } else if (tp->t_flags & TF_ACKNOW) {
12269                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12270                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12271                         return (ret_val);
12272                 } else {
12273                         ctf_do_drop(m, NULL);
12274                         return (0);
12275                 }
12276         }
12277         /*
12278          * Ack processing.
12279          */
12280         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12281                 return (ret_val);
12282         }
12283         if (sbavail(&so->so_snd)) {
12284                 if (ctf_progress_timeout_check(tp, true)) {
12285                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12286                                                 tp, tick, PROGRESS_DROP, __LINE__);
12287                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12288                         return (1);
12289                 }
12290         }
12291         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12292             tiwin, thflags, nxt_pkt));
12293 }
12294
12295 static void inline
12296 rack_clear_rate_sample(struct tcp_rack *rack)
12297 {
12298         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12299         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12300         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12301 }
12302
12303 static void
12304 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12305 {
12306         uint64_t bw_est, rate_wanted;
12307         int chged = 0;
12308         uint32_t user_max, orig_min, orig_max;
12309
12310         orig_min = rack->r_ctl.rc_pace_min_segs;
12311         orig_max = rack->r_ctl.rc_pace_max_segs;
12312         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12313         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12314                 chged = 1;
12315         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12316         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12317                 if (user_max != rack->r_ctl.rc_pace_max_segs)
12318                         chged = 1;
12319         }
12320         if (rack->rc_force_max_seg) {
12321                 rack->r_ctl.rc_pace_max_segs = user_max;
12322         } else if (rack->use_fixed_rate) {
12323                 bw_est = rack_get_bw(rack);
12324                 if ((rack->r_ctl.crte == NULL) ||
12325                     (bw_est != rack->r_ctl.crte->rate)) {
12326                         rack->r_ctl.rc_pace_max_segs = user_max;
12327                 } else {
12328                         /* We are pacing right at the hardware rate */
12329                         uint32_t segsiz;
12330
12331                         segsiz = min(ctf_fixed_maxseg(tp),
12332                                      rack->r_ctl.rc_pace_min_segs);
12333                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12334                                                            tp, bw_est, segsiz, 0,
12335                                                            rack->r_ctl.crte, NULL);
12336                 }
12337         } else if (rack->rc_always_pace) {
12338                 if (rack->r_ctl.gp_bw ||
12339 #ifdef NETFLIX_PEAKRATE
12340                     rack->rc_tp->t_maxpeakrate ||
12341 #endif
12342                     rack->r_ctl.init_rate) {
12343                         /* We have a rate of some sort set */
12344                         uint32_t  orig;
12345
12346                         bw_est = rack_get_bw(rack);
12347                         orig = rack->r_ctl.rc_pace_max_segs;
12348                         if (fill_override)
12349                                 rate_wanted = *fill_override;
12350                         else
12351                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12352                         if (rate_wanted) {
12353                                 /* We have something */
12354                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12355                                                                                    rate_wanted,
12356                                                                                    ctf_fixed_maxseg(rack->rc_tp));
12357                         } else
12358                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12359                         if (orig != rack->r_ctl.rc_pace_max_segs)
12360                                 chged = 1;
12361                 } else if ((rack->r_ctl.gp_bw == 0) &&
12362                            (rack->r_ctl.rc_pace_max_segs == 0)) {
12363                         /*
12364                          * If we have nothing limit us to bursting
12365                          * out IW sized pieces.
12366                          */
12367                         chged = 1;
12368                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12369                 }
12370         }
12371         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12372                 chged = 1;
12373                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12374         }
12375         if (chged)
12376                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12377 }
12378
12379
12380 static void
12381 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12382 {
12383 #ifdef INET6
12384         struct ip6_hdr *ip6 = NULL;
12385 #endif
12386 #ifdef INET
12387         struct ip *ip = NULL;
12388 #endif
12389         struct udphdr *udp = NULL;
12390
12391         /* Ok lets fill in the fast block, it can only be used with no IP options! */
12392 #ifdef INET6
12393         if (rack->r_is_v6) {
12394                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12395                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12396                 if (tp->t_port) {
12397                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12398                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12399                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12400                         udp->uh_dport = tp->t_port;
12401                         rack->r_ctl.fsb.udp = udp;
12402                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12403                 } else
12404                 {
12405                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12406                         rack->r_ctl.fsb.udp = NULL;
12407                 }
12408                 tcpip_fillheaders(rack->rc_inp,
12409                                   tp->t_port,
12410                                   ip6, rack->r_ctl.fsb.th);
12411         } else
12412 #endif                          /* INET6 */
12413         {
12414                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12415                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12416                 if (tp->t_port) {
12417                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12418                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12419                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12420                         udp->uh_dport = tp->t_port;
12421                         rack->r_ctl.fsb.udp = udp;
12422                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12423                 } else
12424                 {
12425                         rack->r_ctl.fsb.udp = NULL;
12426                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12427                 }
12428                 tcpip_fillheaders(rack->rc_inp,
12429                                   tp->t_port,
12430                                   ip, rack->r_ctl.fsb.th);
12431         }
12432         rack->r_fsb_inited = 1;
12433 }
12434
12435 static int
12436 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12437 {
12438         /*
12439          * Allocate the larger of spaces V6 if available else just
12440          * V4 and include udphdr (overbook)
12441          */
12442 #ifdef INET6
12443         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12444 #else
12445         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12446 #endif
12447         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12448                                             M_TCPFSB, M_NOWAIT|M_ZERO);
12449         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12450                 return (ENOMEM);
12451         }
12452         rack->r_fsb_inited = 0;
12453         return (0);
12454 }
12455
12456 static int
12457 rack_init(struct tcpcb *tp)
12458 {
12459         struct inpcb *inp = tptoinpcb(tp);
12460         struct tcp_rack *rack = NULL;
12461 #ifdef INVARIANTS
12462         struct rack_sendmap *insret;
12463 #endif
12464         uint32_t iwin, snt, us_cts;
12465         int err;
12466
12467         tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12468         if (tp->t_fb_ptr == NULL) {
12469                 /*
12470                  * We need to allocate memory but cant. The INP and INP_INFO
12471                  * locks and they are recursive (happens during setup. So a
12472                  * scheme to drop the locks fails :(
12473                  *
12474                  */
12475                 return (ENOMEM);
12476         }
12477         memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12478
12479         rack = (struct tcp_rack *)tp->t_fb_ptr;
12480         RB_INIT(&rack->r_ctl.rc_mtree);
12481         TAILQ_INIT(&rack->r_ctl.rc_free);
12482         TAILQ_INIT(&rack->r_ctl.rc_tmap);
12483         rack->rc_tp = tp;
12484         rack->rc_inp = inp;
12485         /* Set the flag */
12486         rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
12487         /* Probably not needed but lets be sure */
12488         rack_clear_rate_sample(rack);
12489         /*
12490          * Save off the default values, socket options will poke
12491          * at these if pacing is not on or we have not yet
12492          * reached where pacing is on (gp_ready/fixed enabled).
12493          * When they get set into the CC module (when gp_ready
12494          * is enabled or we enable fixed) then we will set these
12495          * values into the CC and place in here the old values
12496          * so we have a restoral. Then we will set the flag
12497          * rc_pacing_cc_set. That way whenever we turn off pacing
12498          * or switch off this stack, we will know to go restore
12499          * the saved values.
12500          */
12501         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12502         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12503         /* We want abe like behavior as well */
12504         rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12505         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12506         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12507         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12508         rack->r_ctl.roundends = tp->snd_max;
12509         if (use_rack_rr)
12510                 rack->use_rack_rr = 1;
12511         if (V_tcp_delack_enabled)
12512                 tp->t_delayed_ack = 1;
12513         else
12514                 tp->t_delayed_ack = 0;
12515 #ifdef TCP_ACCOUNTING
12516         if (rack_tcp_accounting) {
12517                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12518         }
12519 #endif
12520         if (rack_enable_shared_cwnd)
12521                 rack->rack_enable_scwnd = 1;
12522         rack->rc_user_set_max_segs = rack_hptsi_segments;
12523         rack->rc_force_max_seg = 0;
12524         if (rack_use_imac_dack)
12525                 rack->rc_dack_mode = 1;
12526         TAILQ_INIT(&rack->r_ctl.opt_list);
12527         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12528         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12529         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12530         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12531         rack->r_ctl.rc_highest_us_rtt = 0;
12532         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12533         rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12534         if (rack_use_cmp_acks)
12535                 rack->r_use_cmp_ack = 1;
12536         if (rack_disable_prr)
12537                 rack->rack_no_prr = 1;
12538         if (rack_gp_no_rec_chg)
12539                 rack->rc_gp_no_rec_chg = 1;
12540         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12541                 rack->rc_always_pace = 1;
12542                 if (rack->use_fixed_rate || rack->gp_ready)
12543                         rack_set_cc_pacing(rack);
12544         } else
12545                 rack->rc_always_pace = 0;
12546         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12547                 rack->r_mbuf_queue = 1;
12548         else
12549                 rack->r_mbuf_queue = 0;
12550         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12551                 inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12552         else
12553                 inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12554         rack_set_pace_segments(tp, rack, __LINE__, NULL);
12555         if (rack_limits_scwnd)
12556                 rack->r_limit_scw = 1;
12557         else
12558                 rack->r_limit_scw = 0;
12559         rack->rc_labc = V_tcp_abc_l_var;
12560         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12561         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12562         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12563         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12564         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12565         rack->r_ctl.rc_min_to = rack_min_to;
12566         microuptime(&rack->r_ctl.act_rcv_time);
12567         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12568         rack->rc_init_win = rack_default_init_window;
12569         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12570         if (rack_hw_up_only)
12571                 rack->r_up_only = 1;
12572         if (rack_do_dyn_mul) {
12573                 /* When dynamic adjustment is on CA needs to start at 100% */
12574                 rack->rc_gp_dyn_mul = 1;
12575                 if (rack_do_dyn_mul >= 100)
12576                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12577         } else
12578                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12579         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12580         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12581         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12582         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12583                                 rack_probertt_filter_life);
12584         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12585         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12586         rack->r_ctl.rc_time_of_last_probertt = us_cts;
12587         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12588         rack->r_ctl.rc_time_probertt_starts = 0;
12589         if (rack_dsack_std_based & 0x1) {
12590                 /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12591                 rack->rc_rack_tmr_std_based = 1;
12592         }
12593         if (rack_dsack_std_based & 0x2) {
12594                 /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12595                 rack->rc_rack_use_dsack = 1;
12596         }
12597         /* We require at least one measurement, even if the sysctl is 0 */
12598         if (rack_req_measurements)
12599                 rack->r_ctl.req_measurements = rack_req_measurements;
12600         else
12601                 rack->r_ctl.req_measurements = 1;
12602         if (rack_enable_hw_pacing)
12603                 rack->rack_hdw_pace_ena = 1;
12604         if (rack_hw_rate_caps)
12605                 rack->r_rack_hw_rate_caps = 1;
12606         /* Do we force on detection? */
12607 #ifdef NETFLIX_EXP_DETECTION
12608         if (tcp_force_detection)
12609                 rack->do_detection = 1;
12610         else
12611 #endif
12612                 rack->do_detection = 0;
12613         if (rack_non_rxt_use_cr)
12614                 rack->rack_rec_nonrxt_use_cr = 1;
12615         err = rack_init_fsb(tp, rack);
12616         if (err) {
12617                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12618                 tp->t_fb_ptr = NULL;
12619                 return (err);
12620         }
12621         if (tp->snd_una != tp->snd_max) {
12622                 /* Create a send map for the current outstanding data */
12623                 struct rack_sendmap *rsm;
12624
12625                 rsm = rack_alloc(rack);
12626                 if (rsm == NULL) {
12627                         uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12628                         tp->t_fb_ptr = NULL;
12629                         return (ENOMEM);
12630                 }
12631                 rsm->r_no_rtt_allowed = 1;
12632                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12633                 rsm->r_rtr_cnt = 1;
12634                 rsm->r_rtr_bytes = 0;
12635                 if (tp->t_flags & TF_SENTFIN)
12636                         rsm->r_flags |= RACK_HAS_FIN;
12637                 if ((tp->snd_una == tp->iss) &&
12638                     !TCPS_HAVEESTABLISHED(tp->t_state))
12639                         rsm->r_flags |= RACK_HAS_SYN;
12640                 rsm->r_start = tp->snd_una;
12641                 rsm->r_end = tp->snd_max;
12642                 rsm->r_dupack = 0;
12643                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12644                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12645                         if (rsm->m)
12646                                 rsm->orig_m_len = rsm->m->m_len;
12647                         else
12648                                 rsm->orig_m_len = 0;
12649                 } else {
12650                         /*
12651                          * This can happen if we have a stand-alone FIN or
12652                          *  SYN.
12653                          */
12654                         rsm->m = NULL;
12655                         rsm->orig_m_len = 0;
12656                         rsm->soff = 0;
12657                 }
12658 #ifndef INVARIANTS
12659                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12660 #else
12661                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12662                 if (insret != NULL) {
12663                         panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12664                               insret, rack, rsm);
12665                 }
12666 #endif
12667                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12668                 rsm->r_in_tmap = 1;
12669         }
12670         /*
12671          * Timers in Rack are kept in microseconds so lets
12672          * convert any initial incoming variables
12673          * from ticks into usecs. Note that we
12674          * also change the values of t_srtt and t_rttvar, if
12675          * they are non-zero. They are kept with a 5
12676          * bit decimal so we have to carefully convert
12677          * these to get the full precision.
12678          */
12679         rack_convert_rtts(tp);
12680         tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12681         if (rack_do_hystart) {
12682                 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
12683                 if (rack_do_hystart > 1)
12684                         tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
12685                 if (rack_do_hystart > 2)
12686                         tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
12687         }
12688         if (rack_def_profile)
12689                 rack_set_profile(rack, rack_def_profile);
12690         /* Cancel the GP measurement in progress */
12691         tp->t_flags &= ~TF_GPUTINPROG;
12692         if (SEQ_GT(tp->snd_max, tp->iss))
12693                 snt = tp->snd_max - tp->iss;
12694         else
12695                 snt = 0;
12696         iwin = rc_init_window(rack);
12697         if (snt < iwin) {
12698                 /* We are not past the initial window
12699                  * so we need to make sure cwnd is
12700                  * correct.
12701                  */
12702                 if (tp->snd_cwnd < iwin)
12703                         tp->snd_cwnd = iwin;
12704                 /*
12705                  * If we are within the initial window
12706                  * we want ssthresh to be unlimited. Setting
12707                  * it to the rwnd (which the default stack does
12708                  * and older racks) is not really a good idea
12709                  * since we want to be in SS and grow both the
12710                  * cwnd and the rwnd (via dynamic rwnd growth). If
12711                  * we set it to the rwnd then as the peer grows its
12712                  * rwnd we will be stuck in CA and never hit SS.
12713                  *
12714                  * Its far better to raise it up high (this takes the
12715                  * risk that there as been a loss already, probably
12716                  * we should have an indicator in all stacks of loss
12717                  * but we don't), but considering the normal use this
12718                  * is a risk worth taking. The consequences of not
12719                  * hitting SS are far worse than going one more time
12720                  * into it early on (before we have sent even a IW).
12721                  * It is highly unlikely that we will have had a loss
12722                  * before getting the IW out.
12723                  */
12724                 tp->snd_ssthresh = 0xffffffff;
12725         }
12726         rack_stop_all_timers(tp);
12727         /* Lets setup the fsb block */
12728         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12729         rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12730                              __LINE__, RACK_RTTS_INIT);
12731         return (0);
12732 }
12733
12734 static int
12735 rack_handoff_ok(struct tcpcb *tp)
12736 {
12737         if ((tp->t_state == TCPS_CLOSED) ||
12738             (tp->t_state == TCPS_LISTEN)) {
12739                 /* Sure no problem though it may not stick */
12740                 return (0);
12741         }
12742         if ((tp->t_state == TCPS_SYN_SENT) ||
12743             (tp->t_state == TCPS_SYN_RECEIVED)) {
12744                 /*
12745                  * We really don't know if you support sack,
12746                  * you have to get to ESTAB or beyond to tell.
12747                  */
12748                 return (EAGAIN);
12749         }
12750         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12751                 /*
12752                  * Rack will only send a FIN after all data is acknowledged.
12753                  * So in this case we have more data outstanding. We can't
12754                  * switch stacks until either all data and only the FIN
12755                  * is left (in which case rack_init() now knows how
12756                  * to deal with that) <or> all is acknowledged and we
12757                  * are only left with incoming data, though why you
12758                  * would want to switch to rack after all data is acknowledged
12759                  * I have no idea (rrs)!
12760                  */
12761                 return (EAGAIN);
12762         }
12763         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12764                 return (0);
12765         }
12766         /*
12767          * If we reach here we don't do SACK on this connection so we can
12768          * never do rack.
12769          */
12770         return (EINVAL);
12771 }
12772
12773
12774 static void
12775 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12776 {
12777         struct inpcb *inp = tptoinpcb(tp);
12778
12779         if (tp->t_fb_ptr) {
12780                 struct tcp_rack *rack;
12781                 struct rack_sendmap *rsm, *nrsm;
12782 #ifdef INVARIANTS
12783                 struct rack_sendmap *rm;
12784 #endif
12785
12786                 rack = (struct tcp_rack *)tp->t_fb_ptr;
12787                 if (tp->t_in_pkt) {
12788                         /*
12789                          * It is unsafe to process the packets since a
12790                          * reset may be lurking in them (its rare but it
12791                          * can occur). If we were to find a RST, then we
12792                          * would end up dropping the connection and the
12793                          * INP lock, so when we return the caller (tcp_usrreq)
12794                          * will blow up when it trys to unlock the inp.
12795                          */
12796                         struct mbuf *save, *m;
12797
12798                         m = tp->t_in_pkt;
12799                         tp->t_in_pkt = NULL;
12800                         tp->t_tail_pkt = NULL;
12801                         while (m) {
12802                                 save = m->m_nextpkt;
12803                                 m->m_nextpkt = NULL;
12804                                 m_freem(m);
12805                                 m = save;
12806                         }
12807                 }
12808                 tp->t_flags &= ~TF_FORCEDATA;
12809 #ifdef NETFLIX_SHARED_CWND
12810                 if (rack->r_ctl.rc_scw) {
12811                         uint32_t limit;
12812
12813                         if (rack->r_limit_scw)
12814                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12815                         else
12816                                 limit = 0;
12817                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12818                                                   rack->r_ctl.rc_scw_index,
12819                                                   limit);
12820                         rack->r_ctl.rc_scw = NULL;
12821                 }
12822 #endif
12823                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
12824                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12825                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12826                         rack->r_ctl.fsb.th = NULL;
12827                 }
12828                 /* Convert back to ticks, with  */
12829                 if (tp->t_srtt > 1) {
12830                         uint32_t val, frac;
12831
12832                         val = USEC_2_TICKS(tp->t_srtt);
12833                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12834                         tp->t_srtt = val << TCP_RTT_SHIFT;
12835                         /*
12836                          * frac is the fractional part here is left
12837                          * over from converting to hz and shifting.
12838                          * We need to convert this to the 5 bit
12839                          * remainder.
12840                          */
12841                         if (frac) {
12842                                 if (hz == 1000) {
12843                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12844                                 } else {
12845                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12846                                 }
12847                                 tp->t_srtt += frac;
12848                         }
12849                 }
12850                 if (tp->t_rttvar) {
12851                         uint32_t val, frac;
12852
12853                         val = USEC_2_TICKS(tp->t_rttvar);
12854                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12855                         tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12856                         /*
12857                          * frac is the fractional part here is left
12858                          * over from converting to hz and shifting.
12859                          * We need to convert this to the 5 bit
12860                          * remainder.
12861                          */
12862                         if (frac) {
12863                                 if (hz == 1000) {
12864                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12865                                 } else {
12866                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12867                                 }
12868                                 tp->t_rttvar += frac;
12869                         }
12870                 }
12871                 tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12872                 tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12873                 if (rack->rc_always_pace) {
12874                         tcp_decrement_paced_conn();
12875                         rack_undo_cc_pacing(rack);
12876                         rack->rc_always_pace = 0;
12877                 }
12878                 /* Clean up any options if they were not applied */
12879                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12880                         struct deferred_opt_list *dol;
12881
12882                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12883                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12884                         free(dol, M_TCPDO);
12885                 }
12886                 /* rack does not use force data but other stacks may clear it */
12887                 if (rack->r_ctl.crte != NULL) {
12888                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12889                         rack->rack_hdrw_pacing = 0;
12890                         rack->r_ctl.crte = NULL;
12891                 }
12892 #ifdef TCP_BLACKBOX
12893                 tcp_log_flowend(tp);
12894 #endif
12895                 RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12896 #ifndef INVARIANTS
12897                         (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12898 #else
12899                         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12900                         if (rm != rsm) {
12901                                 panic("At fini, rack:%p rsm:%p rm:%p",
12902                                       rack, rsm, rm);
12903                         }
12904 #endif
12905                         uma_zfree(rack_zone, rsm);
12906                 }
12907                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12908                 while (rsm) {
12909                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12910                         uma_zfree(rack_zone, rsm);
12911                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12912                 }
12913                 rack->rc_free_cnt = 0;
12914                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12915                 tp->t_fb_ptr = NULL;
12916         }
12917         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12918         inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12919         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12920         inp->inp_flags2 &= ~INP_MBUF_ACKCMP;
12921         /* Cancel the GP measurement in progress */
12922         tp->t_flags &= ~TF_GPUTINPROG;
12923         inp->inp_flags2 &= ~INP_MBUF_L_ACKS;
12924         /* Make sure snd_nxt is correctly set */
12925         tp->snd_nxt = tp->snd_max;
12926 }
12927
12928 static void
12929 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12930 {
12931         if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12932                 rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
12933         }
12934         switch (tp->t_state) {
12935         case TCPS_SYN_SENT:
12936                 rack->r_state = TCPS_SYN_SENT;
12937                 rack->r_substate = rack_do_syn_sent;
12938                 break;
12939         case TCPS_SYN_RECEIVED:
12940                 rack->r_state = TCPS_SYN_RECEIVED;
12941                 rack->r_substate = rack_do_syn_recv;
12942                 break;
12943         case TCPS_ESTABLISHED:
12944                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12945                 rack->r_state = TCPS_ESTABLISHED;
12946                 rack->r_substate = rack_do_established;
12947                 break;
12948         case TCPS_CLOSE_WAIT:
12949                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12950                 rack->r_state = TCPS_CLOSE_WAIT;
12951                 rack->r_substate = rack_do_close_wait;
12952                 break;
12953         case TCPS_FIN_WAIT_1:
12954                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12955                 rack->r_state = TCPS_FIN_WAIT_1;
12956                 rack->r_substate = rack_do_fin_wait_1;
12957                 break;
12958         case TCPS_CLOSING:
12959                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12960                 rack->r_state = TCPS_CLOSING;
12961                 rack->r_substate = rack_do_closing;
12962                 break;
12963         case TCPS_LAST_ACK:
12964                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12965                 rack->r_state = TCPS_LAST_ACK;
12966                 rack->r_substate = rack_do_lastack;
12967                 break;
12968         case TCPS_FIN_WAIT_2:
12969                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12970                 rack->r_state = TCPS_FIN_WAIT_2;
12971                 rack->r_substate = rack_do_fin_wait_2;
12972                 break;
12973         case TCPS_LISTEN:
12974         case TCPS_CLOSED:
12975         case TCPS_TIME_WAIT:
12976         default:
12977                 break;
12978         };
12979         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12980                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12981
12982 }
12983
12984 static void
12985 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12986 {
12987         /*
12988          * We received an ack, and then did not
12989          * call send or were bounced out due to the
12990          * hpts was running. Now a timer is up as well, is
12991          * it the right timer?
12992          */
12993         struct rack_sendmap *rsm;
12994         int tmr_up;
12995
12996         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12997         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12998                 return;
12999         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
13000         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
13001             (tmr_up == PACE_TMR_RXT)) {
13002                 /* Should be an RXT */
13003                 return;
13004         }
13005         if (rsm == NULL) {
13006                 /* Nothing outstanding? */
13007                 if (tp->t_flags & TF_DELACK) {
13008                         if (tmr_up == PACE_TMR_DELACK)
13009                                 /* We are supposed to have delayed ack up and we do */
13010                                 return;
13011                 } else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
13012                         /*
13013                          * if we hit enobufs then we would expect the possibility
13014                          * of nothing outstanding and the RXT up (and the hptsi timer).
13015                          */
13016                         return;
13017                 } else if (((V_tcp_always_keepalive ||
13018                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13019                             (tp->t_state <= TCPS_CLOSING)) &&
13020                            (tmr_up == PACE_TMR_KEEP) &&
13021                            (tp->snd_max == tp->snd_una)) {
13022                         /* We should have keep alive up and we do */
13023                         return;
13024                 }
13025         }
13026         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
13027                    ((tmr_up == PACE_TMR_TLP) ||
13028                     (tmr_up == PACE_TMR_RACK) ||
13029                     (tmr_up == PACE_TMR_RXT))) {
13030                 /*
13031                  * Either a Rack, TLP or RXT is fine if  we
13032                  * have outstanding data.
13033                  */
13034                 return;
13035         } else if (tmr_up == PACE_TMR_DELACK) {
13036                 /*
13037                  * If the delayed ack was going to go off
13038                  * before the rtx/tlp/rack timer were going to
13039                  * expire, then that would be the timer in control.
13040                  * Note we don't check the time here trusting the
13041                  * code is correct.
13042                  */
13043                 return;
13044         }
13045         /*
13046          * Ok the timer originally started is not what we want now.
13047          * We will force the hpts to be stopped if any, and restart
13048          * with the slot set to what was in the saved slot.
13049          */
13050         if (tcp_in_hpts(rack->rc_inp)) {
13051                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13052                         uint32_t us_cts;
13053
13054                         us_cts = tcp_get_usecs(NULL);
13055                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13056                                 rack->r_early = 1;
13057                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13058                         }
13059                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13060                 }
13061                 tcp_hpts_remove(rack->rc_inp);
13062         }
13063         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13064         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13065 }
13066
13067
13068 static void
13069 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts, uint32_t high_seq)
13070 {
13071         if ((SEQ_LT(tp->snd_wl1, seq) ||
13072             (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
13073             (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
13074                 /* keep track of pure window updates */
13075                 if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
13076                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
13077                 tp->snd_wnd = tiwin;
13078                 rack_validate_fo_sendwin_up(tp, rack);
13079                 tp->snd_wl1 = seq;
13080                 tp->snd_wl2 = ack;
13081                 if (tp->snd_wnd > tp->max_sndwnd)
13082                         tp->max_sndwnd = tp->snd_wnd;
13083             rack->r_wanted_output = 1;
13084         } else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
13085                 tp->snd_wnd = tiwin;
13086                 rack_validate_fo_sendwin_up(tp, rack);
13087                 tp->snd_wl1 = seq;
13088                 tp->snd_wl2 = ack;
13089         } else {
13090                 /* Not a valid win update */
13091                 return;
13092         }
13093         /* Do we exit persists? */
13094         if ((rack->rc_in_persist != 0) &&
13095             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13096                                 rack->r_ctl.rc_pace_min_segs))) {
13097                 rack_exit_persist(tp, rack, cts);
13098         }
13099         /* Do we enter persists? */
13100         if ((rack->rc_in_persist == 0) &&
13101             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13102             TCPS_HAVEESTABLISHED(tp->t_state) &&
13103             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13104             sbavail(&tptosocket(tp)->so_snd) &&
13105             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13106                 /*
13107                  * Here the rwnd is less than
13108                  * the pacing size, we are established,
13109                  * nothing is outstanding, and there is
13110                  * data to send. Enter persists.
13111                  */
13112                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13113         }
13114 }
13115
13116 static void
13117 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13118 {
13119
13120         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13121                 struct inpcb *inp = tptoinpcb(tp);
13122                 union tcp_log_stackspecific log;
13123                 struct timeval ltv;
13124                 char tcp_hdr_buf[60];
13125                 struct tcphdr *th;
13126                 struct timespec ts;
13127                 uint32_t orig_snd_una;
13128                 uint8_t xx = 0;
13129
13130 #ifdef NETFLIX_HTTP_LOGGING
13131                 struct http_sendfile_track *http_req;
13132
13133                 if (SEQ_GT(ae->ack, tp->snd_una)) {
13134                         http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13135                 } else {
13136                         http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13137                 }
13138 #endif
13139                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13140                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
13141                 if (rack->rack_no_prr == 0)
13142                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13143                 else
13144                         log.u_bbr.flex1 = 0;
13145                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13146                 log.u_bbr.use_lt_bw <<= 1;
13147                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13148                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13149                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13150                 log.u_bbr.pkts_out = tp->t_maxseg;
13151                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13152                 log.u_bbr.flex7 = 1;
13153                 log.u_bbr.lost = ae->flags;
13154                 log.u_bbr.cwnd_gain = ackval;
13155                 log.u_bbr.pacing_gain = 0x2;
13156                 if (ae->flags & TSTMP_HDWR) {
13157                         /* Record the hardware timestamp if present */
13158                         log.u_bbr.flex3 = M_TSTMP;
13159                         ts.tv_sec = ae->timestamp / 1000000000;
13160                         ts.tv_nsec = ae->timestamp % 1000000000;
13161                         ltv.tv_sec = ts.tv_sec;
13162                         ltv.tv_usec = ts.tv_nsec / 1000;
13163                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13164                 } else if (ae->flags & TSTMP_LRO) {
13165                         /* Record the LRO the arrival timestamp */
13166                         log.u_bbr.flex3 = M_TSTMP_LRO;
13167                         ts.tv_sec = ae->timestamp / 1000000000;
13168                         ts.tv_nsec = ae->timestamp % 1000000000;
13169                         ltv.tv_sec = ts.tv_sec;
13170                         ltv.tv_usec = ts.tv_nsec / 1000;
13171                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13172                 }
13173                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13174                 /* Log the rcv time */
13175                 log.u_bbr.delRate = ae->timestamp;
13176 #ifdef NETFLIX_HTTP_LOGGING
13177                 log.u_bbr.applimited = tp->t_http_closed;
13178                 log.u_bbr.applimited <<= 8;
13179                 log.u_bbr.applimited |= tp->t_http_open;
13180                 log.u_bbr.applimited <<= 8;
13181                 log.u_bbr.applimited |= tp->t_http_req;
13182                 if (http_req) {
13183                         /* Copy out any client req info */
13184                         /* seconds */
13185                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13186                         /* useconds */
13187                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13188                         log.u_bbr.rttProp = http_req->timestamp;
13189                         log.u_bbr.cur_del_rate = http_req->start;
13190                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13191                                 log.u_bbr.flex8 |= 1;
13192                         } else {
13193                                 log.u_bbr.flex8 |= 2;
13194                                 log.u_bbr.bw_inuse = http_req->end;
13195                         }
13196                         log.u_bbr.flex6 = http_req->start_seq;
13197                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13198                                 log.u_bbr.flex8 |= 4;
13199                                 log.u_bbr.epoch = http_req->end_seq;
13200                         }
13201                 }
13202 #endif
13203                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13204                 th = (struct tcphdr *)tcp_hdr_buf;
13205                 th->th_seq = ae->seq;
13206                 th->th_ack = ae->ack;
13207                 th->th_win = ae->win;
13208                 /* Now fill in the ports */
13209                 th->th_sport = inp->inp_fport;
13210                 th->th_dport = inp->inp_lport;
13211                 tcp_set_flags(th, ae->flags);
13212                 /* Now do we have a timestamp option? */
13213                 if (ae->flags & HAS_TSTMP) {
13214                         u_char *cp;
13215                         uint32_t val;
13216
13217                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13218                         cp = (u_char *)(th + 1);
13219                         *cp = TCPOPT_NOP;
13220                         cp++;
13221                         *cp = TCPOPT_NOP;
13222                         cp++;
13223                         *cp = TCPOPT_TIMESTAMP;
13224                         cp++;
13225                         *cp = TCPOLEN_TIMESTAMP;
13226                         cp++;
13227                         val = htonl(ae->ts_value);
13228                         bcopy((char *)&val,
13229                               (char *)cp, sizeof(uint32_t));
13230                         val = htonl(ae->ts_echo);
13231                         bcopy((char *)&val,
13232                               (char *)(cp + 4), sizeof(uint32_t));
13233                 } else
13234                         th->th_off = (sizeof(struct tcphdr) >> 2);
13235
13236                 /*
13237                  * For sane logging we need to play a little trick.
13238                  * If the ack were fully processed we would have moved
13239                  * snd_una to high_seq, but since compressed acks are
13240                  * processed in two phases, at this point (logging) snd_una
13241                  * won't be advanced. So we would see multiple acks showing
13242                  * the advancement. We can prevent that by "pretending" that
13243                  * snd_una was advanced and then un-advancing it so that the
13244                  * logging code has the right value for tlb_snd_una.
13245                  */
13246                 if (tp->snd_una != high_seq) {
13247                         orig_snd_una = tp->snd_una;
13248                         tp->snd_una = high_seq;
13249                         xx = 1;
13250                 } else
13251                         xx = 0;
13252                 TCP_LOG_EVENTP(tp, th,
13253                                &tptosocket(tp)->so_rcv,
13254                                &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
13255                                0, &log, true, &ltv);
13256                 if (xx) {
13257                         tp->snd_una = orig_snd_una;
13258                 }
13259         }
13260
13261 }
13262
13263 static void
13264 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
13265 {
13266         uint32_t us_rtt;
13267         /*
13268          * A persist or keep-alive was forced out, update our
13269          * min rtt time. Note now worry about lost responses.
13270          * When a subsequent keep-alive or persist times out
13271          * and forced_ack is still on, then the last probe
13272          * was not responded to. In such cases we have a
13273          * sysctl that controls the behavior. Either we apply
13274          * the rtt but with reduced confidence (0). Or we just
13275          * plain don't apply the rtt estimate. Having data flow
13276          * will clear the probe_not_answered flag i.e. cum-ack
13277          * move forward <or> exiting and reentering persists.
13278          */
13279
13280         rack->forced_ack = 0;
13281         rack->rc_tp->t_rxtshift = 0;
13282         if ((rack->rc_in_persist &&
13283              (tiwin == rack->rc_tp->snd_wnd)) ||
13284             (rack->rc_in_persist == 0)) {
13285                 /*
13286                  * In persists only apply the RTT update if this is
13287                  * a response to our window probe. And that
13288                  * means the rwnd sent must match the current
13289                  * snd_wnd. If it does not, then we got a
13290                  * window update ack instead. For keepalive
13291                  * we allow the answer no matter what the window.
13292                  *
13293                  * Note that if the probe_not_answered is set then
13294                  * the forced_ack_ts is the oldest one i.e. the first
13295                  * probe sent that might have been lost. This assures
13296                  * us that if we do calculate an RTT it is longer not
13297                  * some short thing.
13298                  */
13299                 if (rack->rc_in_persist)
13300                         counter_u64_add(rack_persists_acks, 1);
13301                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13302                 if (us_rtt == 0)
13303                         us_rtt = 1;
13304                 if (rack->probe_not_answered == 0) {
13305                         rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13306                         tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13307                 } else {
13308                         /* We have a retransmitted probe here too */
13309                         if (rack_apply_rtt_with_reduced_conf) {
13310                                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13311                                 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
13312                         }
13313                 }
13314         }
13315 }
13316
13317 static int
13318 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13319 {
13320         /*
13321          * Handle a "special" compressed ack mbuf. Each incoming
13322          * ack has only four possible dispositions:
13323          *
13324          * A) It moves the cum-ack forward
13325          * B) It is behind the cum-ack.
13326          * C) It is a window-update ack.
13327          * D) It is a dup-ack.
13328          *
13329          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13330          * in the incoming mbuf. We also need to still pay attention
13331          * to nxt_pkt since there may be another packet after this
13332          * one.
13333          */
13334 #ifdef TCP_ACCOUNTING
13335         uint64_t ts_val;
13336         uint64_t rdstc;
13337 #endif
13338         int segsiz;
13339         struct timespec ts;
13340         struct tcp_rack *rack;
13341         struct tcp_ackent *ae;
13342         uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13343         int cnt, i, did_out, ourfinisacked = 0;
13344         struct tcpopt to_holder, *to = NULL;
13345 #ifdef TCP_ACCOUNTING
13346         int win_up_req = 0;
13347 #endif
13348         int nsegs = 0;
13349         int under_pacing = 1;
13350         int recovery = 0;
13351 #ifdef TCP_ACCOUNTING
13352         sched_pin();
13353 #endif
13354         rack = (struct tcp_rack *)tp->t_fb_ptr;
13355         if (rack->gp_ready &&
13356             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13357                 under_pacing = 0;
13358         else
13359                 under_pacing = 1;
13360
13361         if (rack->r_state != tp->t_state)
13362                 rack_set_state(tp, rack);
13363         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13364             (tp->t_flags & TF_GPUTINPROG)) {
13365                 /*
13366                  * We have a goodput in progress
13367                  * and we have entered a late state.
13368                  * Do we have enough data in the sb
13369                  * to handle the GPUT request?
13370                  */
13371                 uint32_t bytes;
13372
13373                 bytes = tp->gput_ack - tp->gput_seq;
13374                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
13375                         bytes += tp->gput_seq - tp->snd_una;
13376                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
13377                         /*
13378                          * There are not enough bytes in the socket
13379                          * buffer that have been sent to cover this
13380                          * measurement. Cancel it.
13381                          */
13382                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13383                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
13384                                                    tp->gput_seq,
13385                                                    0, 0, 18, __LINE__, NULL, 0);
13386                         tp->t_flags &= ~TF_GPUTINPROG;
13387                 }
13388         }
13389         to = &to_holder;
13390         to->to_flags = 0;
13391         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13392                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13393         cnt = m->m_len / sizeof(struct tcp_ackent);
13394         counter_u64_add(rack_multi_single_eq, cnt);
13395         high_seq = tp->snd_una;
13396         the_win = tp->snd_wnd;
13397         win_seq = tp->snd_wl1;
13398         win_upd_ack = tp->snd_wl2;
13399         cts = tcp_tv_to_usectick(tv);
13400         ms_cts = tcp_tv_to_mssectick(tv);
13401         rack->r_ctl.rc_rcvtime = cts;
13402         segsiz = ctf_fixed_maxseg(tp);
13403         if ((rack->rc_gp_dyn_mul) &&
13404             (rack->use_fixed_rate == 0) &&
13405             (rack->rc_always_pace)) {
13406                 /* Check in on probertt */
13407                 rack_check_probe_rtt(rack, cts);
13408         }
13409         for (i = 0; i < cnt; i++) {
13410 #ifdef TCP_ACCOUNTING
13411                 ts_val = get_cyclecount();
13412 #endif
13413                 rack_clear_rate_sample(rack);
13414                 ae = ((mtod(m, struct tcp_ackent *)) + i);
13415                 /* Setup the window */
13416                 tiwin = ae->win << tp->snd_scale;
13417                 if (tiwin > rack->r_ctl.rc_high_rwnd)
13418                         rack->r_ctl.rc_high_rwnd = tiwin;
13419                 /* figure out the type of ack */
13420                 if (SEQ_LT(ae->ack, high_seq)) {
13421                         /* Case B*/
13422                         ae->ack_val_set = ACK_BEHIND;
13423                 } else if (SEQ_GT(ae->ack, high_seq)) {
13424                         /* Case A */
13425                         ae->ack_val_set = ACK_CUMACK;
13426                 } else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
13427                         /* Case D */
13428                         ae->ack_val_set = ACK_DUPACK;
13429                 } else {
13430                         /* Case C */
13431                         ae->ack_val_set = ACK_RWND;
13432                 }
13433                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13434                 /* Validate timestamp */
13435                 if (ae->flags & HAS_TSTMP) {
13436                         /* Setup for a timestamp */
13437                         to->to_flags = TOF_TS;
13438                         ae->ts_echo -= tp->ts_offset;
13439                         to->to_tsecr = ae->ts_echo;
13440                         to->to_tsval = ae->ts_value;
13441                         /*
13442                          * If echoed timestamp is later than the current time, fall back to
13443                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13444                          * were used when this connection was established.
13445                          */
13446                         if (TSTMP_GT(ae->ts_echo, ms_cts))
13447                                 to->to_tsecr = 0;
13448                         if (tp->ts_recent &&
13449                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13450                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13451 #ifdef TCP_ACCOUNTING
13452                                         rdstc = get_cyclecount();
13453                                         if (rdstc > ts_val) {
13454                                                 counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13455                                                                 (rdstc - ts_val));
13456                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13457                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13458                                                 }
13459                                         }
13460 #endif
13461                                         continue;
13462                                 }
13463                         }
13464                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13465                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13466                                 tp->ts_recent_age = tcp_ts_getticks();
13467                                 tp->ts_recent = ae->ts_value;
13468                         }
13469                 } else {
13470                         /* Setup for a no options */
13471                         to->to_flags = 0;
13472                 }
13473                 /* Update the rcv time and perform idle reduction possibly */
13474                 if  (tp->t_idle_reduce &&
13475                      (tp->snd_max == tp->snd_una) &&
13476                      (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13477                         counter_u64_add(rack_input_idle_reduces, 1);
13478                         rack_cc_after_idle(rack, tp);
13479                 }
13480                 tp->t_rcvtime = ticks;
13481                 /* Now what about ECN of a chain of pure ACKs? */
13482                 if (tcp_ecn_input_segment(tp, ae->flags, 0,
13483                         tcp_packets_this_ack(tp, ae->ack),
13484                         ae->codepoint))
13485                         rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
13486 #ifdef TCP_ACCOUNTING
13487                 /* Count for the specific type of ack in */
13488                 counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13489                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13490                         tp->tcp_cnt_counters[ae->ack_val_set]++;
13491                 }
13492 #endif
13493                 /*
13494                  * Note how we could move up these in the determination
13495                  * above, but we don't so that way the timestamp checks (and ECN)
13496                  * is done first before we do any processing on the ACK.
13497                  * The non-compressed path through the code has this
13498                  * weakness (noted by @jtl) that it actually does some
13499                  * processing before verifying the timestamp information.
13500                  * We don't take that path here which is why we set
13501                  * the ack_val_set first, do the timestamp and ecn
13502                  * processing, and then look at what we have setup.
13503                  */
13504                 if (ae->ack_val_set == ACK_BEHIND) {
13505                         /*
13506                          * Case B flag reordering, if window is not closed
13507                          * or it could be a keep-alive or persists
13508                          */
13509                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13510                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13511                         }
13512                 } else if (ae->ack_val_set == ACK_DUPACK) {
13513                         /* Case D */
13514                         rack_strike_dupack(rack);
13515                 } else if (ae->ack_val_set == ACK_RWND) {
13516                         /* Case C */
13517                         if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13518                                 ts.tv_sec = ae->timestamp / 1000000000;
13519                                 ts.tv_nsec = ae->timestamp % 1000000000;
13520                                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13521                                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13522                         } else {
13523                                 rack->r_ctl.act_rcv_time = *tv;
13524                         }
13525                         if (rack->forced_ack) {
13526                                 rack_handle_probe_response(rack, tiwin,
13527                                                            tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
13528                         }
13529 #ifdef TCP_ACCOUNTING
13530                         win_up_req = 1;
13531 #endif
13532                         win_upd_ack = ae->ack;
13533                         win_seq = ae->seq;
13534                         the_win = tiwin;
13535                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13536                 } else {
13537                         /* Case A */
13538                         if (SEQ_GT(ae->ack, tp->snd_max)) {
13539                                 /*
13540                                  * We just send an ack since the incoming
13541                                  * ack is beyond the largest seq we sent.
13542                                  */
13543                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
13544                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13545                                         if (tp->t_flags && TF_ACKNOW)
13546                                                 rack->r_wanted_output = 1;
13547                                 }
13548                         } else {
13549                                 nsegs++;
13550                                 /* If the window changed setup to update */
13551                                 if (tiwin != tp->snd_wnd) {
13552                                         win_upd_ack = ae->ack;
13553                                         win_seq = ae->seq;
13554                                         the_win = tiwin;
13555                                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13556                                 }
13557 #ifdef TCP_ACCOUNTING
13558                                 /* Account for the acks */
13559                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13560                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13561                                 }
13562                                 counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13563                                                 (((ae->ack - high_seq) + segsiz - 1) / segsiz));
13564 #endif
13565                                 high_seq = ae->ack;
13566                                 if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
13567                                         union tcp_log_stackspecific log;
13568                                         struct timeval tv;
13569
13570                                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13571                                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13572                                         log.u_bbr.flex1 = high_seq;
13573                                         log.u_bbr.flex2 = rack->r_ctl.roundends;
13574                                         log.u_bbr.flex3 = rack->r_ctl.current_round;
13575                                         log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
13576                                         log.u_bbr.flex8 = 8;
13577                                         tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
13578                                                        0, &log, false, NULL, NULL, 0, &tv);
13579                                 }
13580                                 /*
13581                                  * The draft (v3) calls for us to use SEQ_GEQ, but that
13582                                  * causes issues when we are just going app limited. Lets
13583                                  * instead use SEQ_GT <or> where its equal but more data
13584                                  * is outstanding.
13585                                  */
13586                                 if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) ||
13587                                     ((high_seq == rack->r_ctl.roundends) &&
13588                                      SEQ_GT(tp->snd_max, tp->snd_una))) {
13589                                         rack->r_ctl.current_round++;
13590                                         rack->r_ctl.roundends = tp->snd_max;
13591                                         if (CC_ALGO(tp)->newround != NULL) {
13592                                                 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
13593                                         }
13594                                 }
13595                                 /* Setup our act_rcv_time */
13596                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13597                                         ts.tv_sec = ae->timestamp / 1000000000;
13598                                         ts.tv_nsec = ae->timestamp % 1000000000;
13599                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13600                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13601                                 } else {
13602                                         rack->r_ctl.act_rcv_time = *tv;
13603                                 }
13604                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13605                                 if (rack->rc_dsack_round_seen) {
13606                                         /* Is the dsack round over? */
13607                                         if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13608                                                 /* Yes it is */
13609                                                 rack->rc_dsack_round_seen = 0;
13610                                                 rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13611                                         }
13612                                 }
13613                         }
13614                 }
13615                 /* And lets be sure to commit the rtt measurements for this ack */
13616                 tcp_rack_xmit_timer_commit(rack, tp);
13617 #ifdef TCP_ACCOUNTING
13618                 rdstc = get_cyclecount();
13619                 if (rdstc > ts_val) {
13620                         counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13621                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13622                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13623                                 if (ae->ack_val_set == ACK_CUMACK)
13624                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13625                         }
13626                 }
13627 #endif
13628         }
13629 #ifdef TCP_ACCOUNTING
13630         ts_val = get_cyclecount();
13631 #endif
13632         /* Tend to any collapsed window */
13633         if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
13634                 /* The peer collapsed the window */
13635                 rack_collapsed_window(rack, (tp->snd_max - high_seq), __LINE__);
13636         } else if (rack->rc_has_collapsed)
13637                 rack_un_collapse_window(rack, __LINE__);
13638         if ((rack->r_collapse_point_valid) &&
13639             (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
13640                 rack->r_collapse_point_valid = 0;
13641         acked_amount = acked = (high_seq - tp->snd_una);
13642         if (acked) {
13643                 /*
13644                  * Clear the probe not answered flag
13645                  * since cum-ack moved forward.
13646                  */
13647                 rack->probe_not_answered = 0;
13648                 if (rack->sack_attack_disable == 0)
13649                         rack_do_decay(rack);
13650                 if (acked >= segsiz) {
13651                         /*
13652                          * You only get credit for
13653                          * MSS and greater (and you get extra
13654                          * credit for larger cum-ack moves).
13655                          */
13656                         int ac;
13657
13658                         ac = acked / segsiz;
13659                         rack->r_ctl.ack_count += ac;
13660                         counter_u64_add(rack_ack_total, ac);
13661                 }
13662                 if (rack->r_ctl.ack_count > 0xfff00000) {
13663                         /*
13664                          * reduce the number to keep us under
13665                          * a uint32_t.
13666                          */
13667                         rack->r_ctl.ack_count /= 2;
13668                         rack->r_ctl.sack_count /= 2;
13669                 }
13670                 if (tp->t_flags & TF_NEEDSYN) {
13671                         /*
13672                          * T/TCP: Connection was half-synchronized, and our SYN has
13673                          * been ACK'd (so connection is now fully synchronized).  Go
13674                          * to non-starred state, increment snd_una for ACK of SYN,
13675                          * and check if we can do window scaling.
13676                          */
13677                         tp->t_flags &= ~TF_NEEDSYN;
13678                         tp->snd_una++;
13679                         acked_amount = acked = (high_seq - tp->snd_una);
13680                 }
13681                 if (acked > sbavail(&so->so_snd))
13682                         acked_amount = sbavail(&so->so_snd);
13683 #ifdef NETFLIX_EXP_DETECTION
13684                 /*
13685                  * We only care on a cum-ack move if we are in a sack-disabled
13686                  * state. We have already added in to the ack_count, and we never
13687                  * would disable on a cum-ack move, so we only care to do the
13688                  * detection if it may "undo" it, i.e. we were in disabled already.
13689                  */
13690                 if (rack->sack_attack_disable)
13691                         rack_do_detection(tp, rack, acked_amount, segsiz);
13692 #endif
13693                 if (IN_FASTRECOVERY(tp->t_flags) &&
13694                     (rack->rack_no_prr == 0))
13695                         rack_update_prr(tp, rack, acked_amount, high_seq);
13696                 if (IN_RECOVERY(tp->t_flags)) {
13697                         if (SEQ_LT(high_seq, tp->snd_recover) &&
13698                             (SEQ_LT(high_seq, tp->snd_max))) {
13699                                 tcp_rack_partialack(tp);
13700                         } else {
13701                                 rack_post_recovery(tp, high_seq);
13702                                 recovery = 1;
13703                         }
13704                 }
13705                 /* Handle the rack-log-ack part (sendmap) */
13706                 if ((sbused(&so->so_snd) == 0) &&
13707                     (acked > acked_amount) &&
13708                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
13709                     (tp->t_flags & TF_SENTFIN)) {
13710                         /*
13711                          * We must be sure our fin
13712                          * was sent and acked (we can be
13713                          * in FIN_WAIT_1 without having
13714                          * sent the fin).
13715                          */
13716                         ourfinisacked = 1;
13717                         /*
13718                          * Lets make sure snd_una is updated
13719                          * since most likely acked_amount = 0 (it
13720                          * should be).
13721                          */
13722                         tp->snd_una = high_seq;
13723                 }
13724                 /* Did we make a RTO error? */
13725                 if ((tp->t_flags & TF_PREVVALID) &&
13726                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13727                         tp->t_flags &= ~TF_PREVVALID;
13728                         if (tp->t_rxtshift == 1 &&
13729                             (int)(ticks - tp->t_badrxtwin) < 0)
13730                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
13731                 }
13732                 /* Handle the data in the socket buffer */
13733                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13734                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13735                 if (acked_amount > 0) {
13736                         struct mbuf *mfree;
13737
13738                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13739                         SOCKBUF_LOCK(&so->so_snd);
13740                         mfree = sbcut_locked(&so->so_snd, acked_amount);
13741                         tp->snd_una = high_seq;
13742                         /* Note we want to hold the sb lock through the sendmap adjust */
13743                         rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13744                         /* Wake up the socket if we have room to write more */
13745                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13746                         sowwakeup_locked(so);
13747                         m_freem(mfree);
13748                 }
13749                 /* update progress */
13750                 tp->t_acktime = ticks;
13751                 rack_log_progress_event(rack, tp, tp->t_acktime,
13752                                         PROGRESS_UPDATE, __LINE__);
13753                 /* Clear out shifts and such */
13754                 tp->t_rxtshift = 0;
13755                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13756                                    rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13757                 rack->rc_tlp_in_progress = 0;
13758                 rack->r_ctl.rc_tlp_cnt_out = 0;
13759                 /* Send recover and snd_nxt must be dragged along */
13760                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
13761                         tp->snd_recover = tp->snd_una;
13762                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13763                         tp->snd_nxt = tp->snd_una;
13764                 /*
13765                  * If the RXT timer is running we want to
13766                  * stop it, so we can restart a TLP (or new RXT).
13767                  */
13768                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13769                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13770 #ifdef NETFLIX_HTTP_LOGGING
13771                 tcp_http_check_for_comp(rack->rc_tp, high_seq);
13772 #endif
13773                 tp->snd_wl2 = high_seq;
13774                 tp->t_dupacks = 0;
13775                 if (under_pacing &&
13776                     (rack->use_fixed_rate == 0) &&
13777                     (rack->in_probe_rtt == 0) &&
13778                     rack->rc_gp_dyn_mul &&
13779                     rack->rc_always_pace) {
13780                         /* Check if we are dragging bottom */
13781                         rack_check_bottom_drag(tp, rack, so, acked);
13782                 }
13783                 if (tp->snd_una == tp->snd_max) {
13784                         tp->t_flags &= ~TF_PREVVALID;
13785                         rack->r_ctl.retran_during_recovery = 0;
13786                         rack->r_ctl.dsack_byte_cnt = 0;
13787                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13788                         if (rack->r_ctl.rc_went_idle_time == 0)
13789                                 rack->r_ctl.rc_went_idle_time = 1;
13790                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13791                         if (sbavail(&tptosocket(tp)->so_snd) == 0)
13792                                 tp->t_acktime = 0;
13793                         /* Set so we might enter persists... */
13794                         rack->r_wanted_output = 1;
13795                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13796                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13797                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13798                             (sbavail(&so->so_snd) == 0) &&
13799                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13800                                 /*
13801                                  * The socket was gone and the
13802                                  * peer sent data (not now in the past), time to
13803                                  * reset him.
13804                                  */
13805                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13806                                 /* tcp_close will kill the inp pre-log the Reset */
13807                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13808 #ifdef TCP_ACCOUNTING
13809                                 rdstc = get_cyclecount();
13810                                 if (rdstc > ts_val) {
13811                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13812                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13813                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13814                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13815                                         }
13816                                 }
13817 #endif
13818                                 m_freem(m);
13819                                 tp = tcp_close(tp);
13820                                 if (tp == NULL) {
13821 #ifdef TCP_ACCOUNTING
13822                                         sched_unpin();
13823 #endif
13824                                         return (1);
13825                                 }
13826                                 /*
13827                                  * We would normally do drop-with-reset which would
13828                                  * send back a reset. We can't since we don't have
13829                                  * all the needed bits. Instead lets arrange for
13830                                  * a call to tcp_output(). That way since we
13831                                  * are in the closed state we will generate a reset.
13832                                  *
13833                                  * Note if tcp_accounting is on we don't unpin since
13834                                  * we do that after the goto label.
13835                                  */
13836                                 goto send_out_a_rst;
13837                         }
13838                         if ((sbused(&so->so_snd) == 0) &&
13839                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
13840                             (tp->t_flags & TF_SENTFIN)) {
13841                                 /*
13842                                  * If we can't receive any more data, then closing user can
13843                                  * proceed. Starting the timer is contrary to the
13844                                  * specification, but if we don't get a FIN we'll hang
13845                                  * forever.
13846                                  *
13847                                  */
13848                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13849                                         soisdisconnected(so);
13850                                         tcp_timer_activate(tp, TT_2MSL,
13851                                                            (tcp_fast_finwait2_recycle ?
13852                                                             tcp_finwait2_timeout :
13853                                                             TP_MAXIDLE(tp)));
13854                                 }
13855                                 if (ourfinisacked == 0) {
13856                                         /*
13857                                          * We don't change to fin-wait-2 if we have our fin acked
13858                                          * which means we are probably in TCPS_CLOSING.
13859                                          */
13860                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13861                                 }
13862                         }
13863                 }
13864                 /* Wake up the socket if we have room to write more */
13865                 if (sbavail(&so->so_snd)) {
13866                         rack->r_wanted_output = 1;
13867                         if (ctf_progress_timeout_check(tp, true)) {
13868                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13869                                                         tp, tick, PROGRESS_DROP, __LINE__);
13870                                 /*
13871                                  * We cheat here and don't send a RST, we should send one
13872                                  * when the pacer drops the connection.
13873                                  */
13874 #ifdef TCP_ACCOUNTING
13875                                 rdstc = get_cyclecount();
13876                                 if (rdstc > ts_val) {
13877                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13878                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13879                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13880                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13881                                         }
13882                                 }
13883                                 sched_unpin();
13884 #endif
13885                                 (void)tcp_drop(tp, ETIMEDOUT);
13886                                 m_freem(m);
13887                                 return (1);
13888                         }
13889                 }
13890                 if (ourfinisacked) {
13891                         switch(tp->t_state) {
13892                         case TCPS_CLOSING:
13893 #ifdef TCP_ACCOUNTING
13894                                 rdstc = get_cyclecount();
13895                                 if (rdstc > ts_val) {
13896                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13897                                                         (rdstc - ts_val));
13898                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13899                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13900                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13901                                         }
13902                                 }
13903                                 sched_unpin();
13904 #endif
13905                                 tcp_twstart(tp);
13906                                 m_freem(m);
13907                                 return (1);
13908                                 break;
13909                         case TCPS_LAST_ACK:
13910 #ifdef TCP_ACCOUNTING
13911                                 rdstc = get_cyclecount();
13912                                 if (rdstc > ts_val) {
13913                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13914                                                         (rdstc - ts_val));
13915                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13916                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13917                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13918                                         }
13919                                 }
13920                                 sched_unpin();
13921 #endif
13922                                 tp = tcp_close(tp);
13923                                 ctf_do_drop(m, tp);
13924                                 return (1);
13925                                 break;
13926                         case TCPS_FIN_WAIT_1:
13927 #ifdef TCP_ACCOUNTING
13928                                 rdstc = get_cyclecount();
13929                                 if (rdstc > ts_val) {
13930                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13931                                                         (rdstc - ts_val));
13932                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13933                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13934                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13935                                         }
13936                                 }
13937 #endif
13938                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13939                                         soisdisconnected(so);
13940                                         tcp_timer_activate(tp, TT_2MSL,
13941                                                            (tcp_fast_finwait2_recycle ?
13942                                                             tcp_finwait2_timeout :
13943                                                             TP_MAXIDLE(tp)));
13944                                 }
13945                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13946                                 break;
13947                         default:
13948                                 break;
13949                         }
13950                 }
13951                 if (rack->r_fast_output) {
13952                         /*
13953                          * We re doing fast output.. can we expand that?
13954                          */
13955                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13956                 }
13957 #ifdef TCP_ACCOUNTING
13958                 rdstc = get_cyclecount();
13959                 if (rdstc > ts_val) {
13960                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13961                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13962                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13963                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13964                         }
13965                 }
13966
13967         } else if (win_up_req) {
13968                 rdstc = get_cyclecount();
13969                 if (rdstc > ts_val) {
13970                         counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13971                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13972                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13973                         }
13974                 }
13975 #endif
13976         }
13977         /* Now is there a next packet, if so we are done */
13978         m_freem(m);
13979         did_out = 0;
13980         if (nxt_pkt) {
13981 #ifdef TCP_ACCOUNTING
13982                 sched_unpin();
13983 #endif
13984                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13985                 return (0);
13986         }
13987         rack_handle_might_revert(tp, rack);
13988         ctf_calc_rwin(so, tp);
13989         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13990         send_out_a_rst:
13991                 if (tcp_output(tp) < 0) {
13992 #ifdef TCP_ACCOUNTING
13993                         sched_unpin();
13994 #endif
13995                         return (1);
13996                 }
13997                 did_out = 1;
13998         }
13999         rack_free_trim(rack);
14000 #ifdef TCP_ACCOUNTING
14001         sched_unpin();
14002 #endif
14003         rack_timer_audit(tp, rack, &so->so_snd);
14004         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
14005         return (0);
14006 }
14007
14008
14009 static int
14010 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
14011     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
14012     int32_t nxt_pkt, struct timeval *tv)
14013 {
14014         struct inpcb *inp = tptoinpcb(tp);
14015 #ifdef TCP_ACCOUNTING
14016         uint64_t ts_val;
14017 #endif
14018         int32_t thflags, retval, did_out = 0;
14019         int32_t way_out = 0;
14020         /*
14021          * cts - is the current time from tv (caller gets ts) in microseconds.
14022          * ms_cts - is the current time from tv in milliseconds.
14023          * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
14024          */
14025         uint32_t cts, us_cts, ms_cts;
14026         uint32_t tiwin, high_seq;
14027         struct timespec ts;
14028         struct tcpopt to;
14029         struct tcp_rack *rack;
14030         struct rack_sendmap *rsm;
14031         int32_t prev_state = 0;
14032 #ifdef TCP_ACCOUNTING
14033         int ack_val_set = 0xf;
14034 #endif
14035         int nsegs;
14036
14037         NET_EPOCH_ASSERT();
14038         INP_WLOCK_ASSERT(inp);
14039
14040         /*
14041          * tv passed from common code is from either M_TSTMP_LRO or
14042          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
14043          */
14044         rack = (struct tcp_rack *)tp->t_fb_ptr;
14045         if (m->m_flags & M_ACKCMP) {
14046                 /*
14047                  * All compressed ack's are ack's by definition so
14048                  * remove any ack required flag and then do the processing.
14049                  */
14050                 rack->rc_ack_required = 0;
14051                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
14052         }
14053         if (m->m_flags & M_ACKCMP) {
14054                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
14055         }
14056         cts = tcp_tv_to_usectick(tv);
14057         ms_cts =  tcp_tv_to_mssectick(tv);
14058         nsegs = m->m_pkthdr.lro_nsegs;
14059         counter_u64_add(rack_proc_non_comp_ack, 1);
14060         thflags = tcp_get_flags(th);
14061 #ifdef TCP_ACCOUNTING
14062         sched_pin();
14063         if (thflags & TH_ACK)
14064                 ts_val = get_cyclecount();
14065 #endif
14066         if ((m->m_flags & M_TSTMP) ||
14067             (m->m_flags & M_TSTMP_LRO)) {
14068                 mbuf_tstmp2timespec(m, &ts);
14069                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
14070                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
14071         } else
14072                 rack->r_ctl.act_rcv_time = *tv;
14073         kern_prefetch(rack, &prev_state);
14074         prev_state = 0;
14075         /*
14076          * Unscale the window into a 32-bit value. For the SYN_SENT state
14077          * the scale is zero.
14078          */
14079         tiwin = th->th_win << tp->snd_scale;
14080 #ifdef TCP_ACCOUNTING
14081         if (thflags & TH_ACK) {
14082                 /*
14083                  * We have a tradeoff here. We can either do what we are
14084                  * doing i.e. pinning to this CPU and then doing the accounting
14085                  * <or> we could do a critical enter, setup the rdtsc and cpu
14086                  * as in below, and then validate we are on the same CPU on
14087                  * exit. I have choosen to not do the critical enter since
14088                  * that often will gain you a context switch, and instead lock
14089                  * us (line above this if) to the same CPU with sched_pin(). This
14090                  * means we may be context switched out for a higher priority
14091                  * interupt but we won't be moved to another CPU.
14092                  *
14093                  * If this occurs (which it won't very often since we most likely
14094                  * are running this code in interupt context and only a higher
14095                  * priority will bump us ... clock?) we will falsely add in
14096                  * to the time the interupt processing time plus the ack processing
14097                  * time. This is ok since its a rare event.
14098                  */
14099                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
14100                                                     ctf_fixed_maxseg(tp));
14101         }
14102 #endif
14103         /*
14104          * Parse options on any incoming segment.
14105          */
14106         memset(&to, 0, sizeof(to));
14107         tcp_dooptions(&to, (u_char *)(th + 1),
14108             (th->th_off << 2) - sizeof(struct tcphdr),
14109             (thflags & TH_SYN) ? TO_SYN : 0);
14110         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
14111             __func__));
14112         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
14113             __func__));
14114
14115         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
14116             (tp->t_flags & TF_GPUTINPROG)) {
14117                 /*
14118                  * We have a goodput in progress
14119                  * and we have entered a late state.
14120                  * Do we have enough data in the sb
14121                  * to handle the GPUT request?
14122                  */
14123                 uint32_t bytes;
14124
14125                 bytes = tp->gput_ack - tp->gput_seq;
14126                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
14127                         bytes += tp->gput_seq - tp->snd_una;
14128                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
14129                         /*
14130                          * There are not enough bytes in the socket
14131                          * buffer that have been sent to cover this
14132                          * measurement. Cancel it.
14133                          */
14134                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14135                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
14136                                                    tp->gput_seq,
14137                                                    0, 0, 18, __LINE__, NULL, 0);
14138                         tp->t_flags &= ~TF_GPUTINPROG;
14139                 }
14140         }
14141         high_seq = th->th_ack;
14142         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14143                 union tcp_log_stackspecific log;
14144                 struct timeval ltv;
14145 #ifdef NETFLIX_HTTP_LOGGING
14146                 struct http_sendfile_track *http_req;
14147
14148                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
14149                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14150                 } else {
14151                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14152                 }
14153 #endif
14154                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14155                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
14156                 if (rack->rack_no_prr == 0)
14157                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14158                 else
14159                         log.u_bbr.flex1 = 0;
14160                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14161                 log.u_bbr.use_lt_bw <<= 1;
14162                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
14163                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14164                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14165                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14166                 log.u_bbr.flex3 = m->m_flags;
14167                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14168                 log.u_bbr.lost = thflags;
14169                 log.u_bbr.pacing_gain = 0x1;
14170 #ifdef TCP_ACCOUNTING
14171                 log.u_bbr.cwnd_gain = ack_val_set;
14172 #endif
14173                 log.u_bbr.flex7 = 2;
14174                 if (m->m_flags & M_TSTMP) {
14175                         /* Record the hardware timestamp if present */
14176                         mbuf_tstmp2timespec(m, &ts);
14177                         ltv.tv_sec = ts.tv_sec;
14178                         ltv.tv_usec = ts.tv_nsec / 1000;
14179                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14180                 } else if (m->m_flags & M_TSTMP_LRO) {
14181                         /* Record the LRO the arrival timestamp */
14182                         mbuf_tstmp2timespec(m, &ts);
14183                         ltv.tv_sec = ts.tv_sec;
14184                         ltv.tv_usec = ts.tv_nsec / 1000;
14185                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14186                 }
14187                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14188                 /* Log the rcv time */
14189                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14190 #ifdef NETFLIX_HTTP_LOGGING
14191                 log.u_bbr.applimited = tp->t_http_closed;
14192                 log.u_bbr.applimited <<= 8;
14193                 log.u_bbr.applimited |= tp->t_http_open;
14194                 log.u_bbr.applimited <<= 8;
14195                 log.u_bbr.applimited |= tp->t_http_req;
14196                 if (http_req) {
14197                         /* Copy out any client req info */
14198                         /* seconds */
14199                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14200                         /* useconds */
14201                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14202                         log.u_bbr.rttProp = http_req->timestamp;
14203                         log.u_bbr.cur_del_rate = http_req->start;
14204                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14205                                 log.u_bbr.flex8 |= 1;
14206                         } else {
14207                                 log.u_bbr.flex8 |= 2;
14208                                 log.u_bbr.bw_inuse = http_req->end;
14209                         }
14210                         log.u_bbr.flex6 = http_req->start_seq;
14211                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14212                                 log.u_bbr.flex8 |= 4;
14213                                 log.u_bbr.epoch = http_req->end_seq;
14214                         }
14215                 }
14216 #endif
14217                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14218                     tlen, &log, true, &ltv);
14219         }
14220         /* Remove ack required flag if set, we have one  */
14221         if (thflags & TH_ACK)
14222                 rack->rc_ack_required = 0;
14223         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14224                 way_out = 4;
14225                 retval = 0;
14226                 m_freem(m);
14227                 goto done_with_input;
14228         }
14229         /*
14230          * If a segment with the ACK-bit set arrives in the SYN-SENT state
14231          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14232          */
14233         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14234             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14235                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14236                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14237 #ifdef TCP_ACCOUNTING
14238                 sched_unpin();
14239 #endif
14240                 return (1);
14241         }
14242         /*
14243          * If timestamps were negotiated during SYN/ACK and a
14244          * segment without a timestamp is received, silently drop
14245          * the segment, unless it is a RST segment or missing timestamps are
14246          * tolerated.
14247          * See section 3.2 of RFC 7323.
14248          */
14249         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14250             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14251                 way_out = 5;
14252                 retval = 0;
14253                 m_freem(m);
14254                 goto done_with_input;
14255         }
14256
14257         /*
14258          * Segment received on connection. Reset idle time and keep-alive
14259          * timer. XXX: This should be done after segment validation to
14260          * ignore broken/spoofed segs.
14261          */
14262         if  (tp->t_idle_reduce &&
14263              (tp->snd_max == tp->snd_una) &&
14264              (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14265                 counter_u64_add(rack_input_idle_reduces, 1);
14266                 rack_cc_after_idle(rack, tp);
14267         }
14268         tp->t_rcvtime = ticks;
14269 #ifdef STATS
14270         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14271 #endif
14272         if (tiwin > rack->r_ctl.rc_high_rwnd)
14273                 rack->r_ctl.rc_high_rwnd = tiwin;
14274         /*
14275          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14276          * this to occur after we've validated the segment.
14277          */
14278         if (tcp_ecn_input_segment(tp, thflags, tlen,
14279             tcp_packets_this_ack(tp, th->th_ack),
14280             iptos))
14281                 rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
14282
14283         /*
14284          * If echoed timestamp is later than the current time, fall back to
14285          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14286          * were used when this connection was established.
14287          */
14288         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14289                 to.to_tsecr -= tp->ts_offset;
14290                 if (TSTMP_GT(to.to_tsecr, ms_cts))
14291                         to.to_tsecr = 0;
14292         }
14293
14294         /*
14295          * If its the first time in we need to take care of options and
14296          * verify we can do SACK for rack!
14297          */
14298         if (rack->r_state == 0) {
14299                 /* Should be init'd by rack_init() */
14300                 KASSERT(rack->rc_inp != NULL,
14301                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
14302                 if (rack->rc_inp == NULL) {
14303                         rack->rc_inp = inp;
14304                 }
14305
14306                 /*
14307                  * Process options only when we get SYN/ACK back. The SYN
14308                  * case for incoming connections is handled in tcp_syncache.
14309                  * According to RFC1323 the window field in a SYN (i.e., a
14310                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14311                  * this is traditional behavior, may need to be cleaned up.
14312                  */
14313                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14314                         /* Handle parallel SYN for ECN */
14315                         tcp_ecn_input_parallel_syn(tp, thflags, iptos);
14316                         if ((to.to_flags & TOF_SCALE) &&
14317                             (tp->t_flags & TF_REQ_SCALE)) {
14318                                 tp->t_flags |= TF_RCVD_SCALE;
14319                                 tp->snd_scale = to.to_wscale;
14320                         } else
14321                                 tp->t_flags &= ~TF_REQ_SCALE;
14322                         /*
14323                          * Initial send window.  It will be updated with the
14324                          * next incoming segment to the scaled value.
14325                          */
14326                         tp->snd_wnd = th->th_win;
14327                         rack_validate_fo_sendwin_up(tp, rack);
14328                         if ((to.to_flags & TOF_TS) &&
14329                             (tp->t_flags & TF_REQ_TSTMP)) {
14330                                 tp->t_flags |= TF_RCVD_TSTMP;
14331                                 tp->ts_recent = to.to_tsval;
14332                                 tp->ts_recent_age = cts;
14333                         } else
14334                                 tp->t_flags &= ~TF_REQ_TSTMP;
14335                         if (to.to_flags & TOF_MSS) {
14336                                 tcp_mss(tp, to.to_mss);
14337                         }
14338                         if ((tp->t_flags & TF_SACK_PERMIT) &&
14339                             (to.to_flags & TOF_SACKPERM) == 0)
14340                                 tp->t_flags &= ~TF_SACK_PERMIT;
14341                         if (IS_FASTOPEN(tp->t_flags)) {
14342                                 if (to.to_flags & TOF_FASTOPEN) {
14343                                         uint16_t mss;
14344
14345                                         if (to.to_flags & TOF_MSS)
14346                                                 mss = to.to_mss;
14347                                         else
14348                                                 if ((inp->inp_vflag & INP_IPV6) != 0)
14349                                                         mss = TCP6_MSS;
14350                                                 else
14351                                                         mss = TCP_MSS;
14352                                         tcp_fastopen_update_cache(tp, mss,
14353                                             to.to_tfo_len, to.to_tfo_cookie);
14354                                 } else
14355                                         tcp_fastopen_disable_path(tp);
14356                         }
14357                 }
14358                 /*
14359                  * At this point we are at the initial call. Here we decide
14360                  * if we are doing RACK or not. We do this by seeing if
14361                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
14362                  * The code now does do dup-ack counting so if you don't
14363                  * switch back you won't get rack & TLP, but you will still
14364                  * get this stack.
14365                  */
14366
14367                 if ((rack_sack_not_required == 0) &&
14368                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14369                         tcp_switch_back_to_default(tp);
14370                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14371                             tlen, iptos);
14372 #ifdef TCP_ACCOUNTING
14373                         sched_unpin();
14374 #endif
14375                         return (1);
14376                 }
14377                 tcp_set_hpts(inp);
14378                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14379         }
14380         if (thflags & TH_FIN)
14381                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14382         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14383         if ((rack->rc_gp_dyn_mul) &&
14384             (rack->use_fixed_rate == 0) &&
14385             (rack->rc_always_pace)) {
14386                 /* Check in on probertt */
14387                 rack_check_probe_rtt(rack, us_cts);
14388         }
14389         rack_clear_rate_sample(rack);
14390         if ((rack->forced_ack) &&
14391             ((tcp_get_flags(th) & TH_RST) == 0)) {
14392                 rack_handle_probe_response(rack, tiwin, us_cts);
14393         }
14394         /*
14395          * This is the one exception case where we set the rack state
14396          * always. All other times (timers etc) we must have a rack-state
14397          * set (so we assure we have done the checks above for SACK).
14398          */
14399         rack->r_ctl.rc_rcvtime = cts;
14400         if (rack->r_state != tp->t_state)
14401                 rack_set_state(tp, rack);
14402         if (SEQ_GT(th->th_ack, tp->snd_una) &&
14403             (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14404                 kern_prefetch(rsm, &prev_state);
14405         prev_state = rack->r_state;
14406         retval = (*rack->r_substate) (m, th, so,
14407             tp, &to, drop_hdrlen,
14408             tlen, tiwin, thflags, nxt_pkt, iptos);
14409         if (retval == 0) {
14410                 /*
14411                  * If retval is 1 the tcb is unlocked and most likely the tp
14412                  * is gone.
14413                  */
14414                 INP_WLOCK_ASSERT(inp);
14415                 if ((rack->rc_gp_dyn_mul) &&
14416                     (rack->rc_always_pace) &&
14417                     (rack->use_fixed_rate == 0) &&
14418                     rack->in_probe_rtt &&
14419                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
14420                         /*
14421                          * If we are going for target, lets recheck before
14422                          * we output.
14423                          */
14424                         rack_check_probe_rtt(rack, us_cts);
14425                 }
14426                 if (rack->set_pacing_done_a_iw == 0) {
14427                         /* How much has been acked? */
14428                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14429                                 /* We have enough to set in the pacing segment size */
14430                                 rack->set_pacing_done_a_iw = 1;
14431                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
14432                         }
14433                 }
14434                 tcp_rack_xmit_timer_commit(rack, tp);
14435 #ifdef TCP_ACCOUNTING
14436                 /*
14437                  * If we set the ack_val_se to what ack processing we are doing
14438                  * we also want to track how many cycles we burned. Note
14439                  * the bits after tcp_output we let be "free". This is because
14440                  * we are also tracking the tcp_output times as well. Note the
14441                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
14442                  * 0xf cannot be returned and is what we initialize it too to
14443                  * indicate we are not doing the tabulations.
14444                  */
14445                 if (ack_val_set != 0xf) {
14446                         uint64_t crtsc;
14447
14448                         crtsc = get_cyclecount();
14449                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14450                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14451                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14452                         }
14453                 }
14454 #endif
14455                 if (nxt_pkt == 0) {
14456                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14457 do_output_now:
14458                                 if (tcp_output(tp) < 0)
14459                                         return (1);
14460                                 did_out = 1;
14461                         }
14462                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14463                         rack_free_trim(rack);
14464                 }
14465                 /* Update any rounds needed */
14466                 if (rack_verbose_logging &&  (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
14467                         union tcp_log_stackspecific log;
14468                         struct timeval tv;
14469
14470                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14471                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14472                         log.u_bbr.flex1 = high_seq;
14473                         log.u_bbr.flex2 = rack->r_ctl.roundends;
14474                         log.u_bbr.flex3 = rack->r_ctl.current_round;
14475                         log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
14476                         log.u_bbr.flex8 = 9;
14477                         tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
14478                                        0, &log, false, NULL, NULL, 0, &tv);
14479                 }
14480                 /*
14481                  * The draft (v3) calls for us to use SEQ_GEQ, but that
14482                  * causes issues when we are just going app limited. Lets
14483                  * instead use SEQ_GT <or> where its equal but more data
14484                  * is outstanding.
14485                  */
14486                 if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) ||
14487                     ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) {
14488                         rack->r_ctl.current_round++;
14489                         rack->r_ctl.roundends = tp->snd_max;
14490                         if (CC_ALGO(tp)->newround != NULL) {
14491                                 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
14492                         }
14493                 }
14494                 if ((nxt_pkt == 0) &&
14495                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14496                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
14497                      (tp->t_flags & TF_DELACK) ||
14498                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14499                       (tp->t_state <= TCPS_CLOSING)))) {
14500                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
14501                         if ((tp->snd_max == tp->snd_una) &&
14502                             ((tp->t_flags & TF_DELACK) == 0) &&
14503                             (tcp_in_hpts(rack->rc_inp)) &&
14504                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14505                                 /* keep alive not needed if we are hptsi output yet */
14506                                 ;
14507                         } else {
14508                                 int late = 0;
14509                                 if (tcp_in_hpts(inp)) {
14510                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14511                                                 us_cts = tcp_get_usecs(NULL);
14512                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14513                                                         rack->r_early = 1;
14514                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14515                                                 } else
14516                                                         late = 1;
14517                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14518                                         }
14519                                         tcp_hpts_remove(inp);
14520                                 }
14521                                 if (late && (did_out == 0)) {
14522                                         /*
14523                                          * We are late in the sending
14524                                          * and we did not call the output
14525                                          * (this probably should not happen).
14526                                          */
14527                                         goto do_output_now;
14528                                 }
14529                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14530                         }
14531                         way_out = 1;
14532                 } else if (nxt_pkt == 0) {
14533                         /* Do we have the correct timer running? */
14534                         rack_timer_audit(tp, rack, &so->so_snd);
14535                         way_out = 2;
14536                 }
14537         done_with_input:
14538                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14539                 if (did_out)
14540                         rack->r_wanted_output = 0;
14541 #ifdef TCP_ACCOUNTING
14542         } else {
14543                 /*
14544                  * Track the time (see above).
14545                  */
14546                 if (ack_val_set != 0xf) {
14547                         uint64_t crtsc;
14548
14549                         crtsc = get_cyclecount();
14550                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14551                         /*
14552                          * Note we *DO NOT* increment the per-tcb counters since
14553                          * in the else the TP may be gone!!
14554                          */
14555                 }
14556 #endif
14557         }
14558 #ifdef TCP_ACCOUNTING
14559         sched_unpin();
14560 #endif
14561         return (retval);
14562 }
14563
14564 void
14565 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14566     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14567 {
14568         struct timeval tv;
14569
14570         /* First lets see if we have old packets */
14571         if (tp->t_in_pkt) {
14572                 if (ctf_do_queued_segments(so, tp, 1)) {
14573                         m_freem(m);
14574                         return;
14575                 }
14576         }
14577         if (m->m_flags & M_TSTMP_LRO) {
14578                 mbuf_tstmp2timeval(m, &tv);
14579         } else {
14580                 /* Should not be should we kassert instead? */
14581                 tcp_get_usecs(&tv);
14582         }
14583         if (rack_do_segment_nounlock(m, th, so, tp,
14584                                      drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14585                 INP_WUNLOCK(tptoinpcb(tp));
14586         }
14587 }
14588
14589 struct rack_sendmap *
14590 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14591 {
14592         struct rack_sendmap *rsm = NULL;
14593         int32_t idx;
14594         uint32_t srtt = 0, thresh = 0, ts_low = 0;
14595
14596         /* Return the next guy to be re-transmitted */
14597         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14598                 return (NULL);
14599         }
14600         if (tp->t_flags & TF_SENTFIN) {
14601                 /* retran the end FIN? */
14602                 return (NULL);
14603         }
14604         /* ok lets look at this one */
14605         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14606         if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
14607                 return (rsm);
14608         }
14609         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14610                 goto check_it;
14611         }
14612         rsm = rack_find_lowest_rsm(rack);
14613         if (rsm == NULL) {
14614                 return (NULL);
14615         }
14616 check_it:
14617         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14618             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14619                 /*
14620                  * No sack so we automatically do the 3 strikes and
14621                  * retransmit (no rack timer would be started).
14622                  */
14623
14624                 return (rsm);
14625         }
14626         if (rsm->r_flags & RACK_ACKED) {
14627                 return (NULL);
14628         }
14629         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14630             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14631                 /* Its not yet ready */
14632                 return (NULL);
14633         }
14634         srtt = rack_grab_rtt(tp, rack);
14635         idx = rsm->r_rtr_cnt - 1;
14636         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14637         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14638         if ((tsused == ts_low) ||
14639             (TSTMP_LT(tsused, ts_low))) {
14640                 /* No time since sending */
14641                 return (NULL);
14642         }
14643         if ((tsused - ts_low) < thresh) {
14644                 /* It has not been long enough yet */
14645                 return (NULL);
14646         }
14647         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14648             ((rsm->r_flags & RACK_SACK_PASSED) &&
14649              (rack->sack_attack_disable == 0))) {
14650                 /*
14651                  * We have passed the dup-ack threshold <or>
14652                  * a SACK has indicated this is missing.
14653                  * Note that if you are a declared attacker
14654                  * it is only the dup-ack threshold that
14655                  * will cause retransmits.
14656                  */
14657                 /* log retransmit reason */
14658                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14659                 rack->r_fast_output = 0;
14660                 return (rsm);
14661         }
14662         return (NULL);
14663 }
14664
14665 static void
14666 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14667                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14668                            int line, struct rack_sendmap *rsm, uint8_t quality)
14669 {
14670         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14671                 union tcp_log_stackspecific log;
14672                 struct timeval tv;
14673
14674                 memset(&log, 0, sizeof(log));
14675                 log.u_bbr.flex1 = slot;
14676                 log.u_bbr.flex2 = len;
14677                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14678                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14679                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14680                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14681                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14682                 log.u_bbr.use_lt_bw <<= 1;
14683                 log.u_bbr.use_lt_bw |= rack->r_late;
14684                 log.u_bbr.use_lt_bw <<= 1;
14685                 log.u_bbr.use_lt_bw |= rack->r_early;
14686                 log.u_bbr.use_lt_bw <<= 1;
14687                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14688                 log.u_bbr.use_lt_bw <<= 1;
14689                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14690                 log.u_bbr.use_lt_bw <<= 1;
14691                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14692                 log.u_bbr.use_lt_bw <<= 1;
14693                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14694                 log.u_bbr.use_lt_bw <<= 1;
14695                 log.u_bbr.use_lt_bw |= rack->gp_ready;
14696                 log.u_bbr.pkt_epoch = line;
14697                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14698                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14699                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14700                 log.u_bbr.bw_inuse = bw_est;
14701                 log.u_bbr.delRate = bw;
14702                 if (rack->r_ctl.gp_bw == 0)
14703                         log.u_bbr.cur_del_rate = 0;
14704                 else
14705                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
14706                 log.u_bbr.rttProp = len_time;
14707                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14708                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14709                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14710                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14711                         /* We are in slow start */
14712                         log.u_bbr.flex7 = 1;
14713                 } else {
14714                         /* we are on congestion avoidance */
14715                         log.u_bbr.flex7 = 0;
14716                 }
14717                 log.u_bbr.flex8 = method;
14718                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14719                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14720                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14721                 log.u_bbr.cwnd_gain <<= 1;
14722                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14723                 log.u_bbr.cwnd_gain <<= 1;
14724                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14725                 log.u_bbr.bbr_substate = quality;
14726                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14727                     &rack->rc_inp->inp_socket->so_rcv,
14728                     &rack->rc_inp->inp_socket->so_snd,
14729                     BBR_LOG_HPTSI_CALC, 0,
14730                     0, &log, false, &tv);
14731         }
14732 }
14733
14734 static uint32_t
14735 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14736 {
14737         uint32_t new_tso, user_max;
14738
14739         user_max = rack->rc_user_set_max_segs * mss;
14740         if (rack->rc_force_max_seg) {
14741                 return (user_max);
14742         }
14743         if (rack->use_fixed_rate &&
14744             ((rack->r_ctl.crte == NULL) ||
14745              (bw != rack->r_ctl.crte->rate))) {
14746                 /* Use the user mss since we are not exactly matched */
14747                 return (user_max);
14748         }
14749         new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14750         if (new_tso > user_max)
14751                 new_tso = user_max;
14752         return (new_tso);
14753 }
14754
14755 static int32_t
14756 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)
14757 {
14758         uint64_t lentim, fill_bw;
14759
14760         /* Lets first see if we are full, if so continue with normal rate */
14761         rack->r_via_fill_cw = 0;
14762         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14763                 return (slot);
14764         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14765                 return (slot);
14766         if (rack->r_ctl.rc_last_us_rtt == 0)
14767                 return (slot);
14768         if (rack->rc_pace_fill_if_rttin_range &&
14769             (rack->r_ctl.rc_last_us_rtt >=
14770              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14771                 /* The rtt is huge, N * smallest, lets not fill */
14772                 return (slot);
14773         }
14774         /*
14775          * first lets calculate the b/w based on the last us-rtt
14776          * and the sndwnd.
14777          */
14778         fill_bw = rack->r_ctl.cwnd_to_use;
14779         /* Take the rwnd if its smaller */
14780         if (fill_bw > rack->rc_tp->snd_wnd)
14781                 fill_bw = rack->rc_tp->snd_wnd;
14782         if (rack->r_fill_less_agg) {
14783                 /*
14784                  * Now take away the inflight (this will reduce our
14785                  * aggressiveness and yeah, if we get that much out in 1RTT
14786                  * we will have had acks come back and still be behind).
14787                  */
14788                 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14789         }
14790         /* Now lets make it into a b/w */
14791         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14792         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14793         /* We are below the min b/w */
14794         if (non_paced)
14795                 *rate_wanted = fill_bw;
14796         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14797                 return (slot);
14798         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14799                 fill_bw = rack->r_ctl.bw_rate_cap;
14800         rack->r_via_fill_cw = 1;
14801         if (rack->r_rack_hw_rate_caps &&
14802             (rack->r_ctl.crte != NULL)) {
14803                 uint64_t high_rate;
14804
14805                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14806                 if (fill_bw > high_rate) {
14807                         /* We are capping bw at the highest rate table entry */
14808                         if (*rate_wanted > high_rate) {
14809                                 /* The original rate was also capped */
14810                                 rack->r_via_fill_cw = 0;
14811                         }
14812                         rack_log_hdwr_pacing(rack,
14813                                              fill_bw, high_rate, __LINE__,
14814                                              0, 3);
14815                         fill_bw = high_rate;
14816                         if (capped)
14817                                 *capped = 1;
14818                 }
14819         } else if ((rack->r_ctl.crte == NULL) &&
14820                    (rack->rack_hdrw_pacing == 0) &&
14821                    (rack->rack_hdw_pace_ena) &&
14822                    rack->r_rack_hw_rate_caps &&
14823                    (rack->rack_attempt_hdwr_pace == 0) &&
14824                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
14825                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14826                 /*
14827                  * Ok we may have a first attempt that is greater than our top rate
14828                  * lets check.
14829                  */
14830                 uint64_t high_rate;
14831
14832                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14833                 if (high_rate) {
14834                         if (fill_bw > high_rate) {
14835                                 fill_bw = high_rate;
14836                                 if (capped)
14837                                         *capped = 1;
14838                         }
14839                 }
14840         }
14841         /*
14842          * Ok fill_bw holds our mythical b/w to fill the cwnd
14843          * in a rtt, what does that time wise equate too?
14844          */
14845         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14846         lentim /= fill_bw;
14847         *rate_wanted = fill_bw;
14848         if (non_paced || (lentim < slot)) {
14849                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14850                                            0, lentim, 12, __LINE__, NULL, 0);
14851                 return ((int32_t)lentim);
14852         } else
14853                 return (slot);
14854 }
14855
14856 static int32_t
14857 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14858 {
14859         uint64_t srtt;
14860         int32_t slot = 0;
14861         int can_start_hw_pacing = 1;
14862         int err;
14863
14864         if (rack->rc_always_pace == 0) {
14865                 /*
14866                  * We use the most optimistic possible cwnd/srtt for
14867                  * sending calculations. This will make our
14868                  * calculation anticipate getting more through
14869                  * quicker then possible. But thats ok we don't want
14870                  * the peer to have a gap in data sending.
14871                  */
14872                 uint64_t cwnd, tr_perms = 0;
14873                 int32_t reduce = 0;
14874
14875         old_method:
14876                 /*
14877                  * We keep no precise pacing with the old method
14878                  * instead we use the pacer to mitigate bursts.
14879                  */
14880                 if (rack->r_ctl.rc_rack_min_rtt)
14881                         srtt = rack->r_ctl.rc_rack_min_rtt;
14882                 else
14883                         srtt = max(tp->t_srtt, 1);
14884                 if (rack->r_ctl.rc_rack_largest_cwnd)
14885                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14886                 else
14887                         cwnd = rack->r_ctl.cwnd_to_use;
14888                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14889                 tr_perms = (cwnd * 1000) / srtt;
14890                 if (tr_perms == 0) {
14891                         tr_perms = ctf_fixed_maxseg(tp);
14892                 }
14893                 /*
14894                  * Calculate how long this will take to drain, if
14895                  * the calculation comes out to zero, thats ok we
14896                  * will use send_a_lot to possibly spin around for
14897                  * more increasing tot_len_this_send to the point
14898                  * that its going to require a pace, or we hit the
14899                  * cwnd. Which in that case we are just waiting for
14900                  * a ACK.
14901                  */
14902                 slot = len / tr_perms;
14903                 /* Now do we reduce the time so we don't run dry? */
14904                 if (slot && rack_slot_reduction) {
14905                         reduce = (slot / rack_slot_reduction);
14906                         if (reduce < slot) {
14907                                 slot -= reduce;
14908                         } else
14909                                 slot = 0;
14910                 }
14911                 slot *= HPTS_USEC_IN_MSEC;
14912                 if (rack->rc_pace_to_cwnd) {
14913                         uint64_t rate_wanted = 0;
14914
14915                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14916                         rack->rc_ack_can_sendout_data = 1;
14917                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14918                 } else
14919                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14920         } else {
14921                 uint64_t bw_est, res, lentim, rate_wanted;
14922                 uint32_t orig_val, segs, oh;
14923                 int capped = 0;
14924                 int prev_fill;
14925
14926                 if ((rack->r_rr_config == 1) && rsm) {
14927                         return (rack->r_ctl.rc_min_to);
14928                 }
14929                 if (rack->use_fixed_rate) {
14930                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14931                 } else if ((rack->r_ctl.init_rate == 0) &&
14932 #ifdef NETFLIX_PEAKRATE
14933                            (rack->rc_tp->t_maxpeakrate == 0) &&
14934 #endif
14935                            (rack->r_ctl.gp_bw == 0)) {
14936                         /* no way to yet do an estimate */
14937                         bw_est = rate_wanted = 0;
14938                 } else {
14939                         bw_est = rack_get_bw(rack);
14940                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14941                 }
14942                 if ((bw_est == 0) || (rate_wanted == 0) ||
14943                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14944                         /*
14945                          * No way yet to make a b/w estimate or
14946                          * our raise is set incorrectly.
14947                          */
14948                         goto old_method;
14949                 }
14950                 /* We need to account for all the overheads */
14951                 segs = (len + segsiz - 1) / segsiz;
14952                 /*
14953                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
14954                  * and how much data we put in each packet. Yes this
14955                  * means we may be off if we are larger than 1500 bytes
14956                  * or smaller. But this just makes us more conservative.
14957                  */
14958                 if (rack_hw_rate_min &&
14959                     (bw_est < rack_hw_rate_min))
14960                         can_start_hw_pacing = 0;
14961                 if (ETHERNET_SEGMENT_SIZE > segsiz)
14962                         oh = ETHERNET_SEGMENT_SIZE - segsiz;
14963                 else
14964                         oh = 0;
14965                 segs *= oh;
14966                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14967                 res = lentim / rate_wanted;
14968                 slot = (uint32_t)res;
14969                 orig_val = rack->r_ctl.rc_pace_max_segs;
14970                 if (rack->r_ctl.crte == NULL) {
14971                         /*
14972                          * Only do this if we are not hardware pacing
14973                          * since if we are doing hw-pacing below we will
14974                          * set make a call after setting up or changing
14975                          * the rate.
14976                          */
14977                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14978                 } else if (rack->rc_inp->inp_snd_tag == NULL) {
14979                         /*
14980                          * We lost our rate somehow, this can happen
14981                          * if the interface changed underneath us.
14982                          */
14983                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14984                         rack->r_ctl.crte = NULL;
14985                         /* Lets re-allow attempting to setup pacing */
14986                         rack->rack_hdrw_pacing = 0;
14987                         rack->rack_attempt_hdwr_pace = 0;
14988                         rack_log_hdwr_pacing(rack,
14989                                              rate_wanted, bw_est, __LINE__,
14990                                              0, 6);
14991                 }
14992                 /* Did we change the TSO size, if so log it */
14993                 if (rack->r_ctl.rc_pace_max_segs != orig_val)
14994                         rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
14995                 prev_fill = rack->r_via_fill_cw;
14996                 if ((rack->rc_pace_to_cwnd) &&
14997                     (capped == 0) &&
14998                     (rack->use_fixed_rate == 0) &&
14999                     (rack->in_probe_rtt == 0) &&
15000                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
15001                         /*
15002                          * We want to pace at our rate *or* faster to
15003                          * fill the cwnd to the max if its not full.
15004                          */
15005                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
15006                 }
15007                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
15008                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
15009                         if ((rack->rack_hdw_pace_ena) &&
15010                             (can_start_hw_pacing > 0) &&
15011                             (rack->rack_hdrw_pacing == 0) &&
15012                             (rack->rack_attempt_hdwr_pace == 0)) {
15013                                 /*
15014                                  * Lets attempt to turn on hardware pacing
15015                                  * if we can.
15016                                  */
15017                                 rack->rack_attempt_hdwr_pace = 1;
15018                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
15019                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
15020                                                                        rate_wanted,
15021                                                                        RS_PACING_GEQ,
15022                                                                        &err, &rack->r_ctl.crte_prev_rate);
15023                                 if (rack->r_ctl.crte) {
15024                                         rack->rack_hdrw_pacing = 1;
15025                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
15026                                                                                                  0, rack->r_ctl.crte,
15027                                                                                                  NULL);
15028                                         rack_log_hdwr_pacing(rack,
15029                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15030                                                              err, 0);
15031                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
15032                                 } else {
15033                                         counter_u64_add(rack_hw_pace_init_fail, 1);
15034                                 }
15035                         } else if (rack->rack_hdrw_pacing &&
15036                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
15037                                 /* Do we need to adjust our rate? */
15038                                 const struct tcp_hwrate_limit_table *nrte;
15039
15040                                 if (rack->r_up_only &&
15041                                     (rate_wanted < rack->r_ctl.crte->rate)) {
15042                                         /**
15043                                          * We have four possible states here
15044                                          * having to do with the previous time
15045                                          * and this time.
15046                                          *   previous  |  this-time
15047                                          * A)     0      |     0   -- fill_cw not in the picture
15048                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
15049                                          * C)     1      |     1   -- all rates from fill_cw
15050                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
15051                                          *
15052                                          * For case A, C and D we don't allow a drop. But for
15053                                          * case B where we now our on our steady rate we do
15054                                          * allow a drop.
15055                                          *
15056                                          */
15057                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
15058                                                 goto done_w_hdwr;
15059                                 }
15060                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
15061                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
15062                                         if (rack_hw_rate_to_low &&
15063                                             (bw_est < rack_hw_rate_to_low)) {
15064                                                 /*
15065                                                  * The pacing rate is too low for hardware, but
15066                                                  * do allow hardware pacing to be restarted.
15067                                                  */
15068                                                 rack_log_hdwr_pacing(rack,
15069                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
15070                                                              0, 5);
15071                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15072                                                 rack->r_ctl.crte = NULL;
15073                                                 rack->rack_attempt_hdwr_pace = 0;
15074                                                 rack->rack_hdrw_pacing = 0;
15075                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15076                                                 goto done_w_hdwr;
15077                                         }
15078                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
15079                                                                    rack->rc_tp,
15080                                                                    rack->rc_inp->inp_route.ro_nh->nh_ifp,
15081                                                                    rate_wanted,
15082                                                                    RS_PACING_GEQ,
15083                                                                    &err, &rack->r_ctl.crte_prev_rate);
15084                                         if (nrte == NULL) {
15085                                                 /* Lost the rate */
15086                                                 rack->rack_hdrw_pacing = 0;
15087                                                 rack->r_ctl.crte = NULL;
15088                                                 rack_log_hdwr_pacing(rack,
15089                                                                      rate_wanted, 0, __LINE__,
15090                                                                      err, 1);
15091                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15092                                                 counter_u64_add(rack_hw_pace_lost, 1);
15093                                         } else if (nrte != rack->r_ctl.crte) {
15094                                                 rack->r_ctl.crte = nrte;
15095                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
15096                                                                                                          segsiz, 0,
15097                                                                                                          rack->r_ctl.crte,
15098                                                                                                          NULL);
15099                                                 rack_log_hdwr_pacing(rack,
15100                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15101                                                                      err, 2);
15102                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
15103                                         }
15104                                 } else {
15105                                         /* We just need to adjust the segment size */
15106                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15107                                         rack_log_hdwr_pacing(rack,
15108                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15109                                                              0, 4);
15110                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
15111                                 }
15112                         }
15113                 }
15114                 if ((rack->r_ctl.crte != NULL) &&
15115                     (rack->r_ctl.crte->rate == rate_wanted)) {
15116                         /*
15117                          * We need to add a extra if the rates
15118                          * are exactly matched. The idea is
15119                          * we want the software to make sure the
15120                          * queue is empty before adding more, this
15121                          * gives us N MSS extra pace times where
15122                          * N is our sysctl
15123                          */
15124                         slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
15125                 }
15126 done_w_hdwr:
15127                 if (rack_limit_time_with_srtt &&
15128                     (rack->use_fixed_rate == 0) &&
15129 #ifdef NETFLIX_PEAKRATE
15130                     (rack->rc_tp->t_maxpeakrate == 0) &&
15131 #endif
15132                     (rack->rack_hdrw_pacing == 0)) {
15133                         /*
15134                          * Sanity check, we do not allow the pacing delay
15135                          * to be longer than the SRTT of the path. If it is
15136                          * a slow path, then adding a packet should increase
15137                          * the RTT and compensate for this i.e. the srtt will
15138                          * be greater so the allowed pacing time will be greater.
15139                          *
15140                          * Note this restriction is not for where a peak rate
15141                          * is set, we are doing fixed pacing or hardware pacing.
15142                          */
15143                         if (rack->rc_tp->t_srtt)
15144                                 srtt = rack->rc_tp->t_srtt;
15145                         else
15146                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
15147                         if (srtt < (uint64_t)slot) {
15148                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15149                                 slot = srtt;
15150                         }
15151                 }
15152                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15153         }
15154         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15155                 /*
15156                  * If this rate is seeing enobufs when it
15157                  * goes to send then either the nic is out
15158                  * of gas or we are mis-estimating the time
15159                  * somehow and not letting the queue empty
15160                  * completely. Lets add to the pacing time.
15161                  */
15162                 int hw_boost_delay;
15163
15164                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15165                 if (hw_boost_delay > rack_enobuf_hw_max)
15166                         hw_boost_delay = rack_enobuf_hw_max;
15167                 else if (hw_boost_delay < rack_enobuf_hw_min)
15168                         hw_boost_delay = rack_enobuf_hw_min;
15169                 slot += hw_boost_delay;
15170         }
15171         return (slot);
15172 }
15173
15174 static void
15175 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15176     tcp_seq startseq, uint32_t sb_offset)
15177 {
15178         struct rack_sendmap *my_rsm = NULL;
15179         struct rack_sendmap fe;
15180
15181         if (tp->t_state < TCPS_ESTABLISHED) {
15182                 /*
15183                  * We don't start any measurements if we are
15184                  * not at least established.
15185                  */
15186                 return;
15187         }
15188         if (tp->t_state >= TCPS_FIN_WAIT_1) {
15189                 /*
15190                  * We will get no more data into the SB
15191                  * this means we need to have the data available
15192                  * before we start a measurement.
15193                  */
15194
15195                 if (sbavail(&tptosocket(tp)->so_snd) <
15196                     max(rc_init_window(rack),
15197                         (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15198                         /* Nope not enough data */
15199                         return;
15200                 }
15201         }
15202         tp->t_flags |= TF_GPUTINPROG;
15203         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15204         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15205         tp->gput_seq = startseq;
15206         rack->app_limited_needs_set = 0;
15207         if (rack->in_probe_rtt)
15208                 rack->measure_saw_probe_rtt = 1;
15209         else if ((rack->measure_saw_probe_rtt) &&
15210                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15211                 rack->measure_saw_probe_rtt = 0;
15212         if (rack->rc_gp_filled)
15213                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15214         else {
15215                 /* Special case initial measurement */
15216                 struct timeval tv;
15217
15218                 tp->gput_ts = tcp_get_usecs(&tv);
15219                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15220         }
15221         /*
15222          * We take a guess out into the future,
15223          * if we have no measurement and no
15224          * initial rate, we measure the first
15225          * initial-windows worth of data to
15226          * speed up getting some GP measurement and
15227          * thus start pacing.
15228          */
15229         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15230                 rack->app_limited_needs_set = 1;
15231                 tp->gput_ack = startseq + max(rc_init_window(rack),
15232                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15233                 rack_log_pacing_delay_calc(rack,
15234                                            tp->gput_seq,
15235                                            tp->gput_ack,
15236                                            0,
15237                                            tp->gput_ts,
15238                                            rack->r_ctl.rc_app_limited_cnt,
15239                                            9,
15240                                            __LINE__, NULL, 0);
15241                 return;
15242         }
15243         if (sb_offset) {
15244                 /*
15245                  * We are out somewhere in the sb
15246                  * can we use the already outstanding data?
15247                  */
15248                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
15249                         /*
15250                          * Yes first one is good and in this case
15251                          * the tp->gput_ts is correctly set based on
15252                          * the last ack that arrived (no need to
15253                          * set things up when an ack comes in).
15254                          */
15255                         my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15256                         if ((my_rsm == NULL) ||
15257                             (my_rsm->r_rtr_cnt != 1)) {
15258                                 /* retransmission? */
15259                                 goto use_latest;
15260                         }
15261                 } else {
15262                         if (rack->r_ctl.rc_first_appl == NULL) {
15263                                 /*
15264                                  * If rc_first_appl is NULL
15265                                  * then the cnt should be 0.
15266                                  * This is probably an error, maybe
15267                                  * a KASSERT would be approprate.
15268                                  */
15269                                 goto use_latest;
15270                         }
15271                         /*
15272                          * If we have a marker pointer to the last one that is
15273                          * app limited we can use that, but we need to set
15274                          * things up so that when it gets ack'ed we record
15275                          * the ack time (if its not already acked).
15276                          */
15277                         rack->app_limited_needs_set = 1;
15278                         /*
15279                          * We want to get to the rsm that is either
15280                          * next with space i.e. over 1 MSS or the one
15281                          * after that (after the app-limited).
15282                          */
15283                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15284                                          rack->r_ctl.rc_first_appl);
15285                         if (my_rsm) {
15286                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15287                                         /* Have to use the next one */
15288                                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15289                                                          my_rsm);
15290                                 else {
15291                                         /* Use after the first MSS of it is acked */
15292                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15293                                         goto start_set;
15294                                 }
15295                         }
15296                         if ((my_rsm == NULL) ||
15297                             (my_rsm->r_rtr_cnt != 1)) {
15298                                 /*
15299                                  * Either its a retransmit or
15300                                  * the last is the app-limited one.
15301                                  */
15302                                 goto use_latest;
15303                         }
15304                 }
15305                 tp->gput_seq = my_rsm->r_start;
15306 start_set:
15307                 if (my_rsm->r_flags & RACK_ACKED) {
15308                         /*
15309                          * This one has been acked use the arrival ack time
15310                          */
15311                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15312                         rack->app_limited_needs_set = 0;
15313                 }
15314                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15315                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15316                 rack_log_pacing_delay_calc(rack,
15317                                            tp->gput_seq,
15318                                            tp->gput_ack,
15319                                            (uint64_t)my_rsm,
15320                                            tp->gput_ts,
15321                                            rack->r_ctl.rc_app_limited_cnt,
15322                                            9,
15323                                            __LINE__, NULL, 0);
15324                 return;
15325         }
15326
15327 use_latest:
15328         /*
15329          * We don't know how long we may have been
15330          * idle or if this is the first-send. Lets
15331          * setup the flag so we will trim off
15332          * the first ack'd data so we get a true
15333          * measurement.
15334          */
15335         rack->app_limited_needs_set = 1;
15336         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15337         /* Find this guy so we can pull the send time */
15338         fe.r_start = startseq;
15339         my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15340         if (my_rsm) {
15341                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15342                 if (my_rsm->r_flags & RACK_ACKED) {
15343                         /*
15344                          * Unlikely since its probably what was
15345                          * just transmitted (but I am paranoid).
15346                          */
15347                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15348                         rack->app_limited_needs_set = 0;
15349                 }
15350                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15351                         /* This also is unlikely */
15352                         tp->gput_seq = my_rsm->r_start;
15353                 }
15354         } else {
15355                 /*
15356                  * TSNH unless we have some send-map limit,
15357                  * and even at that it should not be hitting
15358                  * that limit (we should have stopped sending).
15359                  */
15360                 struct timeval tv;
15361
15362                 microuptime(&tv);
15363                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15364         }
15365         rack_log_pacing_delay_calc(rack,
15366                                    tp->gput_seq,
15367                                    tp->gput_ack,
15368                                    (uint64_t)my_rsm,
15369                                    tp->gput_ts,
15370                                    rack->r_ctl.rc_app_limited_cnt,
15371                                    9, __LINE__, NULL, 0);
15372 }
15373
15374 static inline uint32_t
15375 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15376     uint32_t avail, int32_t sb_offset)
15377 {
15378         uint32_t len;
15379         uint32_t sendwin;
15380
15381         if (tp->snd_wnd > cwnd_to_use)
15382                 sendwin = cwnd_to_use;
15383         else
15384                 sendwin = tp->snd_wnd;
15385         if (ctf_outstanding(tp) >= tp->snd_wnd) {
15386                 /* We never want to go over our peers rcv-window */
15387                 len = 0;
15388         } else {
15389                 uint32_t flight;
15390
15391                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15392                 if (flight >= sendwin) {
15393                         /*
15394                          * We have in flight what we are allowed by cwnd (if
15395                          * it was rwnd blocking it would have hit above out
15396                          * >= tp->snd_wnd).
15397                          */
15398                         return (0);
15399                 }
15400                 len = sendwin - flight;
15401                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15402                         /* We would send too much (beyond the rwnd) */
15403                         len = tp->snd_wnd - ctf_outstanding(tp);
15404                 }
15405                 if ((len + sb_offset) > avail) {
15406                         /*
15407                          * We don't have that much in the SB, how much is
15408                          * there?
15409                          */
15410                         len = avail - sb_offset;
15411                 }
15412         }
15413         return (len);
15414 }
15415
15416 static void
15417 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15418              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15419              int rsm_is_null, int optlen, int line, uint16_t mode)
15420 {
15421         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15422                 union tcp_log_stackspecific log;
15423                 struct timeval tv;
15424
15425                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15426                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15427                 log.u_bbr.flex1 = error;
15428                 log.u_bbr.flex2 = flags;
15429                 log.u_bbr.flex3 = rsm_is_null;
15430                 log.u_bbr.flex4 = ipoptlen;
15431                 log.u_bbr.flex5 = tp->rcv_numsacks;
15432                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15433                 log.u_bbr.flex7 = optlen;
15434                 log.u_bbr.flex8 = rack->r_fsb_inited;
15435                 log.u_bbr.applimited = rack->r_fast_output;
15436                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15437                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15438                 log.u_bbr.cwnd_gain = mode;
15439                 log.u_bbr.pkts_out = orig_len;
15440                 log.u_bbr.lt_epoch = len;
15441                 log.u_bbr.delivered = line;
15442                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15443                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15444                 tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15445                                len, &log, false, NULL, NULL, 0, &tv);
15446         }
15447 }
15448
15449
15450 static struct mbuf *
15451 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15452                    struct rack_fast_send_blk *fsb,
15453                    int32_t seglimit, int32_t segsize, int hw_tls)
15454 {
15455 #ifdef KERN_TLS
15456         struct ktls_session *tls, *ntls;
15457 #ifdef INVARIANTS
15458         struct mbuf *start;
15459 #endif
15460 #endif
15461         struct mbuf *m, *n, **np, *smb;
15462         struct mbuf *top;
15463         int32_t off, soff;
15464         int32_t len = *plen;
15465         int32_t fragsize;
15466         int32_t len_cp = 0;
15467         uint32_t mlen, frags;
15468
15469         soff = off = the_off;
15470         smb = m = the_m;
15471         np = &top;
15472         top = NULL;
15473 #ifdef KERN_TLS
15474         if (hw_tls && (m->m_flags & M_EXTPG))
15475                 tls = m->m_epg_tls;
15476         else
15477                 tls = NULL;
15478 #ifdef INVARIANTS
15479         start = m;
15480 #endif
15481 #endif
15482         while (len > 0) {
15483                 if (m == NULL) {
15484                         *plen = len_cp;
15485                         break;
15486                 }
15487 #ifdef KERN_TLS
15488                 if (hw_tls) {
15489                         if (m->m_flags & M_EXTPG)
15490                                 ntls = m->m_epg_tls;
15491                         else
15492                                 ntls = NULL;
15493
15494                         /*
15495                          * Avoid mixing TLS records with handshake
15496                          * data or TLS records from different
15497                          * sessions.
15498                          */
15499                         if (tls != ntls) {
15500                                 MPASS(m != start);
15501                                 *plen = len_cp;
15502                                 break;
15503                         }
15504                 }
15505 #endif
15506                 mlen = min(len, m->m_len - off);
15507                 if (seglimit) {
15508                         /*
15509                          * For M_EXTPG mbufs, add 3 segments
15510                          * + 1 in case we are crossing page boundaries
15511                          * + 2 in case the TLS hdr/trailer are used
15512                          * It is cheaper to just add the segments
15513                          * than it is to take the cache miss to look
15514                          * at the mbuf ext_pgs state in detail.
15515                          */
15516                         if (m->m_flags & M_EXTPG) {
15517                                 fragsize = min(segsize, PAGE_SIZE);
15518                                 frags = 3;
15519                         } else {
15520                                 fragsize = segsize;
15521                                 frags = 0;
15522                         }
15523
15524                         /* Break if we really can't fit anymore. */
15525                         if ((frags + 1) >= seglimit) {
15526                                 *plen = len_cp;
15527                                 break;
15528                         }
15529
15530                         /*
15531                          * Reduce size if you can't copy the whole
15532                          * mbuf. If we can't copy the whole mbuf, also
15533                          * adjust len so the loop will end after this
15534                          * mbuf.
15535                          */
15536                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15537                                 mlen = (seglimit - frags - 1) * fragsize;
15538                                 len = mlen;
15539                                 *plen = len_cp + len;
15540                         }
15541                         frags += howmany(mlen, fragsize);
15542                         if (frags == 0)
15543                                 frags++;
15544                         seglimit -= frags;
15545                         KASSERT(seglimit > 0,
15546                             ("%s: seglimit went too low", __func__));
15547                 }
15548                 n = m_get(M_NOWAIT, m->m_type);
15549                 *np = n;
15550                 if (n == NULL)
15551                         goto nospace;
15552                 n->m_len = mlen;
15553                 soff += mlen;
15554                 len_cp += n->m_len;
15555                 if (m->m_flags & (M_EXT|M_EXTPG)) {
15556                         n->m_data = m->m_data + off;
15557                         mb_dupcl(n, m);
15558                 } else {
15559                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15560                             (u_int)n->m_len);
15561                 }
15562                 len -= n->m_len;
15563                 off = 0;
15564                 m = m->m_next;
15565                 np = &n->m_next;
15566                 if (len || (soff == smb->m_len)) {
15567                         /*
15568                          * We have more so we move forward  or
15569                          * we have consumed the entire mbuf and
15570                          * len has fell to 0.
15571                          */
15572                         soff = 0;
15573                         smb = m;
15574                 }
15575
15576         }
15577         if (fsb != NULL) {
15578                 fsb->m = smb;
15579                 fsb->off = soff;
15580                 if (smb) {
15581                         /*
15582                          * Save off the size of the mbuf. We do
15583                          * this so that we can recognize when it
15584                          * has been trimmed by sbcut() as acks
15585                          * come in.
15586                          */
15587                         fsb->o_m_len = smb->m_len;
15588                 } else {
15589                         /*
15590                          * This is the case where the next mbuf went to NULL. This
15591                          * means with this copy we have sent everything in the sb.
15592                          * In theory we could clear the fast_output flag, but lets
15593                          * not since its possible that we could get more added
15594                          * and acks that call the extend function which would let
15595                          * us send more.
15596                          */
15597                         fsb->o_m_len = 0;
15598                 }
15599         }
15600         return (top);
15601 nospace:
15602         if (top)
15603                 m_freem(top);
15604         return (NULL);
15605
15606 }
15607
15608 /*
15609  * This is a copy of m_copym(), taking the TSO segment size/limit
15610  * constraints into account, and advancing the sndptr as it goes.
15611  */
15612 static struct mbuf *
15613 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15614                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15615 {
15616         struct mbuf *m, *n;
15617         int32_t soff;
15618
15619         soff = rack->r_ctl.fsb.off;
15620         m = rack->r_ctl.fsb.m;
15621         if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15622                 /*
15623                  * The mbuf had the front of it chopped off by an ack
15624                  * we need to adjust the soff/off by that difference.
15625                  */
15626                 uint32_t delta;
15627
15628                 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15629                 soff -= delta;
15630         } else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15631                 /*
15632                  * The mbuf was expanded probably by
15633                  * a m_compress. Just update o_m_len.
15634                  */
15635                 rack->r_ctl.fsb.o_m_len = m->m_len;
15636         }
15637         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15638         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15639         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15640                                  __FUNCTION__,
15641                                  rack, *plen, m, m->m_len));
15642         /* Save off the right location before we copy and advance */
15643         *s_soff = soff;
15644         *s_mb = rack->r_ctl.fsb.m;
15645         n = rack_fo_base_copym(m, soff, plen,
15646                                &rack->r_ctl.fsb,
15647                                seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15648         return (n);
15649 }
15650
15651 static int
15652 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15653                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15654 {
15655         /*
15656          * Enter the fast retransmit path. We are given that a sched_pin is
15657          * in place (if accounting is compliled in) and the cycle count taken
15658          * at the entry is in the ts_val. The concept her is that the rsm
15659          * now holds the mbuf offsets and such so we can directly transmit
15660          * without a lot of overhead, the len field is already set for
15661          * us to prohibit us from sending too much (usually its 1MSS).
15662          */
15663         struct ip *ip = NULL;
15664         struct udphdr *udp = NULL;
15665         struct tcphdr *th = NULL;
15666         struct mbuf *m = NULL;
15667         struct inpcb *inp;
15668         uint8_t *cpto;
15669         struct tcp_log_buffer *lgb;
15670 #ifdef TCP_ACCOUNTING
15671         uint64_t crtsc;
15672         int cnt_thru = 1;
15673 #endif
15674         struct tcpopt to;
15675         u_char opt[TCP_MAXOLEN];
15676         uint32_t hdrlen, optlen;
15677         int32_t slot, segsiz, max_val, tso = 0, error, ulen = 0;
15678         uint16_t flags;
15679         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15680         uint32_t if_hw_tsomaxsegsize;
15681
15682 #ifdef INET6
15683         struct ip6_hdr *ip6 = NULL;
15684
15685         if (rack->r_is_v6) {
15686                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15687                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15688         } else
15689 #endif                          /* INET6 */
15690         {
15691                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15692                 hdrlen = sizeof(struct tcpiphdr);
15693         }
15694         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15695                 goto failed;
15696         }
15697         if (doing_tlp) {
15698                 /* Its a TLP add the flag, it may already be there but be sure */
15699                 rsm->r_flags |= RACK_TLP;
15700         } else {
15701                 /* If it was a TLP it is not not on this retransmit */
15702                 rsm->r_flags &= ~RACK_TLP;
15703         }
15704         startseq = rsm->r_start;
15705         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15706         inp = rack->rc_inp;
15707         to.to_flags = 0;
15708         flags = tcp_outflags[tp->t_state];
15709         if (flags & (TH_SYN|TH_RST)) {
15710                 goto failed;
15711         }
15712         if (rsm->r_flags & RACK_HAS_FIN) {
15713                 /* We can't send a FIN here */
15714                 goto failed;
15715         }
15716         if (flags & TH_FIN) {
15717                 /* We never send a FIN */
15718                 flags &= ~TH_FIN;
15719         }
15720         if (tp->t_flags & TF_RCVD_TSTMP) {
15721                 to.to_tsval = ms_cts + tp->ts_offset;
15722                 to.to_tsecr = tp->ts_recent;
15723                 to.to_flags = TOF_TS;
15724         }
15725         optlen = tcp_addoptions(&to, opt);
15726         hdrlen += optlen;
15727         udp = rack->r_ctl.fsb.udp;
15728         if (udp)
15729                 hdrlen += sizeof(struct udphdr);
15730         if (rack->r_ctl.rc_pace_max_segs)
15731                 max_val = rack->r_ctl.rc_pace_max_segs;
15732         else if (rack->rc_user_set_max_segs)
15733                 max_val = rack->rc_user_set_max_segs * segsiz;
15734         else
15735                 max_val = len;
15736         if ((tp->t_flags & TF_TSO) &&
15737             V_tcp_do_tso &&
15738             (len > segsiz) &&
15739             (tp->t_port == 0))
15740                 tso = 1;
15741 #ifdef INET6
15742         if (MHLEN < hdrlen + max_linkhdr)
15743                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15744         else
15745 #endif
15746                 m = m_gethdr(M_NOWAIT, MT_DATA);
15747         if (m == NULL)
15748                 goto failed;
15749         m->m_data += max_linkhdr;
15750         m->m_len = hdrlen;
15751         th = rack->r_ctl.fsb.th;
15752         /* Establish the len to send */
15753         if (len > max_val)
15754                 len = max_val;
15755         if ((tso) && (len + optlen > tp->t_maxseg)) {
15756                 uint32_t if_hw_tsomax;
15757                 int32_t max_len;
15758
15759                 /* extract TSO information */
15760                 if_hw_tsomax = tp->t_tsomax;
15761                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15762                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15763                 /*
15764                  * Check if we should limit by maximum payload
15765                  * length:
15766                  */
15767                 if (if_hw_tsomax != 0) {
15768                         /* compute maximum TSO length */
15769                         max_len = (if_hw_tsomax - hdrlen -
15770                                    max_linkhdr);
15771                         if (max_len <= 0) {
15772                                 goto failed;
15773                         } else if (len > max_len) {
15774                                 len = max_len;
15775                         }
15776                 }
15777                 if (len <= segsiz) {
15778                         /*
15779                          * In case there are too many small fragments don't
15780                          * use TSO:
15781                          */
15782                         tso = 0;
15783                 }
15784         } else {
15785                 tso = 0;
15786         }
15787         if ((tso == 0) && (len > segsiz))
15788                 len = segsiz;
15789         if ((len == 0) ||
15790             (len <= MHLEN - hdrlen - max_linkhdr)) {
15791                 goto failed;
15792         }
15793         th->th_seq = htonl(rsm->r_start);
15794         th->th_ack = htonl(tp->rcv_nxt);
15795         /*
15796          * The PUSH bit should only be applied
15797          * if the full retransmission is made. If
15798          * we are sending less than this is the
15799          * left hand edge and should not have
15800          * the PUSH bit.
15801          */
15802         if ((rsm->r_flags & RACK_HAD_PUSH) &&
15803             (len == (rsm->r_end - rsm->r_start)))
15804                 flags |= TH_PUSH;
15805         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15806         if (th->th_win == 0) {
15807                 tp->t_sndzerowin++;
15808                 tp->t_flags |= TF_RXWIN0SENT;
15809         } else
15810                 tp->t_flags &= ~TF_RXWIN0SENT;
15811         if (rsm->r_flags & RACK_TLP) {
15812                 /*
15813                  * TLP should not count in retran count, but
15814                  * in its own bin
15815                  */
15816                 counter_u64_add(rack_tlp_retran, 1);
15817                 counter_u64_add(rack_tlp_retran_bytes, len);
15818         } else {
15819                 tp->t_sndrexmitpack++;
15820                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15821                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15822         }
15823 #ifdef STATS
15824         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15825                                  len);
15826 #endif
15827         if (rsm->m == NULL)
15828                 goto failed;
15829         if (rsm->orig_m_len != rsm->m->m_len) {
15830                 /* Fix up the orig_m_len and possibly the mbuf offset */
15831                 rack_adjust_orig_mlen(rsm);
15832         }
15833         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15834         if (len <= segsiz) {
15835                 /*
15836                  * Must have ran out of mbufs for the copy
15837                  * shorten it to no longer need tso. Lets
15838                  * not put on sendalot since we are low on
15839                  * mbufs.
15840                  */
15841                 tso = 0;
15842         }
15843         if ((m->m_next == NULL) || (len <= 0)){
15844                 goto failed;
15845         }
15846         if (udp) {
15847                 if (rack->r_is_v6)
15848                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15849                 else
15850                         ulen = hdrlen + len - sizeof(struct ip);
15851                 udp->uh_ulen = htons(ulen);
15852         }
15853         m->m_pkthdr.rcvif = (struct ifnet *)0;
15854         if (TCPS_HAVERCVDSYN(tp->t_state) &&
15855             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
15856                 int ect = tcp_ecn_output_established(tp, &flags, len, true);
15857                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
15858                     (tp->t_flags2 & TF2_ECN_SND_ECE))
15859                     tp->t_flags2 &= ~TF2_ECN_SND_ECE;
15860 #ifdef INET6
15861                 if (rack->r_is_v6) {
15862                     ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
15863                     ip6->ip6_flow |= htonl(ect << 20);
15864                 }
15865                 else
15866 #endif
15867                 {
15868                     ip->ip_tos &= ~IPTOS_ECN_MASK;
15869                     ip->ip_tos |= ect;
15870                 }
15871         }
15872         tcp_set_flags(th, flags);
15873         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15874 #ifdef INET6
15875         if (rack->r_is_v6) {
15876                 if (tp->t_port) {
15877                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15878                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15879                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15880                         th->th_sum = htons(0);
15881                         UDPSTAT_INC(udps_opackets);
15882                 } else {
15883                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15884                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15885                         th->th_sum = in6_cksum_pseudo(ip6,
15886                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15887                                                       0);
15888                 }
15889         }
15890 #endif
15891 #if defined(INET6) && defined(INET)
15892         else
15893 #endif
15894 #ifdef INET
15895         {
15896                 if (tp->t_port) {
15897                         m->m_pkthdr.csum_flags = CSUM_UDP;
15898                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15899                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15900                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15901                         th->th_sum = htons(0);
15902                         UDPSTAT_INC(udps_opackets);
15903                 } else {
15904                         m->m_pkthdr.csum_flags = CSUM_TCP;
15905                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15906                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15907                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15908                                                                         IPPROTO_TCP + len + optlen));
15909                 }
15910                 /* IP version must be set here for ipv4/ipv6 checking later */
15911                 KASSERT(ip->ip_v == IPVERSION,
15912                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15913         }
15914 #endif
15915         if (tso) {
15916                 KASSERT(len > tp->t_maxseg - optlen,
15917                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15918                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15919                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15920         }
15921 #ifdef INET6
15922         if (rack->r_is_v6) {
15923                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15924                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15925                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15926                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15927                 else
15928                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15929         }
15930 #endif
15931 #if defined(INET) && defined(INET6)
15932         else
15933 #endif
15934 #ifdef INET
15935         {
15936                 ip->ip_len = htons(m->m_pkthdr.len);
15937                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15938                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15939                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15940                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15941                                 ip->ip_off |= htons(IP_DF);
15942                         }
15943                 } else {
15944                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15945                 }
15946         }
15947 #endif
15948         /* Time to copy in our header */
15949         cpto = mtod(m, uint8_t *);
15950         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15951         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15952         if (optlen) {
15953                 bcopy(opt, th + 1, optlen);
15954                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15955         } else {
15956                 th->th_off = sizeof(struct tcphdr) >> 2;
15957         }
15958         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15959                 union tcp_log_stackspecific log;
15960
15961                 if (rsm->r_flags & RACK_RWND_COLLAPSED) {
15962                         rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
15963                         counter_u64_add(rack_collapsed_win_rxt, 1);
15964                         counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
15965                 }
15966                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15967                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15968                 if (rack->rack_no_prr)
15969                         log.u_bbr.flex1 = 0;
15970                 else
15971                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15972                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15973                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15974                 log.u_bbr.flex4 = max_val;
15975                 log.u_bbr.flex5 = 0;
15976                 /* Save off the early/late values */
15977                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15978                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15979                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15980                 if (doing_tlp == 0)
15981                         log.u_bbr.flex8 = 1;
15982                 else
15983                         log.u_bbr.flex8 = 2;
15984                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15985                 log.u_bbr.flex7 = 55;
15986                 log.u_bbr.pkts_out = tp->t_maxseg;
15987                 log.u_bbr.timeStamp = cts;
15988                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15989                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15990                 log.u_bbr.delivered = 0;
15991                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15992                                      len, &log, false, NULL, NULL, 0, tv);
15993         } else
15994                 lgb = NULL;
15995 #ifdef INET6
15996         if (rack->r_is_v6) {
15997                 error = ip6_output(m, NULL,
15998                                    &inp->inp_route6,
15999                                    0, NULL, NULL, inp);
16000         }
16001 #endif
16002 #if defined(INET) && defined(INET6)
16003         else
16004 #endif
16005 #ifdef INET
16006         {
16007                 error = ip_output(m, NULL,
16008                                   &inp->inp_route,
16009                                   0, 0, inp);
16010         }
16011 #endif
16012         m = NULL;
16013         if (lgb) {
16014                 lgb->tlb_errno = error;
16015                 lgb = NULL;
16016         }
16017         if (error) {
16018                 goto failed;
16019         }
16020         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
16021                         rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
16022         if (doing_tlp && (rack->fast_rsm_hack == 0)) {
16023                 rack->rc_tlp_in_progress = 1;
16024                 rack->r_ctl.rc_tlp_cnt_out++;
16025         }
16026         if (error == 0) {
16027                 tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
16028                 if (doing_tlp) {
16029                         rack->rc_last_sent_tlp_past_cumack = 0;
16030                         rack->rc_last_sent_tlp_seq_valid = 1;
16031                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
16032                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
16033                 }
16034         }
16035         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16036         rack->forced_ack = 0;   /* If we send something zap the FA flag */
16037         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
16038                 rack->r_ctl.retran_during_recovery += len;
16039         {
16040                 int idx;
16041
16042                 idx = (len / segsiz) + 3;
16043                 if (idx >= TCP_MSS_ACCT_ATIMER)
16044                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16045                 else
16046                         counter_u64_add(rack_out_size[idx], 1);
16047         }
16048         if (tp->t_rtttime == 0) {
16049                 tp->t_rtttime = ticks;
16050                 tp->t_rtseq = startseq;
16051                 KMOD_TCPSTAT_INC(tcps_segstimed);
16052         }
16053         counter_u64_add(rack_fto_rsm_send, 1);
16054         if (error && (error == ENOBUFS)) {
16055                 if (rack->r_ctl.crte != NULL) {
16056                         rack_trace_point(rack, RACK_TP_HWENOBUF);
16057                 } else
16058                         rack_trace_point(rack, RACK_TP_ENOBUF);
16059                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
16060                 if (rack->rc_enobuf < 0x7f)
16061                         rack->rc_enobuf++;
16062                 if (slot < (10 * HPTS_USEC_IN_MSEC))
16063                         slot = 10 * HPTS_USEC_IN_MSEC;
16064         } else
16065                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
16066         if ((slot == 0) ||
16067             (rack->rc_always_pace == 0) ||
16068             (rack->r_rr_config == 1)) {
16069                 /*
16070                  * We have no pacing set or we
16071                  * are using old-style rack or
16072                  * we are overridden to use the old 1ms pacing.
16073                  */
16074                 slot = rack->r_ctl.rc_min_to;
16075         }
16076         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
16077 #ifdef TCP_ACCOUNTING
16078         crtsc = get_cyclecount();
16079         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16080                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16081         }
16082         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16083         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16084                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16085         }
16086         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16087         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16088                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
16089         }
16090         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
16091         sched_unpin();
16092 #endif
16093         return (0);
16094 failed:
16095         if (m)
16096                 m_free(m);
16097         return (-1);
16098 }
16099
16100 static void
16101 rack_sndbuf_autoscale(struct tcp_rack *rack)
16102 {
16103         /*
16104          * Automatic sizing of send socket buffer.  Often the send buffer
16105          * size is not optimally adjusted to the actual network conditions
16106          * at hand (delay bandwidth product).  Setting the buffer size too
16107          * small limits throughput on links with high bandwidth and high
16108          * delay (eg. trans-continental/oceanic links).  Setting the
16109          * buffer size too big consumes too much real kernel memory,
16110          * especially with many connections on busy servers.
16111          *
16112          * The criteria to step up the send buffer one notch are:
16113          *  1. receive window of remote host is larger than send buffer
16114          *     (with a fudge factor of 5/4th);
16115          *  2. send buffer is filled to 7/8th with data (so we actually
16116          *     have data to make use of it);
16117          *  3. send buffer fill has not hit maximal automatic size;
16118          *  4. our send window (slow start and cogestion controlled) is
16119          *     larger than sent but unacknowledged data in send buffer.
16120          *
16121          * Note that the rack version moves things much faster since
16122          * we want to avoid hitting cache lines in the rack_fast_output()
16123          * path so this is called much less often and thus moves
16124          * the SB forward by a percentage.
16125          */
16126         struct socket *so;
16127         struct tcpcb *tp;
16128         uint32_t sendwin, scaleup;
16129
16130         tp = rack->rc_tp;
16131         so = rack->rc_inp->inp_socket;
16132         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
16133         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
16134                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
16135                     sbused(&so->so_snd) >=
16136                     (so->so_snd.sb_hiwat / 8 * 7) &&
16137                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
16138                     sendwin >= (sbused(&so->so_snd) -
16139                     (tp->snd_nxt - tp->snd_una))) {
16140                         if (rack_autosndbuf_inc)
16141                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
16142                         else
16143                                 scaleup = V_tcp_autosndbuf_inc;
16144                         if (scaleup < V_tcp_autosndbuf_inc)
16145                                 scaleup = V_tcp_autosndbuf_inc;
16146                         scaleup += so->so_snd.sb_hiwat;
16147                         if (scaleup > V_tcp_autosndbuf_max)
16148                                 scaleup = V_tcp_autosndbuf_max;
16149                         if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
16150                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16151                 }
16152         }
16153 }
16154
16155 static int
16156 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16157                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16158 {
16159         /*
16160          * Enter to do fast output. We are given that the sched_pin is
16161          * in place (if accounting is compiled in) and the cycle count taken
16162          * at entry is in place in ts_val. The idea here is that
16163          * we know how many more bytes needs to be sent (presumably either
16164          * during pacing or to fill the cwnd and that was greater than
16165          * the max-burst). We have how much to send and all the info we
16166          * need to just send.
16167          */
16168         struct ip *ip = NULL;
16169         struct udphdr *udp = NULL;
16170         struct tcphdr *th = NULL;
16171         struct mbuf *m, *s_mb;
16172         struct inpcb *inp;
16173         uint8_t *cpto;
16174         struct tcp_log_buffer *lgb;
16175 #ifdef TCP_ACCOUNTING
16176         uint64_t crtsc;
16177 #endif
16178         struct tcpopt to;
16179         u_char opt[TCP_MAXOLEN];
16180         uint32_t hdrlen, optlen;
16181 #ifdef TCP_ACCOUNTING
16182         int cnt_thru = 1;
16183 #endif
16184         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
16185         uint16_t flags;
16186         uint32_t s_soff;
16187         uint32_t if_hw_tsomaxsegcount = 0, startseq;
16188         uint32_t if_hw_tsomaxsegsize;
16189         uint16_t add_flag = RACK_SENT_FP;
16190 #ifdef INET6
16191         struct ip6_hdr *ip6 = NULL;
16192
16193         if (rack->r_is_v6) {
16194                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16195                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16196         } else
16197 #endif                          /* INET6 */
16198         {
16199                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16200                 hdrlen = sizeof(struct tcpiphdr);
16201         }
16202         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16203                 m = NULL;
16204                 goto failed;
16205         }
16206         startseq = tp->snd_max;
16207         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16208         inp = rack->rc_inp;
16209         len = rack->r_ctl.fsb.left_to_send;
16210         to.to_flags = 0;
16211         flags = rack->r_ctl.fsb.tcp_flags;
16212         if (tp->t_flags & TF_RCVD_TSTMP) {
16213                 to.to_tsval = ms_cts + tp->ts_offset;
16214                 to.to_tsecr = tp->ts_recent;
16215                 to.to_flags = TOF_TS;
16216         }
16217         optlen = tcp_addoptions(&to, opt);
16218         hdrlen += optlen;
16219         udp = rack->r_ctl.fsb.udp;
16220         if (udp)
16221                 hdrlen += sizeof(struct udphdr);
16222         if (rack->r_ctl.rc_pace_max_segs)
16223                 max_val = rack->r_ctl.rc_pace_max_segs;
16224         else if (rack->rc_user_set_max_segs)
16225                 max_val = rack->rc_user_set_max_segs * segsiz;
16226         else
16227                 max_val = len;
16228         if ((tp->t_flags & TF_TSO) &&
16229             V_tcp_do_tso &&
16230             (len > segsiz) &&
16231             (tp->t_port == 0))
16232                 tso = 1;
16233 again:
16234 #ifdef INET6
16235         if (MHLEN < hdrlen + max_linkhdr)
16236                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16237         else
16238 #endif
16239                 m = m_gethdr(M_NOWAIT, MT_DATA);
16240         if (m == NULL)
16241                 goto failed;
16242         m->m_data += max_linkhdr;
16243         m->m_len = hdrlen;
16244         th = rack->r_ctl.fsb.th;
16245         /* Establish the len to send */
16246         if (len > max_val)
16247                 len = max_val;
16248         if ((tso) && (len + optlen > tp->t_maxseg)) {
16249                 uint32_t if_hw_tsomax;
16250                 int32_t max_len;
16251
16252                 /* extract TSO information */
16253                 if_hw_tsomax = tp->t_tsomax;
16254                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16255                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16256                 /*
16257                  * Check if we should limit by maximum payload
16258                  * length:
16259                  */
16260                 if (if_hw_tsomax != 0) {
16261                         /* compute maximum TSO length */
16262                         max_len = (if_hw_tsomax - hdrlen -
16263                                    max_linkhdr);
16264                         if (max_len <= 0) {
16265                                 goto failed;
16266                         } else if (len > max_len) {
16267                                 len = max_len;
16268                         }
16269                 }
16270                 if (len <= segsiz) {
16271                         /*
16272                          * In case there are too many small fragments don't
16273                          * use TSO:
16274                          */
16275                         tso = 0;
16276                 }
16277         } else {
16278                 tso = 0;
16279         }
16280         if ((tso == 0) && (len > segsiz))
16281                 len = segsiz;
16282         if ((len == 0) ||
16283             (len <= MHLEN - hdrlen - max_linkhdr)) {
16284                 goto failed;
16285         }
16286         sb_offset = tp->snd_max - tp->snd_una;
16287         th->th_seq = htonl(tp->snd_max);
16288         th->th_ack = htonl(tp->rcv_nxt);
16289         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16290         if (th->th_win == 0) {
16291                 tp->t_sndzerowin++;
16292                 tp->t_flags |= TF_RXWIN0SENT;
16293         } else
16294                 tp->t_flags &= ~TF_RXWIN0SENT;
16295         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
16296         KMOD_TCPSTAT_INC(tcps_sndpack);
16297         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16298 #ifdef STATS
16299         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16300                                  len);
16301 #endif
16302         if (rack->r_ctl.fsb.m == NULL)
16303                 goto failed;
16304
16305         /* s_mb and s_soff are saved for rack_log_output */
16306         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16307                                     &s_mb, &s_soff);
16308         if (len <= segsiz) {
16309                 /*
16310                  * Must have ran out of mbufs for the copy
16311                  * shorten it to no longer need tso. Lets
16312                  * not put on sendalot since we are low on
16313                  * mbufs.
16314                  */
16315                 tso = 0;
16316         }
16317         if (rack->r_ctl.fsb.rfo_apply_push &&
16318             (len == rack->r_ctl.fsb.left_to_send)) {
16319                 flags |= TH_PUSH;
16320                 add_flag |= RACK_HAD_PUSH;
16321         }
16322         if ((m->m_next == NULL) || (len <= 0)){
16323                 goto failed;
16324         }
16325         if (udp) {
16326                 if (rack->r_is_v6)
16327                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
16328                 else
16329                         ulen = hdrlen + len - sizeof(struct ip);
16330                 udp->uh_ulen = htons(ulen);
16331         }
16332         m->m_pkthdr.rcvif = (struct ifnet *)0;
16333         if (TCPS_HAVERCVDSYN(tp->t_state) &&
16334             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
16335                 int ect = tcp_ecn_output_established(tp, &flags, len, false);
16336                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
16337                     (tp->t_flags2 & TF2_ECN_SND_ECE))
16338                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
16339 #ifdef INET6
16340                 if (rack->r_is_v6) {
16341                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
16342                         ip6->ip6_flow |= htonl(ect << 20);
16343                 }
16344                 else
16345 #endif
16346                 {
16347                         ip->ip_tos &= ~IPTOS_ECN_MASK;
16348                         ip->ip_tos |= ect;
16349                 }
16350         }
16351         tcp_set_flags(th, flags);
16352         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
16353 #ifdef INET6
16354         if (rack->r_is_v6) {
16355                 if (tp->t_port) {
16356                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16357                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16358                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16359                         th->th_sum = htons(0);
16360                         UDPSTAT_INC(udps_opackets);
16361                 } else {
16362                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16363                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16364                         th->th_sum = in6_cksum_pseudo(ip6,
16365                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16366                                                       0);
16367                 }
16368         }
16369 #endif
16370 #if defined(INET6) && defined(INET)
16371         else
16372 #endif
16373 #ifdef INET
16374         {
16375                 if (tp->t_port) {
16376                         m->m_pkthdr.csum_flags = CSUM_UDP;
16377                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16378                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16379                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16380                         th->th_sum = htons(0);
16381                         UDPSTAT_INC(udps_opackets);
16382                 } else {
16383                         m->m_pkthdr.csum_flags = CSUM_TCP;
16384                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16385                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
16386                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16387                                                                         IPPROTO_TCP + len + optlen));
16388                 }
16389                 /* IP version must be set here for ipv4/ipv6 checking later */
16390                 KASSERT(ip->ip_v == IPVERSION,
16391                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
16392         }
16393 #endif
16394         if (tso) {
16395                 KASSERT(len > tp->t_maxseg - optlen,
16396                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
16397                 m->m_pkthdr.csum_flags |= CSUM_TSO;
16398                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16399         }
16400 #ifdef INET6
16401         if (rack->r_is_v6) {
16402                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16403                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16404                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16405                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16406                 else
16407                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16408         }
16409 #endif
16410 #if defined(INET) && defined(INET6)
16411         else
16412 #endif
16413 #ifdef INET
16414         {
16415                 ip->ip_len = htons(m->m_pkthdr.len);
16416                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16417                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16418                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16419                         if (tp->t_port == 0 || len < V_tcp_minmss) {
16420                                 ip->ip_off |= htons(IP_DF);
16421                         }
16422                 } else {
16423                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16424                 }
16425         }
16426 #endif
16427         /* Time to copy in our header */
16428         cpto = mtod(m, uint8_t *);
16429         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16430         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16431         if (optlen) {
16432                 bcopy(opt, th + 1, optlen);
16433                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16434         } else {
16435                 th->th_off = sizeof(struct tcphdr) >> 2;
16436         }
16437         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16438                 union tcp_log_stackspecific log;
16439
16440                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16441                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16442                 if (rack->rack_no_prr)
16443                         log.u_bbr.flex1 = 0;
16444                 else
16445                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16446                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16447                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16448                 log.u_bbr.flex4 = max_val;
16449                 log.u_bbr.flex5 = 0;
16450                 /* Save off the early/late values */
16451                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16452                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16453                 log.u_bbr.bw_inuse = rack_get_bw(rack);
16454                 log.u_bbr.flex8 = 0;
16455                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16456                 log.u_bbr.flex7 = 44;
16457                 log.u_bbr.pkts_out = tp->t_maxseg;
16458                 log.u_bbr.timeStamp = cts;
16459                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16460                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16461                 log.u_bbr.delivered = 0;
16462                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16463                                      len, &log, false, NULL, NULL, 0, tv);
16464         } else
16465                 lgb = NULL;
16466 #ifdef INET6
16467         if (rack->r_is_v6) {
16468                 error = ip6_output(m, NULL,
16469                                    &inp->inp_route6,
16470                                    0, NULL, NULL, inp);
16471         }
16472 #endif
16473 #if defined(INET) && defined(INET6)
16474         else
16475 #endif
16476 #ifdef INET
16477         {
16478                 error = ip_output(m, NULL,
16479                                   &inp->inp_route,
16480                                   0, 0, inp);
16481         }
16482 #endif
16483         if (lgb) {
16484                 lgb->tlb_errno = error;
16485                 lgb = NULL;
16486         }
16487         if (error) {
16488                 *send_err = error;
16489                 m = NULL;
16490                 goto failed;
16491         }
16492         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16493                         NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16494         m = NULL;
16495         if (tp->snd_una == tp->snd_max) {
16496                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
16497                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16498                 tp->t_acktime = ticks;
16499         }
16500         if (error == 0)
16501                 tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16502
16503         rack->forced_ack = 0;   /* If we send something zap the FA flag */
16504         tot_len += len;
16505         if ((tp->t_flags & TF_GPUTINPROG) == 0)
16506                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16507         tp->snd_max += len;
16508         tp->snd_nxt = tp->snd_max;
16509         {
16510                 int idx;
16511
16512                 idx = (len / segsiz) + 3;
16513                 if (idx >= TCP_MSS_ACCT_ATIMER)
16514                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16515                 else
16516                         counter_u64_add(rack_out_size[idx], 1);
16517         }
16518         if (len <= rack->r_ctl.fsb.left_to_send)
16519                 rack->r_ctl.fsb.left_to_send -= len;
16520         else
16521                 rack->r_ctl.fsb.left_to_send = 0;
16522         if (rack->r_ctl.fsb.left_to_send < segsiz) {
16523                 rack->r_fast_output = 0;
16524                 rack->r_ctl.fsb.left_to_send = 0;
16525                 /* At the end of fast_output scale up the sb */
16526                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16527                 rack_sndbuf_autoscale(rack);
16528                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16529         }
16530         if (tp->t_rtttime == 0) {
16531                 tp->t_rtttime = ticks;
16532                 tp->t_rtseq = startseq;
16533                 KMOD_TCPSTAT_INC(tcps_segstimed);
16534         }
16535         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16536             (max_val > len) &&
16537             (tso == 0)) {
16538                 max_val -= len;
16539                 len = segsiz;
16540                 th = rack->r_ctl.fsb.th;
16541 #ifdef TCP_ACCOUNTING
16542                 cnt_thru++;
16543 #endif
16544                 goto again;
16545         }
16546         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16547         counter_u64_add(rack_fto_send, 1);
16548         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16549         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16550 #ifdef TCP_ACCOUNTING
16551         crtsc = get_cyclecount();
16552         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16553                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16554         }
16555         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16556         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16557                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16558         }
16559         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16560         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16561                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16562         }
16563         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16564         sched_unpin();
16565 #endif
16566         return (0);
16567 failed:
16568         if (m)
16569                 m_free(m);
16570         rack->r_fast_output = 0;
16571         return (-1);
16572 }
16573
16574 static struct rack_sendmap *
16575 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
16576 {
16577         struct rack_sendmap *rsm = NULL;
16578         struct rack_sendmap fe;
16579         int thresh;
16580
16581 restart:
16582         fe.r_start = rack->r_ctl.last_collapse_point;
16583         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
16584         if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
16585                 /* Nothing, strange turn off validity  */
16586                 rack->r_collapse_point_valid = 0;
16587                 return (NULL);
16588         }
16589         /* Can we send it yet? */
16590         if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
16591                 /*
16592                  * Receiver window has not grown enough for
16593                  * the segment to be put on the wire.
16594                  */
16595                 return (NULL);
16596         }
16597         if (rsm->r_flags & RACK_ACKED) {
16598                 /*
16599                  * It has been sacked, lets move to the
16600                  * next one if possible.
16601                  */
16602                 rack->r_ctl.last_collapse_point = rsm->r_end;
16603                 /* Are we done? */
16604                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16605                             rack->r_ctl.high_collapse_point)) {
16606                         rack->r_collapse_point_valid = 0;
16607                         return (NULL);
16608                 }
16609                 goto restart;
16610         }
16611         /* Now has it been long enough ? */
16612         thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
16613         if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
16614                 rack_log_collapse(rack, rsm->r_start,
16615                                   (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16616                                   thresh, __LINE__, 6, rsm->r_flags, rsm);
16617                 return (rsm);
16618         }
16619         /* Not enough time */
16620         rack_log_collapse(rack, rsm->r_start,
16621                           (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16622                           thresh, __LINE__, 7, rsm->r_flags, rsm);
16623         return (NULL);
16624 }
16625
16626 static int
16627 rack_output(struct tcpcb *tp)
16628 {
16629         struct socket *so;
16630         uint32_t recwin;
16631         uint32_t sb_offset, s_moff = 0;
16632         int32_t len, error = 0;
16633         uint16_t flags;
16634         struct mbuf *m, *s_mb = NULL;
16635         struct mbuf *mb;
16636         uint32_t if_hw_tsomaxsegcount = 0;
16637         uint32_t if_hw_tsomaxsegsize;
16638         int32_t segsiz, minseg;
16639         long tot_len_this_send = 0;
16640 #ifdef INET
16641         struct ip *ip = NULL;
16642 #endif
16643         struct udphdr *udp = NULL;
16644         struct tcp_rack *rack;
16645         struct tcphdr *th;
16646         uint8_t pass = 0;
16647         uint8_t mark = 0;
16648         uint8_t wanted_cookie = 0;
16649         u_char opt[TCP_MAXOLEN];
16650         unsigned ipoptlen, optlen, hdrlen, ulen=0;
16651         uint32_t rack_seq;
16652
16653 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16654         unsigned ipsec_optlen = 0;
16655
16656 #endif
16657         int32_t idle, sendalot;
16658         int32_t sub_from_prr = 0;
16659         volatile int32_t sack_rxmit;
16660         struct rack_sendmap *rsm = NULL;
16661         int32_t tso, mtu;
16662         struct tcpopt to;
16663         int32_t slot = 0;
16664         int32_t sup_rack = 0;
16665         uint32_t cts, ms_cts, delayed, early;
16666         uint16_t add_flag = RACK_SENT_SP;
16667         /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16668         uint8_t hpts_calling,  doing_tlp = 0;
16669         uint32_t cwnd_to_use, pace_max_seg;
16670         int32_t do_a_prefetch = 0;
16671         int32_t prefetch_rsm = 0;
16672         int32_t orig_len = 0;
16673         struct timeval tv;
16674         int32_t prefetch_so_done = 0;
16675         struct tcp_log_buffer *lgb;
16676         struct inpcb *inp = tptoinpcb(tp);
16677         struct sockbuf *sb;
16678         uint64_t ts_val = 0;
16679 #ifdef TCP_ACCOUNTING
16680         uint64_t crtsc;
16681 #endif
16682 #ifdef INET6
16683         struct ip6_hdr *ip6 = NULL;
16684         int32_t isipv6;
16685 #endif
16686         bool hw_tls = false;
16687
16688         NET_EPOCH_ASSERT();
16689         INP_WLOCK_ASSERT(inp);
16690
16691         /* setup and take the cache hits here */
16692         rack = (struct tcp_rack *)tp->t_fb_ptr;
16693 #ifdef TCP_ACCOUNTING
16694         sched_pin();
16695         ts_val = get_cyclecount();
16696 #endif
16697         hpts_calling = inp->inp_hpts_calls;
16698 #ifdef TCP_OFFLOAD
16699         if (tp->t_flags & TF_TOE) {
16700 #ifdef TCP_ACCOUNTING
16701                 sched_unpin();
16702 #endif
16703                 return (tcp_offload_output(tp));
16704         }
16705 #endif
16706         /*
16707          * For TFO connections in SYN_RECEIVED, only allow the initial
16708          * SYN|ACK and those sent by the retransmit timer.
16709          */
16710         if (IS_FASTOPEN(tp->t_flags) &&
16711             (tp->t_state == TCPS_SYN_RECEIVED) &&
16712             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16713             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16714 #ifdef TCP_ACCOUNTING
16715                 sched_unpin();
16716 #endif
16717                 return (0);
16718         }
16719 #ifdef INET6
16720         if (rack->r_state) {
16721                 /* Use the cache line loaded if possible */
16722                 isipv6 = rack->r_is_v6;
16723         } else {
16724                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16725         }
16726 #endif
16727         early = 0;
16728         cts = tcp_get_usecs(&tv);
16729         ms_cts = tcp_tv_to_mssectick(&tv);
16730         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16731             tcp_in_hpts(rack->rc_inp)) {
16732                 /*
16733                  * We are on the hpts for some timer but not hptsi output.
16734                  * Remove from the hpts unconditionally.
16735                  */
16736                 rack_timer_cancel(tp, rack, cts, __LINE__);
16737         }
16738         /* Are we pacing and late? */
16739         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16740             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16741                 /* We are delayed */
16742                 delayed = cts - rack->r_ctl.rc_last_output_to;
16743         } else {
16744                 delayed = 0;
16745         }
16746         /* Do the timers, which may override the pacer */
16747         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16748                 int retval;
16749
16750                 retval = rack_process_timers(tp, rack, cts, hpts_calling,
16751                     &doing_tlp);
16752                 if (retval != 0) {
16753                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16754 #ifdef TCP_ACCOUNTING
16755                         sched_unpin();
16756 #endif
16757                         /*
16758                          * If timers want tcp_drop(), then pass error out,
16759                          * otherwise suppress it.
16760                          */
16761                         return (retval < 0 ? retval : 0);
16762                 }
16763         }
16764         if (rack->rc_in_persist) {
16765                 if (tcp_in_hpts(rack->rc_inp) == 0) {
16766                         /* Timer is not running */
16767                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16768                 }
16769 #ifdef TCP_ACCOUNTING
16770                 sched_unpin();
16771 #endif
16772                 return (0);
16773         }
16774         if ((rack->rc_ack_required == 1) &&
16775             (rack->r_timer_override == 0)){
16776                 /* A timeout occurred and no ack has arrived */
16777                 if (tcp_in_hpts(rack->rc_inp) == 0) {
16778                         /* Timer is not running */
16779                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16780                 }
16781 #ifdef TCP_ACCOUNTING
16782                 sched_unpin();
16783 #endif
16784                 return (0);
16785         }
16786         if ((rack->r_timer_override) ||
16787             (rack->rc_ack_can_sendout_data) ||
16788             (delayed) ||
16789             (tp->t_state < TCPS_ESTABLISHED)) {
16790                 rack->rc_ack_can_sendout_data = 0;
16791                 if (tcp_in_hpts(rack->rc_inp))
16792                         tcp_hpts_remove(rack->rc_inp);
16793         } else if (tcp_in_hpts(rack->rc_inp)) {
16794                 /*
16795                  * On the hpts you can't pass even if ACKNOW is on, we will
16796                  * when the hpts fires.
16797                  */
16798 #ifdef TCP_ACCOUNTING
16799                 crtsc = get_cyclecount();
16800                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16801                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16802                 }
16803                 counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16804                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16805                         tp->tcp_cnt_counters[SND_BLOCKED]++;
16806                 }
16807                 counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16808                 sched_unpin();
16809 #endif
16810                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16811                 return (0);
16812         }
16813         rack->rc_inp->inp_hpts_calls = 0;
16814         /* Finish out both pacing early and late accounting */
16815         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16816             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16817                 early = rack->r_ctl.rc_last_output_to - cts;
16818         } else
16819                 early = 0;
16820         if (delayed) {
16821                 rack->r_ctl.rc_agg_delayed += delayed;
16822                 rack->r_late = 1;
16823         } else if (early) {
16824                 rack->r_ctl.rc_agg_early += early;
16825                 rack->r_early = 1;
16826         }
16827         /* Now that early/late accounting is done turn off the flag */
16828         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16829         rack->r_wanted_output = 0;
16830         rack->r_timer_override = 0;
16831         if ((tp->t_state != rack->r_state) &&
16832             TCPS_HAVEESTABLISHED(tp->t_state)) {
16833                 rack_set_state(tp, rack);
16834         }
16835         if ((rack->r_fast_output) &&
16836             (doing_tlp == 0) &&
16837             (tp->rcv_numsacks == 0)) {
16838                 int ret;
16839
16840                 error = 0;
16841                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16842                 if (ret >= 0)
16843                         return(ret);
16844                 else if (error) {
16845                         inp = rack->rc_inp;
16846                         so = inp->inp_socket;
16847                         sb = &so->so_snd;
16848                         goto nomore;
16849                 }
16850         }
16851         inp = rack->rc_inp;
16852         /*
16853          * For TFO connections in SYN_SENT or SYN_RECEIVED,
16854          * only allow the initial SYN or SYN|ACK and those sent
16855          * by the retransmit timer.
16856          */
16857         if (IS_FASTOPEN(tp->t_flags) &&
16858             ((tp->t_state == TCPS_SYN_RECEIVED) ||
16859              (tp->t_state == TCPS_SYN_SENT)) &&
16860             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16861             (tp->t_rxtshift == 0)) {              /* not a retransmit */
16862                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16863                 so = inp->inp_socket;
16864                 sb = &so->so_snd;
16865                 goto just_return_nolock;
16866         }
16867         /*
16868          * Determine length of data that should be transmitted, and flags
16869          * that will be used. If there is some data or critical controls
16870          * (SYN, RST) to send, then transmit; otherwise, investigate
16871          * further.
16872          */
16873         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16874         if (tp->t_idle_reduce) {
16875                 if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16876                         rack_cc_after_idle(rack, tp);
16877         }
16878         tp->t_flags &= ~TF_LASTIDLE;
16879         if (idle) {
16880                 if (tp->t_flags & TF_MORETOCOME) {
16881                         tp->t_flags |= TF_LASTIDLE;
16882                         idle = 0;
16883                 }
16884         }
16885         if ((tp->snd_una == tp->snd_max) &&
16886             rack->r_ctl.rc_went_idle_time &&
16887             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16888                 idle = cts - rack->r_ctl.rc_went_idle_time;
16889                 if (idle > rack_min_probertt_hold) {
16890                         /* Count as a probe rtt */
16891                         if (rack->in_probe_rtt == 0) {
16892                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16893                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16894                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16895                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16896                         } else {
16897                                 rack_exit_probertt(rack, cts);
16898                         }
16899                 }
16900                 idle = 0;
16901         }
16902         if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16903                 rack_init_fsb_block(tp, rack);
16904 again:
16905         /*
16906          * If we've recently taken a timeout, snd_max will be greater than
16907          * snd_nxt.  There may be SACK information that allows us to avoid
16908          * resending already delivered data.  Adjust snd_nxt accordingly.
16909          */
16910         sendalot = 0;
16911         cts = tcp_get_usecs(&tv);
16912         ms_cts = tcp_tv_to_mssectick(&tv);
16913         tso = 0;
16914         mtu = 0;
16915         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16916         minseg = segsiz;
16917         if (rack->r_ctl.rc_pace_max_segs == 0)
16918                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16919         else
16920                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16921         sb_offset = tp->snd_max - tp->snd_una;
16922         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16923         flags = tcp_outflags[tp->t_state];
16924         while (rack->rc_free_cnt < rack_free_cache) {
16925                 rsm = rack_alloc(rack);
16926                 if (rsm == NULL) {
16927                         if (inp->inp_hpts_calls)
16928                                 /* Retry in a ms */
16929                                 slot = (1 * HPTS_USEC_IN_MSEC);
16930                         so = inp->inp_socket;
16931                         sb = &so->so_snd;
16932                         goto just_return_nolock;
16933                 }
16934                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16935                 rack->rc_free_cnt++;
16936                 rsm = NULL;
16937         }
16938         if (inp->inp_hpts_calls)
16939                 inp->inp_hpts_calls = 0;
16940         sack_rxmit = 0;
16941         len = 0;
16942         rsm = NULL;
16943         if (flags & TH_RST) {
16944                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16945                 so = inp->inp_socket;
16946                 sb = &so->so_snd;
16947                 goto send;
16948         }
16949         if (rack->r_ctl.rc_resend) {
16950                 /* Retransmit timer */
16951                 rsm = rack->r_ctl.rc_resend;
16952                 rack->r_ctl.rc_resend = NULL;
16953                 len = rsm->r_end - rsm->r_start;
16954                 sack_rxmit = 1;
16955                 sendalot = 0;
16956                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16957                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16958                          __func__, __LINE__,
16959                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16960                 sb_offset = rsm->r_start - tp->snd_una;
16961                 if (len >= segsiz)
16962                         len = segsiz;
16963         } else if (rack->r_collapse_point_valid &&
16964                    ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
16965                 /*
16966                  * If an RSM is returned then enough time has passed
16967                  * for us to retransmit it. Move up the collapse point,
16968                  * since this rsm has its chance to retransmit now.
16969                  */
16970                 rack_trace_point(rack, RACK_TP_COLLAPSED_RXT);
16971                 rack->r_ctl.last_collapse_point = rsm->r_end;
16972                 /* Are we done? */
16973                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16974                             rack->r_ctl.high_collapse_point))
16975                         rack->r_collapse_point_valid = 0;
16976                 sack_rxmit = 1;
16977                 /* We are not doing a TLP */
16978                 doing_tlp = 0;
16979                 len = rsm->r_end - rsm->r_start;
16980                 sb_offset = rsm->r_start - tp->snd_una;
16981                 sendalot = 0;
16982                 if ((rack->full_size_rxt == 0) &&
16983                     (rack->shape_rxt_to_pacing_min == 0) &&
16984                     (len >= segsiz))
16985                         len = segsiz;
16986         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16987                 /* We have a retransmit that takes precedence */
16988                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16989                     ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
16990                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16991                         /* Enter recovery if not induced by a time-out */
16992                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
16993                 }
16994 #ifdef INVARIANTS
16995                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16996                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16997                               tp, rack, rsm, rsm->r_start, tp->snd_una);
16998                 }
16999 #endif
17000                 len = rsm->r_end - rsm->r_start;
17001                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17002                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17003                          __func__, __LINE__,
17004                          rsm->r_start, tp->snd_una, tp, rack, rsm));
17005                 sb_offset = rsm->r_start - tp->snd_una;
17006                 sendalot = 0;
17007                 if (len >= segsiz)
17008                         len = segsiz;
17009                 if (len > 0) {
17010                         sack_rxmit = 1;
17011                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
17012                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
17013                             min(len, segsiz));
17014                 }
17015         } else if (rack->r_ctl.rc_tlpsend) {
17016                 /* Tail loss probe */
17017                 long cwin;
17018                 long tlen;
17019
17020                 /*
17021                  * Check if we can do a TLP with a RACK'd packet
17022                  * this can happen if we are not doing the rack
17023                  * cheat and we skipped to a TLP and it
17024                  * went off.
17025                  */
17026                 rsm = rack->r_ctl.rc_tlpsend;
17027                 /* We are doing a TLP make sure the flag is preent */
17028                 rsm->r_flags |= RACK_TLP;
17029                 rack->r_ctl.rc_tlpsend = NULL;
17030                 sack_rxmit = 1;
17031                 tlen = rsm->r_end - rsm->r_start;
17032                 if (tlen > segsiz)
17033                         tlen = segsiz;
17034                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17035                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17036                          __func__, __LINE__,
17037                          rsm->r_start, tp->snd_una, tp, rack, rsm));
17038                 sb_offset = rsm->r_start - tp->snd_una;
17039                 cwin = min(tp->snd_wnd, tlen);
17040                 len = cwin;
17041         }
17042         if (rack->r_must_retran &&
17043             (doing_tlp == 0) &&
17044             (SEQ_GT(tp->snd_max, tp->snd_una)) &&
17045             (rsm == NULL)) {
17046                 /*
17047                  * There are two different ways that we
17048                  * can get into this block:
17049                  * a) This is a non-sack connection, we had a time-out
17050                  *    and thus r_must_retran was set and everything
17051                  *    left outstanding as been marked for retransmit.
17052                  * b) The MTU of the path shrank, so that everything
17053                  *    was marked to be retransmitted with the smaller
17054                  *    mtu and r_must_retran was set.
17055                  *
17056                  * This means that we expect the sendmap (outstanding)
17057                  * to all be marked must. We can use the tmap to
17058                  * look at them.
17059                  *
17060                  */
17061                 int sendwin, flight;
17062
17063                 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
17064                 flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
17065                 if (flight >= sendwin) {
17066                         /*
17067                          * We can't send yet.
17068                          */
17069                         so = inp->inp_socket;
17070                         sb = &so->so_snd;
17071                         goto just_return_nolock;
17072                 }
17073                 /*
17074                  * This is the case a/b mentioned above. All
17075                  * outstanding/not-acked should be marked.
17076                  * We can use the tmap to find them.
17077                  */
17078                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17079                 if (rsm == NULL) {
17080                         /* TSNH */
17081                         rack->r_must_retran = 0;
17082                         rack->r_ctl.rc_out_at_rto = 0;
17083                         so = inp->inp_socket;
17084                         sb = &so->so_snd;
17085                         goto just_return_nolock;
17086                 }
17087                 if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
17088                         /*
17089                          * The first one does not have the flag, did we collapse
17090                          * further up in our list?
17091                          */
17092                         rack->r_must_retran = 0;
17093                         rack->r_ctl.rc_out_at_rto = 0;
17094                         rsm = NULL;
17095                         sack_rxmit = 0;
17096                 } else {
17097                         sack_rxmit = 1;
17098                         len = rsm->r_end - rsm->r_start;
17099                         sb_offset = rsm->r_start - tp->snd_una;
17100                         sendalot = 0;
17101                         if ((rack->full_size_rxt == 0) &&
17102                             (rack->shape_rxt_to_pacing_min == 0) &&
17103                             (len >= segsiz))
17104                                 len = segsiz;
17105                         /*
17106                          * Delay removing the flag RACK_MUST_RXT so
17107                          * that the fastpath for retransmit will
17108                          * work with this rsm.
17109                          */
17110                 }
17111         }
17112         /*
17113          * Enforce a connection sendmap count limit if set
17114          * as long as we are not retransmiting.
17115          */
17116         if ((rsm == NULL) &&
17117             (rack->do_detection == 0) &&
17118             (V_tcp_map_entries_limit > 0) &&
17119             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
17120                 counter_u64_add(rack_to_alloc_limited, 1);
17121                 if (!rack->alloc_limit_reported) {
17122                         rack->alloc_limit_reported = 1;
17123                         counter_u64_add(rack_alloc_limited_conns, 1);
17124                 }
17125                 so = inp->inp_socket;
17126                 sb = &so->so_snd;
17127                 goto just_return_nolock;
17128         }
17129         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
17130                 /* we are retransmitting the fin */
17131                 len--;
17132                 if (len) {
17133                         /*
17134                          * When retransmitting data do *not* include the
17135                          * FIN. This could happen from a TLP probe.
17136                          */
17137                         flags &= ~TH_FIN;
17138                 }
17139         }
17140         if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
17141             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
17142                 int ret;
17143
17144                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
17145                 if (ret == 0)
17146                         return (0);
17147         }
17148         so = inp->inp_socket;
17149         sb = &so->so_snd;
17150         if (do_a_prefetch == 0) {
17151                 kern_prefetch(sb, &do_a_prefetch);
17152                 do_a_prefetch = 1;
17153         }
17154 #ifdef NETFLIX_SHARED_CWND
17155         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
17156             rack->rack_enable_scwnd) {
17157                 /* We are doing cwnd sharing */
17158                 if (rack->gp_ready &&
17159                     (rack->rack_attempted_scwnd == 0) &&
17160                     (rack->r_ctl.rc_scw == NULL) &&
17161                     tp->t_lib) {
17162                         /* The pcbid is in, lets make an attempt */
17163                         counter_u64_add(rack_try_scwnd, 1);
17164                         rack->rack_attempted_scwnd = 1;
17165                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
17166                                                                    &rack->r_ctl.rc_scw_index,
17167                                                                    segsiz);
17168                 }
17169                 if (rack->r_ctl.rc_scw &&
17170                     (rack->rack_scwnd_is_idle == 1) &&
17171                     sbavail(&so->so_snd)) {
17172                         /* we are no longer out of data */
17173                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17174                         rack->rack_scwnd_is_idle = 0;
17175                 }
17176                 if (rack->r_ctl.rc_scw) {
17177                         /* First lets update and get the cwnd */
17178                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
17179                                                                     rack->r_ctl.rc_scw_index,
17180                                                                     tp->snd_cwnd, tp->snd_wnd, segsiz);
17181                 }
17182         }
17183 #endif
17184         /*
17185          * Get standard flags, and add SYN or FIN if requested by 'hidden'
17186          * state flags.
17187          */
17188         if (tp->t_flags & TF_NEEDFIN)
17189                 flags |= TH_FIN;
17190         if (tp->t_flags & TF_NEEDSYN)
17191                 flags |= TH_SYN;
17192         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
17193                 void *end_rsm;
17194                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17195                 if (end_rsm)
17196                         kern_prefetch(end_rsm, &prefetch_rsm);
17197                 prefetch_rsm = 1;
17198         }
17199         SOCKBUF_LOCK(sb);
17200         /*
17201          * If snd_nxt == snd_max and we have transmitted a FIN, the
17202          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17203          * negative length.  This can also occur when TCP opens up its
17204          * congestion window while receiving additional duplicate acks after
17205          * fast-retransmit because TCP will reset snd_nxt to snd_max after
17206          * the fast-retransmit.
17207          *
17208          * In the normal retransmit-FIN-only case, however, snd_nxt will be
17209          * set to snd_una, the sb_offset will be 0, and the length may wind
17210          * up 0.
17211          *
17212          * If sack_rxmit is true we are retransmitting from the scoreboard
17213          * in which case len is already set.
17214          */
17215         if ((sack_rxmit == 0) &&
17216             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17217                 uint32_t avail;
17218
17219                 avail = sbavail(sb);
17220                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17221                         sb_offset = tp->snd_nxt - tp->snd_una;
17222                 else
17223                         sb_offset = 0;
17224                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17225                         if (rack->r_ctl.rc_tlp_new_data) {
17226                                 /* TLP is forcing out new data */
17227                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17228                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17229                                 }
17230                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17231                                         if (tp->snd_wnd > sb_offset)
17232                                                 len = tp->snd_wnd - sb_offset;
17233                                         else
17234                                                 len = 0;
17235                                 } else {
17236                                         len = rack->r_ctl.rc_tlp_new_data;
17237                                 }
17238                                 rack->r_ctl.rc_tlp_new_data = 0;
17239                         }  else {
17240                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17241                         }
17242                         if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17243                                 /*
17244                                  * For prr=off, we need to send only 1 MSS
17245                                  * at a time. We do this because another sack could
17246                                  * be arriving that causes us to send retransmits and
17247                                  * we don't want to be on a long pace due to a larger send
17248                                  * that keeps us from sending out the retransmit.
17249                                  */
17250                                 len = segsiz;
17251                         }
17252                 } else {
17253                         uint32_t outstanding;
17254                         /*
17255                          * We are inside of a Fast recovery episode, this
17256                          * is caused by a SACK or 3 dup acks. At this point
17257                          * we have sent all the retransmissions and we rely
17258                          * on PRR to dictate what we will send in the form of
17259                          * new data.
17260                          */
17261
17262                         outstanding = tp->snd_max - tp->snd_una;
17263                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17264                                 if (tp->snd_wnd > outstanding) {
17265                                         len = tp->snd_wnd - outstanding;
17266                                         /* Check to see if we have the data */
17267                                         if ((sb_offset + len) > avail) {
17268                                                 /* It does not all fit */
17269                                                 if (avail > sb_offset)
17270                                                         len = avail - sb_offset;
17271                                                 else
17272                                                         len = 0;
17273                                         }
17274                                 } else {
17275                                         len = 0;
17276                                 }
17277                         } else if (avail > sb_offset) {
17278                                 len = avail - sb_offset;
17279                         } else {
17280                                 len = 0;
17281                         }
17282                         if (len > 0) {
17283                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
17284                                         len = rack->r_ctl.rc_prr_sndcnt;
17285                                 }
17286                                 if (len > 0) {
17287                                         sub_from_prr = 1;
17288                                 }
17289                         }
17290                         if (len > segsiz) {
17291                                 /*
17292                                  * We should never send more than a MSS when
17293                                  * retransmitting or sending new data in prr
17294                                  * mode unless the override flag is on. Most
17295                                  * likely the PRR algorithm is not going to
17296                                  * let us send a lot as well :-)
17297                                  */
17298                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
17299                                         len = segsiz;
17300                                 }
17301                         } else if (len < segsiz) {
17302                                 /*
17303                                  * Do we send any? The idea here is if the
17304                                  * send empty's the socket buffer we want to
17305                                  * do it. However if not then lets just wait
17306                                  * for our prr_sndcnt to get bigger.
17307                                  */
17308                                 long leftinsb;
17309
17310                                 leftinsb = sbavail(sb) - sb_offset;
17311                                 if (leftinsb > len) {
17312                                         /* This send does not empty the sb */
17313                                         len = 0;
17314                                 }
17315                         }
17316                 }
17317         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17318                 /*
17319                  * If you have not established
17320                  * and are not doing FAST OPEN
17321                  * no data please.
17322                  */
17323                 if ((sack_rxmit == 0) &&
17324                     (!IS_FASTOPEN(tp->t_flags))){
17325                         len = 0;
17326                         sb_offset = 0;
17327                 }
17328         }
17329         if (prefetch_so_done == 0) {
17330                 kern_prefetch(so, &prefetch_so_done);
17331                 prefetch_so_done = 1;
17332         }
17333         /*
17334          * Lop off SYN bit if it has already been sent.  However, if this is
17335          * SYN-SENT state and if segment contains data and if we don't know
17336          * that foreign host supports TAO, suppress sending segment.
17337          */
17338         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17339             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17340                 /*
17341                  * When sending additional segments following a TFO SYN|ACK,
17342                  * do not include the SYN bit.
17343                  */
17344                 if (IS_FASTOPEN(tp->t_flags) &&
17345                     (tp->t_state == TCPS_SYN_RECEIVED))
17346                         flags &= ~TH_SYN;
17347         }
17348         /*
17349          * Be careful not to send data and/or FIN on SYN segments. This
17350          * measure is needed to prevent interoperability problems with not
17351          * fully conformant TCP implementations.
17352          */
17353         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17354                 len = 0;
17355                 flags &= ~TH_FIN;
17356         }
17357         /*
17358          * On TFO sockets, ensure no data is sent in the following cases:
17359          *
17360          *  - When retransmitting SYN|ACK on a passively-created socket
17361          *
17362          *  - When retransmitting SYN on an actively created socket
17363          *
17364          *  - When sending a zero-length cookie (cookie request) on an
17365          *    actively created socket
17366          *
17367          *  - When the socket is in the CLOSED state (RST is being sent)
17368          */
17369         if (IS_FASTOPEN(tp->t_flags) &&
17370             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17371              ((tp->t_state == TCPS_SYN_SENT) &&
17372               (tp->t_tfo_client_cookie_len == 0)) ||
17373              (flags & TH_RST))) {
17374                 sack_rxmit = 0;
17375                 len = 0;
17376         }
17377         /* Without fast-open there should never be data sent on a SYN */
17378         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17379                 tp->snd_nxt = tp->iss;
17380                 len = 0;
17381         }
17382         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17383                 /* We only send 1 MSS if we have a DSACK block */
17384                 add_flag |= RACK_SENT_W_DSACK;
17385                 len = segsiz;
17386         }
17387         orig_len = len;
17388         if (len <= 0) {
17389                 /*
17390                  * If FIN has been sent but not acked, but we haven't been
17391                  * called to retransmit, len will be < 0.  Otherwise, window
17392                  * shrank after we sent into it.  If window shrank to 0,
17393                  * cancel pending retransmit, pull snd_nxt back to (closed)
17394                  * window, and set the persist timer if it isn't already
17395                  * going.  If the window didn't close completely, just wait
17396                  * for an ACK.
17397                  *
17398                  * We also do a general check here to ensure that we will
17399                  * set the persist timer when we have data to send, but a
17400                  * 0-byte window. This makes sure the persist timer is set
17401                  * even if the packet hits one of the "goto send" lines
17402                  * below.
17403                  */
17404                 len = 0;
17405                 if ((tp->snd_wnd == 0) &&
17406                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17407                     (tp->snd_una == tp->snd_max) &&
17408                     (sb_offset < (int)sbavail(sb))) {
17409                         rack_enter_persist(tp, rack, cts);
17410                 }
17411         } else if ((rsm == NULL) &&
17412                    (doing_tlp == 0) &&
17413                    (len < pace_max_seg)) {
17414                 /*
17415                  * We are not sending a maximum sized segment for
17416                  * some reason. Should we not send anything (think
17417                  * sws or persists)?
17418                  */
17419                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17420                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17421                     (len < minseg) &&
17422                     (len < (int)(sbavail(sb) - sb_offset))) {
17423                         /*
17424                          * Here the rwnd is less than
17425                          * the minimum pacing size, this is not a retransmit,
17426                          * we are established and
17427                          * the send is not the last in the socket buffer
17428                          * we send nothing, and we may enter persists
17429                          * if nothing is outstanding.
17430                          */
17431                         len = 0;
17432                         if (tp->snd_max == tp->snd_una) {
17433                                 /*
17434                                  * Nothing out we can
17435                                  * go into persists.
17436                                  */
17437                                 rack_enter_persist(tp, rack, cts);
17438                         }
17439                      } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17440                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17441                            (len < (int)(sbavail(sb) - sb_offset)) &&
17442                            (len < minseg)) {
17443                         /*
17444                          * Here we are not retransmitting, and
17445                          * the cwnd is not so small that we could
17446                          * not send at least a min size (rxt timer
17447                          * not having gone off), We have 2 segments or
17448                          * more already in flight, its not the tail end
17449                          * of the socket buffer  and the cwnd is blocking
17450                          * us from sending out a minimum pacing segment size.
17451                          * Lets not send anything.
17452                          */
17453                         len = 0;
17454                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17455                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17456                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17457                            (len < (int)(sbavail(sb) - sb_offset)) &&
17458                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
17459                         /*
17460                          * Here we have a send window but we have
17461                          * filled it up and we can't send another pacing segment.
17462                          * We also have in flight more than 2 segments
17463                          * and we are not completing the sb i.e. we allow
17464                          * the last bytes of the sb to go out even if
17465                          * its not a full pacing segment.
17466                          */
17467                         len = 0;
17468                 } else if ((rack->r_ctl.crte != NULL) &&
17469                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17470                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17471                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17472                            (len < (int)(sbavail(sb) - sb_offset))) {
17473                         /*
17474                          * Here we are doing hardware pacing, this is not a TLP,
17475                          * we are not sending a pace max segment size, there is rwnd
17476                          * room to send at least N pace_max_seg, the cwnd is greater
17477                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
17478                          * more segments in flight and its not the tail of the socket buffer.
17479                          *
17480                          * We don't want to send instead we need to get more ack's in to
17481                          * allow us to send a full pacing segment. Normally, if we are pacing
17482                          * about the right speed, we should have finished our pacing
17483                          * send as most of the acks have come back if we are at the
17484                          * right rate. This is a bit fuzzy since return path delay
17485                          * can delay the acks, which is why we want to make sure we
17486                          * have cwnd space to have a bit more than a max pace segments in flight.
17487                          *
17488                          * If we have not gotten our acks back we are pacing at too high a
17489                          * rate delaying will not hurt and will bring our GP estimate down by
17490                          * injecting the delay. If we don't do this we will send
17491                          * 2 MSS out in response to the acks being clocked in which
17492                          * defeats the point of hw-pacing (i.e. to help us get
17493                          * larger TSO's out).
17494                          */
17495                         len = 0;
17496
17497                 }
17498
17499         }
17500         /* len will be >= 0 after this point. */
17501         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17502         rack_sndbuf_autoscale(rack);
17503         /*
17504          * Decide if we can use TCP Segmentation Offloading (if supported by
17505          * hardware).
17506          *
17507          * TSO may only be used if we are in a pure bulk sending state.  The
17508          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17509          * options prevent using TSO.  With TSO the TCP header is the same
17510          * (except for the sequence number) for all generated packets.  This
17511          * makes it impossible to transmit any options which vary per
17512          * generated segment or packet.
17513          *
17514          * IPv4 handling has a clear separation of ip options and ip header
17515          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17516          * the right thing below to provide length of just ip options and thus
17517          * checking for ipoptlen is enough to decide if ip options are present.
17518          */
17519         ipoptlen = 0;
17520 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17521         /*
17522          * Pre-calculate here as we save another lookup into the darknesses
17523          * of IPsec that way and can actually decide if TSO is ok.
17524          */
17525 #ifdef INET6
17526         if (isipv6 && IPSEC_ENABLED(ipv6))
17527                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
17528 #ifdef INET
17529         else
17530 #endif
17531 #endif                          /* INET6 */
17532 #ifdef INET
17533                 if (IPSEC_ENABLED(ipv4))
17534                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
17535 #endif                          /* INET */
17536 #endif
17537
17538 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17539         ipoptlen += ipsec_optlen;
17540 #endif
17541         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17542             (tp->t_port == 0) &&
17543             ((tp->t_flags & TF_SIGNATURE) == 0) &&
17544             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17545             ipoptlen == 0)
17546                 tso = 1;
17547         {
17548                 uint32_t outstanding __unused;
17549
17550                 outstanding = tp->snd_max - tp->snd_una;
17551                 if (tp->t_flags & TF_SENTFIN) {
17552                         /*
17553                          * If we sent a fin, snd_max is 1 higher than
17554                          * snd_una
17555                          */
17556                         outstanding--;
17557                 }
17558                 if (sack_rxmit) {
17559                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17560                                 flags &= ~TH_FIN;
17561                 } else {
17562                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17563                                    sbused(sb)))
17564                                 flags &= ~TH_FIN;
17565                 }
17566         }
17567         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17568             (long)TCP_MAXWIN << tp->rcv_scale);
17569
17570         /*
17571          * Sender silly window avoidance.   We transmit under the following
17572          * conditions when len is non-zero:
17573          *
17574          * - We have a full segment (or more with TSO) - This is the last
17575          * buffer in a write()/send() and we are either idle or running
17576          * NODELAY - we've timed out (e.g. persist timer) - we have more
17577          * then 1/2 the maximum send window's worth of data (receiver may be
17578          * limited the window size) - we need to retransmit
17579          */
17580         if (len) {
17581                 if (len >= segsiz) {
17582                         goto send;
17583                 }
17584                 /*
17585                  * NOTE! on localhost connections an 'ack' from the remote
17586                  * end may occur synchronously with the output and cause us
17587                  * to flush a buffer queued with moretocome.  XXX
17588                  *
17589                  */
17590                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
17591                     (idle || (tp->t_flags & TF_NODELAY)) &&
17592                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17593                     (tp->t_flags & TF_NOPUSH) == 0) {
17594                         pass = 2;
17595                         goto send;
17596                 }
17597                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
17598                         pass = 22;
17599                         goto send;
17600                 }
17601                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17602                         pass = 4;
17603                         goto send;
17604                 }
17605                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
17606                         pass = 5;
17607                         goto send;
17608                 }
17609                 if (sack_rxmit) {
17610                         pass = 6;
17611                         goto send;
17612                 }
17613                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17614                     (ctf_outstanding(tp) < (segsiz * 2))) {
17615                         /*
17616                          * We have less than two MSS outstanding (delayed ack)
17617                          * and our rwnd will not let us send a full sized
17618                          * MSS. Lets go ahead and let this small segment
17619                          * out because we want to try to have at least two
17620                          * packets inflight to not be caught by delayed ack.
17621                          */
17622                         pass = 12;
17623                         goto send;
17624                 }
17625         }
17626         /*
17627          * Sending of standalone window updates.
17628          *
17629          * Window updates are important when we close our window due to a
17630          * full socket buffer and are opening it again after the application
17631          * reads data from it.  Once the window has opened again and the
17632          * remote end starts to send again the ACK clock takes over and
17633          * provides the most current window information.
17634          *
17635          * We must avoid the silly window syndrome whereas every read from
17636          * the receive buffer, no matter how small, causes a window update
17637          * to be sent.  We also should avoid sending a flurry of window
17638          * updates when the socket buffer had queued a lot of data and the
17639          * application is doing small reads.
17640          *
17641          * Prevent a flurry of pointless window updates by only sending an
17642          * update when we can increase the advertized window by more than
17643          * 1/4th of the socket buffer capacity.  When the buffer is getting
17644          * full or is very small be more aggressive and send an update
17645          * whenever we can increase by two mss sized segments. In all other
17646          * situations the ACK's to new incoming data will carry further
17647          * window increases.
17648          *
17649          * Don't send an independent window update if a delayed ACK is
17650          * pending (it will get piggy-backed on it) or the remote side
17651          * already has done a half-close and won't send more data.  Skip
17652          * this if the connection is in T/TCP half-open state.
17653          */
17654         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17655             !(tp->t_flags & TF_DELACK) &&
17656             !TCPS_HAVERCVDFIN(tp->t_state)) {
17657                 /*
17658                  * "adv" is the amount we could increase the window, taking
17659                  * into account that we are limited by TCP_MAXWIN <<
17660                  * tp->rcv_scale.
17661                  */
17662                 int32_t adv;
17663                 int oldwin;
17664
17665                 adv = recwin;
17666                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17667                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
17668                         if (adv > oldwin)
17669                             adv -= oldwin;
17670                         else {
17671                                 /* We can't increase the window */
17672                                 adv = 0;
17673                         }
17674                 } else
17675                         oldwin = 0;
17676
17677                 /*
17678                  * If the new window size ends up being the same as or less
17679                  * than the old size when it is scaled, then don't force
17680                  * a window update.
17681                  */
17682                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17683                         goto dontupdate;
17684
17685                 if (adv >= (int32_t)(2 * segsiz) &&
17686                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17687                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17688                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17689                         pass = 7;
17690                         goto send;
17691                 }
17692                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17693                         pass = 23;
17694                         goto send;
17695                 }
17696         }
17697 dontupdate:
17698
17699         /*
17700          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17701          * is also a catch-all for the retransmit timer timeout case.
17702          */
17703         if (tp->t_flags & TF_ACKNOW) {
17704                 pass = 8;
17705                 goto send;
17706         }
17707         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17708                 pass = 9;
17709                 goto send;
17710         }
17711         /*
17712          * If our state indicates that FIN should be sent and we have not
17713          * yet done so, then we need to send.
17714          */
17715         if ((flags & TH_FIN) &&
17716             (tp->snd_nxt == tp->snd_una)) {
17717                 pass = 11;
17718                 goto send;
17719         }
17720         /*
17721          * No reason to send a segment, just return.
17722          */
17723 just_return:
17724         SOCKBUF_UNLOCK(sb);
17725 just_return_nolock:
17726         {
17727                 int app_limited = CTF_JR_SENT_DATA;
17728
17729                 if (tot_len_this_send > 0) {
17730                         /* Make sure snd_nxt is up to max */
17731                         rack->r_ctl.fsb.recwin = recwin;
17732                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17733                         if ((error == 0) &&
17734                             rack_use_rfo &&
17735                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
17736                             (ipoptlen == 0) &&
17737                             (tp->snd_nxt == tp->snd_max) &&
17738                             (tp->rcv_numsacks == 0) &&
17739                             rack->r_fsb_inited &&
17740                             TCPS_HAVEESTABLISHED(tp->t_state) &&
17741                             (rack->r_must_retran == 0) &&
17742                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
17743                             (len > 0) && (orig_len > 0) &&
17744                             (orig_len > len) &&
17745                             ((orig_len - len) >= segsiz) &&
17746                             ((optlen == 0) ||
17747                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17748                                 /* We can send at least one more MSS using our fsb */
17749
17750                                 rack->r_fast_output = 1;
17751                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17752                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17753                                 rack->r_ctl.fsb.tcp_flags = flags;
17754                                 rack->r_ctl.fsb.left_to_send = orig_len - len;
17755                                 if (hw_tls)
17756                                         rack->r_ctl.fsb.hw_tls = 1;
17757                                 else
17758                                         rack->r_ctl.fsb.hw_tls = 0;
17759                                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17760                                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
17761                                         rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17762                                          (tp->snd_max - tp->snd_una)));
17763                                 if (rack->r_ctl.fsb.left_to_send < segsiz)
17764                                         rack->r_fast_output = 0;
17765                                 else {
17766                                         if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17767                                                 rack->r_ctl.fsb.rfo_apply_push = 1;
17768                                         else
17769                                                 rack->r_ctl.fsb.rfo_apply_push = 0;
17770                                 }
17771                         } else
17772                                 rack->r_fast_output = 0;
17773
17774
17775                         rack_log_fsb(rack, tp, so, flags,
17776                                      ipoptlen, orig_len, len, 0,
17777                                      1, optlen, __LINE__, 1);
17778                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17779                                 tp->snd_nxt = tp->snd_max;
17780                 } else {
17781                         int end_window = 0;
17782                         uint32_t seq = tp->gput_ack;
17783
17784                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17785                         if (rsm) {
17786                                 /*
17787                                  * Mark the last sent that we just-returned (hinting
17788                                  * that delayed ack may play a role in any rtt measurement).
17789                                  */
17790                                 rsm->r_just_ret = 1;
17791                         }
17792                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17793                         rack->r_ctl.rc_agg_delayed = 0;
17794                         rack->r_early = 0;
17795                         rack->r_late = 0;
17796                         rack->r_ctl.rc_agg_early = 0;
17797                         if ((ctf_outstanding(tp) +
17798                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17799                                  minseg)) >= tp->snd_wnd) {
17800                                 /* We are limited by the rwnd */
17801                                 app_limited = CTF_JR_RWND_LIMITED;
17802                                 if (IN_FASTRECOVERY(tp->t_flags))
17803                                     rack->r_ctl.rc_prr_sndcnt = 0;
17804                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17805                                 /* We are limited by whats available -- app limited */
17806                                 app_limited = CTF_JR_APP_LIMITED;
17807                                 if (IN_FASTRECOVERY(tp->t_flags))
17808                                     rack->r_ctl.rc_prr_sndcnt = 0;
17809                         } else if ((idle == 0) &&
17810                                    ((tp->t_flags & TF_NODELAY) == 0) &&
17811                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17812                                    (len < segsiz)) {
17813                                 /*
17814                                  * No delay is not on and the
17815                                  * user is sending less than 1MSS. This
17816                                  * brings out SWS avoidance so we
17817                                  * don't send. Another app-limited case.
17818                                  */
17819                                 app_limited = CTF_JR_APP_LIMITED;
17820                         } else if (tp->t_flags & TF_NOPUSH) {
17821                                 /*
17822                                  * The user has requested no push of
17823                                  * the last segment and we are
17824                                  * at the last segment. Another app
17825                                  * limited case.
17826                                  */
17827                                 app_limited = CTF_JR_APP_LIMITED;
17828                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17829                                 /* Its the cwnd */
17830                                 app_limited = CTF_JR_CWND_LIMITED;
17831                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
17832                                    (rack->rack_no_prr == 0) &&
17833                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17834                                 app_limited = CTF_JR_PRR;
17835                         } else {
17836                                 /* Now why here are we not sending? */
17837 #ifdef NOW
17838 #ifdef INVARIANTS
17839                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17840 #endif
17841 #endif
17842                                 app_limited = CTF_JR_ASSESSING;
17843                         }
17844                         /*
17845                          * App limited in some fashion, for our pacing GP
17846                          * measurements we don't want any gap (even cwnd).
17847                          * Close  down the measurement window.
17848                          */
17849                         if (rack_cwnd_block_ends_measure &&
17850                             ((app_limited == CTF_JR_CWND_LIMITED) ||
17851                              (app_limited == CTF_JR_PRR))) {
17852                                 /*
17853                                  * The reason we are not sending is
17854                                  * the cwnd (or prr). We have been configured
17855                                  * to end the measurement window in
17856                                  * this case.
17857                                  */
17858                                 end_window = 1;
17859                         } else if (rack_rwnd_block_ends_measure &&
17860                                    (app_limited == CTF_JR_RWND_LIMITED)) {
17861                                 /*
17862                                  * We are rwnd limited and have been
17863                                  * configured to end the measurement
17864                                  * window in this case.
17865                                  */
17866                                 end_window = 1;
17867                         } else if (app_limited == CTF_JR_APP_LIMITED) {
17868                                 /*
17869                                  * A true application limited period, we have
17870                                  * ran out of data.
17871                                  */
17872                                 end_window = 1;
17873                         } else if (app_limited == CTF_JR_ASSESSING) {
17874                                 /*
17875                                  * In the assessing case we hit the end of
17876                                  * the if/else and had no known reason
17877                                  * This will panic us under invariants..
17878                                  *
17879                                  * If we get this out in logs we need to
17880                                  * investagate which reason we missed.
17881                                  */
17882                                 end_window = 1;
17883                         }
17884                         if (end_window) {
17885                                 uint8_t log = 0;
17886
17887                                 /* Adjust the Gput measurement */
17888                                 if ((tp->t_flags & TF_GPUTINPROG) &&
17889                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
17890                                         tp->gput_ack = tp->snd_max;
17891                                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17892                                                 /*
17893                                                  * There is not enough to measure.
17894                                                  */
17895                                                 tp->t_flags &= ~TF_GPUTINPROG;
17896                                                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17897                                                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
17898                                                                            tp->gput_seq,
17899                                                                            0, 0, 18, __LINE__, NULL, 0);
17900                                         } else
17901                                                 log = 1;
17902                                 }
17903                                 /* Mark the last packet has app limited */
17904                                 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17905                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17906                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
17907                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17908                                         else {
17909                                                 /*
17910                                                  * Go out to the end app limited and mark
17911                                                  * this new one as next and move the end_appl up
17912                                                  * to this guy.
17913                                                  */
17914                                                 if (rack->r_ctl.rc_end_appl)
17915                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17916                                                 rack->r_ctl.rc_end_appl = rsm;
17917                                         }
17918                                         rsm->r_flags |= RACK_APP_LIMITED;
17919                                         rack->r_ctl.rc_app_limited_cnt++;
17920                                 }
17921                                 if (log)
17922                                         rack_log_pacing_delay_calc(rack,
17923                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
17924                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17925                         }
17926                 }
17927                 /* Check if we need to go into persists or not */
17928                 if ((tp->snd_max == tp->snd_una) &&
17929                     TCPS_HAVEESTABLISHED(tp->t_state) &&
17930                     sbavail(sb) &&
17931                     (sbavail(sb) > tp->snd_wnd) &&
17932                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17933                         /* Yes lets make sure to move to persist before timer-start */
17934                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17935                 }
17936                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17937                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17938         }
17939 #ifdef NETFLIX_SHARED_CWND
17940         if ((sbavail(sb) == 0) &&
17941             rack->r_ctl.rc_scw) {
17942                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17943                 rack->rack_scwnd_is_idle = 1;
17944         }
17945 #endif
17946 #ifdef TCP_ACCOUNTING
17947         if (tot_len_this_send > 0) {
17948                 crtsc = get_cyclecount();
17949                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17950                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
17951                 }
17952                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17953                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17954                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17955                 }
17956                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17957                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17958                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17959                 }
17960                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17961         } else {
17962                 crtsc = get_cyclecount();
17963                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17964                         tp->tcp_cnt_counters[SND_LIMITED]++;
17965                 }
17966                 counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17967                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17968                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17969                 }
17970                 counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17971         }
17972         sched_unpin();
17973 #endif
17974         return (0);
17975
17976 send:
17977         if (rsm || sack_rxmit)
17978                 counter_u64_add(rack_nfto_resend, 1);
17979         else
17980                 counter_u64_add(rack_non_fto_send, 1);
17981         if ((flags & TH_FIN) &&
17982             sbavail(sb)) {
17983                 /*
17984                  * We do not transmit a FIN
17985                  * with data outstanding. We
17986                  * need to make it so all data
17987                  * is acked first.
17988                  */
17989                 flags &= ~TH_FIN;
17990         }
17991         /* Enforce stack imposed max seg size if we have one */
17992         if (rack->r_ctl.rc_pace_max_segs &&
17993             (len > rack->r_ctl.rc_pace_max_segs)) {
17994                 mark = 1;
17995                 len = rack->r_ctl.rc_pace_max_segs;
17996         }
17997         SOCKBUF_LOCK_ASSERT(sb);
17998         if (len > 0) {
17999                 if (len >= segsiz)
18000                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
18001                 else
18002                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
18003         }
18004         /*
18005          * Before ESTABLISHED, force sending of initial options unless TCP
18006          * set not to do any options. NOTE: we assume that the IP/TCP header
18007          * plus TCP options always fit in a single mbuf, leaving room for a
18008          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
18009          * + optlen <= MCLBYTES
18010          */
18011         optlen = 0;
18012 #ifdef INET6
18013         if (isipv6)
18014                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18015         else
18016 #endif
18017                 hdrlen = sizeof(struct tcpiphdr);
18018
18019         /*
18020          * Compute options for segment. We only have to care about SYN and
18021          * established connection segments.  Options for SYN-ACK segments
18022          * are handled in TCP syncache.
18023          */
18024         to.to_flags = 0;
18025         if ((tp->t_flags & TF_NOOPT) == 0) {
18026                 /* Maximum segment size. */
18027                 if (flags & TH_SYN) {
18028                         tp->snd_nxt = tp->iss;
18029                         to.to_mss = tcp_mssopt(&inp->inp_inc);
18030                         if (tp->t_port)
18031                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
18032                         to.to_flags |= TOF_MSS;
18033
18034                         /*
18035                          * On SYN or SYN|ACK transmits on TFO connections,
18036                          * only include the TFO option if it is not a
18037                          * retransmit, as the presence of the TFO option may
18038                          * have caused the original SYN or SYN|ACK to have
18039                          * been dropped by a middlebox.
18040                          */
18041                         if (IS_FASTOPEN(tp->t_flags) &&
18042                             (tp->t_rxtshift == 0)) {
18043                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
18044                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
18045                                         to.to_tfo_cookie =
18046                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
18047                                         to.to_flags |= TOF_FASTOPEN;
18048                                         wanted_cookie = 1;
18049                                 } else if (tp->t_state == TCPS_SYN_SENT) {
18050                                         to.to_tfo_len =
18051                                                 tp->t_tfo_client_cookie_len;
18052                                         to.to_tfo_cookie =
18053                                                 tp->t_tfo_cookie.client;
18054                                         to.to_flags |= TOF_FASTOPEN;
18055                                         wanted_cookie = 1;
18056                                         /*
18057                                          * If we wind up having more data to
18058                                          * send with the SYN than can fit in
18059                                          * one segment, don't send any more
18060                                          * until the SYN|ACK comes back from
18061                                          * the other end.
18062                                          */
18063                                         sendalot = 0;
18064                                 }
18065                         }
18066                 }
18067                 /* Window scaling. */
18068                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
18069                         to.to_wscale = tp->request_r_scale;
18070                         to.to_flags |= TOF_SCALE;
18071                 }
18072                 /* Timestamps. */
18073                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
18074                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
18075                         to.to_tsval = ms_cts + tp->ts_offset;
18076                         to.to_tsecr = tp->ts_recent;
18077                         to.to_flags |= TOF_TS;
18078                 }
18079                 /* Set receive buffer autosizing timestamp. */
18080                 if (tp->rfbuf_ts == 0 &&
18081                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
18082                         tp->rfbuf_ts = tcp_ts_getticks();
18083                 /* Selective ACK's. */
18084                 if (tp->t_flags & TF_SACK_PERMIT) {
18085                         if (flags & TH_SYN)
18086                                 to.to_flags |= TOF_SACKPERM;
18087                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18088                                  tp->rcv_numsacks > 0) {
18089                                 to.to_flags |= TOF_SACK;
18090                                 to.to_nsacks = tp->rcv_numsacks;
18091                                 to.to_sacks = (u_char *)tp->sackblks;
18092                         }
18093                 }
18094 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18095                 /* TCP-MD5 (RFC2385). */
18096                 if (tp->t_flags & TF_SIGNATURE)
18097                         to.to_flags |= TOF_SIGNATURE;
18098 #endif                          /* TCP_SIGNATURE */
18099
18100                 /* Processing the options. */
18101                 hdrlen += optlen = tcp_addoptions(&to, opt);
18102                 /*
18103                  * If we wanted a TFO option to be added, but it was unable
18104                  * to fit, ensure no data is sent.
18105                  */
18106                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
18107                     !(to.to_flags & TOF_FASTOPEN))
18108                         len = 0;
18109         }
18110         if (tp->t_port) {
18111                 if (V_tcp_udp_tunneling_port == 0) {
18112                         /* The port was removed?? */
18113                         SOCKBUF_UNLOCK(&so->so_snd);
18114 #ifdef TCP_ACCOUNTING
18115                         crtsc = get_cyclecount();
18116                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18117                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18118                         }
18119                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18120                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18121                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18122                         }
18123                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18124                         sched_unpin();
18125 #endif
18126                         return (EHOSTUNREACH);
18127                 }
18128                 hdrlen += sizeof(struct udphdr);
18129         }
18130 #ifdef INET6
18131         if (isipv6)
18132                 ipoptlen = ip6_optlen(inp);
18133         else
18134 #endif
18135                 if (inp->inp_options)
18136                         ipoptlen = inp->inp_options->m_len -
18137                                 offsetof(struct ipoption, ipopt_list);
18138                 else
18139                         ipoptlen = 0;
18140 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18141         ipoptlen += ipsec_optlen;
18142 #endif
18143
18144         /*
18145          * Adjust data length if insertion of options will bump the packet
18146          * length beyond the t_maxseg length. Clear the FIN bit because we
18147          * cut off the tail of the segment.
18148          */
18149         if (len + optlen + ipoptlen > tp->t_maxseg) {
18150                 if (tso) {
18151                         uint32_t if_hw_tsomax;
18152                         uint32_t moff;
18153                         int32_t max_len;
18154
18155                         /* extract TSO information */
18156                         if_hw_tsomax = tp->t_tsomax;
18157                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18158                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18159                         KASSERT(ipoptlen == 0,
18160                                 ("%s: TSO can't do IP options", __func__));
18161
18162                         /*
18163                          * Check if we should limit by maximum payload
18164                          * length:
18165                          */
18166                         if (if_hw_tsomax != 0) {
18167                                 /* compute maximum TSO length */
18168                                 max_len = (if_hw_tsomax - hdrlen -
18169                                            max_linkhdr);
18170                                 if (max_len <= 0) {
18171                                         len = 0;
18172                                 } else if (len > max_len) {
18173                                         sendalot = 1;
18174                                         len = max_len;
18175                                         mark = 2;
18176                                 }
18177                         }
18178                         /*
18179                          * Prevent the last segment from being fractional
18180                          * unless the send sockbuf can be emptied:
18181                          */
18182                         max_len = (tp->t_maxseg - optlen);
18183                         if ((sb_offset + len) < sbavail(sb)) {
18184                                 moff = len % (u_int)max_len;
18185                                 if (moff != 0) {
18186                                         mark = 3;
18187                                         len -= moff;
18188                                 }
18189                         }
18190                         /*
18191                          * In case there are too many small fragments don't
18192                          * use TSO:
18193                          */
18194                         if (len <= segsiz) {
18195                                 mark = 4;
18196                                 tso = 0;
18197                         }
18198                         /*
18199                          * Send the FIN in a separate segment after the bulk
18200                          * sending is done. We don't trust the TSO
18201                          * implementations to clear the FIN flag on all but
18202                          * the last segment.
18203                          */
18204                         if (tp->t_flags & TF_NEEDFIN) {
18205                                 sendalot = 4;
18206                         }
18207                 } else {
18208                         mark = 5;
18209                         if (optlen + ipoptlen >= tp->t_maxseg) {
18210                                 /*
18211                                  * Since we don't have enough space to put
18212                                  * the IP header chain and the TCP header in
18213                                  * one packet as required by RFC 7112, don't
18214                                  * send it. Also ensure that at least one
18215                                  * byte of the payload can be put into the
18216                                  * TCP segment.
18217                                  */
18218                                 SOCKBUF_UNLOCK(&so->so_snd);
18219                                 error = EMSGSIZE;
18220                                 sack_rxmit = 0;
18221                                 goto out;
18222                         }
18223                         len = tp->t_maxseg - optlen - ipoptlen;
18224                         sendalot = 5;
18225                 }
18226         } else {
18227                 tso = 0;
18228                 mark = 6;
18229         }
18230         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18231                 ("%s: len > IP_MAXPACKET", __func__));
18232 #ifdef DIAGNOSTIC
18233 #ifdef INET6
18234         if (max_linkhdr + hdrlen > MCLBYTES)
18235 #else
18236                 if (max_linkhdr + hdrlen > MHLEN)
18237 #endif
18238                         panic("tcphdr too big");
18239 #endif
18240
18241         /*
18242          * This KASSERT is here to catch edge cases at a well defined place.
18243          * Before, those had triggered (random) panic conditions further
18244          * down.
18245          */
18246         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18247         if ((len == 0) &&
18248             (flags & TH_FIN) &&
18249             (sbused(sb))) {
18250                 /*
18251                  * We have outstanding data, don't send a fin by itself!.
18252                  */
18253                 goto just_return;
18254         }
18255         /*
18256          * Grab a header mbuf, attaching a copy of data to be transmitted,
18257          * and initialize the header from the template for sends on this
18258          * connection.
18259          */
18260         hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
18261         if (len) {
18262                 uint32_t max_val;
18263                 uint32_t moff;
18264
18265                 if (rack->r_ctl.rc_pace_max_segs)
18266                         max_val = rack->r_ctl.rc_pace_max_segs;
18267                 else if (rack->rc_user_set_max_segs)
18268                         max_val = rack->rc_user_set_max_segs * segsiz;
18269                 else
18270                         max_val = len;
18271                 /*
18272                  * We allow a limit on sending with hptsi.
18273                  */
18274                 if (len > max_val) {
18275                         mark = 7;
18276                         len = max_val;
18277                 }
18278 #ifdef INET6
18279                 if (MHLEN < hdrlen + max_linkhdr)
18280                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18281                 else
18282 #endif
18283                         m = m_gethdr(M_NOWAIT, MT_DATA);
18284
18285                 if (m == NULL) {
18286                         SOCKBUF_UNLOCK(sb);
18287                         error = ENOBUFS;
18288                         sack_rxmit = 0;
18289                         goto out;
18290                 }
18291                 m->m_data += max_linkhdr;
18292                 m->m_len = hdrlen;
18293
18294                 /*
18295                  * Start the m_copy functions from the closest mbuf to the
18296                  * sb_offset in the socket buffer chain.
18297                  */
18298                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
18299                 s_mb = mb;
18300                 s_moff = moff;
18301                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18302                         m_copydata(mb, moff, (int)len,
18303                                    mtod(m, caddr_t)+hdrlen);
18304                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18305                                 sbsndptr_adv(sb, mb, len);
18306                         m->m_len += len;
18307                 } else {
18308                         struct sockbuf *msb;
18309
18310                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18311                                 msb = NULL;
18312                         else
18313                                 msb = sb;
18314                         m->m_next = tcp_m_copym(
18315                                 mb, moff, &len,
18316                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18317                                 ((rsm == NULL) ? hw_tls : 0)
18318 #ifdef NETFLIX_COPY_ARGS
18319                                 , &s_mb, &s_moff
18320 #endif
18321                                 );
18322                         if (len <= (tp->t_maxseg - optlen)) {
18323                                 /*
18324                                  * Must have ran out of mbufs for the copy
18325                                  * shorten it to no longer need tso. Lets
18326                                  * not put on sendalot since we are low on
18327                                  * mbufs.
18328                                  */
18329                                 tso = 0;
18330                         }
18331                         if (m->m_next == NULL) {
18332                                 SOCKBUF_UNLOCK(sb);
18333                                 (void)m_free(m);
18334                                 error = ENOBUFS;
18335                                 sack_rxmit = 0;
18336                                 goto out;
18337                         }
18338                 }
18339                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18340                         if (rsm && (rsm->r_flags & RACK_TLP)) {
18341                                 /*
18342                                  * TLP should not count in retran count, but
18343                                  * in its own bin
18344                                  */
18345                                 counter_u64_add(rack_tlp_retran, 1);
18346                                 counter_u64_add(rack_tlp_retran_bytes, len);
18347                         } else {
18348                                 tp->t_sndrexmitpack++;
18349                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18350                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18351                         }
18352 #ifdef STATS
18353                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18354                                                  len);
18355 #endif
18356                 } else {
18357                         KMOD_TCPSTAT_INC(tcps_sndpack);
18358                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18359 #ifdef STATS
18360                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18361                                                  len);
18362 #endif
18363                 }
18364                 /*
18365                  * If we're sending everything we've got, set PUSH. (This
18366                  * will keep happy those implementations which only give
18367                  * data to the user when a buffer fills or a PUSH comes in.)
18368                  */
18369                 if (sb_offset + len == sbused(sb) &&
18370                     sbused(sb) &&
18371                     !(flags & TH_SYN)) {
18372                         flags |= TH_PUSH;
18373                         add_flag |= RACK_HAD_PUSH;
18374                 }
18375
18376                 SOCKBUF_UNLOCK(sb);
18377         } else {
18378                 SOCKBUF_UNLOCK(sb);
18379                 if (tp->t_flags & TF_ACKNOW)
18380                         KMOD_TCPSTAT_INC(tcps_sndacks);
18381                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
18382                         KMOD_TCPSTAT_INC(tcps_sndctrl);
18383                 else
18384                         KMOD_TCPSTAT_INC(tcps_sndwinup);
18385
18386                 m = m_gethdr(M_NOWAIT, MT_DATA);
18387                 if (m == NULL) {
18388                         error = ENOBUFS;
18389                         sack_rxmit = 0;
18390                         goto out;
18391                 }
18392 #ifdef INET6
18393                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18394                     MHLEN >= hdrlen) {
18395                         M_ALIGN(m, hdrlen);
18396                 } else
18397 #endif
18398                         m->m_data += max_linkhdr;
18399                 m->m_len = hdrlen;
18400         }
18401         SOCKBUF_UNLOCK_ASSERT(sb);
18402         m->m_pkthdr.rcvif = (struct ifnet *)0;
18403 #ifdef MAC
18404         mac_inpcb_create_mbuf(inp, m);
18405 #endif
18406         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18407 #ifdef INET6
18408                 if (isipv6)
18409                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18410                 else
18411 #endif                          /* INET6 */
18412                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18413                 th = rack->r_ctl.fsb.th;
18414                 udp = rack->r_ctl.fsb.udp;
18415                 if (udp) {
18416 #ifdef INET6
18417                         if (isipv6)
18418                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
18419                         else
18420 #endif                          /* INET6 */
18421                                 ulen = hdrlen + len - sizeof(struct ip);
18422                         udp->uh_ulen = htons(ulen);
18423                 }
18424         } else {
18425 #ifdef INET6
18426                 if (isipv6) {
18427                         ip6 = mtod(m, struct ip6_hdr *);
18428                         if (tp->t_port) {
18429                                 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18430                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18431                                 udp->uh_dport = tp->t_port;
18432                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
18433                                 udp->uh_ulen = htons(ulen);
18434                                 th = (struct tcphdr *)(udp + 1);
18435                         } else
18436                                 th = (struct tcphdr *)(ip6 + 1);
18437                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
18438                 } else
18439 #endif                          /* INET6 */
18440                 {
18441                         ip = mtod(m, struct ip *);
18442                         if (tp->t_port) {
18443                                 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18444                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18445                                 udp->uh_dport = tp->t_port;
18446                                 ulen = hdrlen + len - sizeof(struct ip);
18447                                 udp->uh_ulen = htons(ulen);
18448                                 th = (struct tcphdr *)(udp + 1);
18449                         } else
18450                                 th = (struct tcphdr *)(ip + 1);
18451                         tcpip_fillheaders(inp, tp->t_port, ip, th);
18452                 }
18453         }
18454         /*
18455          * Fill in fields, remembering maximum advertised window for use in
18456          * delaying messages about window sizes. If resending a FIN, be sure
18457          * not to use a new sequence number.
18458          */
18459         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18460             tp->snd_nxt == tp->snd_max)
18461                 tp->snd_nxt--;
18462         /*
18463          * If we are starting a connection, send ECN setup SYN packet. If we
18464          * are on a retransmit, we may resend those bits a number of times
18465          * as per RFC 3168.
18466          */
18467         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
18468                 flags |= tcp_ecn_output_syn_sent(tp);
18469         }
18470         /* Also handle parallel SYN for ECN */
18471         if (TCPS_HAVERCVDSYN(tp->t_state) &&
18472             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18473                 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
18474                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18475                     (tp->t_flags2 & TF2_ECN_SND_ECE))
18476                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18477 #ifdef INET6
18478                 if (isipv6) {
18479                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18480                         ip6->ip6_flow |= htonl(ect << 20);
18481                 }
18482                 else
18483 #endif
18484                 {
18485                         ip->ip_tos &= ~IPTOS_ECN_MASK;
18486                         ip->ip_tos |= ect;
18487                 }
18488         }
18489         /*
18490          * If we are doing retransmissions, then snd_nxt will not reflect
18491          * the first unsent octet.  For ACK only packets, we do not want the
18492          * sequence number of the retransmitted packet, we want the sequence
18493          * number of the next unsent octet.  So, if there is no data (and no
18494          * SYN or FIN), use snd_max instead of snd_nxt when filling in
18495          * ti_seq.  But if we are in persist state, snd_max might reflect
18496          * one byte beyond the right edge of the window, so use snd_nxt in
18497          * that case, since we know we aren't doing a retransmission.
18498          * (retransmit and persist are mutually exclusive...)
18499          */
18500         if (sack_rxmit == 0) {
18501                 if (len || (flags & (TH_SYN | TH_FIN))) {
18502                         th->th_seq = htonl(tp->snd_nxt);
18503                         rack_seq = tp->snd_nxt;
18504                 } else {
18505                         th->th_seq = htonl(tp->snd_max);
18506                         rack_seq = tp->snd_max;
18507                 }
18508         } else {
18509                 th->th_seq = htonl(rsm->r_start);
18510                 rack_seq = rsm->r_start;
18511         }
18512         th->th_ack = htonl(tp->rcv_nxt);
18513         tcp_set_flags(th, flags);
18514         /*
18515          * Calculate receive window.  Don't shrink window, but avoid silly
18516          * window syndrome.
18517          * If a RST segment is sent, advertise a window of zero.
18518          */
18519         if (flags & TH_RST) {
18520                 recwin = 0;
18521         } else {
18522                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18523                     recwin < (long)segsiz) {
18524                         recwin = 0;
18525                 }
18526                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18527                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18528                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18529         }
18530
18531         /*
18532          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18533          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18534          * handled in syncache.
18535          */
18536         if (flags & TH_SYN)
18537                 th->th_win = htons((u_short)
18538                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18539         else {
18540                 /* Avoid shrinking window with window scaling. */
18541                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
18542                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18543         }
18544         /*
18545          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18546          * window.  This may cause the remote transmitter to stall.  This
18547          * flag tells soreceive() to disable delayed acknowledgements when
18548          * draining the buffer.  This can occur if the receiver is
18549          * attempting to read more data than can be buffered prior to
18550          * transmitting on the connection.
18551          */
18552         if (th->th_win == 0) {
18553                 tp->t_sndzerowin++;
18554                 tp->t_flags |= TF_RXWIN0SENT;
18555         } else
18556                 tp->t_flags &= ~TF_RXWIN0SENT;
18557         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
18558         /* Now are we using fsb?, if so copy the template data to the mbuf */
18559         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18560                 uint8_t *cpto;
18561
18562                 cpto = mtod(m, uint8_t *);
18563                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18564                 /*
18565                  * We have just copied in:
18566                  * IP/IP6
18567                  * <optional udphdr>
18568                  * tcphdr (no options)
18569                  *
18570                  * We need to grab the correct pointers into the mbuf
18571                  * for both the tcp header, and possibly the udp header (if tunneling).
18572                  * We do this by using the offset in the copy buffer and adding it
18573                  * to the mbuf base pointer (cpto).
18574                  */
18575 #ifdef INET6
18576                 if (isipv6)
18577                         ip6 = mtod(m, struct ip6_hdr *);
18578                 else
18579 #endif                          /* INET6 */
18580                         ip = mtod(m, struct ip *);
18581                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18582                 /* If we have a udp header lets set it into the mbuf as well */
18583                 if (udp)
18584                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18585         }
18586 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18587         if (to.to_flags & TOF_SIGNATURE) {
18588                 /*
18589                  * Calculate MD5 signature and put it into the place
18590                  * determined before.
18591                  * NOTE: since TCP options buffer doesn't point into
18592                  * mbuf's data, calculate offset and use it.
18593                  */
18594                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18595                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18596                         /*
18597                          * Do not send segment if the calculation of MD5
18598                          * digest has failed.
18599                          */
18600                         goto out;
18601                 }
18602         }
18603 #endif
18604         if (optlen) {
18605                 bcopy(opt, th + 1, optlen);
18606                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18607         }
18608         /*
18609          * Put TCP length in extended header, and then checksum extended
18610          * header and data.
18611          */
18612         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
18613 #ifdef INET6
18614         if (isipv6) {
18615                 /*
18616                  * ip6_plen is not need to be filled now, and will be filled
18617                  * in ip6_output.
18618                  */
18619                 if (tp->t_port) {
18620                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18621                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18622                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18623                         th->th_sum = htons(0);
18624                         UDPSTAT_INC(udps_opackets);
18625                 } else {
18626                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18627                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18628                         th->th_sum = in6_cksum_pseudo(ip6,
18629                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18630                                                       0);
18631                 }
18632         }
18633 #endif
18634 #if defined(INET6) && defined(INET)
18635         else
18636 #endif
18637 #ifdef INET
18638         {
18639                 if (tp->t_port) {
18640                         m->m_pkthdr.csum_flags = CSUM_UDP;
18641                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18642                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18643                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18644                         th->th_sum = htons(0);
18645                         UDPSTAT_INC(udps_opackets);
18646                 } else {
18647                         m->m_pkthdr.csum_flags = CSUM_TCP;
18648                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18649                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
18650                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18651                                                                         IPPROTO_TCP + len + optlen));
18652                 }
18653                 /* IP version must be set here for ipv4/ipv6 checking later */
18654                 KASSERT(ip->ip_v == IPVERSION,
18655                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
18656         }
18657 #endif
18658         /*
18659          * Enable TSO and specify the size of the segments. The TCP pseudo
18660          * header checksum is always provided. XXX: Fixme: This is currently
18661          * not the case for IPv6.
18662          */
18663         if (tso) {
18664                 KASSERT(len > tp->t_maxseg - optlen,
18665                         ("%s: len <= tso_segsz", __func__));
18666                 m->m_pkthdr.csum_flags |= CSUM_TSO;
18667                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18668         }
18669         KASSERT(len + hdrlen == m_length(m, NULL),
18670                 ("%s: mbuf chain different than expected: %d + %u != %u",
18671                  __func__, len, hdrlen, m_length(m, NULL)));
18672
18673 #ifdef TCP_HHOOK
18674         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18675         hhook_run_tcp_est_out(tp, th, &to, len, tso);
18676 #endif
18677         /* We're getting ready to send; log now. */
18678         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18679                 union tcp_log_stackspecific log;
18680
18681                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18682                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18683                 if (rack->rack_no_prr)
18684                         log.u_bbr.flex1 = 0;
18685                 else
18686                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18687                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18688                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18689                 log.u_bbr.flex4 = orig_len;
18690                 /* Save off the early/late values */
18691                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18692                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18693                 log.u_bbr.bw_inuse = rack_get_bw(rack);
18694                 log.u_bbr.flex8 = 0;
18695                 if (rsm) {
18696                         if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18697                                 rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18698                                 counter_u64_add(rack_collapsed_win_rxt, 1);
18699                                 counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18700                         }
18701                         if (doing_tlp)
18702                                 log.u_bbr.flex8 = 2;
18703                         else
18704                                 log.u_bbr.flex8 = 1;
18705                 } else {
18706                         if (doing_tlp)
18707                                 log.u_bbr.flex8 = 3;
18708                         else
18709                                 log.u_bbr.flex8 = 0;
18710                 }
18711                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18712                 log.u_bbr.flex7 = mark;
18713                 log.u_bbr.flex7 <<= 8;
18714                 log.u_bbr.flex7 |= pass;
18715                 log.u_bbr.pkts_out = tp->t_maxseg;
18716                 log.u_bbr.timeStamp = cts;
18717                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18718                 log.u_bbr.lt_epoch = cwnd_to_use;
18719                 log.u_bbr.delivered = sendalot;
18720                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18721                                      len, &log, false, NULL, NULL, 0, &tv);
18722         } else
18723                 lgb = NULL;
18724
18725         /*
18726          * Fill in IP length and desired time to live and send to IP level.
18727          * There should be a better way to handle ttl and tos; we could keep
18728          * them in the template, but need a way to checksum without them.
18729          */
18730         /*
18731          * m->m_pkthdr.len should have been set before cksum calcuration,
18732          * because in6_cksum() need it.
18733          */
18734 #ifdef INET6
18735         if (isipv6) {
18736                 /*
18737                  * we separately set hoplimit for every segment, since the
18738                  * user might want to change the value via setsockopt. Also,
18739                  * desired default hop limit might be changed via Neighbor
18740                  * Discovery.
18741                  */
18742                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18743
18744                 /*
18745                  * Set the packet size here for the benefit of DTrace
18746                  * probes. ip6_output() will set it properly; it's supposed
18747                  * to include the option header lengths as well.
18748                  */
18749                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18750
18751                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18752                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18753                 else
18754                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18755
18756                 if (tp->t_state == TCPS_SYN_SENT)
18757                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18758
18759                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18760                 /* TODO: IPv6 IP6TOS_ECT bit on */
18761                 error = ip6_output(m,
18762 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18763                                    inp->in6p_outputopts,
18764 #else
18765                                    NULL,
18766 #endif
18767                                    &inp->inp_route6,
18768                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18769                                    NULL, NULL, inp);
18770
18771                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18772                         mtu = inp->inp_route6.ro_nh->nh_mtu;
18773         }
18774 #endif                          /* INET6 */
18775 #if defined(INET) && defined(INET6)
18776         else
18777 #endif
18778 #ifdef INET
18779         {
18780                 ip->ip_len = htons(m->m_pkthdr.len);
18781 #ifdef INET6
18782                 if (inp->inp_vflag & INP_IPV6PROTO)
18783                         ip->ip_ttl = in6_selecthlim(inp, NULL);
18784 #endif                          /* INET6 */
18785                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18786                 /*
18787                  * If we do path MTU discovery, then we set DF on every
18788                  * packet. This might not be the best thing to do according
18789                  * to RFC3390 Section 2. However the tcp hostcache migitates
18790                  * the problem so it affects only the first tcp connection
18791                  * with a host.
18792                  *
18793                  * NB: Don't set DF on small MTU/MSS to have a safe
18794                  * fallback.
18795                  */
18796                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18797                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18798                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18799                                 ip->ip_off |= htons(IP_DF);
18800                         }
18801                 } else {
18802                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18803                 }
18804
18805                 if (tp->t_state == TCPS_SYN_SENT)
18806                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18807
18808                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18809
18810                 error = ip_output(m,
18811 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18812                                   inp->inp_options,
18813 #else
18814                                   NULL,
18815 #endif
18816                                   &inp->inp_route,
18817                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18818                                   inp);
18819                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18820                         mtu = inp->inp_route.ro_nh->nh_mtu;
18821         }
18822 #endif                          /* INET */
18823
18824 out:
18825         if (lgb) {
18826                 lgb->tlb_errno = error;
18827                 lgb = NULL;
18828         }
18829         /*
18830          * In transmit state, time the transmission and arrange for the
18831          * retransmit.  In persist state, just set snd_max.
18832          */
18833         if (error == 0) {
18834                 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18835                 if (rsm && doing_tlp) {
18836                         rack->rc_last_sent_tlp_past_cumack = 0;
18837                         rack->rc_last_sent_tlp_seq_valid = 1;
18838                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18839                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18840                 }
18841                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
18842                 if (rsm && (doing_tlp == 0)) {
18843                         /* Set we retransmitted */
18844                         rack->rc_gp_saw_rec = 1;
18845                 } else {
18846                         if (cwnd_to_use > tp->snd_ssthresh) {
18847                                 /* Set we sent in CA */
18848                                 rack->rc_gp_saw_ca = 1;
18849                         } else {
18850                                 /* Set we sent in SS */
18851                                 rack->rc_gp_saw_ss = 1;
18852                         }
18853                 }
18854                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18855                     (tp->t_flags & TF_SACK_PERMIT) &&
18856                     tp->rcv_numsacks > 0)
18857                         tcp_clean_dsack_blocks(tp);
18858                 tot_len_this_send += len;
18859                 if (len == 0)
18860                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18861                 else if (len == 1) {
18862                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18863                 } else if (len > 1) {
18864                         int idx;
18865
18866                         idx = (len / segsiz) + 3;
18867                         if (idx >= TCP_MSS_ACCT_ATIMER)
18868                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18869                         else
18870                                 counter_u64_add(rack_out_size[idx], 1);
18871                 }
18872         }
18873         if ((rack->rack_no_prr == 0) &&
18874             sub_from_prr &&
18875             (error == 0)) {
18876                 if (rack->r_ctl.rc_prr_sndcnt >= len)
18877                         rack->r_ctl.rc_prr_sndcnt -= len;
18878                 else
18879                         rack->r_ctl.rc_prr_sndcnt = 0;
18880         }
18881         sub_from_prr = 0;
18882         if (doing_tlp) {
18883                 /* Make sure the TLP is added */
18884                 add_flag |= RACK_TLP;
18885         } else if (rsm) {
18886                 /* If its a resend without TLP then it must not have the flag */
18887                 rsm->r_flags &= ~RACK_TLP;
18888         }
18889         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18890                         rack_to_usec_ts(&tv),
18891                         rsm, add_flag, s_mb, s_moff, hw_tls);
18892
18893
18894         if ((error == 0) &&
18895             (len > 0) &&
18896             (tp->snd_una == tp->snd_max))
18897                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18898         {
18899                 tcp_seq startseq = tp->snd_nxt;
18900
18901                 /* Track our lost count */
18902                 if (rsm && (doing_tlp == 0))
18903                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18904                 /*
18905                  * Advance snd_nxt over sequence space of this segment.
18906                  */
18907                 if (error)
18908                         /* We don't log or do anything with errors */
18909                         goto nomore;
18910                 if (doing_tlp == 0) {
18911                         if (rsm == NULL) {
18912                                 /*
18913                                  * Not a retransmission of some
18914                                  * sort, new data is going out so
18915                                  * clear our TLP count and flag.
18916                                  */
18917                                 rack->rc_tlp_in_progress = 0;
18918                                 rack->r_ctl.rc_tlp_cnt_out = 0;
18919                         }
18920                 } else {
18921                         /*
18922                          * We have just sent a TLP, mark that it is true
18923                          * and make sure our in progress is set so we
18924                          * continue to check the count.
18925                          */
18926                         rack->rc_tlp_in_progress = 1;
18927                         rack->r_ctl.rc_tlp_cnt_out++;
18928                 }
18929                 if (flags & (TH_SYN | TH_FIN)) {
18930                         if (flags & TH_SYN)
18931                                 tp->snd_nxt++;
18932                         if (flags & TH_FIN) {
18933                                 tp->snd_nxt++;
18934                                 tp->t_flags |= TF_SENTFIN;
18935                         }
18936                 }
18937                 /* In the ENOBUFS case we do *not* update snd_max */
18938                 if (sack_rxmit)
18939                         goto nomore;
18940
18941                 tp->snd_nxt += len;
18942                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18943                         if (tp->snd_una == tp->snd_max) {
18944                                 /*
18945                                  * Update the time we just added data since
18946                                  * none was outstanding.
18947                                  */
18948                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18949                                 tp->t_acktime = ticks;
18950                         }
18951                         tp->snd_max = tp->snd_nxt;
18952                         /*
18953                          * Time this transmission if not a retransmission and
18954                          * not currently timing anything.
18955                          * This is only relevant in case of switching back to
18956                          * the base stack.
18957                          */
18958                         if (tp->t_rtttime == 0) {
18959                                 tp->t_rtttime = ticks;
18960                                 tp->t_rtseq = startseq;
18961                                 KMOD_TCPSTAT_INC(tcps_segstimed);
18962                         }
18963                         if (len &&
18964                             ((tp->t_flags & TF_GPUTINPROG) == 0))
18965                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18966                 }
18967                 /*
18968                  * If we are doing FO we need to update the mbuf position and subtract
18969                  * this happens when the peer sends us duplicate information and
18970                  * we thus want to send a DSACK.
18971                  *
18972                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18973                  * turned off? If not then we are going to echo multiple DSACK blocks
18974                  * out (with the TSO), which we should not be doing.
18975                  */
18976                 if (rack->r_fast_output && len) {
18977                         if (rack->r_ctl.fsb.left_to_send > len)
18978                                 rack->r_ctl.fsb.left_to_send -= len;
18979                         else
18980                                 rack->r_ctl.fsb.left_to_send = 0;
18981                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18982                                 rack->r_fast_output = 0;
18983                         if (rack->r_fast_output) {
18984                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18985                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18986                         }
18987                 }
18988         }
18989 nomore:
18990         if (error) {
18991                 rack->r_ctl.rc_agg_delayed = 0;
18992                 rack->r_early = 0;
18993                 rack->r_late = 0;
18994                 rack->r_ctl.rc_agg_early = 0;
18995                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
18996                 /*
18997                  * Failures do not advance the seq counter above. For the
18998                  * case of ENOBUFS we will fall out and retry in 1ms with
18999                  * the hpts. Everything else will just have to retransmit
19000                  * with the timer.
19001                  *
19002                  * In any case, we do not want to loop around for another
19003                  * send without a good reason.
19004                  */
19005                 sendalot = 0;
19006                 switch (error) {
19007                 case EPERM:
19008                         tp->t_softerror = error;
19009 #ifdef TCP_ACCOUNTING
19010                         crtsc = get_cyclecount();
19011                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19012                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19013                         }
19014                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19015                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19016                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19017                         }
19018                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19019                         sched_unpin();
19020 #endif
19021                         return (error);
19022                 case ENOBUFS:
19023                         /*
19024                          * Pace us right away to retry in a some
19025                          * time
19026                          */
19027                         if (rack->r_ctl.crte != NULL) {
19028                                 rack_trace_point(rack, RACK_TP_HWENOBUF);
19029                         } else
19030                                 rack_trace_point(rack, RACK_TP_ENOBUF);
19031                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
19032                         if (rack->rc_enobuf < 0x7f)
19033                                 rack->rc_enobuf++;
19034                         if (slot < (10 * HPTS_USEC_IN_MSEC))
19035                                 slot = 10 * HPTS_USEC_IN_MSEC;
19036                         if (rack->r_ctl.crte != NULL) {
19037                                 counter_u64_add(rack_saw_enobuf_hw, 1);
19038                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
19039                         }
19040                         counter_u64_add(rack_saw_enobuf, 1);
19041                         goto enobufs;
19042                 case EMSGSIZE:
19043                         /*
19044                          * For some reason the interface we used initially
19045                          * to send segments changed to another or lowered
19046                          * its MTU. If TSO was active we either got an
19047                          * interface without TSO capabilits or TSO was
19048                          * turned off. If we obtained mtu from ip_output()
19049                          * then update it and try again.
19050                          */
19051                         if (tso)
19052                                 tp->t_flags &= ~TF_TSO;
19053                         if (mtu != 0) {
19054                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
19055                                 goto again;
19056                         }
19057                         slot = 10 * HPTS_USEC_IN_MSEC;
19058                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19059 #ifdef TCP_ACCOUNTING
19060                         crtsc = get_cyclecount();
19061                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19062                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19063                         }
19064                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19065                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19066                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19067                         }
19068                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19069                         sched_unpin();
19070 #endif
19071                         return (error);
19072                 case ENETUNREACH:
19073                         counter_u64_add(rack_saw_enetunreach, 1);
19074                 case EHOSTDOWN:
19075                 case EHOSTUNREACH:
19076                 case ENETDOWN:
19077                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
19078                                 tp->t_softerror = error;
19079                         }
19080                         /* FALLTHROUGH */
19081                 default:
19082                         slot = 10 * HPTS_USEC_IN_MSEC;
19083                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19084 #ifdef TCP_ACCOUNTING
19085                         crtsc = get_cyclecount();
19086                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19087                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19088                         }
19089                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19090                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19091                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19092                         }
19093                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19094                         sched_unpin();
19095 #endif
19096                         return (error);
19097                 }
19098         } else {
19099                 rack->rc_enobuf = 0;
19100                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19101                         rack->r_ctl.retran_during_recovery += len;
19102         }
19103         KMOD_TCPSTAT_INC(tcps_sndtotal);
19104
19105         /*
19106          * Data sent (as far as we can tell). If this advertises a larger
19107          * window than any other segment, then remember the size of the
19108          * advertised window. Any pending ACK has now been sent.
19109          */
19110         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
19111                 tp->rcv_adv = tp->rcv_nxt + recwin;
19112
19113         tp->last_ack_sent = tp->rcv_nxt;
19114         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19115 enobufs:
19116         if (sendalot) {
19117                 /* Do we need to turn off sendalot? */
19118                 if (rack->r_ctl.rc_pace_max_segs &&
19119                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
19120                         /* We hit our max. */
19121                         sendalot = 0;
19122                 } else if ((rack->rc_user_set_max_segs) &&
19123                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
19124                         /* We hit the user defined max */
19125                         sendalot = 0;
19126                 }
19127         }
19128         if ((error == 0) && (flags & TH_FIN))
19129                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
19130         if (flags & TH_RST) {
19131                 /*
19132                  * We don't send again after sending a RST.
19133                  */
19134                 slot = 0;
19135                 sendalot = 0;
19136                 if (error == 0)
19137                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
19138         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
19139                 /*
19140                  * Get our pacing rate, if an error
19141                  * occurred in sending (ENOBUF) we would
19142                  * hit the else if with slot preset. Other
19143                  * errors return.
19144                  */
19145                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
19146         }
19147         if (rsm &&
19148             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
19149             rack->use_rack_rr) {
19150                 /* Its a retransmit and we use the rack cheat? */
19151                 if ((slot == 0) ||
19152                     (rack->rc_always_pace == 0) ||
19153                     (rack->r_rr_config == 1)) {
19154                         /*
19155                          * We have no pacing set or we
19156                          * are using old-style rack or
19157                          * we are overridden to use the old 1ms pacing.
19158                          */
19159                         slot = rack->r_ctl.rc_min_to;
19160                 }
19161         }
19162         /* We have sent clear the flag */
19163         rack->r_ent_rec_ns = 0;
19164         if (rack->r_must_retran) {
19165                 if (rsm) {
19166                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
19167                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
19168                                 /*
19169                                  * We have retransmitted all.
19170                                  */
19171                                 rack->r_must_retran = 0;
19172                                 rack->r_ctl.rc_out_at_rto = 0;
19173                         }
19174                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19175                         /*
19176                          * Sending new data will also kill
19177                          * the loop.
19178                          */
19179                         rack->r_must_retran = 0;
19180                         rack->r_ctl.rc_out_at_rto = 0;
19181                 }
19182         }
19183         rack->r_ctl.fsb.recwin = recwin;
19184         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
19185             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19186                 /*
19187                  * We hit an RTO and now have past snd_max at the RTO
19188                  * clear all the WAS flags.
19189                  */
19190                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
19191         }
19192         if (slot) {
19193                 /* set the rack tcb into the slot N */
19194                 if ((error == 0) &&
19195                     rack_use_rfo &&
19196                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
19197                     (rsm == NULL) &&
19198                     (tp->snd_nxt == tp->snd_max) &&
19199                     (ipoptlen == 0) &&
19200                     (tp->rcv_numsacks == 0) &&
19201                     rack->r_fsb_inited &&
19202                     TCPS_HAVEESTABLISHED(tp->t_state) &&
19203                     (rack->r_must_retran == 0) &&
19204                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
19205                     (len > 0) && (orig_len > 0) &&
19206                     (orig_len > len) &&
19207                     ((orig_len - len) >= segsiz) &&
19208                     ((optlen == 0) ||
19209                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19210                         /* We can send at least one more MSS using our fsb */
19211
19212                         rack->r_fast_output = 1;
19213                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19214                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19215                         rack->r_ctl.fsb.tcp_flags = flags;
19216                         rack->r_ctl.fsb.left_to_send = orig_len - len;
19217                         if (hw_tls)
19218                                 rack->r_ctl.fsb.hw_tls = 1;
19219                         else
19220                                 rack->r_ctl.fsb.hw_tls = 0;
19221                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19222                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19223                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19224                                  (tp->snd_max - tp->snd_una)));
19225                         if (rack->r_ctl.fsb.left_to_send < segsiz)
19226                                 rack->r_fast_output = 0;
19227                         else {
19228                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19229                                         rack->r_ctl.fsb.rfo_apply_push = 1;
19230                                 else
19231                                         rack->r_ctl.fsb.rfo_apply_push = 0;
19232                         }
19233                 } else
19234                         rack->r_fast_output = 0;
19235                 rack_log_fsb(rack, tp, so, flags,
19236                              ipoptlen, orig_len, len, error,
19237                              (rsm == NULL), optlen, __LINE__, 2);
19238         } else if (sendalot) {
19239                 int ret;
19240
19241                 sack_rxmit = 0;
19242                 if ((error == 0) &&
19243                     rack_use_rfo &&
19244                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
19245                     (rsm == NULL) &&
19246                     (ipoptlen == 0) &&
19247                     (tp->rcv_numsacks == 0) &&
19248                     (tp->snd_nxt == tp->snd_max) &&
19249                     (rack->r_must_retran == 0) &&
19250                     rack->r_fsb_inited &&
19251                     TCPS_HAVEESTABLISHED(tp->t_state) &&
19252                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
19253                     (len > 0) && (orig_len > 0) &&
19254                     (orig_len > len) &&
19255                     ((orig_len - len) >= segsiz) &&
19256                     ((optlen == 0) ||
19257                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19258                         /* we can use fast_output for more */
19259
19260                         rack->r_fast_output = 1;
19261                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19262                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19263                         rack->r_ctl.fsb.tcp_flags = flags;
19264                         rack->r_ctl.fsb.left_to_send = orig_len - len;
19265                         if (hw_tls)
19266                                 rack->r_ctl.fsb.hw_tls = 1;
19267                         else
19268                                 rack->r_ctl.fsb.hw_tls = 0;
19269                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19270                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19271                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19272                                  (tp->snd_max - tp->snd_una)));
19273                         if (rack->r_ctl.fsb.left_to_send < segsiz) {
19274                                 rack->r_fast_output = 0;
19275                         }
19276                         if (rack->r_fast_output) {
19277                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19278                                         rack->r_ctl.fsb.rfo_apply_push = 1;
19279                                 else
19280                                         rack->r_ctl.fsb.rfo_apply_push = 0;
19281                                 rack_log_fsb(rack, tp, so, flags,
19282                                              ipoptlen, orig_len, len, error,
19283                                              (rsm == NULL), optlen, __LINE__, 3);
19284                                 error = 0;
19285                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19286                                 if (ret >= 0)
19287                                         return (ret);
19288                                 else if (error)
19289                                         goto nomore;
19290
19291                         }
19292                 }
19293                 goto again;
19294         }
19295         /* Assure when we leave that snd_nxt will point to top */
19296         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19297                 tp->snd_nxt = tp->snd_max;
19298         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19299 #ifdef TCP_ACCOUNTING
19300         crtsc = get_cyclecount() - ts_val;
19301         if (tot_len_this_send) {
19302                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19303                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
19304                 }
19305                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
19306                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19307                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19308                 }
19309                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
19310                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19311                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19312                 }
19313                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
19314         } else {
19315                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19316                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
19317                 }
19318                 counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
19319                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19320                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19321                 }
19322                 counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
19323         }
19324         sched_unpin();
19325 #endif
19326         if (error == ENOBUFS)
19327                 error = 0;
19328         return (error);
19329 }
19330
19331 static void
19332 rack_update_seg(struct tcp_rack *rack)
19333 {
19334         uint32_t orig_val;
19335
19336         orig_val = rack->r_ctl.rc_pace_max_segs;
19337         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19338         if (orig_val != rack->r_ctl.rc_pace_max_segs)
19339                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19340 }
19341
19342 static void
19343 rack_mtu_change(struct tcpcb *tp)
19344 {
19345         /*
19346          * The MSS may have changed
19347          */
19348         struct tcp_rack *rack;
19349         struct rack_sendmap *rsm;
19350
19351         rack = (struct tcp_rack *)tp->t_fb_ptr;
19352         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19353                 /*
19354                  * The MTU has changed we need to resend everything
19355                  * since all we have sent is lost. We first fix
19356                  * up the mtu though.
19357                  */
19358                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19359                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
19360                 rack_remxt_tmr(tp);
19361                 rack->r_fast_output = 0;
19362                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19363                                                 rack->r_ctl.rc_sacked);
19364                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19365                 rack->r_must_retran = 1;
19366                 /* Mark all inflight to needing to be rxt'd */
19367                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19368                         rsm->r_flags |= RACK_MUST_RXT;
19369                 }
19370         }
19371         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19372         /* We don't use snd_nxt to retransmit */
19373         tp->snd_nxt = tp->snd_max;
19374 }
19375
19376 static int
19377 rack_set_profile(struct tcp_rack *rack, int prof)
19378 {
19379         int err = EINVAL;
19380         if (prof == 1) {
19381                 /* pace_always=1 */
19382                 if (rack->rc_always_pace == 0) {
19383                         if (tcp_can_enable_pacing() == 0)
19384                                 return (EBUSY);
19385                 }
19386                 rack->rc_always_pace = 1;
19387                 if (rack->use_fixed_rate || rack->gp_ready)
19388                         rack_set_cc_pacing(rack);
19389                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19390                 rack->rack_attempt_hdwr_pace = 0;
19391                 /* cmpack=1 */
19392                 if (rack_use_cmp_acks)
19393                         rack->r_use_cmp_ack = 1;
19394                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19395                     rack->r_use_cmp_ack)
19396                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19397                 /* scwnd=1 */
19398                 rack->rack_enable_scwnd = 1;
19399                 /* dynamic=100 */
19400                 rack->rc_gp_dyn_mul = 1;
19401                 /* gp_inc_ca */
19402                 rack->r_ctl.rack_per_of_gp_ca = 100;
19403                 /* rrr_conf=3 */
19404                 rack->r_rr_config = 3;
19405                 /* npush=2 */
19406                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19407                 /* fillcw=1 */
19408                 rack->rc_pace_to_cwnd = 1;
19409                 rack->rc_pace_fill_if_rttin_range = 0;
19410                 rack->rtt_limit_mul = 0;
19411                 /* noprr=1 */
19412                 rack->rack_no_prr = 1;
19413                 /* lscwnd=1 */
19414                 rack->r_limit_scw = 1;
19415                 /* gp_inc_rec */
19416                 rack->r_ctl.rack_per_of_gp_rec = 90;
19417                 err = 0;
19418
19419         } else if (prof == 3) {
19420                 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19421                 /* pace_always=1 */
19422                 if (rack->rc_always_pace == 0) {
19423                         if (tcp_can_enable_pacing() == 0)
19424                                 return (EBUSY);
19425                 }
19426                 rack->rc_always_pace = 1;
19427                 if (rack->use_fixed_rate || rack->gp_ready)
19428                         rack_set_cc_pacing(rack);
19429                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19430                 rack->rack_attempt_hdwr_pace = 0;
19431                 /* cmpack=1 */
19432                 if (rack_use_cmp_acks)
19433                         rack->r_use_cmp_ack = 1;
19434                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19435                     rack->r_use_cmp_ack)
19436                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19437                 /* scwnd=1 */
19438                 rack->rack_enable_scwnd = 1;
19439                 /* dynamic=100 */
19440                 rack->rc_gp_dyn_mul = 1;
19441                 /* gp_inc_ca */
19442                 rack->r_ctl.rack_per_of_gp_ca = 100;
19443                 /* rrr_conf=3 */
19444                 rack->r_rr_config = 3;
19445                 /* npush=2 */
19446                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19447                 /* fillcw=2 */
19448                 rack->rc_pace_to_cwnd = 1;
19449                 rack->r_fill_less_agg = 1;
19450                 rack->rc_pace_fill_if_rttin_range = 0;
19451                 rack->rtt_limit_mul = 0;
19452                 /* noprr=1 */
19453                 rack->rack_no_prr = 1;
19454                 /* lscwnd=1 */
19455                 rack->r_limit_scw = 1;
19456                 /* gp_inc_rec */
19457                 rack->r_ctl.rack_per_of_gp_rec = 90;
19458                 err = 0;
19459
19460
19461         } else if (prof == 2) {
19462                 /* cmpack=1 */
19463                 if (rack->rc_always_pace == 0) {
19464                         if (tcp_can_enable_pacing() == 0)
19465                                 return (EBUSY);
19466                 }
19467                 rack->rc_always_pace = 1;
19468                 if (rack->use_fixed_rate || rack->gp_ready)
19469                         rack_set_cc_pacing(rack);
19470                 rack->r_use_cmp_ack = 1;
19471                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19472                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19473                 /* pace_always=1 */
19474                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19475                 /* scwnd=1 */
19476                 rack->rack_enable_scwnd = 1;
19477                 /* dynamic=100 */
19478                 rack->rc_gp_dyn_mul = 1;
19479                 rack->r_ctl.rack_per_of_gp_ca = 100;
19480                 /* rrr_conf=3 */
19481                 rack->r_rr_config = 3;
19482                 /* npush=2 */
19483                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19484                 /* fillcw=1 */
19485                 rack->rc_pace_to_cwnd = 1;
19486                 rack->rc_pace_fill_if_rttin_range = 0;
19487                 rack->rtt_limit_mul = 0;
19488                 /* noprr=1 */
19489                 rack->rack_no_prr = 1;
19490                 /* lscwnd=0 */
19491                 rack->r_limit_scw = 0;
19492                 err = 0;
19493         } else if (prof == 0) {
19494                 /* This changes things back to the default settings */
19495                 err = 0;
19496                 if (rack->rc_always_pace) {
19497                         tcp_decrement_paced_conn();
19498                         rack_undo_cc_pacing(rack);
19499                         rack->rc_always_pace = 0;
19500                 }
19501                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19502                         rack->rc_always_pace = 1;
19503                         if (rack->use_fixed_rate || rack->gp_ready)
19504                                 rack_set_cc_pacing(rack);
19505                 } else
19506                         rack->rc_always_pace = 0;
19507                 if (rack_dsack_std_based & 0x1) {
19508                         /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19509                         rack->rc_rack_tmr_std_based = 1;
19510                 }
19511                 if (rack_dsack_std_based & 0x2) {
19512                         /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19513                         rack->rc_rack_use_dsack = 1;
19514                 }
19515                 if (rack_use_cmp_acks)
19516                         rack->r_use_cmp_ack = 1;
19517                 else
19518                         rack->r_use_cmp_ack = 0;
19519                 if (rack_disable_prr)
19520                         rack->rack_no_prr = 1;
19521                 else
19522                         rack->rack_no_prr = 0;
19523                 if (rack_gp_no_rec_chg)
19524                         rack->rc_gp_no_rec_chg = 1;
19525                 else
19526                         rack->rc_gp_no_rec_chg = 0;
19527                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19528                         rack->r_mbuf_queue = 1;
19529                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19530                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19531                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19532                 } else {
19533                         rack->r_mbuf_queue = 0;
19534                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19535                 }
19536                 if (rack_enable_shared_cwnd)
19537                         rack->rack_enable_scwnd = 1;
19538                 else
19539                         rack->rack_enable_scwnd = 0;
19540                 if (rack_do_dyn_mul) {
19541                         /* When dynamic adjustment is on CA needs to start at 100% */
19542                         rack->rc_gp_dyn_mul = 1;
19543                         if (rack_do_dyn_mul >= 100)
19544                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19545                 } else {
19546                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19547                         rack->rc_gp_dyn_mul = 0;
19548                 }
19549                 rack->r_rr_config = 0;
19550                 rack->r_ctl.rc_no_push_at_mrtt = 0;
19551                 rack->rc_pace_to_cwnd = 0;
19552                 rack->rc_pace_fill_if_rttin_range = 0;
19553                 rack->rtt_limit_mul = 0;
19554
19555                 if (rack_enable_hw_pacing)
19556                         rack->rack_hdw_pace_ena = 1;
19557                 else
19558                         rack->rack_hdw_pace_ena = 0;
19559                 if (rack_disable_prr)
19560                         rack->rack_no_prr = 1;
19561                 else
19562                         rack->rack_no_prr = 0;
19563                 if (rack_limits_scwnd)
19564                         rack->r_limit_scw  = 1;
19565                 else
19566                         rack->r_limit_scw  = 0;
19567                 err = 0;
19568         }
19569         return (err);
19570 }
19571
19572 static int
19573 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19574 {
19575         struct deferred_opt_list *dol;
19576
19577         dol = malloc(sizeof(struct deferred_opt_list),
19578                      M_TCPFSB, M_NOWAIT|M_ZERO);
19579         if (dol == NULL) {
19580                 /*
19581                  * No space yikes -- fail out..
19582                  */
19583                 return (0);
19584         }
19585         dol->optname = sopt_name;
19586         dol->optval = loptval;
19587         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19588         return (1);
19589 }
19590
19591 static int
19592 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19593                     uint32_t optval, uint64_t loptval)
19594 {
19595         struct epoch_tracker et;
19596         struct sockopt sopt;
19597         struct cc_newreno_opts opt;
19598         struct inpcb *inp = tptoinpcb(tp);
19599         uint64_t val;
19600         int error = 0;
19601         uint16_t ca, ss;
19602
19603         switch (sopt_name) {
19604
19605         case TCP_RACK_DSACK_OPT:
19606                 RACK_OPTS_INC(tcp_rack_dsack_opt);
19607                 if (optval & 0x1) {
19608                         rack->rc_rack_tmr_std_based = 1;
19609                 } else {
19610                         rack->rc_rack_tmr_std_based = 0;
19611                 }
19612                 if (optval & 0x2) {
19613                         rack->rc_rack_use_dsack = 1;
19614                 } else {
19615                         rack->rc_rack_use_dsack = 0;
19616                 }
19617                 rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19618                 break;
19619         case TCP_RACK_PACING_BETA:
19620                 RACK_OPTS_INC(tcp_rack_beta);
19621                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
19622                         /* This only works for newreno. */
19623                         error = EINVAL;
19624                         break;
19625                 }
19626                 if (rack->rc_pacing_cc_set) {
19627                         /*
19628                          * Set them into the real CC module
19629                          * whats in the rack pcb is the old values
19630                          * to be used on restoral/
19631                          */
19632                         sopt.sopt_dir = SOPT_SET;
19633                         opt.name = CC_NEWRENO_BETA;
19634                         opt.val = optval;
19635                         if (CC_ALGO(tp)->ctl_output != NULL)
19636                                 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
19637                         else {
19638                                 error = ENOENT;
19639                                 break;
19640                         }
19641                 } else {
19642                         /*
19643                          * Not pacing yet so set it into our local
19644                          * rack pcb storage.
19645                          */
19646                         rack->r_ctl.rc_saved_beta.beta = optval;
19647                 }
19648                 break;
19649         case TCP_RACK_TIMER_SLOP:
19650                 RACK_OPTS_INC(tcp_rack_timer_slop);
19651                 rack->r_ctl.timer_slop = optval;
19652                 if (rack->rc_tp->t_srtt) {
19653                         /*
19654                          * If we have an SRTT lets update t_rxtcur
19655                          * to have the new slop.
19656                          */
19657                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19658                                            rack_rto_min, rack_rto_max,
19659                                            rack->r_ctl.timer_slop);
19660                 }
19661                 break;
19662         case TCP_RACK_PACING_BETA_ECN:
19663                 RACK_OPTS_INC(tcp_rack_beta_ecn);
19664                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
19665                         /* This only works for newreno. */
19666                         error = EINVAL;
19667                         break;
19668                 }
19669                 if (rack->rc_pacing_cc_set) {
19670                         /*
19671                          * Set them into the real CC module
19672                          * whats in the rack pcb is the old values
19673                          * to be used on restoral/
19674                          */
19675                         sopt.sopt_dir = SOPT_SET;
19676                         opt.name = CC_NEWRENO_BETA_ECN;
19677                         opt.val = optval;
19678                         if (CC_ALGO(tp)->ctl_output != NULL)
19679                                 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
19680                         else
19681                                 error = ENOENT;
19682                 } else {
19683                         /*
19684                          * Not pacing yet so set it into our local
19685                          * rack pcb storage.
19686                          */
19687                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19688                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19689                 }
19690                 break;
19691         case TCP_DEFER_OPTIONS:
19692                 RACK_OPTS_INC(tcp_defer_opt);
19693                 if (optval) {
19694                         if (rack->gp_ready) {
19695                                 /* Too late */
19696                                 error = EINVAL;
19697                                 break;
19698                         }
19699                         rack->defer_options = 1;
19700                 } else
19701                         rack->defer_options = 0;
19702                 break;
19703         case TCP_RACK_MEASURE_CNT:
19704                 RACK_OPTS_INC(tcp_rack_measure_cnt);
19705                 if (optval && (optval <= 0xff)) {
19706                         rack->r_ctl.req_measurements = optval;
19707                 } else
19708                         error = EINVAL;
19709                 break;
19710         case TCP_REC_ABC_VAL:
19711                 RACK_OPTS_INC(tcp_rec_abc_val);
19712                 if (optval > 0)
19713                         rack->r_use_labc_for_rec = 1;
19714                 else
19715                         rack->r_use_labc_for_rec = 0;
19716                 break;
19717         case TCP_RACK_ABC_VAL:
19718                 RACK_OPTS_INC(tcp_rack_abc_val);
19719                 if ((optval > 0) && (optval < 255))
19720                         rack->rc_labc = optval;
19721                 else
19722                         error = EINVAL;
19723                 break;
19724         case TCP_HDWR_UP_ONLY:
19725                 RACK_OPTS_INC(tcp_pacing_up_only);
19726                 if (optval)
19727                         rack->r_up_only = 1;
19728                 else
19729                         rack->r_up_only = 0;
19730                 break;
19731         case TCP_PACING_RATE_CAP:
19732                 RACK_OPTS_INC(tcp_pacing_rate_cap);
19733                 rack->r_ctl.bw_rate_cap = loptval;
19734                 break;
19735         case TCP_RACK_PROFILE:
19736                 RACK_OPTS_INC(tcp_profile);
19737                 error = rack_set_profile(rack, optval);
19738                 break;
19739         case TCP_USE_CMP_ACKS:
19740                 RACK_OPTS_INC(tcp_use_cmp_acks);
19741                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19742                         /* You can't turn it off once its on! */
19743                         error = EINVAL;
19744                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19745                         rack->r_use_cmp_ack = 1;
19746                         rack->r_mbuf_queue = 1;
19747                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19748                 }
19749                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19750                         inp->inp_flags2 |= INP_MBUF_ACKCMP;
19751                 break;
19752         case TCP_SHARED_CWND_TIME_LIMIT:
19753                 RACK_OPTS_INC(tcp_lscwnd);
19754                 if (optval)
19755                         rack->r_limit_scw = 1;
19756                 else
19757                         rack->r_limit_scw = 0;
19758                 break;
19759         case TCP_RACK_PACE_TO_FILL:
19760                 RACK_OPTS_INC(tcp_fillcw);
19761                 if (optval == 0)
19762                         rack->rc_pace_to_cwnd = 0;
19763                 else {
19764                         rack->rc_pace_to_cwnd = 1;
19765                         if (optval > 1)
19766                                 rack->r_fill_less_agg = 1;
19767                 }
19768                 if ((optval >= rack_gp_rtt_maxmul) &&
19769                     rack_gp_rtt_maxmul &&
19770                     (optval < 0xf)) {
19771                         rack->rc_pace_fill_if_rttin_range = 1;
19772                         rack->rtt_limit_mul = optval;
19773                 } else {
19774                         rack->rc_pace_fill_if_rttin_range = 0;
19775                         rack->rtt_limit_mul = 0;
19776                 }
19777                 break;
19778         case TCP_RACK_NO_PUSH_AT_MAX:
19779                 RACK_OPTS_INC(tcp_npush);
19780                 if (optval == 0)
19781                         rack->r_ctl.rc_no_push_at_mrtt = 0;
19782                 else if (optval < 0xff)
19783                         rack->r_ctl.rc_no_push_at_mrtt = optval;
19784                 else
19785                         error = EINVAL;
19786                 break;
19787         case TCP_SHARED_CWND_ENABLE:
19788                 RACK_OPTS_INC(tcp_rack_scwnd);
19789                 if (optval == 0)
19790                         rack->rack_enable_scwnd = 0;
19791                 else
19792                         rack->rack_enable_scwnd = 1;
19793                 break;
19794         case TCP_RACK_MBUF_QUEUE:
19795                 /* Now do we use the LRO mbuf-queue feature */
19796                 RACK_OPTS_INC(tcp_rack_mbufq);
19797                 if (optval || rack->r_use_cmp_ack)
19798                         rack->r_mbuf_queue = 1;
19799                 else
19800                         rack->r_mbuf_queue = 0;
19801                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19802                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19803                 else
19804                         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19805                 break;
19806         case TCP_RACK_NONRXT_CFG_RATE:
19807                 RACK_OPTS_INC(tcp_rack_cfg_rate);
19808                 if (optval == 0)
19809                         rack->rack_rec_nonrxt_use_cr = 0;
19810                 else
19811                         rack->rack_rec_nonrxt_use_cr = 1;
19812                 break;
19813         case TCP_NO_PRR:
19814                 RACK_OPTS_INC(tcp_rack_noprr);
19815                 if (optval == 0)
19816                         rack->rack_no_prr = 0;
19817                 else if (optval == 1)
19818                         rack->rack_no_prr = 1;
19819                 else if (optval == 2)
19820                         rack->no_prr_addback = 1;
19821                 else
19822                         error = EINVAL;
19823                 break;
19824         case TCP_TIMELY_DYN_ADJ:
19825                 RACK_OPTS_INC(tcp_timely_dyn);
19826                 if (optval == 0)
19827                         rack->rc_gp_dyn_mul = 0;
19828                 else {
19829                         rack->rc_gp_dyn_mul = 1;
19830                         if (optval >= 100) {
19831                                 /*
19832                                  * If the user sets something 100 or more
19833                                  * its the gp_ca value.
19834                                  */
19835                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
19836                         }
19837                 }
19838                 break;
19839         case TCP_RACK_DO_DETECTION:
19840                 RACK_OPTS_INC(tcp_rack_do_detection);
19841                 if (optval == 0)
19842                         rack->do_detection = 0;
19843                 else
19844                         rack->do_detection = 1;
19845                 break;
19846         case TCP_RACK_TLP_USE:
19847                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19848                         error = EINVAL;
19849                         break;
19850                 }
19851                 RACK_OPTS_INC(tcp_tlp_use);
19852                 rack->rack_tlp_threshold_use = optval;
19853                 break;
19854         case TCP_RACK_TLP_REDUCE:
19855                 /* RACK TLP cwnd reduction (bool) */
19856                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19857                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19858                 break;
19859         /*  Pacing related ones */
19860         case TCP_RACK_PACE_ALWAYS:
19861                 /*
19862                  * zero is old rack method, 1 is new
19863                  * method using a pacing rate.
19864                  */
19865                 RACK_OPTS_INC(tcp_rack_pace_always);
19866                 if (optval > 0) {
19867                         if (rack->rc_always_pace) {
19868                                 error = EALREADY;
19869                                 break;
19870                         } else if (tcp_can_enable_pacing()) {
19871                                 rack->rc_always_pace = 1;
19872                                 if (rack->use_fixed_rate || rack->gp_ready)
19873                                         rack_set_cc_pacing(rack);
19874                         }
19875                         else {
19876                                 error = ENOSPC;
19877                                 break;
19878                         }
19879                 } else {
19880                         if (rack->rc_always_pace) {
19881                                 tcp_decrement_paced_conn();
19882                                 rack->rc_always_pace = 0;
19883                                 rack_undo_cc_pacing(rack);
19884                         }
19885                 }
19886                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19887                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19888                 else
19889                         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19890                 /* A rate may be set irate or other, if so set seg size */
19891                 rack_update_seg(rack);
19892                 break;
19893         case TCP_BBR_RACK_INIT_RATE:
19894                 RACK_OPTS_INC(tcp_initial_rate);
19895                 val = optval;
19896                 /* Change from kbits per second to bytes per second */
19897                 val *= 1000;
19898                 val /= 8;
19899                 rack->r_ctl.init_rate = val;
19900                 if (rack->rc_init_win != rack_default_init_window) {
19901                         uint32_t win, snt;
19902
19903                         /*
19904                          * Options don't always get applied
19905                          * in the order you think. So in order
19906                          * to assure we update a cwnd we need
19907                          * to check and see if we are still
19908                          * where we should raise the cwnd.
19909                          */
19910                         win = rc_init_window(rack);
19911                         if (SEQ_GT(tp->snd_max, tp->iss))
19912                                 snt = tp->snd_max - tp->iss;
19913                         else
19914                                 snt = 0;
19915                         if ((snt < win) &&
19916                             (tp->snd_cwnd < win))
19917                                 tp->snd_cwnd = win;
19918                 }
19919                 if (rack->rc_always_pace)
19920                         rack_update_seg(rack);
19921                 break;
19922         case TCP_BBR_IWINTSO:
19923                 RACK_OPTS_INC(tcp_initial_win);
19924                 if (optval && (optval <= 0xff)) {
19925                         uint32_t win, snt;
19926
19927                         rack->rc_init_win = optval;
19928                         win = rc_init_window(rack);
19929                         if (SEQ_GT(tp->snd_max, tp->iss))
19930                                 snt = tp->snd_max - tp->iss;
19931                         else
19932                                 snt = 0;
19933                         if ((snt < win) &&
19934                             (tp->t_srtt |
19935 #ifdef NETFLIX_PEAKRATE
19936                              tp->t_maxpeakrate |
19937 #endif
19938                              rack->r_ctl.init_rate)) {
19939                                 /*
19940                                  * We are not past the initial window
19941                                  * and we have some bases for pacing,
19942                                  * so we need to possibly adjust up
19943                                  * the cwnd. Note even if we don't set
19944                                  * the cwnd, its still ok to raise the rc_init_win
19945                                  * which can be used coming out of idle when we
19946                                  * would have a rate.
19947                                  */
19948                                 if (tp->snd_cwnd < win)
19949                                         tp->snd_cwnd = win;
19950                         }
19951                         if (rack->rc_always_pace)
19952                                 rack_update_seg(rack);
19953                 } else
19954                         error = EINVAL;
19955                 break;
19956         case TCP_RACK_FORCE_MSEG:
19957                 RACK_OPTS_INC(tcp_rack_force_max_seg);
19958                 if (optval)
19959                         rack->rc_force_max_seg = 1;
19960                 else
19961                         rack->rc_force_max_seg = 0;
19962                 break;
19963         case TCP_RACK_PACE_MAX_SEG:
19964                 /* Max segments size in a pace in bytes */
19965                 RACK_OPTS_INC(tcp_rack_max_seg);
19966                 rack->rc_user_set_max_segs = optval;
19967                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19968                 break;
19969         case TCP_RACK_PACE_RATE_REC:
19970                 /* Set the fixed pacing rate in Bytes per second ca */
19971                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19972                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19973                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19974                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19975                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19976                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19977                 rack->use_fixed_rate = 1;
19978                 if (rack->rc_always_pace)
19979                         rack_set_cc_pacing(rack);
19980                 rack_log_pacing_delay_calc(rack,
19981                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19982                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19983                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19984                                            __LINE__, NULL,0);
19985                 break;
19986
19987         case TCP_RACK_PACE_RATE_SS:
19988                 /* Set the fixed pacing rate in Bytes per second ca */
19989                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19990                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19991                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19992                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19993                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19994                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19995                 rack->use_fixed_rate = 1;
19996                 if (rack->rc_always_pace)
19997                         rack_set_cc_pacing(rack);
19998                 rack_log_pacing_delay_calc(rack,
19999                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
20000                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
20001                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20002                                            __LINE__, NULL, 0);
20003                 break;
20004
20005         case TCP_RACK_PACE_RATE_CA:
20006                 /* Set the fixed pacing rate in Bytes per second ca */
20007                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
20008                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20009                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
20010                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20011                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
20012                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20013                 rack->use_fixed_rate = 1;
20014                 if (rack->rc_always_pace)
20015                         rack_set_cc_pacing(rack);
20016                 rack_log_pacing_delay_calc(rack,
20017                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
20018                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
20019                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20020                                            __LINE__, NULL, 0);
20021                 break;
20022         case TCP_RACK_GP_INCREASE_REC:
20023                 RACK_OPTS_INC(tcp_gp_inc_rec);
20024                 rack->r_ctl.rack_per_of_gp_rec = optval;
20025                 rack_log_pacing_delay_calc(rack,
20026                                            rack->r_ctl.rack_per_of_gp_ss,
20027                                            rack->r_ctl.rack_per_of_gp_ca,
20028                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20029                                            __LINE__, NULL, 0);
20030                 break;
20031         case TCP_RACK_GP_INCREASE_CA:
20032                 RACK_OPTS_INC(tcp_gp_inc_ca);
20033                 ca = optval;
20034                 if (ca < 100) {
20035                         /*
20036                          * We don't allow any reduction
20037                          * over the GP b/w.
20038                          */
20039                         error = EINVAL;
20040                         break;
20041                 }
20042                 rack->r_ctl.rack_per_of_gp_ca = ca;
20043                 rack_log_pacing_delay_calc(rack,
20044                                            rack->r_ctl.rack_per_of_gp_ss,
20045                                            rack->r_ctl.rack_per_of_gp_ca,
20046                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20047                                            __LINE__, NULL, 0);
20048                 break;
20049         case TCP_RACK_GP_INCREASE_SS:
20050                 RACK_OPTS_INC(tcp_gp_inc_ss);
20051                 ss = optval;
20052                 if (ss < 100) {
20053                         /*
20054                          * We don't allow any reduction
20055                          * over the GP b/w.
20056                          */
20057                         error = EINVAL;
20058                         break;
20059                 }
20060                 rack->r_ctl.rack_per_of_gp_ss = ss;
20061                 rack_log_pacing_delay_calc(rack,
20062                                            rack->r_ctl.rack_per_of_gp_ss,
20063                                            rack->r_ctl.rack_per_of_gp_ca,
20064                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20065                                            __LINE__, NULL, 0);
20066                 break;
20067         case TCP_RACK_RR_CONF:
20068                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
20069                 if (optval && optval <= 3)
20070                         rack->r_rr_config = optval;
20071                 else
20072                         rack->r_rr_config = 0;
20073                 break;
20074         case TCP_HDWR_RATE_CAP:
20075                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
20076                 if (optval) {
20077                         if (rack->r_rack_hw_rate_caps == 0)
20078                                 rack->r_rack_hw_rate_caps = 1;
20079                         else
20080                                 error = EALREADY;
20081                 } else {
20082                         rack->r_rack_hw_rate_caps = 0;
20083                 }
20084                 break;
20085         case TCP_BBR_HDWR_PACE:
20086                 RACK_OPTS_INC(tcp_hdwr_pacing);
20087                 if (optval){
20088                         if (rack->rack_hdrw_pacing == 0) {
20089                                 rack->rack_hdw_pace_ena = 1;
20090                                 rack->rack_attempt_hdwr_pace = 0;
20091                         } else
20092                                 error = EALREADY;
20093                 } else {
20094                         rack->rack_hdw_pace_ena = 0;
20095 #ifdef RATELIMIT
20096                         if (rack->r_ctl.crte != NULL) {
20097                                 rack->rack_hdrw_pacing = 0;
20098                                 rack->rack_attempt_hdwr_pace = 0;
20099                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
20100                                 rack->r_ctl.crte = NULL;
20101                         }
20102 #endif
20103                 }
20104                 break;
20105         /*  End Pacing related ones */
20106         case TCP_RACK_PRR_SENDALOT:
20107                 /* Allow PRR to send more than one seg */
20108                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
20109                 rack->r_ctl.rc_prr_sendalot = optval;
20110                 break;
20111         case TCP_RACK_MIN_TO:
20112                 /* Minimum time between rack t-o's in ms */
20113                 RACK_OPTS_INC(tcp_rack_min_to);
20114                 rack->r_ctl.rc_min_to = optval;
20115                 break;
20116         case TCP_RACK_EARLY_SEG:
20117                 /* If early recovery max segments */
20118                 RACK_OPTS_INC(tcp_rack_early_seg);
20119                 rack->r_ctl.rc_early_recovery_segs = optval;
20120                 break;
20121         case TCP_RACK_ENABLE_HYSTART:
20122         {
20123                 if (optval) {
20124                         tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
20125                         if (rack_do_hystart > RACK_HYSTART_ON)
20126                                 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
20127                         if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
20128                                 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
20129                 } else {
20130                         tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
20131                 }
20132         }
20133         break;
20134         case TCP_RACK_REORD_THRESH:
20135                 /* RACK reorder threshold (shift amount) */
20136                 RACK_OPTS_INC(tcp_rack_reord_thresh);
20137                 if ((optval > 0) && (optval < 31))
20138                         rack->r_ctl.rc_reorder_shift = optval;
20139                 else
20140                         error = EINVAL;
20141                 break;
20142         case TCP_RACK_REORD_FADE:
20143                 /* Does reordering fade after ms time */
20144                 RACK_OPTS_INC(tcp_rack_reord_fade);
20145                 rack->r_ctl.rc_reorder_fade = optval;
20146                 break;
20147         case TCP_RACK_TLP_THRESH:
20148                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
20149                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
20150                 if (optval)
20151                         rack->r_ctl.rc_tlp_threshold = optval;
20152                 else
20153                         error = EINVAL;
20154                 break;
20155         case TCP_BBR_USE_RACK_RR:
20156                 RACK_OPTS_INC(tcp_rack_rr);
20157                 if (optval)
20158                         rack->use_rack_rr = 1;
20159                 else
20160                         rack->use_rack_rr = 0;
20161                 break;
20162         case TCP_FAST_RSM_HACK:
20163                 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
20164                 if (optval)
20165                         rack->fast_rsm_hack = 1;
20166                 else
20167                         rack->fast_rsm_hack = 0;
20168                 break;
20169         case TCP_RACK_PKT_DELAY:
20170                 /* RACK added ms i.e. rack-rtt + reord + N */
20171                 RACK_OPTS_INC(tcp_rack_pkt_delay);
20172                 rack->r_ctl.rc_pkt_delay = optval;
20173                 break;
20174         case TCP_DELACK:
20175                 RACK_OPTS_INC(tcp_rack_delayed_ack);
20176                 if (optval == 0)
20177                         tp->t_delayed_ack = 0;
20178                 else
20179                         tp->t_delayed_ack = 1;
20180                 if (tp->t_flags & TF_DELACK) {
20181                         tp->t_flags &= ~TF_DELACK;
20182                         tp->t_flags |= TF_ACKNOW;
20183                         NET_EPOCH_ENTER(et);
20184                         rack_output(tp);
20185                         NET_EPOCH_EXIT(et);
20186                 }
20187                 break;
20188
20189         case TCP_BBR_RACK_RTT_USE:
20190                 RACK_OPTS_INC(tcp_rack_rtt_use);
20191                 if ((optval != USE_RTT_HIGH) &&
20192                     (optval != USE_RTT_LOW) &&
20193                     (optval != USE_RTT_AVG))
20194                         error = EINVAL;
20195                 else
20196                         rack->r_ctl.rc_rate_sample_method = optval;
20197                 break;
20198         case TCP_DATA_AFTER_CLOSE:
20199                 RACK_OPTS_INC(tcp_data_after_close);
20200                 if (optval)
20201                         rack->rc_allow_data_af_clo = 1;
20202                 else
20203                         rack->rc_allow_data_af_clo = 0;
20204                 break;
20205         default:
20206                 break;
20207         }
20208 #ifdef NETFLIX_STATS
20209         tcp_log_socket_option(tp, sopt_name, optval, error);
20210 #endif
20211         return (error);
20212 }
20213
20214
20215 static void
20216 rack_apply_deferred_options(struct tcp_rack *rack)
20217 {
20218         struct deferred_opt_list *dol, *sdol;
20219         uint32_t s_optval;
20220
20221         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20222                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20223                 /* Disadvantage of deferal is you loose the error return */
20224                 s_optval = (uint32_t)dol->optval;
20225                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20226                 free(dol, M_TCPDO);
20227         }
20228 }
20229
20230 static void
20231 rack_hw_tls_change(struct tcpcb *tp, int chg)
20232 {
20233         /*
20234          * HW tls state has changed.. fix all
20235          * rsm's in flight.
20236          */
20237         struct tcp_rack *rack;
20238         struct rack_sendmap *rsm;
20239
20240         rack = (struct tcp_rack *)tp->t_fb_ptr;
20241         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
20242                 if (chg)
20243                         rsm->r_hw_tls = 1;
20244                 else
20245                         rsm->r_hw_tls = 0;
20246         }
20247         if (chg)
20248                 rack->r_ctl.fsb.hw_tls = 1;
20249         else
20250                 rack->r_ctl.fsb.hw_tls = 0;
20251 }
20252
20253 static int
20254 rack_pru_options(struct tcpcb *tp, int flags)
20255 {
20256         if (flags & PRUS_OOB)
20257                 return (EOPNOTSUPP);
20258         return (0);
20259 }
20260
20261 static struct tcp_function_block __tcp_rack = {
20262         .tfb_tcp_block_name = __XSTRING(STACKNAME),
20263         .tfb_tcp_output = rack_output,
20264         .tfb_do_queued_segments = ctf_do_queued_segments,
20265         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
20266         .tfb_tcp_do_segment = rack_do_segment,
20267         .tfb_tcp_ctloutput = rack_ctloutput,
20268         .tfb_tcp_fb_init = rack_init,
20269         .tfb_tcp_fb_fini = rack_fini,
20270         .tfb_tcp_timer_stop_all = rack_stopall,
20271         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20272         .tfb_tcp_handoff_ok = rack_handoff_ok,
20273         .tfb_tcp_mtu_chg = rack_mtu_change,
20274         .tfb_pru_options = rack_pru_options,
20275         .tfb_hwtls_change = rack_hw_tls_change,
20276         .tfb_compute_pipe = rack_compute_pipe,
20277         .tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
20278 };
20279
20280 /*
20281  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20282  * socket option arguments.  When it re-acquires the lock after the copy, it
20283  * has to revalidate that the connection is still valid for the socket
20284  * option.
20285  */
20286 static int
20287 rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt)
20288 {
20289 #ifdef INET6
20290         struct ip6_hdr *ip6;
20291 #endif
20292 #ifdef INET
20293         struct ip *ip;
20294 #endif
20295         struct tcpcb *tp;
20296         struct tcp_rack *rack;
20297         uint64_t loptval;
20298         int32_t error = 0, optval;
20299
20300         tp = intotcpcb(inp);
20301         rack = (struct tcp_rack *)tp->t_fb_ptr;
20302         if (rack == NULL) {
20303                 INP_WUNLOCK(inp);
20304                 return (EINVAL);
20305         }
20306 #ifdef INET6
20307         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20308 #endif
20309 #ifdef INET
20310         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20311 #endif
20312
20313         switch (sopt->sopt_level) {
20314 #ifdef INET6
20315         case IPPROTO_IPV6:
20316                 MPASS(inp->inp_vflag & INP_IPV6PROTO);
20317                 switch (sopt->sopt_name) {
20318                 case IPV6_USE_MIN_MTU:
20319                         tcp6_use_min_mtu(tp);
20320                         break;
20321                 case IPV6_TCLASS:
20322                         /*
20323                          * The DSCP codepoint has changed, update the fsb.
20324                          */
20325                         ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20326                             (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20327                         break;
20328                 }
20329                 INP_WUNLOCK(inp);
20330                 return (0);
20331 #endif
20332 #ifdef INET
20333         case IPPROTO_IP:
20334                 switch (sopt->sopt_name) {
20335                 case IP_TOS:
20336                         /*
20337                          * The DSCP codepoint has changed, update the fsb.
20338                          */
20339                         ip->ip_tos = rack->rc_inp->inp_ip_tos;
20340                         break;
20341                 case IP_TTL:
20342                         /*
20343                          * The TTL has changed, update the fsb.
20344                          */
20345                         ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20346                         break;
20347                 }
20348                 INP_WUNLOCK(inp);
20349                 return (0);
20350 #endif
20351         }
20352
20353         switch (sopt->sopt_name) {
20354         case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
20355         /*  Pacing related ones */
20356         case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
20357         case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
20358         case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
20359         case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
20360         case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
20361         case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
20362         case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
20363         case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
20364         case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
20365         case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
20366         case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
20367         case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
20368         case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
20369         case TCP_HDWR_RATE_CAP:                 /*  URL:hdwrcap boolean */
20370         case TCP_PACING_RATE_CAP:               /*  URL:cap  -- used by side-channel */
20371         case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
20372        /* End pacing related */
20373         case TCP_FAST_RSM_HACK:                 /*  URL:frsm_hack */
20374         case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20375         case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
20376         case TCP_RACK_MIN_TO:                   /*  URL:min_to */
20377         case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
20378         case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
20379         case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
20380         case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
20381         case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
20382         case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
20383         case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
20384         case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
20385         case TCP_RACK_DO_DETECTION:             /*  URL:detect */
20386         case TCP_NO_PRR:                        /*  URL:noprr */
20387         case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
20388         case TCP_DATA_AFTER_CLOSE:              /*  no URL */
20389         case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
20390         case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
20391         case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
20392         case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
20393         case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
20394         case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
20395         case TCP_RACK_PROFILE:                  /*  URL:profile */
20396         case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
20397         case TCP_RACK_ABC_VAL:                  /*  URL:labc */
20398         case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
20399         case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
20400         case TCP_DEFER_OPTIONS:                 /*  URL:defer */
20401         case TCP_RACK_DSACK_OPT:                /*  URL:dsack */
20402         case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
20403         case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
20404         case TCP_RACK_TIMER_SLOP:               /*  URL:timer_slop */
20405         case TCP_RACK_ENABLE_HYSTART:           /*  URL:hystart */
20406                 break;
20407         default:
20408                 /* Filter off all unknown options to the base stack */
20409                 return (tcp_default_ctloutput(inp, sopt));
20410                 break;
20411         }
20412         INP_WUNLOCK(inp);
20413         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20414                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20415                 /*
20416                  * We truncate it down to 32 bits for the socket-option trace this
20417                  * means rates > 34Gbps won't show right, but thats probably ok.
20418                  */
20419                 optval = (uint32_t)loptval;
20420         } else {
20421                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20422                 /* Save it in 64 bit form too */
20423                 loptval = optval;
20424         }
20425         if (error)
20426                 return (error);
20427         INP_WLOCK(inp);
20428         if (inp->inp_flags & INP_DROPPED) {
20429                 INP_WUNLOCK(inp);
20430                 return (ECONNRESET);
20431         }
20432         if (tp->t_fb != &__tcp_rack) {
20433                 INP_WUNLOCK(inp);
20434                 return (ENOPROTOOPT);
20435         }
20436         if (rack->defer_options && (rack->gp_ready == 0) &&
20437             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20438             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20439             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20440             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20441                 /* Options are beind deferred */
20442                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20443                         INP_WUNLOCK(inp);
20444                         return (0);
20445                 } else {
20446                         /* No memory to defer, fail */
20447                         INP_WUNLOCK(inp);
20448                         return (ENOMEM);
20449                 }
20450         }
20451         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20452         INP_WUNLOCK(inp);
20453         return (error);
20454 }
20455
20456 static void
20457 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20458 {
20459
20460         INP_WLOCK_ASSERT(tptoinpcb(tp));
20461         bzero(ti, sizeof(*ti));
20462
20463         ti->tcpi_state = tp->t_state;
20464         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20465                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20466         if (tp->t_flags & TF_SACK_PERMIT)
20467                 ti->tcpi_options |= TCPI_OPT_SACK;
20468         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20469                 ti->tcpi_options |= TCPI_OPT_WSCALE;
20470                 ti->tcpi_snd_wscale = tp->snd_scale;
20471                 ti->tcpi_rcv_wscale = tp->rcv_scale;
20472         }
20473         if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
20474                 ti->tcpi_options |= TCPI_OPT_ECN;
20475         if (tp->t_flags & TF_FASTOPEN)
20476                 ti->tcpi_options |= TCPI_OPT_TFO;
20477         /* still kept in ticks is t_rcvtime */
20478         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20479         /* Since we hold everything in precise useconds this is easy */
20480         ti->tcpi_rtt = tp->t_srtt;
20481         ti->tcpi_rttvar = tp->t_rttvar;
20482         ti->tcpi_rto = tp->t_rxtcur;
20483         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20484         ti->tcpi_snd_cwnd = tp->snd_cwnd;
20485         /*
20486          * FreeBSD-specific extension fields for tcp_info.
20487          */
20488         ti->tcpi_rcv_space = tp->rcv_wnd;
20489         ti->tcpi_rcv_nxt = tp->rcv_nxt;
20490         ti->tcpi_snd_wnd = tp->snd_wnd;
20491         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
20492         ti->tcpi_snd_nxt = tp->snd_nxt;
20493         ti->tcpi_snd_mss = tp->t_maxseg;
20494         ti->tcpi_rcv_mss = tp->t_maxseg;
20495         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20496         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20497         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20498 #ifdef NETFLIX_STATS
20499         ti->tcpi_total_tlp = tp->t_sndtlppack;
20500         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20501         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20502 #endif
20503 #ifdef TCP_OFFLOAD
20504         if (tp->t_flags & TF_TOE) {
20505                 ti->tcpi_options |= TCPI_OPT_TOE;
20506                 tcp_offload_tcp_info(tp, ti);
20507         }
20508 #endif
20509 }
20510
20511 static int
20512 rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt)
20513 {
20514         struct tcpcb *tp;
20515         struct tcp_rack *rack;
20516         int32_t error, optval;
20517         uint64_t val, loptval;
20518         struct  tcp_info ti;
20519         /*
20520          * Because all our options are either boolean or an int, we can just
20521          * pull everything into optval and then unlock and copy. If we ever
20522          * add a option that is not a int, then this will have quite an
20523          * impact to this routine.
20524          */
20525         error = 0;
20526         tp = intotcpcb(inp);
20527         rack = (struct tcp_rack *)tp->t_fb_ptr;
20528         if (rack == NULL) {
20529                 INP_WUNLOCK(inp);
20530                 return (EINVAL);
20531         }
20532         switch (sopt->sopt_name) {
20533         case TCP_INFO:
20534                 /* First get the info filled */
20535                 rack_fill_info(tp, &ti);
20536                 /* Fix up the rtt related fields if needed */
20537                 INP_WUNLOCK(inp);
20538                 error = sooptcopyout(sopt, &ti, sizeof ti);
20539                 return (error);
20540         /*
20541          * Beta is the congestion control value for NewReno that influences how
20542          * much of a backoff happens when loss is detected. It is normally set
20543          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20544          * when you exit recovery.
20545          */
20546         case TCP_RACK_PACING_BETA:
20547                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
20548                         error = EINVAL;
20549                 else if (rack->rc_pacing_cc_set == 0)
20550                         optval = rack->r_ctl.rc_saved_beta.beta;
20551                 else {
20552                         /*
20553                          * Reach out into the CC data and report back what
20554                          * I have previously set. Yeah it looks hackish but
20555                          * we don't want to report the saved values.
20556                          */
20557                         if (tp->t_ccv.cc_data)
20558                                 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
20559                         else
20560                                 error = EINVAL;
20561                 }
20562                 break;
20563                 /*
20564                  * Beta_ecn is the congestion control value for NewReno that influences how
20565                  * much of a backoff happens when a ECN mark is detected. It is normally set
20566                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20567                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
20568                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
20569                  */
20570
20571         case TCP_RACK_PACING_BETA_ECN:
20572                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
20573                         error = EINVAL;
20574                 else if (rack->rc_pacing_cc_set == 0)
20575                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20576                 else {
20577                         /*
20578                          * Reach out into the CC data and report back what
20579                          * I have previously set. Yeah it looks hackish but
20580                          * we don't want to report the saved values.
20581                          */
20582                         if (tp->t_ccv.cc_data)
20583                                 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
20584                         else
20585                                 error = EINVAL;
20586                 }
20587                 break;
20588         case TCP_RACK_DSACK_OPT:
20589                 optval = 0;
20590                 if (rack->rc_rack_tmr_std_based) {
20591                         optval |= 1;
20592                 }
20593                 if (rack->rc_rack_use_dsack) {
20594                         optval |= 2;
20595                 }
20596                 break;
20597         case TCP_RACK_ENABLE_HYSTART:
20598         {
20599                 if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
20600                         optval = RACK_HYSTART_ON;
20601                         if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
20602                                 optval = RACK_HYSTART_ON_W_SC;
20603                         if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
20604                                 optval = RACK_HYSTART_ON_W_SC_C;
20605                 } else {
20606                         optval = RACK_HYSTART_OFF;
20607                 }
20608         }
20609         break;
20610         case TCP_FAST_RSM_HACK:
20611                 optval = rack->fast_rsm_hack;
20612                 break;
20613         case TCP_DEFER_OPTIONS:
20614                 optval = rack->defer_options;
20615                 break;
20616         case TCP_RACK_MEASURE_CNT:
20617                 optval = rack->r_ctl.req_measurements;
20618                 break;
20619         case TCP_REC_ABC_VAL:
20620                 optval = rack->r_use_labc_for_rec;
20621                 break;
20622         case TCP_RACK_ABC_VAL:
20623                 optval = rack->rc_labc;
20624                 break;
20625         case TCP_HDWR_UP_ONLY:
20626                 optval= rack->r_up_only;
20627                 break;
20628         case TCP_PACING_RATE_CAP:
20629                 loptval = rack->r_ctl.bw_rate_cap;
20630                 break;
20631         case TCP_RACK_PROFILE:
20632                 /* You cannot retrieve a profile, its write only */
20633                 error = EINVAL;
20634                 break;
20635         case TCP_USE_CMP_ACKS:
20636                 optval = rack->r_use_cmp_ack;
20637                 break;
20638         case TCP_RACK_PACE_TO_FILL:
20639                 optval = rack->rc_pace_to_cwnd;
20640                 if (optval && rack->r_fill_less_agg)
20641                         optval++;
20642                 break;
20643         case TCP_RACK_NO_PUSH_AT_MAX:
20644                 optval = rack->r_ctl.rc_no_push_at_mrtt;
20645                 break;
20646         case TCP_SHARED_CWND_ENABLE:
20647                 optval = rack->rack_enable_scwnd;
20648                 break;
20649         case TCP_RACK_NONRXT_CFG_RATE:
20650                 optval = rack->rack_rec_nonrxt_use_cr;
20651                 break;
20652         case TCP_NO_PRR:
20653                 if (rack->rack_no_prr  == 1)
20654                         optval = 1;
20655                 else if (rack->no_prr_addback == 1)
20656                         optval = 2;
20657                 else
20658                         optval = 0;
20659                 break;
20660         case TCP_RACK_DO_DETECTION:
20661                 optval = rack->do_detection;
20662                 break;
20663         case TCP_RACK_MBUF_QUEUE:
20664                 /* Now do we use the LRO mbuf-queue feature */
20665                 optval = rack->r_mbuf_queue;
20666                 break;
20667         case TCP_TIMELY_DYN_ADJ:
20668                 optval = rack->rc_gp_dyn_mul;
20669                 break;
20670         case TCP_BBR_IWINTSO:
20671                 optval = rack->rc_init_win;
20672                 break;
20673         case TCP_RACK_TLP_REDUCE:
20674                 /* RACK TLP cwnd reduction (bool) */
20675                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20676                 break;
20677         case TCP_BBR_RACK_INIT_RATE:
20678                 val = rack->r_ctl.init_rate;
20679                 /* convert to kbits per sec */
20680                 val *= 8;
20681                 val /= 1000;
20682                 optval = (uint32_t)val;
20683                 break;
20684         case TCP_RACK_FORCE_MSEG:
20685                 optval = rack->rc_force_max_seg;
20686                 break;
20687         case TCP_RACK_PACE_MAX_SEG:
20688                 /* Max segments in a pace */
20689                 optval = rack->rc_user_set_max_segs;
20690                 break;
20691         case TCP_RACK_PACE_ALWAYS:
20692                 /* Use the always pace method */
20693                 optval = rack->rc_always_pace;
20694                 break;
20695         case TCP_RACK_PRR_SENDALOT:
20696                 /* Allow PRR to send more than one seg */
20697                 optval = rack->r_ctl.rc_prr_sendalot;
20698                 break;
20699         case TCP_RACK_MIN_TO:
20700                 /* Minimum time between rack t-o's in ms */
20701                 optval = rack->r_ctl.rc_min_to;
20702                 break;
20703         case TCP_RACK_EARLY_SEG:
20704                 /* If early recovery max segments */
20705                 optval = rack->r_ctl.rc_early_recovery_segs;
20706                 break;
20707         case TCP_RACK_REORD_THRESH:
20708                 /* RACK reorder threshold (shift amount) */
20709                 optval = rack->r_ctl.rc_reorder_shift;
20710                 break;
20711         case TCP_RACK_REORD_FADE:
20712                 /* Does reordering fade after ms time */
20713                 optval = rack->r_ctl.rc_reorder_fade;
20714                 break;
20715         case TCP_BBR_USE_RACK_RR:
20716                 /* Do we use the rack cheat for rxt */
20717                 optval = rack->use_rack_rr;
20718                 break;
20719         case TCP_RACK_RR_CONF:
20720                 optval = rack->r_rr_config;
20721                 break;
20722         case TCP_HDWR_RATE_CAP:
20723                 optval = rack->r_rack_hw_rate_caps;
20724                 break;
20725         case TCP_BBR_HDWR_PACE:
20726                 optval = rack->rack_hdw_pace_ena;
20727                 break;
20728         case TCP_RACK_TLP_THRESH:
20729                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
20730                 optval = rack->r_ctl.rc_tlp_threshold;
20731                 break;
20732         case TCP_RACK_PKT_DELAY:
20733                 /* RACK added ms i.e. rack-rtt + reord + N */
20734                 optval = rack->r_ctl.rc_pkt_delay;
20735                 break;
20736         case TCP_RACK_TLP_USE:
20737                 optval = rack->rack_tlp_threshold_use;
20738                 break;
20739         case TCP_RACK_PACE_RATE_CA:
20740                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20741                 break;
20742         case TCP_RACK_PACE_RATE_SS:
20743                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20744                 break;
20745         case TCP_RACK_PACE_RATE_REC:
20746                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20747                 break;
20748         case TCP_RACK_GP_INCREASE_SS:
20749                 optval = rack->r_ctl.rack_per_of_gp_ca;
20750                 break;
20751         case TCP_RACK_GP_INCREASE_CA:
20752                 optval = rack->r_ctl.rack_per_of_gp_ss;
20753                 break;
20754         case TCP_BBR_RACK_RTT_USE:
20755                 optval = rack->r_ctl.rc_rate_sample_method;
20756                 break;
20757         case TCP_DELACK:
20758                 optval = tp->t_delayed_ack;
20759                 break;
20760         case TCP_DATA_AFTER_CLOSE:
20761                 optval = rack->rc_allow_data_af_clo;
20762                 break;
20763         case TCP_SHARED_CWND_TIME_LIMIT:
20764                 optval = rack->r_limit_scw;
20765                 break;
20766         case TCP_RACK_TIMER_SLOP:
20767                 optval = rack->r_ctl.timer_slop;
20768                 break;
20769         default:
20770                 return (tcp_default_ctloutput(inp, sopt));
20771                 break;
20772         }
20773         INP_WUNLOCK(inp);
20774         if (error == 0) {
20775                 if (TCP_PACING_RATE_CAP)
20776                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
20777                 else
20778                         error = sooptcopyout(sopt, &optval, sizeof optval);
20779         }
20780         return (error);
20781 }
20782
20783 static int
20784 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt)
20785 {
20786         if (sopt->sopt_dir == SOPT_SET) {
20787                 return (rack_set_sockopt(inp, sopt));
20788         } else if (sopt->sopt_dir == SOPT_GET) {
20789                 return (rack_get_sockopt(inp, sopt));
20790         } else {
20791                 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
20792         }
20793 }
20794
20795 static const char *rack_stack_names[] = {
20796         __XSTRING(STACKNAME),
20797 #ifdef STACKALIAS
20798         __XSTRING(STACKALIAS),
20799 #endif
20800 };
20801
20802 static int
20803 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20804 {
20805         memset(mem, 0, size);
20806         return (0);
20807 }
20808
20809 static void
20810 rack_dtor(void *mem, int32_t size, void *arg)
20811 {
20812
20813 }
20814
20815 static bool rack_mod_inited = false;
20816
20817 static int
20818 tcp_addrack(module_t mod, int32_t type, void *data)
20819 {
20820         int32_t err = 0;
20821         int num_stacks;
20822
20823         switch (type) {
20824         case MOD_LOAD:
20825                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20826                     sizeof(struct rack_sendmap),
20827                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20828
20829                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20830                     sizeof(struct tcp_rack),
20831                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20832
20833                 sysctl_ctx_init(&rack_sysctl_ctx);
20834                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20835                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20836                     OID_AUTO,
20837 #ifdef STACKALIAS
20838                     __XSTRING(STACKALIAS),
20839 #else
20840                     __XSTRING(STACKNAME),
20841 #endif
20842                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20843                     "");
20844                 if (rack_sysctl_root == NULL) {
20845                         printf("Failed to add sysctl node\n");
20846                         err = EFAULT;
20847                         goto free_uma;
20848                 }
20849                 rack_init_sysctls();
20850                 num_stacks = nitems(rack_stack_names);
20851                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20852                     rack_stack_names, &num_stacks);
20853                 if (err) {
20854                         printf("Failed to register %s stack name for "
20855                             "%s module\n", rack_stack_names[num_stacks],
20856                             __XSTRING(MODNAME));
20857                         sysctl_ctx_free(&rack_sysctl_ctx);
20858 free_uma:
20859                         uma_zdestroy(rack_zone);
20860                         uma_zdestroy(rack_pcb_zone);
20861                         rack_counter_destroy();
20862                         printf("Failed to register rack module -- err:%d\n", err);
20863                         return (err);
20864                 }
20865                 tcp_lro_reg_mbufq();
20866                 rack_mod_inited = true;
20867                 break;
20868         case MOD_QUIESCE:
20869                 err = deregister_tcp_functions(&__tcp_rack, true, false);
20870                 break;
20871         case MOD_UNLOAD:
20872                 err = deregister_tcp_functions(&__tcp_rack, false, true);
20873                 if (err == EBUSY)
20874                         break;
20875                 if (rack_mod_inited) {
20876                         uma_zdestroy(rack_zone);
20877                         uma_zdestroy(rack_pcb_zone);
20878                         sysctl_ctx_free(&rack_sysctl_ctx);
20879                         rack_counter_destroy();
20880                         rack_mod_inited = false;
20881                 }
20882                 tcp_lro_dereg_mbufq();
20883                 err = 0;
20884                 break;
20885         default:
20886                 return (EOPNOTSUPP);
20887         }
20888         return (err);
20889 }
20890
20891 static moduledata_t tcp_rack = {
20892         .name = __XSTRING(MODNAME),
20893         .evhand = tcp_addrack,
20894         .priv = 0
20895 };
20896
20897 MODULE_VERSION(MODNAME, 1);
20898 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20899 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);