sys/netinet/tcp_stacks/rack.c

   1 /*-
   2  * Copyright (c) 2016-2020 Netflix, Inc.
   3  *
   4  * Redistribution and use in source and binary forms, with or without
   5  * modification, are permitted provided that the following conditions
   6  * are met:
   7  * 1. Redistributions of source code must retain the above copyright
   8  *    notice, this list of conditions and the following disclaimer.
   9  * 2. Redistributions in binary form must reproduce the above copyright
  10  *    notice, this list of conditions and the following disclaimer in the
  11  *    documentation and/or other materials provided with the distribution.
  12  *
  13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  23  * SUCH DAMAGE.
  24  *
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include "opt_inet.h"
  31 #include "opt_inet6.h"
  32 #include "opt_ipsec.h"
  33 #include "opt_ratelimit.h"
  34 #include "opt_kern_tls.h"
  35 #if defined(INET) || defined(INET6)
  36 #include <sys/param.h>
  37 #include <sys/arb.h>
  38 #include <sys/module.h>
  39 #include <sys/kernel.h>
  40 #ifdef TCP_HHOOK
  41 #include <sys/hhook.h>
  42 #endif
  43 #include <sys/lock.h>
  44 #include <sys/malloc.h>
  45 #include <sys/lock.h>
  46 #include <sys/mutex.h>
  47 #include <sys/mbuf.h>
  48 #include <sys/proc.h>           /* for proc0 declaration */
  49 #include <sys/socket.h>
  50 #include <sys/socketvar.h>
  51 #include <sys/sysctl.h>
  52 #include <sys/systm.h>
  53 #ifdef STATS
  54 #include <sys/qmath.h>
  55 #include <sys/tree.h>
  56 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
  57 #else
  58 #include <sys/tree.h>
  59 #endif
  60 #include <sys/refcount.h>
  61 #include <sys/queue.h>
  62 #include <sys/tim_filter.h>
  63 #include <sys/smp.h>
  64 #include <sys/kthread.h>
  65 #include <sys/kern_prefetch.h>
  66 #include <sys/protosw.h>
  67 #ifdef TCP_ACCOUNTING
  68 #include <sys/sched.h>
  69 #include <machine/cpu.h>
  70 #endif
  71 #include <vm/uma.h>
  72
  73 #include <net/route.h>
  74 #include <net/route/nhop.h>
  75 #include <net/vnet.h>
  76
  77 #define TCPSTATES               /* for logging */
  78
  79 #include <netinet/in.h>
  80 #include <netinet/in_kdtrace.h>
  81 #include <netinet/in_pcb.h>
  82 #include <netinet/ip.h>
  83 #include <netinet/ip_icmp.h>    /* required for icmp_var.h */
  84 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
  85 #include <netinet/ip_var.h>
  86 #include <netinet/ip6.h>
  87 #include <netinet6/in6_pcb.h>
  88 #include <netinet6/ip6_var.h>
  89 #include <netinet/tcp.h>
  90 #define TCPOUTFLAGS
  91 #include <netinet/tcp_fsm.h>
  92 #include <netinet/tcp_seq.h>
  93 #include <netinet/tcp_timer.h>
  94 #include <netinet/tcp_var.h>
  95 #include <netinet/tcp_log_buf.h>
  96 #include <netinet/tcp_syncache.h>
  97 #include <netinet/tcp_hpts.h>
  98 #include <netinet/tcp_ratelimit.h>
  99 #include <netinet/tcp_accounting.h>
 100 #include <netinet/tcpip.h>
 101 #include <netinet/cc/cc.h>
 102 #include <netinet/cc/cc_newreno.h>
 103 #include <netinet/tcp_fastopen.h>
 104 #include <netinet/tcp_lro.h>
 105 #ifdef NETFLIX_SHARED_CWND
 106 #include <netinet/tcp_shared_cwnd.h>
 107 #endif
 108 #ifdef TCP_OFFLOAD
 109 #include <netinet/tcp_offload.h>
 110 #endif
 111 #ifdef INET6
 112 #include <netinet6/tcp6_var.h>
 113 #endif
 114 #include <netinet/tcp_ecn.h>
 115
 116 #include <netipsec/ipsec_support.h>
 117
 118 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 119 #include <netipsec/ipsec.h>
 120 #include <netipsec/ipsec6.h>
 121 #endif                          /* IPSEC */
 122
 123 #include <netinet/udp.h>
 124 #include <netinet/udp_var.h>
 125 #include <machine/in_cksum.h>
 126
 127 #ifdef MAC
 128 #include <security/mac/mac_framework.h>
 129 #endif
 130 #include "sack_filter.h"
 131 #include "tcp_rack.h"
 132 #include "rack_bbr_common.h"
 133
 134 uma_zone_t rack_zone;
 135 uma_zone_t rack_pcb_zone;
 136
 137 #ifndef TICKS2SBT
 138 #define TICKS2SBT(__t)  (tick_sbt * ((sbintime_t)(__t)))
 139 #endif
 140
 141 VNET_DECLARE(uint32_t, newreno_beta);
 142 VNET_DECLARE(uint32_t, newreno_beta_ecn);
 143 #define V_newreno_beta VNET(newreno_beta)
 144 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
 145
 146
 147 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
 148 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
 149
 150 struct sysctl_ctx_list rack_sysctl_ctx;
 151 struct sysctl_oid *rack_sysctl_root;
 152
 153 #define CUM_ACKED 1
 154 #define SACKED 2
 155
 156 /*
 157  * The RACK module incorporates a number of
 158  * TCP ideas that have been put out into the IETF
 159  * over the last few years:
 160  * - Matt Mathis's Rate Halving which slowly drops
 161  *    the congestion window so that the ack clock can
 162  *    be maintained during a recovery.
 163  * - Yuchung Cheng's RACK TCP (for which its named) that
 164  *    will stop us using the number of dup acks and instead
 165  *    use time as the gage of when we retransmit.
 166  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
 167  *    of Dukkipati et.al.
 168  * RACK depends on SACK, so if an endpoint arrives that
 169  * cannot do SACK the state machine below will shuttle the
 170  * connection back to using the "default" TCP stack that is
 171  * in FreeBSD.
 172  *
 173  * To implement RACK the original TCP stack was first decomposed
 174  * into a functional state machine with individual states
 175  * for each of the possible TCP connection states. The do_segment
 176  * functions role in life is to mandate the connection supports SACK
 177  * initially and then assure that the RACK state matches the conenction
 178  * state before calling the states do_segment function. Each
 179  * state is simplified due to the fact that the original do_segment
 180  * has been decomposed and we *know* what state we are in (no
 181  * switches on the state) and all tests for SACK are gone. This
 182  * greatly simplifies what each state does.
 183  *
 184  * TCP output is also over-written with a new version since it
 185  * must maintain the new rack scoreboard.
 186  *
 187  */
 188 static int32_t rack_tlp_thresh = 1;
 189 static int32_t rack_tlp_limit = 2;      /* No more than 2 TLPs w-out new data */
 190 static int32_t rack_tlp_use_greater = 1;
 191 static int32_t rack_reorder_thresh = 2;
 192 static int32_t rack_reorder_fade = 60000000;    /* 0 - never fade, def 60,000,000
 193                                                  * - 60 seconds */
 194 static uint8_t rack_req_measurements = 1;
 195 /* Attack threshold detections */
 196 static uint32_t rack_highest_sack_thresh_seen = 0;
 197 static uint32_t rack_highest_move_thresh_seen = 0;
 198 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
 199 static int32_t rack_hw_pace_extra_slots = 2;    /* 2 extra MSS time betweens */
 200 static int32_t rack_hw_rate_caps = 1; /* 1; */
 201 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
 202 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
 203 static int32_t rack_hw_up_only = 1;
 204 static int32_t rack_stats_gets_ms_rtt = 1;
 205 static int32_t rack_prr_addbackmax = 2;
 206 static int32_t rack_do_hystart = 0;
 207 static int32_t rack_apply_rtt_with_reduced_conf = 0;
 208
 209 static int32_t rack_pkt_delay = 1000;
 210 static int32_t rack_send_a_lot_in_prr = 1;
 211 static int32_t rack_min_to = 1000;      /* Number of microsecond  min timeout */
 212 static int32_t rack_verbose_logging = 0;
 213 static int32_t rack_ignore_data_after_close = 1;
 214 static int32_t rack_enable_shared_cwnd = 1;
 215 static int32_t rack_use_cmp_acks = 1;
 216 static int32_t rack_use_fsb = 1;
 217 static int32_t rack_use_rfo = 1;
 218 static int32_t rack_use_rsm_rfo = 1;
 219 static int32_t rack_max_abc_post_recovery = 2;
 220 static int32_t rack_client_low_buf = 0;
 221 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 */
 222 #ifdef TCP_ACCOUNTING
 223 static int32_t rack_tcp_accounting = 0;
 224 #endif
 225 static int32_t rack_limits_scwnd = 1;
 226 static int32_t rack_enable_mqueue_for_nonpaced = 0;
 227 static int32_t rack_disable_prr = 0;
 228 static int32_t use_rack_rr = 1;
 229 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
 230 static int32_t rack_persist_min = 250000;       /* 250usec */
 231 static int32_t rack_persist_max = 2000000;      /* 2 Second in usec's */
 232 static int32_t rack_sack_not_required = 1;      /* set to one to allow non-sack to use rack */
 233 static int32_t rack_default_init_window = 0;    /* Use system default */
 234 static int32_t rack_limit_time_with_srtt = 0;
 235 static int32_t rack_autosndbuf_inc = 20;        /* In percentage form */
 236 static int32_t rack_enobuf_hw_boost_mult = 2;   /* How many times the hw rate we boost slot using time_between */
 237 static int32_t rack_enobuf_hw_max = 12000;      /* 12 ms in usecs */
 238 static int32_t rack_enobuf_hw_min = 10000;      /* 10 ms in usecs */
 239 static int32_t rack_hw_rwnd_factor = 2;         /* How many max_segs the rwnd must be before we hold off sending */
 240
 241 /*
 242  * Currently regular tcp has a rto_min of 30ms
 243  * the backoff goes 12 times so that ends up
 244  * being a total of 122.850 seconds before a
 245  * connection is killed.
 246  */
 247 static uint32_t rack_def_data_window = 20;
 248 static uint32_t rack_goal_bdp = 2;
 249 static uint32_t rack_min_srtts = 1;
 250 static uint32_t rack_min_measure_usec = 0;
 251 static int32_t rack_tlp_min = 10000;    /* 10ms */
 252 static int32_t rack_rto_min = 30000;    /* 30,000 usec same as main freebsd */
 253 static int32_t rack_rto_max = 4000000;  /* 4 seconds in usec's */
 254 static const int32_t rack_free_cache = 2;
 255 static int32_t rack_hptsi_segments = 40;
 256 static int32_t rack_rate_sample_method = USE_RTT_LOW;
 257 static int32_t rack_pace_every_seg = 0;
 258 static int32_t rack_delayed_ack_time = 40000;   /* 40ms in usecs */
 259 static int32_t rack_slot_reduction = 4;
 260 static int32_t rack_wma_divisor = 8;            /* For WMA calculation */
 261 static int32_t rack_cwnd_block_ends_measure = 0;
 262 static int32_t rack_rwnd_block_ends_measure = 0;
 263 static int32_t rack_def_profile = 0;
 264
 265 static int32_t rack_lower_cwnd_at_tlp = 0;
 266 static int32_t rack_limited_retran = 0;
 267 static int32_t rack_always_send_oldest = 0;
 268 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
 269
 270 static uint16_t rack_per_of_gp_ss = 250;        /* 250 % slow-start */
 271 static uint16_t rack_per_of_gp_ca = 200;        /* 200 % congestion-avoidance */
 272 static uint16_t rack_per_of_gp_rec = 200;       /* 200 % of bw */
 273
 274 /* Probertt */
 275 static uint16_t rack_per_of_gp_probertt = 60;   /* 60% of bw */
 276 static uint16_t rack_per_of_gp_lowthresh = 40;  /* 40% is bottom */
 277 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
 278 static uint16_t rack_atexit_prtt_hbp = 130;     /* Clamp to 130% on exit prtt if highly buffered path */
 279 static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */
 280
 281 static uint32_t rack_max_drain_wait = 2;        /* How man gp srtt's before we give up draining */
 282 static uint32_t rack_must_drain = 1;            /* How many GP srtt's we *must* wait */
 283 static uint32_t rack_probertt_use_min_rtt_entry = 1;    /* Use the min to calculate the goal else gp_srtt */
 284 static uint32_t rack_probertt_use_min_rtt_exit = 0;
 285 static uint32_t rack_probe_rtt_sets_cwnd = 0;
 286 static uint32_t rack_probe_rtt_safety_val = 2000000;    /* No more than 2 sec in probe-rtt */
 287 static uint32_t rack_time_between_probertt = 9600000;   /* 9.6 sec in usecs */
 288 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;       /* How many srtt periods does probe-rtt last top fraction */
 289 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;       /* How many srtt periods does probe-rtt last bottom fraction */
 290 static uint32_t rack_min_probertt_hold = 40000;         /* Equal to delayed ack time */
 291 static uint32_t rack_probertt_filter_life = 10000000;
 292 static uint32_t rack_probertt_lower_within = 10;
 293 static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds)  to count as a lowering */
 294 static int32_t rack_pace_one_seg = 0;           /* Shall we pace for less than 1.4Meg 1MSS at a time */
 295 static int32_t rack_probertt_clear_is = 1;
 296 static int32_t rack_max_drain_hbp = 1;          /* Extra drain times gpsrtt for highly buffered paths */
 297 static int32_t rack_hbp_thresh = 3;             /* what is the divisor max_rtt/min_rtt to decided a hbp */
 298
 299 /* Part of pacing */
 300 static int32_t rack_max_per_above = 30;         /* When we go to increment stop if above 100+this% */
 301
 302 /* Timely information */
 303 /* Combine these two gives the range of 'no change' to bw */
 304 /* ie the up/down provide the upper and lower bound */
 305 static int32_t rack_gp_per_bw_mul_up = 2;       /* 2% */
 306 static int32_t rack_gp_per_bw_mul_down = 4;     /* 4% */
 307 static int32_t rack_gp_rtt_maxmul = 3;          /* 3 x maxmin */
 308 static int32_t rack_gp_rtt_minmul = 1;          /* minrtt + (minrtt/mindiv) is lower rtt */
 309 static int32_t rack_gp_rtt_mindiv = 4;          /* minrtt + (minrtt * minmul/mindiv) is lower rtt */
 310 static int32_t rack_gp_decrease_per = 20;       /* 20% decrease in multiplier */
 311 static int32_t rack_gp_increase_per = 2;        /* 2% increase in multiplier */
 312 static int32_t rack_per_lower_bound = 50;       /* Don't allow to drop below this multiplier */
 313 static int32_t rack_per_upper_bound_ss = 0;     /* Don't allow SS to grow above this */
 314 static int32_t rack_per_upper_bound_ca = 0;     /* Don't allow CA to grow above this */
 315 static int32_t rack_do_dyn_mul = 0;             /* Are the rack gp multipliers dynamic */
 316 static int32_t rack_gp_no_rec_chg = 1;          /* Prohibit recovery from reducing it's multiplier */
 317 static int32_t rack_timely_dec_clear = 6;       /* Do we clear decrement count at a value (6)? */
 318 static int32_t rack_timely_max_push_rise = 3;   /* One round of pushing */
 319 static int32_t rack_timely_max_push_drop = 3;   /* Three round of pushing */
 320 static int32_t rack_timely_min_segs = 4;        /* 4 segment minimum */
 321 static int32_t rack_use_max_for_nobackoff = 0;
 322 static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */
 323 static int32_t rack_timely_no_stopping = 0;
 324 static int32_t rack_down_raise_thresh = 100;
 325 static int32_t rack_req_segs = 1;
 326 static uint64_t rack_bw_rate_cap = 0;
 327
 328
 329 /* Weird delayed ack mode */
 330 static int32_t rack_use_imac_dack = 0;
 331 /* Rack specific counters */
 332 counter_u64_t rack_saw_enobuf;
 333 counter_u64_t rack_saw_enobuf_hw;
 334 counter_u64_t rack_saw_enetunreach;
 335 counter_u64_t rack_persists_sends;
 336 counter_u64_t rack_persists_acks;
 337 counter_u64_t rack_persists_loss;
 338 counter_u64_t rack_persists_lost_ends;
 339 #ifdef INVARIANTS
 340 counter_u64_t rack_adjust_map_bw;
 341 #endif
 342 /* Tail loss probe counters */
 343 counter_u64_t rack_tlp_tot;
 344 counter_u64_t rack_tlp_newdata;
 345 counter_u64_t rack_tlp_retran;
 346 counter_u64_t rack_tlp_retran_bytes;
 347 counter_u64_t rack_to_tot;
 348 counter_u64_t rack_hot_alloc;
 349 counter_u64_t rack_to_alloc;
 350 counter_u64_t rack_to_alloc_hard;
 351 counter_u64_t rack_to_alloc_emerg;
 352 counter_u64_t rack_to_alloc_limited;
 353 counter_u64_t rack_alloc_limited_conns;
 354 counter_u64_t rack_split_limited;
 355
 356 counter_u64_t rack_multi_single_eq;
 357 counter_u64_t rack_proc_non_comp_ack;
 358
 359 counter_u64_t rack_fto_send;
 360 counter_u64_t rack_fto_rsm_send;
 361 counter_u64_t rack_nfto_resend;
 362 counter_u64_t rack_non_fto_send;
 363 counter_u64_t rack_extended_rfo;
 364
 365 counter_u64_t rack_sack_proc_all;
 366 counter_u64_t rack_sack_proc_short;
 367 counter_u64_t rack_sack_proc_restart;
 368 counter_u64_t rack_sack_attacks_detected;
 369 counter_u64_t rack_sack_attacks_reversed;
 370 counter_u64_t rack_sack_used_next_merge;
 371 counter_u64_t rack_sack_splits;
 372 counter_u64_t rack_sack_used_prev_merge;
 373 counter_u64_t rack_sack_skipped_acked;
 374 counter_u64_t rack_ack_total;
 375 counter_u64_t rack_express_sack;
 376 counter_u64_t rack_sack_total;
 377 counter_u64_t rack_move_none;
 378 counter_u64_t rack_move_some;
 379
 380 counter_u64_t rack_input_idle_reduces;
 381 counter_u64_t rack_collapsed_win;
 382 counter_u64_t rack_collapsed_win_seen;
 383 counter_u64_t rack_collapsed_win_rxt;
 384 counter_u64_t rack_collapsed_win_rxt_bytes;
 385 counter_u64_t rack_try_scwnd;
 386 counter_u64_t rack_hw_pace_init_fail;
 387 counter_u64_t rack_hw_pace_lost;
 388
 389 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
 390 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
 391
 392
 393 #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
 394
 395 #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {  \
 396         (tv) = (value) + slop;   \
 397         if ((u_long)(tv) < (u_long)(tvmin)) \
 398                 (tv) = (tvmin); \
 399         if ((u_long)(tv) > (u_long)(tvmax)) \
 400                 (tv) = (tvmax); \
 401 } while (0)
 402
 403 static void
 404 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
 405
 406 static int
 407 rack_process_ack(struct mbuf *m, struct tcphdr *th,
 408     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
 409     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
 410 static int
 411 rack_process_data(struct mbuf *m, struct tcphdr *th,
 412     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 413     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
 414 static void
 415 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
 416    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
 417 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
 418 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
 419     uint8_t limit_type);
 420 static struct rack_sendmap *
 421 rack_check_recovery_mode(struct tcpcb *tp,
 422     uint32_t tsused);
 423 static void
 424 rack_cong_signal(struct tcpcb *tp,
 425                  uint32_t type, uint32_t ack, int );
 426 static void rack_counter_destroy(void);
 427 static int
 428 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt);
 429 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
 430 static void
 431 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
 432 static void
 433 rack_do_segment(struct mbuf *m, struct tcphdr *th,
 434     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 435     uint8_t iptos);
 436 static void rack_dtor(void *mem, int32_t size, void *arg);
 437 static void
 438 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
 439     uint32_t flex1, uint32_t flex2,
 440     uint32_t flex3, uint32_t flex4,
 441     uint32_t flex5, uint32_t flex6,
 442     uint16_t flex7, uint8_t mod);
 443
 444 static void
 445 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
 446    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
 447    struct rack_sendmap *rsm, uint8_t quality);
 448 static struct rack_sendmap *
 449 rack_find_high_nonack(struct tcp_rack *rack,
 450     struct rack_sendmap *rsm);
 451 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
 452 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
 453 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
 454 static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt);
 455 static void
 456 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
 457                             tcp_seq th_ack, int line, uint8_t quality);
 458 static uint32_t
 459 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
 460 static int32_t rack_handoff_ok(struct tcpcb *tp);
 461 static int32_t rack_init(struct tcpcb *tp);
 462 static void rack_init_sysctls(void);
 463 static void
 464 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
 465     struct tcphdr *th, int entered_rec, int dup_ack_struck);
 466 static void
 467 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
 468     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
 469     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
 470
 471 static void
 472 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
 473     struct rack_sendmap *rsm);
 474 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
 475 static int32_t rack_output(struct tcpcb *tp);
 476
 477 static uint32_t
 478 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
 479     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
 480     uint32_t cts, int *moved_two);
 481 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
 482 static void rack_remxt_tmr(struct tcpcb *tp);
 483 static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt);
 484 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
 485 static int32_t rack_stopall(struct tcpcb *tp);
 486 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
 487 static uint32_t
 488 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
 489     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
 490 static void
 491 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
 492     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
 493 static int
 494 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
 495     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
 496 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
 497 static int
 498 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
 499     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 500     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 501 static int
 502 rack_do_closing(struct mbuf *m, struct tcphdr *th,
 503     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 504     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 505 static int
 506 rack_do_established(struct mbuf *m, struct tcphdr *th,
 507     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 508     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 509 static int
 510 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
 511     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 512     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
 513 static int
 514 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
 515     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 516     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 517 static int
 518 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
 519     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 520     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 521 static int
 522 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
 523     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 524     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 525 static int
 526 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
 527     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 528     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 529 static int
 530 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
 531     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 532     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 533 struct rack_sendmap *
 534 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
 535     uint32_t tsused);
 536 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
 537     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
 538 static void
 539      tcp_rack_partialack(struct tcpcb *tp);
 540 static int
 541 rack_set_profile(struct tcp_rack *rack, int prof);
 542 static void
 543 rack_apply_deferred_options(struct tcp_rack *rack);
 544
 545 int32_t rack_clear_counter=0;
 546
 547 static void
 548 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
 549 {
 550         struct sockopt sopt;
 551         struct cc_newreno_opts opt;
 552         struct newreno old;
 553         struct tcpcb *tp;
 554         int error, failed = 0;
 555
 556         tp = rack->rc_tp;
 557         if (tp->t_cc == NULL) {
 558                 /* Tcb is leaving */
 559                 return;
 560         }
 561         rack->rc_pacing_cc_set = 1;
 562         if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
 563                 /* Not new-reno we can't play games with beta! */
 564                 failed = 1;
 565                 goto out;
 566
 567         }
 568         if (CC_ALGO(tp)->ctl_output == NULL)  {
 569                 /* Huh, not using new-reno so no swaps.? */
 570                 failed = 2;
 571                 goto out;
 572         }
 573         /* Get the current values out */
 574         sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
 575         sopt.sopt_dir = SOPT_GET;
 576         opt.name = CC_NEWRENO_BETA;
 577         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 578         if (error)  {
 579                 failed = 3;
 580                 goto out;
 581         }
 582         old.beta = opt.val;
 583         opt.name = CC_NEWRENO_BETA_ECN;
 584         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 585         if (error)  {
 586                 failed = 4;
 587                 goto out;
 588         }
 589         old.beta_ecn = opt.val;
 590
 591         /* Now lets set in the values we have stored */
 592         sopt.sopt_dir = SOPT_SET;
 593         opt.name = CC_NEWRENO_BETA;
 594         opt.val = rack->r_ctl.rc_saved_beta.beta;
 595         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 596         if (error)  {
 597                 failed = 5;
 598                 goto out;
 599         }
 600         opt.name = CC_NEWRENO_BETA_ECN;
 601         opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
 602         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 603         if (error) {
 604                 failed = 6;
 605                 goto out;
 606         }
 607         /* Save off the values for restoral */
 608         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 609 out:
 610         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
 611                 union tcp_log_stackspecific log;
 612                 struct timeval tv;
 613                 struct newreno *ptr;
 614
 615                 ptr = ((struct newreno *)tp->t_ccv.cc_data);
 616                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 617                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 618                 log.u_bbr.flex1 = ptr->beta;
 619                 log.u_bbr.flex2 = ptr->beta_ecn;
 620                 log.u_bbr.flex3 = ptr->newreno_flags;
 621                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 622                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 623                 log.u_bbr.flex6 = failed;
 624                 log.u_bbr.flex7 = rack->gp_ready;
 625                 log.u_bbr.flex7 <<= 1;
 626                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 627                 log.u_bbr.flex7 <<= 1;
 628                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 629                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 630                 log.u_bbr.flex8 = flex8;
 631                 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
 632                                0, &log, false, NULL, NULL, 0, &tv);
 633         }
 634 }
 635
 636 static void
 637 rack_set_cc_pacing(struct tcp_rack *rack)
 638 {
 639         if (rack->rc_pacing_cc_set)
 640                 return;
 641         /*
 642          * Use the swap utility placing in 3 for flex8 to id a
 643          * set of a new set of values.
 644          */
 645         rack->rc_pacing_cc_set = 1;
 646         rack_swap_beta_values(rack, 3);
 647 }
 648
 649 static void
 650 rack_undo_cc_pacing(struct tcp_rack *rack)
 651 {
 652         if (rack->rc_pacing_cc_set == 0)
 653                 return;
 654         /*
 655          * Use the swap utility placing in 4 for flex8 to id a
 656          * restoral of the old values.
 657          */
 658         rack->rc_pacing_cc_set = 0;
 659         rack_swap_beta_values(rack, 4);
 660 }
 661
 662 #ifdef NETFLIX_PEAKRATE
 663 static inline void
 664 rack_update_peakrate_thr(struct tcpcb *tp)
 665 {
 666         /* Keep in mind that t_maxpeakrate is in B/s. */
 667         uint64_t peak;
 668         peak = uqmax((tp->t_maxseg * 2),
 669                      (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
 670         tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
 671 }
 672 #endif
 673
 674 static int
 675 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
 676 {
 677         uint32_t stat;
 678         int32_t error;
 679
 680         error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
 681         if (error || req->newptr == NULL)
 682                 return error;
 683
 684         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
 685         if (error)
 686                 return (error);
 687         if (stat == 1) {
 688 #ifdef INVARIANTS
 689                 printf("Clearing RACK counters\n");
 690 #endif
 691                 counter_u64_zero(rack_tlp_tot);
 692                 counter_u64_zero(rack_tlp_newdata);
 693                 counter_u64_zero(rack_tlp_retran);
 694                 counter_u64_zero(rack_tlp_retran_bytes);
 695                 counter_u64_zero(rack_to_tot);
 696                 counter_u64_zero(rack_saw_enobuf);
 697                 counter_u64_zero(rack_saw_enobuf_hw);
 698                 counter_u64_zero(rack_saw_enetunreach);
 699                 counter_u64_zero(rack_persists_sends);
 700                 counter_u64_zero(rack_persists_acks);
 701                 counter_u64_zero(rack_persists_loss);
 702                 counter_u64_zero(rack_persists_lost_ends);
 703 #ifdef INVARIANTS
 704                 counter_u64_zero(rack_adjust_map_bw);
 705 #endif
 706                 counter_u64_zero(rack_to_alloc_hard);
 707                 counter_u64_zero(rack_to_alloc_emerg);
 708                 counter_u64_zero(rack_sack_proc_all);
 709                 counter_u64_zero(rack_fto_send);
 710                 counter_u64_zero(rack_fto_rsm_send);
 711                 counter_u64_zero(rack_extended_rfo);
 712                 counter_u64_zero(rack_hw_pace_init_fail);
 713                 counter_u64_zero(rack_hw_pace_lost);
 714                 counter_u64_zero(rack_non_fto_send);
 715                 counter_u64_zero(rack_nfto_resend);
 716                 counter_u64_zero(rack_sack_proc_short);
 717                 counter_u64_zero(rack_sack_proc_restart);
 718                 counter_u64_zero(rack_to_alloc);
 719                 counter_u64_zero(rack_to_alloc_limited);
 720                 counter_u64_zero(rack_alloc_limited_conns);
 721                 counter_u64_zero(rack_split_limited);
 722                 counter_u64_zero(rack_multi_single_eq);
 723                 counter_u64_zero(rack_proc_non_comp_ack);
 724                 counter_u64_zero(rack_sack_attacks_detected);
 725                 counter_u64_zero(rack_sack_attacks_reversed);
 726                 counter_u64_zero(rack_sack_used_next_merge);
 727                 counter_u64_zero(rack_sack_used_prev_merge);
 728                 counter_u64_zero(rack_sack_splits);
 729                 counter_u64_zero(rack_sack_skipped_acked);
 730                 counter_u64_zero(rack_ack_total);
 731                 counter_u64_zero(rack_express_sack);
 732                 counter_u64_zero(rack_sack_total);
 733                 counter_u64_zero(rack_move_none);
 734                 counter_u64_zero(rack_move_some);
 735                 counter_u64_zero(rack_try_scwnd);
 736                 counter_u64_zero(rack_collapsed_win);
 737                 counter_u64_zero(rack_collapsed_win_rxt);
 738                 counter_u64_zero(rack_collapsed_win_seen);
 739                 counter_u64_zero(rack_collapsed_win_rxt_bytes);
 740         }
 741         rack_clear_counter = 0;
 742         return (0);
 743 }
 744
 745 static void
 746 rack_init_sysctls(void)
 747 {
 748         struct sysctl_oid *rack_counters;
 749         struct sysctl_oid *rack_attack;
 750         struct sysctl_oid *rack_pacing;
 751         struct sysctl_oid *rack_timely;
 752         struct sysctl_oid *rack_timers;
 753         struct sysctl_oid *rack_tlp;
 754         struct sysctl_oid *rack_misc;
 755         struct sysctl_oid *rack_features;
 756         struct sysctl_oid *rack_measure;
 757         struct sysctl_oid *rack_probertt;
 758         struct sysctl_oid *rack_hw_pacing;
 759
 760         rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 761             SYSCTL_CHILDREN(rack_sysctl_root),
 762             OID_AUTO,
 763             "sack_attack",
 764             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 765             "Rack Sack Attack Counters and Controls");
 766         rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 767             SYSCTL_CHILDREN(rack_sysctl_root),
 768             OID_AUTO,
 769             "stats",
 770             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 771             "Rack Counters");
 772         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 773             SYSCTL_CHILDREN(rack_sysctl_root),
 774             OID_AUTO, "rate_sample_method", CTLFLAG_RW,
 775             &rack_rate_sample_method , USE_RTT_LOW,
 776             "What method should we use for rate sampling 0=high, 1=low ");
 777         /* Probe rtt related controls */
 778         rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 779             SYSCTL_CHILDREN(rack_sysctl_root),
 780             OID_AUTO,
 781             "probertt",
 782             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 783             "ProbeRTT related Controls");
 784         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 785             SYSCTL_CHILDREN(rack_probertt),
 786             OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
 787             &rack_atexit_prtt_hbp, 130,
 788             "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
 789         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 790             SYSCTL_CHILDREN(rack_probertt),
 791             OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
 792             &rack_atexit_prtt, 130,
 793             "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
 794         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 795             SYSCTL_CHILDREN(rack_probertt),
 796             OID_AUTO, "gp_per_mul", CTLFLAG_RW,
 797             &rack_per_of_gp_probertt, 60,
 798             "What percentage of goodput do we pace at in probertt");
 799         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 800             SYSCTL_CHILDREN(rack_probertt),
 801             OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
 802             &rack_per_of_gp_probertt_reduce, 10,
 803             "What percentage of goodput do we reduce every gp_srtt");
 804         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 805             SYSCTL_CHILDREN(rack_probertt),
 806             OID_AUTO, "gp_per_low", CTLFLAG_RW,
 807             &rack_per_of_gp_lowthresh, 40,
 808             "What percentage of goodput do we allow the multiplier to fall to");
 809         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 810             SYSCTL_CHILDREN(rack_probertt),
 811             OID_AUTO, "time_between", CTLFLAG_RW,
 812             & rack_time_between_probertt, 96000000,
 813             "How many useconds between the lowest rtt falling must past before we enter probertt");
 814         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 815             SYSCTL_CHILDREN(rack_probertt),
 816             OID_AUTO, "safety", CTLFLAG_RW,
 817             &rack_probe_rtt_safety_val, 2000000,
 818             "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
 819         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 820             SYSCTL_CHILDREN(rack_probertt),
 821             OID_AUTO, "sets_cwnd", CTLFLAG_RW,
 822             &rack_probe_rtt_sets_cwnd, 0,
 823             "Do we set the cwnd too (if always_lower is on)");
 824         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 825             SYSCTL_CHILDREN(rack_probertt),
 826             OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
 827             &rack_max_drain_wait, 2,
 828             "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
 829         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 830             SYSCTL_CHILDREN(rack_probertt),
 831             OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
 832             &rack_must_drain, 1,
 833             "We must drain this many gp_srtt's waiting for flight to reach goal");
 834         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 835             SYSCTL_CHILDREN(rack_probertt),
 836             OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
 837             &rack_probertt_use_min_rtt_entry, 1,
 838             "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
 839         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 840             SYSCTL_CHILDREN(rack_probertt),
 841             OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
 842             &rack_probertt_use_min_rtt_exit, 0,
 843             "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
 844         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 845             SYSCTL_CHILDREN(rack_probertt),
 846             OID_AUTO, "length_div", CTLFLAG_RW,
 847             &rack_probertt_gpsrtt_cnt_div, 0,
 848             "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
 849         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 850             SYSCTL_CHILDREN(rack_probertt),
 851             OID_AUTO, "length_mul", CTLFLAG_RW,
 852             &rack_probertt_gpsrtt_cnt_mul, 0,
 853             "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
 854         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 855             SYSCTL_CHILDREN(rack_probertt),
 856             OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
 857             &rack_min_probertt_hold, 200000,
 858             "What is the minimum time we hold probertt at target");
 859         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 860             SYSCTL_CHILDREN(rack_probertt),
 861             OID_AUTO, "filter_life", CTLFLAG_RW,
 862             &rack_probertt_filter_life, 10000000,
 863             "What is the time for the filters life in useconds");
 864         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 865             SYSCTL_CHILDREN(rack_probertt),
 866             OID_AUTO, "lower_within", CTLFLAG_RW,
 867             &rack_probertt_lower_within, 10,
 868             "If the rtt goes lower within this percentage of the time, go into probe-rtt");
 869         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 870             SYSCTL_CHILDREN(rack_probertt),
 871             OID_AUTO, "must_move", CTLFLAG_RW,
 872             &rack_min_rtt_movement, 250,
 873             "How much is the minimum movement in rtt to count as a drop for probertt purposes");
 874         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 875             SYSCTL_CHILDREN(rack_probertt),
 876             OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
 877             &rack_probertt_clear_is, 1,
 878             "Do we clear I/S counts on exiting probe-rtt");
 879         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 880             SYSCTL_CHILDREN(rack_probertt),
 881             OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
 882             &rack_max_drain_hbp, 1,
 883             "How many extra drain gpsrtt's do we get in highly buffered paths");
 884         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 885             SYSCTL_CHILDREN(rack_probertt),
 886             OID_AUTO, "hbp_threshold", CTLFLAG_RW,
 887             &rack_hbp_thresh, 3,
 888             "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
 889         /* Pacing related sysctls */
 890         rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 891             SYSCTL_CHILDREN(rack_sysctl_root),
 892             OID_AUTO,
 893             "pacing",
 894             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 895             "Pacing related Controls");
 896         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 897             SYSCTL_CHILDREN(rack_pacing),
 898             OID_AUTO, "max_pace_over", CTLFLAG_RW,
 899             &rack_max_per_above, 30,
 900             "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
 901         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 902             SYSCTL_CHILDREN(rack_pacing),
 903             OID_AUTO, "pace_to_one", CTLFLAG_RW,
 904             &rack_pace_one_seg, 0,
 905             "Do we allow low b/w pacing of 1MSS instead of two");
 906         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 907             SYSCTL_CHILDREN(rack_pacing),
 908             OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
 909             &rack_limit_time_with_srtt, 0,
 910             "Do we limit pacing time based on srtt");
 911         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 912             SYSCTL_CHILDREN(rack_pacing),
 913             OID_AUTO, "init_win", CTLFLAG_RW,
 914             &rack_default_init_window, 0,
 915             "Do we have a rack initial window 0 = system default");
 916         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 917             SYSCTL_CHILDREN(rack_pacing),
 918             OID_AUTO, "gp_per_ss", CTLFLAG_RW,
 919             &rack_per_of_gp_ss, 250,
 920             "If non zero, what percentage of goodput to pace at in slow start");
 921         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 922             SYSCTL_CHILDREN(rack_pacing),
 923             OID_AUTO, "gp_per_ca", CTLFLAG_RW,
 924             &rack_per_of_gp_ca, 150,
 925             "If non zero, what percentage of goodput to pace at in congestion avoidance");
 926         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 927             SYSCTL_CHILDREN(rack_pacing),
 928             OID_AUTO, "gp_per_rec", CTLFLAG_RW,
 929             &rack_per_of_gp_rec, 200,
 930             "If non zero, what percentage of goodput to pace at in recovery");
 931         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 932             SYSCTL_CHILDREN(rack_pacing),
 933             OID_AUTO, "pace_max_seg", CTLFLAG_RW,
 934             &rack_hptsi_segments, 40,
 935             "What size is the max for TSO segments in pacing and burst mitigation");
 936         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 937             SYSCTL_CHILDREN(rack_pacing),
 938             OID_AUTO, "burst_reduces", CTLFLAG_RW,
 939             &rack_slot_reduction, 4,
 940             "When doing only burst mitigation what is the reduce divisor");
 941         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 942             SYSCTL_CHILDREN(rack_sysctl_root),
 943             OID_AUTO, "use_pacing", CTLFLAG_RW,
 944             &rack_pace_every_seg, 0,
 945             "If set we use pacing, if clear we use only the original burst mitigation");
 946         SYSCTL_ADD_U64(&rack_sysctl_ctx,
 947             SYSCTL_CHILDREN(rack_pacing),
 948             OID_AUTO, "rate_cap", CTLFLAG_RW,
 949             &rack_bw_rate_cap, 0,
 950             "If set we apply this value to the absolute rate cap used by pacing");
 951         SYSCTL_ADD_U8(&rack_sysctl_ctx,
 952             SYSCTL_CHILDREN(rack_sysctl_root),
 953             OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
 954             &rack_req_measurements, 1,
 955             "If doing dynamic pacing, how many measurements must be in before we start pacing?");
 956         /* Hardware pacing */
 957         rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 958             SYSCTL_CHILDREN(rack_sysctl_root),
 959             OID_AUTO,
 960             "hdwr_pacing",
 961             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 962             "Pacing related Controls");
 963         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 964             SYSCTL_CHILDREN(rack_hw_pacing),
 965             OID_AUTO, "rwnd_factor", CTLFLAG_RW,
 966             &rack_hw_rwnd_factor, 2,
 967             "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
 968         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 969             SYSCTL_CHILDREN(rack_hw_pacing),
 970             OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
 971             &rack_enobuf_hw_boost_mult, 2,
 972             "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
 973         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 974             SYSCTL_CHILDREN(rack_hw_pacing),
 975             OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
 976             &rack_enobuf_hw_max, 2,
 977             "What is the max boost the pacing time if we see a ENOBUFS?");
 978         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 979             SYSCTL_CHILDREN(rack_hw_pacing),
 980             OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
 981             &rack_enobuf_hw_min, 2,
 982             "What is the min boost the pacing time if we see a ENOBUFS?");
 983         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 984             SYSCTL_CHILDREN(rack_hw_pacing),
 985             OID_AUTO, "enable", CTLFLAG_RW,
 986             &rack_enable_hw_pacing, 0,
 987             "Should RACK attempt to use hw pacing?");
 988         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 989             SYSCTL_CHILDREN(rack_hw_pacing),
 990             OID_AUTO, "rate_cap", CTLFLAG_RW,
 991             &rack_hw_rate_caps, 1,
 992             "Does the highest hardware pacing rate cap the rate we will send at??");
 993         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 994             SYSCTL_CHILDREN(rack_hw_pacing),
 995             OID_AUTO, "rate_min", CTLFLAG_RW,
 996             &rack_hw_rate_min, 0,
 997             "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
 998         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 999             SYSCTL_CHILDREN(rack_hw_pacing),
1000             OID_AUTO, "rate_to_low", CTLFLAG_RW,
1001             &rack_hw_rate_to_low, 0,
1002             "If we fall below this rate, dis-engage hw pacing?");
1003         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1004             SYSCTL_CHILDREN(rack_hw_pacing),
1005             OID_AUTO, "up_only", CTLFLAG_RW,
1006             &rack_hw_up_only, 1,
1007             "Do we allow hw pacing to lower the rate selected?");
1008         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1009             SYSCTL_CHILDREN(rack_hw_pacing),
1010             OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1011             &rack_hw_pace_extra_slots, 2,
1012             "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1013         rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1014             SYSCTL_CHILDREN(rack_sysctl_root),
1015             OID_AUTO,
1016             "timely",
1017             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1018             "Rack Timely RTT Controls");
1019         /* Timely based GP dynmics */
1020         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1021             SYSCTL_CHILDREN(rack_timely),
1022             OID_AUTO, "upper", CTLFLAG_RW,
1023             &rack_gp_per_bw_mul_up, 2,
1024             "Rack timely upper range for equal b/w (in percentage)");
1025         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1026             SYSCTL_CHILDREN(rack_timely),
1027             OID_AUTO, "lower", CTLFLAG_RW,
1028             &rack_gp_per_bw_mul_down, 4,
1029             "Rack timely lower range for equal b/w (in percentage)");
1030         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1031             SYSCTL_CHILDREN(rack_timely),
1032             OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1033             &rack_gp_rtt_maxmul, 3,
1034             "Rack timely multiplier of lowest rtt for rtt_max");
1035         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1036             SYSCTL_CHILDREN(rack_timely),
1037             OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1038             &rack_gp_rtt_mindiv, 4,
1039             "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1040         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1041             SYSCTL_CHILDREN(rack_timely),
1042             OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1043             &rack_gp_rtt_minmul, 1,
1044             "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1045         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1046             SYSCTL_CHILDREN(rack_timely),
1047             OID_AUTO, "decrease", CTLFLAG_RW,
1048             &rack_gp_decrease_per, 20,
1049             "Rack timely decrease percentage of our GP multiplication factor");
1050         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1051             SYSCTL_CHILDREN(rack_timely),
1052             OID_AUTO, "increase", CTLFLAG_RW,
1053             &rack_gp_increase_per, 2,
1054             "Rack timely increase perentage of our GP multiplication factor");
1055         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1056             SYSCTL_CHILDREN(rack_timely),
1057             OID_AUTO, "lowerbound", CTLFLAG_RW,
1058             &rack_per_lower_bound, 50,
1059             "Rack timely lowest percentage we allow GP multiplier to fall to");
1060         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1061             SYSCTL_CHILDREN(rack_timely),
1062             OID_AUTO, "upperboundss", CTLFLAG_RW,
1063             &rack_per_upper_bound_ss, 0,
1064             "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1065         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1066             SYSCTL_CHILDREN(rack_timely),
1067             OID_AUTO, "upperboundca", CTLFLAG_RW,
1068             &rack_per_upper_bound_ca, 0,
1069             "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1070         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1071             SYSCTL_CHILDREN(rack_timely),
1072             OID_AUTO, "dynamicgp", CTLFLAG_RW,
1073             &rack_do_dyn_mul, 0,
1074             "Rack timely do we enable dynmaic timely goodput by default");
1075         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1076             SYSCTL_CHILDREN(rack_timely),
1077             OID_AUTO, "no_rec_red", CTLFLAG_RW,
1078             &rack_gp_no_rec_chg, 1,
1079             "Rack timely do we prohibit the recovery multiplier from being lowered");
1080         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1081             SYSCTL_CHILDREN(rack_timely),
1082             OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1083             &rack_timely_dec_clear, 6,
1084             "Rack timely what threshold do we count to before another boost during b/w decent");
1085         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1086             SYSCTL_CHILDREN(rack_timely),
1087             OID_AUTO, "max_push_rise", CTLFLAG_RW,
1088             &rack_timely_max_push_rise, 3,
1089             "Rack timely how many times do we push up with b/w increase");
1090         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1091             SYSCTL_CHILDREN(rack_timely),
1092             OID_AUTO, "max_push_drop", CTLFLAG_RW,
1093             &rack_timely_max_push_drop, 3,
1094             "Rack timely how many times do we push back on b/w decent");
1095         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1096             SYSCTL_CHILDREN(rack_timely),
1097             OID_AUTO, "min_segs", CTLFLAG_RW,
1098             &rack_timely_min_segs, 4,
1099             "Rack timely when setting the cwnd what is the min num segments");
1100         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1101             SYSCTL_CHILDREN(rack_timely),
1102             OID_AUTO, "noback_max", CTLFLAG_RW,
1103             &rack_use_max_for_nobackoff, 0,
1104             "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1105         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1106             SYSCTL_CHILDREN(rack_timely),
1107             OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1108             &rack_timely_int_timely_only, 0,
1109             "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1110         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1111             SYSCTL_CHILDREN(rack_timely),
1112             OID_AUTO, "nonstop", CTLFLAG_RW,
1113             &rack_timely_no_stopping, 0,
1114             "Rack timely don't stop increase");
1115         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1116             SYSCTL_CHILDREN(rack_timely),
1117             OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1118             &rack_down_raise_thresh, 100,
1119             "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1120         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1121             SYSCTL_CHILDREN(rack_timely),
1122             OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1123             &rack_req_segs, 1,
1124             "Bottom dragging if not these many segments outstanding and room");
1125
1126         /* TLP and Rack related parameters */
1127         rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1128             SYSCTL_CHILDREN(rack_sysctl_root),
1129             OID_AUTO,
1130             "tlp",
1131             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1132             "TLP and Rack related Controls");
1133         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1134             SYSCTL_CHILDREN(rack_tlp),
1135             OID_AUTO, "use_rrr", CTLFLAG_RW,
1136             &use_rack_rr, 1,
1137             "Do we use Rack Rapid Recovery");
1138         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1139             SYSCTL_CHILDREN(rack_tlp),
1140             OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1141             &rack_max_abc_post_recovery, 2,
1142             "Since we do early recovery, do we override the l_abc to a value, if so what?");
1143         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1144             SYSCTL_CHILDREN(rack_tlp),
1145             OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1146             &rack_non_rxt_use_cr, 0,
1147             "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1148         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1149             SYSCTL_CHILDREN(rack_tlp),
1150             OID_AUTO, "tlpmethod", CTLFLAG_RW,
1151             &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1152             "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1153         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1154             SYSCTL_CHILDREN(rack_tlp),
1155             OID_AUTO, "limit", CTLFLAG_RW,
1156             &rack_tlp_limit, 2,
1157             "How many TLP's can be sent without sending new data");
1158         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1159             SYSCTL_CHILDREN(rack_tlp),
1160             OID_AUTO, "use_greater", CTLFLAG_RW,
1161             &rack_tlp_use_greater, 1,
1162             "Should we use the rack_rtt time if its greater than srtt");
1163         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1164             SYSCTL_CHILDREN(rack_tlp),
1165             OID_AUTO, "tlpminto", CTLFLAG_RW,
1166             &rack_tlp_min, 10000,
1167             "TLP minimum timeout per the specification (in microseconds)");
1168         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1169             SYSCTL_CHILDREN(rack_tlp),
1170             OID_AUTO, "send_oldest", CTLFLAG_RW,
1171             &rack_always_send_oldest, 0,
1172             "Should we always send the oldest TLP and RACK-TLP");
1173         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1174             SYSCTL_CHILDREN(rack_tlp),
1175             OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1176             &rack_limited_retran, 0,
1177             "How many times can a rack timeout drive out sends");
1178         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1179             SYSCTL_CHILDREN(rack_tlp),
1180             OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1181             &rack_lower_cwnd_at_tlp, 0,
1182             "When a TLP completes a retran should we enter recovery");
1183         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1184             SYSCTL_CHILDREN(rack_tlp),
1185             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1186             &rack_reorder_thresh, 2,
1187             "What factor for rack will be added when seeing reordering (shift right)");
1188         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1189             SYSCTL_CHILDREN(rack_tlp),
1190             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1191             &rack_tlp_thresh, 1,
1192             "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1193         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1194             SYSCTL_CHILDREN(rack_tlp),
1195             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1196             &rack_reorder_fade, 60000000,
1197             "Does reorder detection fade, if so how many microseconds (0 means never)");
1198         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1199             SYSCTL_CHILDREN(rack_tlp),
1200             OID_AUTO, "pktdelay", CTLFLAG_RW,
1201             &rack_pkt_delay, 1000,
1202             "Extra RACK time (in microseconds) besides reordering thresh");
1203
1204         /* Timer related controls */
1205         rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1206             SYSCTL_CHILDREN(rack_sysctl_root),
1207             OID_AUTO,
1208             "timers",
1209             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1210             "Timer related controls");
1211         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1212             SYSCTL_CHILDREN(rack_timers),
1213             OID_AUTO, "persmin", CTLFLAG_RW,
1214             &rack_persist_min, 250000,
1215             "What is the minimum time in microseconds between persists");
1216         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1217             SYSCTL_CHILDREN(rack_timers),
1218             OID_AUTO, "persmax", CTLFLAG_RW,
1219             &rack_persist_max, 2000000,
1220             "What is the largest delay in microseconds between persists");
1221         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1222             SYSCTL_CHILDREN(rack_timers),
1223             OID_AUTO, "delayed_ack", CTLFLAG_RW,
1224             &rack_delayed_ack_time, 40000,
1225             "Delayed ack time (40ms in microseconds)");
1226         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1227             SYSCTL_CHILDREN(rack_timers),
1228             OID_AUTO, "minrto", CTLFLAG_RW,
1229             &rack_rto_min, 30000,
1230             "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1231         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1232             SYSCTL_CHILDREN(rack_timers),
1233             OID_AUTO, "maxrto", CTLFLAG_RW,
1234             &rack_rto_max, 4000000,
1235             "Maximum RTO in microseconds -- should be at least as large as min_rto");
1236         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1237             SYSCTL_CHILDREN(rack_timers),
1238             OID_AUTO, "minto", CTLFLAG_RW,
1239             &rack_min_to, 1000,
1240             "Minimum rack timeout in microseconds");
1241         /* Measure controls */
1242         rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1243             SYSCTL_CHILDREN(rack_sysctl_root),
1244             OID_AUTO,
1245             "measure",
1246             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1247             "Measure related controls");
1248         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1249             SYSCTL_CHILDREN(rack_measure),
1250             OID_AUTO, "wma_divisor", CTLFLAG_RW,
1251             &rack_wma_divisor, 8,
1252             "When doing b/w calculation what is the  divisor for the WMA");
1253         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1254             SYSCTL_CHILDREN(rack_measure),
1255             OID_AUTO, "end_cwnd", CTLFLAG_RW,
1256             &rack_cwnd_block_ends_measure, 0,
1257             "Does a cwnd just-return end the measurement window (app limited)");
1258         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1259             SYSCTL_CHILDREN(rack_measure),
1260             OID_AUTO, "end_rwnd", CTLFLAG_RW,
1261             &rack_rwnd_block_ends_measure, 0,
1262             "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1263         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1264             SYSCTL_CHILDREN(rack_measure),
1265             OID_AUTO, "min_target", CTLFLAG_RW,
1266             &rack_def_data_window, 20,
1267             "What is the minimum target window (in mss) for a GP measurements");
1268         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1269             SYSCTL_CHILDREN(rack_measure),
1270             OID_AUTO, "goal_bdp", CTLFLAG_RW,
1271             &rack_goal_bdp, 2,
1272             "What is the goal BDP to measure");
1273         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1274             SYSCTL_CHILDREN(rack_measure),
1275             OID_AUTO, "min_srtts", CTLFLAG_RW,
1276             &rack_min_srtts, 1,
1277             "What is the goal BDP to measure");
1278         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1279             SYSCTL_CHILDREN(rack_measure),
1280             OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1281             &rack_min_measure_usec, 0,
1282             "What is the Minimum time time for a measurement if 0, this is off");
1283         /* Features */
1284         rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1285             SYSCTL_CHILDREN(rack_sysctl_root),
1286             OID_AUTO,
1287             "features",
1288             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1289             "Feature controls");
1290         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1291             SYSCTL_CHILDREN(rack_features),
1292             OID_AUTO, "cmpack", CTLFLAG_RW,
1293             &rack_use_cmp_acks, 1,
1294             "Should RACK have LRO send compressed acks");
1295         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1296             SYSCTL_CHILDREN(rack_features),
1297             OID_AUTO, "fsb", CTLFLAG_RW,
1298             &rack_use_fsb, 1,
1299             "Should RACK use the fast send block?");
1300         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1301             SYSCTL_CHILDREN(rack_features),
1302             OID_AUTO, "rfo", CTLFLAG_RW,
1303             &rack_use_rfo, 1,
1304             "Should RACK use rack_fast_output()?");
1305         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1306             SYSCTL_CHILDREN(rack_features),
1307             OID_AUTO, "rsmrfo", CTLFLAG_RW,
1308             &rack_use_rsm_rfo, 1,
1309             "Should RACK use rack_fast_rsm_output()?");
1310         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1311             SYSCTL_CHILDREN(rack_features),
1312             OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1313             &rack_enable_mqueue_for_nonpaced, 0,
1314             "Should RACK use mbuf queuing for non-paced connections");
1315         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1316             SYSCTL_CHILDREN(rack_features),
1317             OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1318             &rack_do_hystart, 0,
1319             "Should RACK enable HyStart++ on connections?");
1320         /* Misc rack controls */
1321         rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1322             SYSCTL_CHILDREN(rack_sysctl_root),
1323             OID_AUTO,
1324             "misc",
1325             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1326             "Misc related controls");
1327 #ifdef TCP_ACCOUNTING
1328         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1329             SYSCTL_CHILDREN(rack_misc),
1330             OID_AUTO, "tcp_acct", CTLFLAG_RW,
1331             &rack_tcp_accounting, 0,
1332             "Should we turn on TCP accounting for all rack sessions?");
1333 #endif
1334         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1335             SYSCTL_CHILDREN(rack_misc),
1336             OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1337             &rack_apply_rtt_with_reduced_conf, 0,
1338             "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1339         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1340             SYSCTL_CHILDREN(rack_misc),
1341             OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1342             &rack_dsack_std_based, 3,
1343             "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1344         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1345             SYSCTL_CHILDREN(rack_misc),
1346             OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1347             &rack_prr_addbackmax, 2,
1348             "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1349         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1350             SYSCTL_CHILDREN(rack_misc),
1351             OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1352             &rack_stats_gets_ms_rtt, 1,
1353             "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1354         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1355             SYSCTL_CHILDREN(rack_misc),
1356             OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1357             &rack_client_low_buf, 0,
1358             "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1359         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1360             SYSCTL_CHILDREN(rack_misc),
1361             OID_AUTO, "defprofile", CTLFLAG_RW,
1362             &rack_def_profile, 0,
1363             "Should RACK use a default profile (0=no, num == profile num)?");
1364         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1365             SYSCTL_CHILDREN(rack_misc),
1366             OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1367             &rack_enable_shared_cwnd, 1,
1368             "Should RACK try to use the shared cwnd on connections where allowed");
1369         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1370             SYSCTL_CHILDREN(rack_misc),
1371             OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1372             &rack_limits_scwnd, 1,
1373             "Should RACK place low end time limits on the shared cwnd feature");
1374         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1375             SYSCTL_CHILDREN(rack_misc),
1376             OID_AUTO, "iMac_dack", CTLFLAG_RW,
1377             &rack_use_imac_dack, 0,
1378             "Should RACK try to emulate iMac delayed ack");
1379         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1380             SYSCTL_CHILDREN(rack_misc),
1381             OID_AUTO, "no_prr", CTLFLAG_RW,
1382             &rack_disable_prr, 0,
1383             "Should RACK not use prr and only pace (must have pacing on)");
1384         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1385             SYSCTL_CHILDREN(rack_misc),
1386             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1387             &rack_verbose_logging, 0,
1388             "Should RACK black box logging be verbose");
1389         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1390             SYSCTL_CHILDREN(rack_misc),
1391             OID_AUTO, "data_after_close", CTLFLAG_RW,
1392             &rack_ignore_data_after_close, 1,
1393             "Do we hold off sending a RST until all pending data is ack'd");
1394         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1395             SYSCTL_CHILDREN(rack_misc),
1396             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1397             &rack_sack_not_required, 1,
1398             "Do we allow rack to run on connections not supporting SACK");
1399         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1400             SYSCTL_CHILDREN(rack_misc),
1401             OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1402             &rack_send_a_lot_in_prr, 1,
1403             "Send a lot in prr");
1404         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1405             SYSCTL_CHILDREN(rack_misc),
1406             OID_AUTO, "autoscale", CTLFLAG_RW,
1407             &rack_autosndbuf_inc, 20,
1408             "What percentage should rack scale up its snd buffer by?");
1409         /* Sack Attacker detection stuff */
1410         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1411             SYSCTL_CHILDREN(rack_attack),
1412             OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1413             &rack_highest_sack_thresh_seen, 0,
1414             "Highest sack to ack ratio seen");
1415         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1416             SYSCTL_CHILDREN(rack_attack),
1417             OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1418             &rack_highest_move_thresh_seen, 0,
1419             "Highest move to non-move ratio seen");
1420         rack_ack_total = counter_u64_alloc(M_WAITOK);
1421         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1422             SYSCTL_CHILDREN(rack_attack),
1423             OID_AUTO, "acktotal", CTLFLAG_RD,
1424             &rack_ack_total,
1425             "Total number of Ack's");
1426         rack_express_sack = counter_u64_alloc(M_WAITOK);
1427         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1428             SYSCTL_CHILDREN(rack_attack),
1429             OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1430             &rack_express_sack,
1431             "Total expresss number of Sack's");
1432         rack_sack_total = counter_u64_alloc(M_WAITOK);
1433         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1434             SYSCTL_CHILDREN(rack_attack),
1435             OID_AUTO, "sacktotal", CTLFLAG_RD,
1436             &rack_sack_total,
1437             "Total number of SACKs");
1438         rack_move_none = counter_u64_alloc(M_WAITOK);
1439         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1440             SYSCTL_CHILDREN(rack_attack),
1441             OID_AUTO, "move_none", CTLFLAG_RD,
1442             &rack_move_none,
1443             "Total number of SACK index reuse of positions under threshold");
1444         rack_move_some = counter_u64_alloc(M_WAITOK);
1445         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1446             SYSCTL_CHILDREN(rack_attack),
1447             OID_AUTO, "move_some", CTLFLAG_RD,
1448             &rack_move_some,
1449             "Total number of SACK index reuse of positions over threshold");
1450         rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1451         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1452             SYSCTL_CHILDREN(rack_attack),
1453             OID_AUTO, "attacks", CTLFLAG_RD,
1454             &rack_sack_attacks_detected,
1455             "Total number of SACK attackers that had sack disabled");
1456         rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1457         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1458             SYSCTL_CHILDREN(rack_attack),
1459             OID_AUTO, "reversed", CTLFLAG_RD,
1460             &rack_sack_attacks_reversed,
1461             "Total number of SACK attackers that were later determined false positive");
1462         rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1463         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1464             SYSCTL_CHILDREN(rack_attack),
1465             OID_AUTO, "nextmerge", CTLFLAG_RD,
1466             &rack_sack_used_next_merge,
1467             "Total number of times we used the next merge");
1468         rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1469         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1470             SYSCTL_CHILDREN(rack_attack),
1471             OID_AUTO, "prevmerge", CTLFLAG_RD,
1472             &rack_sack_used_prev_merge,
1473             "Total number of times we used the prev merge");
1474         /* Counters */
1475         rack_fto_send = counter_u64_alloc(M_WAITOK);
1476         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1477             SYSCTL_CHILDREN(rack_counters),
1478             OID_AUTO, "fto_send", CTLFLAG_RD,
1479             &rack_fto_send, "Total number of rack_fast_output sends");
1480         rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1481         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1482             SYSCTL_CHILDREN(rack_counters),
1483             OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1484             &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1485         rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1486         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1487             SYSCTL_CHILDREN(rack_counters),
1488             OID_AUTO, "nfto_resend", CTLFLAG_RD,
1489             &rack_nfto_resend, "Total number of rack_output retransmissions");
1490         rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1491         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1492             SYSCTL_CHILDREN(rack_counters),
1493             OID_AUTO, "nfto_send", CTLFLAG_RD,
1494             &rack_non_fto_send, "Total number of rack_output first sends");
1495         rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1496         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1497             SYSCTL_CHILDREN(rack_counters),
1498             OID_AUTO, "rfo_extended", CTLFLAG_RD,
1499             &rack_extended_rfo, "Total number of times we extended rfo");
1500
1501         rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1502         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1503             SYSCTL_CHILDREN(rack_counters),
1504             OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1505             &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1506         rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1507
1508         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1509             SYSCTL_CHILDREN(rack_counters),
1510             OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1511             &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1512         rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1513         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1514             SYSCTL_CHILDREN(rack_counters),
1515             OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1516             &rack_tlp_tot,
1517             "Total number of tail loss probe expirations");
1518         rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1519         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1520             SYSCTL_CHILDREN(rack_counters),
1521             OID_AUTO, "tlp_new", CTLFLAG_RD,
1522             &rack_tlp_newdata,
1523             "Total number of tail loss probe sending new data");
1524         rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1525         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1526             SYSCTL_CHILDREN(rack_counters),
1527             OID_AUTO, "tlp_retran", CTLFLAG_RD,
1528             &rack_tlp_retran,
1529             "Total number of tail loss probe sending retransmitted data");
1530         rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1531         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1532             SYSCTL_CHILDREN(rack_counters),
1533             OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1534             &rack_tlp_retran_bytes,
1535             "Total bytes of tail loss probe sending retransmitted data");
1536         rack_to_tot = counter_u64_alloc(M_WAITOK);
1537         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1538             SYSCTL_CHILDREN(rack_counters),
1539             OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1540             &rack_to_tot,
1541             "Total number of times the rack to expired");
1542         rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1543         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1544             SYSCTL_CHILDREN(rack_counters),
1545             OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1546             &rack_saw_enobuf,
1547             "Total number of times a sends returned enobuf for non-hdwr paced connections");
1548         rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1549         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1550             SYSCTL_CHILDREN(rack_counters),
1551             OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1552             &rack_saw_enobuf_hw,
1553             "Total number of times a send returned enobuf for hdwr paced connections");
1554         rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1555         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1556             SYSCTL_CHILDREN(rack_counters),
1557             OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1558             &rack_saw_enetunreach,
1559             "Total number of times a send received a enetunreachable");
1560         rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1561         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1562             SYSCTL_CHILDREN(rack_counters),
1563             OID_AUTO, "alloc_hot", CTLFLAG_RD,
1564             &rack_hot_alloc,
1565             "Total allocations from the top of our list");
1566         rack_to_alloc = counter_u64_alloc(M_WAITOK);
1567         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1568             SYSCTL_CHILDREN(rack_counters),
1569             OID_AUTO, "allocs", CTLFLAG_RD,
1570             &rack_to_alloc,
1571             "Total allocations of tracking structures");
1572         rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1573         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1574             SYSCTL_CHILDREN(rack_counters),
1575             OID_AUTO, "allochard", CTLFLAG_RD,
1576             &rack_to_alloc_hard,
1577             "Total allocations done with sleeping the hard way");
1578         rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1579         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1580             SYSCTL_CHILDREN(rack_counters),
1581             OID_AUTO, "allocemerg", CTLFLAG_RD,
1582             &rack_to_alloc_emerg,
1583             "Total allocations done from emergency cache");
1584         rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1585         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1586             SYSCTL_CHILDREN(rack_counters),
1587             OID_AUTO, "alloc_limited", CTLFLAG_RD,
1588             &rack_to_alloc_limited,
1589             "Total allocations dropped due to limit");
1590         rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1591         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1592             SYSCTL_CHILDREN(rack_counters),
1593             OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1594             &rack_alloc_limited_conns,
1595             "Connections with allocations dropped due to limit");
1596         rack_split_limited = counter_u64_alloc(M_WAITOK);
1597         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1598             SYSCTL_CHILDREN(rack_counters),
1599             OID_AUTO, "split_limited", CTLFLAG_RD,
1600             &rack_split_limited,
1601             "Split allocations dropped due to limit");
1602         rack_persists_sends = counter_u64_alloc(M_WAITOK);
1603         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1604             SYSCTL_CHILDREN(rack_counters),
1605             OID_AUTO, "persist_sends", CTLFLAG_RD,
1606             &rack_persists_sends,
1607             "Number of times we sent a persist probe");
1608         rack_persists_acks = counter_u64_alloc(M_WAITOK);
1609         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1610             SYSCTL_CHILDREN(rack_counters),
1611             OID_AUTO, "persist_acks", CTLFLAG_RD,
1612             &rack_persists_acks,
1613             "Number of times a persist probe was acked");
1614         rack_persists_loss = counter_u64_alloc(M_WAITOK);
1615         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1616             SYSCTL_CHILDREN(rack_counters),
1617             OID_AUTO, "persist_loss", CTLFLAG_RD,
1618             &rack_persists_loss,
1619             "Number of times we detected a lost persist probe (no ack)");
1620         rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1621         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1622             SYSCTL_CHILDREN(rack_counters),
1623             OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1624             &rack_persists_lost_ends,
1625             "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1626 #ifdef INVARIANTS
1627         rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1628         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1629             SYSCTL_CHILDREN(rack_counters),
1630             OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1631             &rack_adjust_map_bw,
1632             "Number of times we hit the case where the sb went up and down on a sendmap entry");
1633 #endif
1634         rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1635         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1636             SYSCTL_CHILDREN(rack_counters),
1637             OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1638             &rack_multi_single_eq,
1639             "Number of compressed acks total represented");
1640         rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1641         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1642             SYSCTL_CHILDREN(rack_counters),
1643             OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1644             &rack_proc_non_comp_ack,
1645             "Number of non compresseds acks that we processed");
1646
1647
1648         rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1649         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1650             SYSCTL_CHILDREN(rack_counters),
1651             OID_AUTO, "sack_long", CTLFLAG_RD,
1652             &rack_sack_proc_all,
1653             "Total times we had to walk whole list for sack processing");
1654         rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1655         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1656             SYSCTL_CHILDREN(rack_counters),
1657             OID_AUTO, "sack_restart", CTLFLAG_RD,
1658             &rack_sack_proc_restart,
1659             "Total times we had to walk whole list due to a restart");
1660         rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1661         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1662             SYSCTL_CHILDREN(rack_counters),
1663             OID_AUTO, "sack_short", CTLFLAG_RD,
1664             &rack_sack_proc_short,
1665             "Total times we took shortcut for sack processing");
1666         rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1667         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1668             SYSCTL_CHILDREN(rack_attack),
1669             OID_AUTO, "skipacked", CTLFLAG_RD,
1670             &rack_sack_skipped_acked,
1671             "Total number of times we skipped previously sacked");
1672         rack_sack_splits = counter_u64_alloc(M_WAITOK);
1673         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1674             SYSCTL_CHILDREN(rack_attack),
1675             OID_AUTO, "ofsplit", CTLFLAG_RD,
1676             &rack_sack_splits,
1677             "Total number of times we did the old fashion tree split");
1678         rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1679         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1680             SYSCTL_CHILDREN(rack_counters),
1681             OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1682             &rack_input_idle_reduces,
1683             "Total number of idle reductions on input");
1684         rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1685         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1686             SYSCTL_CHILDREN(rack_counters),
1687             OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1688             &rack_collapsed_win_seen,
1689             "Total number of collapsed window events seen (where our window shrinks)");
1690
1691         rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1692         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1693             SYSCTL_CHILDREN(rack_counters),
1694             OID_AUTO, "collapsed_win", CTLFLAG_RD,
1695             &rack_collapsed_win,
1696             "Total number of collapsed window events where we mark packets");
1697         rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1698         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1699             SYSCTL_CHILDREN(rack_counters),
1700             OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1701             &rack_collapsed_win_rxt,
1702             "Total number of packets that were retransmitted");
1703         rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1704         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1705             SYSCTL_CHILDREN(rack_counters),
1706             OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1707             &rack_collapsed_win_rxt_bytes,
1708             "Total number of bytes that were retransmitted");
1709         rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1710         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1711             SYSCTL_CHILDREN(rack_counters),
1712             OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1713             &rack_try_scwnd,
1714             "Total number of scwnd attempts");
1715         COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1716         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1717             OID_AUTO, "outsize", CTLFLAG_RD,
1718             rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1719         COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1720         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1721             OID_AUTO, "opts", CTLFLAG_RD,
1722             rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1723         SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1724             SYSCTL_CHILDREN(rack_sysctl_root),
1725             OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1726             &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1727 }
1728
1729 static __inline int
1730 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1731 {
1732         if (SEQ_GEQ(b->r_start, a->r_start) &&
1733             SEQ_LT(b->r_start, a->r_end)) {
1734                 /*
1735                  * The entry b is within the
1736                  * block a. i.e.:
1737                  * a --   |-------------|
1738                  * b --   |----|
1739                  * <or>
1740                  * b --       |------|
1741                  * <or>
1742                  * b --       |-----------|
1743                  */
1744                 return (0);
1745         } else if (SEQ_GEQ(b->r_start, a->r_end)) {
1746                 /*
1747                  * b falls as either the next
1748                  * sequence block after a so a
1749                  * is said to be smaller than b.
1750                  * i.e:
1751                  * a --   |------|
1752                  * b --          |--------|
1753                  * or
1754                  * b --              |-----|
1755                  */
1756                 return (1);
1757         }
1758         /*
1759          * Whats left is where a is
1760          * larger than b. i.e:
1761          * a --         |-------|
1762          * b --  |---|
1763          * or even possibly
1764          * b --   |--------------|
1765          */
1766         return (-1);
1767 }
1768
1769 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1770 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1771
1772 static uint32_t
1773 rc_init_window(struct tcp_rack *rack)
1774 {
1775         uint32_t win;
1776
1777         if (rack->rc_init_win == 0) {
1778                 /*
1779                  * Nothing set by the user, use the system stack
1780                  * default.
1781                  */
1782                 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1783         }
1784         win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1785         return (win);
1786 }
1787
1788 static uint64_t
1789 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1790 {
1791         if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1792                 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1793         else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1794                 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1795         else
1796                 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1797 }
1798
1799 static uint64_t
1800 rack_get_bw(struct tcp_rack *rack)
1801 {
1802         if (rack->use_fixed_rate) {
1803                 /* Return the fixed pacing rate */
1804                 return (rack_get_fixed_pacing_bw(rack));
1805         }
1806         if (rack->r_ctl.gp_bw == 0) {
1807                 /*
1808                  * We have yet no b/w measurement,
1809                  * if we have a user set initial bw
1810                  * return it. If we don't have that and
1811                  * we have an srtt, use the tcp IW (10) to
1812                  * calculate a fictional b/w over the SRTT
1813                  * which is more or less a guess. Note
1814                  * we don't use our IW from rack on purpose
1815                  * so if we have like IW=30, we are not
1816                  * calculating a "huge" b/w.
1817                  */
1818                 uint64_t bw, srtt;
1819                 if (rack->r_ctl.init_rate)
1820                         return (rack->r_ctl.init_rate);
1821
1822                 /* Has the user set a max peak rate? */
1823 #ifdef NETFLIX_PEAKRATE
1824                 if (rack->rc_tp->t_maxpeakrate)
1825                         return (rack->rc_tp->t_maxpeakrate);
1826 #endif
1827                 /* Ok lets come up with the IW guess, if we have a srtt */
1828                 if (rack->rc_tp->t_srtt == 0) {
1829                         /*
1830                          * Go with old pacing method
1831                          * i.e. burst mitigation only.
1832                          */
1833                         return (0);
1834                 }
1835                 /* Ok lets get the initial TCP win (not racks) */
1836                 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1837                 srtt = (uint64_t)rack->rc_tp->t_srtt;
1838                 bw *= (uint64_t)USECS_IN_SECOND;
1839                 bw /= srtt;
1840                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1841                         bw = rack->r_ctl.bw_rate_cap;
1842                 return (bw);
1843         } else {
1844                 uint64_t bw;
1845
1846                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
1847                         /* Averaging is done, we can return the value */
1848                         bw = rack->r_ctl.gp_bw;
1849                 } else {
1850                         /* Still doing initial average must calculate */
1851                         bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
1852                 }
1853 #ifdef NETFLIX_PEAKRATE
1854                 if ((rack->rc_tp->t_maxpeakrate) &&
1855                     (bw > rack->rc_tp->t_maxpeakrate)) {
1856                         /* The user has set a peak rate to pace at
1857                          * don't allow us to pace faster than that.
1858                          */
1859                         return (rack->rc_tp->t_maxpeakrate);
1860                 }
1861 #endif
1862                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1863                         bw = rack->r_ctl.bw_rate_cap;
1864                 return (bw);
1865         }
1866 }
1867
1868 static uint16_t
1869 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1870 {
1871         if (rack->use_fixed_rate) {
1872                 return (100);
1873         } else if (rack->in_probe_rtt && (rsm == NULL))
1874                 return (rack->r_ctl.rack_per_of_gp_probertt);
1875         else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
1876                   rack->r_ctl.rack_per_of_gp_rec)) {
1877                 if (rsm) {
1878                         /* a retransmission always use the recovery rate */
1879                         return (rack->r_ctl.rack_per_of_gp_rec);
1880                 } else if (rack->rack_rec_nonrxt_use_cr) {
1881                         /* Directed to use the configured rate */
1882                         goto configured_rate;
1883                 } else if (rack->rack_no_prr &&
1884                            (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1885                         /* No PRR, lets just use the b/w estimate only */
1886                         return (100);
1887                 } else {
1888                         /*
1889                          * Here we may have a non-retransmit but we
1890                          * have no overrides, so just use the recovery
1891                          * rate (prr is in effect).
1892                          */
1893                         return (rack->r_ctl.rack_per_of_gp_rec);
1894                 }
1895         }
1896 configured_rate:
1897         /* For the configured rate we look at our cwnd vs the ssthresh */
1898         if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1899                 return (rack->r_ctl.rack_per_of_gp_ss);
1900         else
1901                 return (rack->r_ctl.rack_per_of_gp_ca);
1902 }
1903
1904 static void
1905 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
1906 {
1907         /*
1908          * Types of logs (mod value)
1909          * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
1910          * 2 = a dsack round begins, persist is reset to 16.
1911          * 3 = a dsack round ends
1912          * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
1913          * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
1914          * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
1915          */
1916         if (tcp_bblogging_on(rack->rc_tp)) {
1917                 union tcp_log_stackspecific log;
1918                 struct timeval tv;
1919
1920                 memset(&log, 0, sizeof(log));
1921                 log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
1922                 log.u_bbr.flex1 <<= 1;
1923                 log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
1924                 log.u_bbr.flex1 <<= 1;
1925                 log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
1926                 log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
1927                 log.u_bbr.flex3 = rack->r_ctl.num_dsack;
1928                 log.u_bbr.flex4 = flex4;
1929                 log.u_bbr.flex5 = flex5;
1930                 log.u_bbr.flex6 = flex6;
1931                 log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
1932                 log.u_bbr.flex8 = mod;
1933                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1934                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
1935                     &rack->rc_inp->inp_socket->so_rcv,
1936                     &rack->rc_inp->inp_socket->so_snd,
1937                     RACK_DSACK_HANDLING, 0,
1938                     0, &log, false, &tv);
1939         }
1940 }
1941
1942 static void
1943 rack_log_hdwr_pacing(struct tcp_rack *rack,
1944                      uint64_t rate, uint64_t hw_rate, int line,
1945                      int error, uint16_t mod)
1946 {
1947         if (tcp_bblogging_on(rack->rc_tp)) {
1948                 union tcp_log_stackspecific log;
1949                 struct timeval tv;
1950                 const struct ifnet *ifp;
1951
1952                 memset(&log, 0, sizeof(log));
1953                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
1954                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
1955                 if (rack->r_ctl.crte) {
1956                         ifp = rack->r_ctl.crte->ptbl->rs_ifp;
1957                 } else if (rack->rc_inp->inp_route.ro_nh &&
1958                            rack->rc_inp->inp_route.ro_nh->nh_ifp) {
1959                         ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
1960                 } else
1961                         ifp = NULL;
1962                 if (ifp) {
1963                         log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
1964                         log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
1965                 }
1966                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1967                 log.u_bbr.bw_inuse = rate;
1968                 log.u_bbr.flex5 = line;
1969                 log.u_bbr.flex6 = error;
1970                 log.u_bbr.flex7 = mod;
1971                 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
1972                 log.u_bbr.flex8 = rack->use_fixed_rate;
1973                 log.u_bbr.flex8 <<= 1;
1974                 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
1975                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
1976                 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
1977                 if (rack->r_ctl.crte)
1978                         log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
1979                 else
1980                         log.u_bbr.cur_del_rate = 0;
1981                 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
1982                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
1983                     &rack->rc_inp->inp_socket->so_rcv,
1984                     &rack->rc_inp->inp_socket->so_snd,
1985                     BBR_LOG_HDWR_PACE, 0,
1986                     0, &log, false, &tv);
1987         }
1988 }
1989
1990 static uint64_t
1991 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
1992 {
1993         /*
1994          * We allow rack_per_of_gp_xx to dictate our bw rate we want.
1995          */
1996         uint64_t bw_est, high_rate;
1997         uint64_t gain;
1998
1999         gain = (uint64_t)rack_get_output_gain(rack, rsm);
2000         bw_est = bw * gain;
2001         bw_est /= (uint64_t)100;
2002         /* Never fall below the minimum (def 64kbps) */
2003         if (bw_est < RACK_MIN_BW)
2004                 bw_est = RACK_MIN_BW;
2005         if (rack->r_rack_hw_rate_caps) {
2006                 /* Rate caps are in place */
2007                 if (rack->r_ctl.crte != NULL) {
2008                         /* We have a hdwr rate already */
2009                         high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2010                         if (bw_est >= high_rate) {
2011                                 /* We are capping bw at the highest rate table entry */
2012                                 rack_log_hdwr_pacing(rack,
2013                                                      bw_est, high_rate, __LINE__,
2014                                                      0, 3);
2015                                 bw_est = high_rate;
2016                                 if (capped)
2017                                         *capped = 1;
2018                         }
2019                 } else if ((rack->rack_hdrw_pacing == 0) &&
2020                            (rack->rack_hdw_pace_ena) &&
2021                            (rack->rack_attempt_hdwr_pace == 0) &&
2022                            (rack->rc_inp->inp_route.ro_nh != NULL) &&
2023                            (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2024                         /*
2025                          * Special case, we have not yet attempted hardware
2026                          * pacing, and yet we may, when we do, find out if we are
2027                          * above the highest rate. We need to know the maxbw for the interface
2028                          * in question (if it supports ratelimiting). We get back
2029                          * a 0, if the interface is not found in the RL lists.
2030                          */
2031                         high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2032                         if (high_rate) {
2033                                 /* Yep, we have a rate is it above this rate? */
2034                                 if (bw_est > high_rate) {
2035                                         bw_est = high_rate;
2036                                         if (capped)
2037                                                 *capped = 1;
2038                                 }
2039                         }
2040                 }
2041         }
2042         return (bw_est);
2043 }
2044
2045 static void
2046 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2047 {
2048         if (tcp_bblogging_on(rack->rc_tp)) {
2049                 union tcp_log_stackspecific log;
2050                 struct timeval tv;
2051
2052                 if ((mod != 1) && (rack_verbose_logging == 0)) {
2053                         /*
2054                          * We get 3 values currently for mod
2055                          * 1 - We are retransmitting and this tells the reason.
2056                          * 2 - We are clearing a dup-ack count.
2057                          * 3 - We are incrementing a dup-ack count.
2058                          *
2059                          * The clear/increment are only logged
2060                          * if you have BBverbose on.
2061                          */
2062                         return;
2063                 }
2064                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2065                 log.u_bbr.flex1 = tsused;
2066                 log.u_bbr.flex2 = thresh;
2067                 log.u_bbr.flex3 = rsm->r_flags;
2068                 log.u_bbr.flex4 = rsm->r_dupack;
2069                 log.u_bbr.flex5 = rsm->r_start;
2070                 log.u_bbr.flex6 = rsm->r_end;
2071                 log.u_bbr.flex8 = mod;
2072                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2073                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2074                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2075                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2076                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2077                 log.u_bbr.pacing_gain = rack->r_must_retran;
2078                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2079                     &rack->rc_inp->inp_socket->so_rcv,
2080                     &rack->rc_inp->inp_socket->so_snd,
2081                     BBR_LOG_SETTINGS_CHG, 0,
2082                     0, &log, false, &tv);
2083         }
2084 }
2085
2086 static void
2087 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2088 {
2089         if (tcp_bblogging_on(rack->rc_tp)) {
2090                 union tcp_log_stackspecific log;
2091                 struct timeval tv;
2092
2093                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2094                 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2095                 log.u_bbr.flex2 = to;
2096                 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2097                 log.u_bbr.flex4 = slot;
2098                 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2099                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2100                 log.u_bbr.flex7 = rack->rc_in_persist;
2101                 log.u_bbr.flex8 = which;
2102                 if (rack->rack_no_prr)
2103                         log.u_bbr.pkts_out = 0;
2104                 else
2105                         log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2106                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2107                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2108                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2109                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2110                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2111                 log.u_bbr.pacing_gain = rack->r_must_retran;
2112                 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2113                 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2114                 log.u_bbr.lost = rack_rto_min;
2115                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2116                     &rack->rc_inp->inp_socket->so_rcv,
2117                     &rack->rc_inp->inp_socket->so_snd,
2118                     BBR_LOG_TIMERSTAR, 0,
2119                     0, &log, false, &tv);
2120         }
2121 }
2122
2123 static void
2124 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2125 {
2126         if (tcp_bblogging_on(rack->rc_tp)) {
2127                 union tcp_log_stackspecific log;
2128                 struct timeval tv;
2129
2130                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2131                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2132                 log.u_bbr.flex8 = to_num;
2133                 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2134                 log.u_bbr.flex2 = rack->rc_rack_rtt;
2135                 if (rsm == NULL)
2136                         log.u_bbr.flex3 = 0;
2137                 else
2138                         log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2139                 if (rack->rack_no_prr)
2140                         log.u_bbr.flex5 = 0;
2141                 else
2142                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2143                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2144                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2145                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2146                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2147                 log.u_bbr.pacing_gain = rack->r_must_retran;
2148                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2149                     &rack->rc_inp->inp_socket->so_rcv,
2150                     &rack->rc_inp->inp_socket->so_snd,
2151                     BBR_LOG_RTO, 0,
2152                     0, &log, false, &tv);
2153         }
2154 }
2155
2156 static void
2157 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2158                  struct rack_sendmap *prev,
2159                  struct rack_sendmap *rsm,
2160                  struct rack_sendmap *next,
2161                  int flag, uint32_t th_ack, int line)
2162 {
2163         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2164                 union tcp_log_stackspecific log;
2165                 struct timeval tv;
2166
2167                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2168                 log.u_bbr.flex8 = flag;
2169                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2170                 log.u_bbr.cur_del_rate = (uint64_t)prev;
2171                 log.u_bbr.delRate = (uint64_t)rsm;
2172                 log.u_bbr.rttProp = (uint64_t)next;
2173                 log.u_bbr.flex7 = 0;
2174                 if (prev) {
2175                         log.u_bbr.flex1 = prev->r_start;
2176                         log.u_bbr.flex2 = prev->r_end;
2177                         log.u_bbr.flex7 |= 0x4;
2178                 }
2179                 if (rsm) {
2180                         log.u_bbr.flex3 = rsm->r_start;
2181                         log.u_bbr.flex4 = rsm->r_end;
2182                         log.u_bbr.flex7 |= 0x2;
2183                 }
2184                 if (next) {
2185                         log.u_bbr.flex5 = next->r_start;
2186                         log.u_bbr.flex6 = next->r_end;
2187                         log.u_bbr.flex7 |= 0x1;
2188                 }
2189                 log.u_bbr.applimited = line;
2190                 log.u_bbr.pkts_out = th_ack;
2191                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2192                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2193                 if (rack->rack_no_prr)
2194                         log.u_bbr.lost = 0;
2195                 else
2196                         log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2197                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2198                     &rack->rc_inp->inp_socket->so_rcv,
2199                     &rack->rc_inp->inp_socket->so_snd,
2200                     TCP_LOG_MAPCHG, 0,
2201                     0, &log, false, &tv);
2202         }
2203 }
2204
2205 static void
2206 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2207                  struct rack_sendmap *rsm, int conf)
2208 {
2209         if (tcp_bblogging_on(tp)) {
2210                 union tcp_log_stackspecific log;
2211                 struct timeval tv;
2212                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2213                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2214                 log.u_bbr.flex1 = t;
2215                 log.u_bbr.flex2 = len;
2216                 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2217                 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2218                 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2219                 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2220                 log.u_bbr.flex7 = conf;
2221                 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2222                 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2223                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2224                 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2225                 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2226                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2227                 if (rsm) {
2228                         log.u_bbr.pkt_epoch = rsm->r_start;
2229                         log.u_bbr.lost = rsm->r_end;
2230                         log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2231                         /* We loose any upper of the 24 bits */
2232                         log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2233                 } else {
2234                         /* Its a SYN */
2235                         log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2236                         log.u_bbr.lost = 0;
2237                         log.u_bbr.cwnd_gain = 0;
2238                         log.u_bbr.pacing_gain = 0;
2239                 }
2240                 /* Write out general bits of interest rrs here */
2241                 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2242                 log.u_bbr.use_lt_bw <<= 1;
2243                 log.u_bbr.use_lt_bw |= rack->forced_ack;
2244                 log.u_bbr.use_lt_bw <<= 1;
2245                 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2246                 log.u_bbr.use_lt_bw <<= 1;
2247                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2248                 log.u_bbr.use_lt_bw <<= 1;
2249                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2250                 log.u_bbr.use_lt_bw <<= 1;
2251                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2252                 log.u_bbr.use_lt_bw <<= 1;
2253                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2254                 log.u_bbr.use_lt_bw <<= 1;
2255                 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2256                 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2257                 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2258                 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2259                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2260                 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2261                 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2262                 log.u_bbr.bw_inuse <<= 32;
2263                 if (rsm)
2264                         log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2265                 TCP_LOG_EVENTP(tp, NULL,
2266                     &rack->rc_inp->inp_socket->so_rcv,
2267                     &rack->rc_inp->inp_socket->so_snd,
2268                     BBR_LOG_BBRRTT, 0,
2269                     0, &log, false, &tv);
2270
2271
2272         }
2273 }
2274
2275 static void
2276 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2277 {
2278         /*
2279          * Log the rtt sample we are
2280          * applying to the srtt algorithm in
2281          * useconds.
2282          */
2283         if (tcp_bblogging_on(rack->rc_tp)) {
2284                 union tcp_log_stackspecific log;
2285                 struct timeval tv;
2286
2287                 /* Convert our ms to a microsecond */
2288                 memset(&log, 0, sizeof(log));
2289                 log.u_bbr.flex1 = rtt;
2290                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2291                 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2292                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2293                 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2294                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2295                 log.u_bbr.flex7 = 1;
2296                 log.u_bbr.flex8 = rack->sack_attack_disable;
2297                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2298                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2299                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2300                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2301                 log.u_bbr.pacing_gain = rack->r_must_retran;
2302                 /*
2303                  * We capture in delRate the upper 32 bits as
2304                  * the confidence level we had declared, and the
2305                  * lower 32 bits as the actual RTT using the arrival
2306                  * timestamp.
2307                  */
2308                 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2309                 log.u_bbr.delRate <<= 32;
2310                 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2311                 /* Lets capture all the things that make up t_rtxcur */
2312                 log.u_bbr.applimited = rack_rto_min;
2313                 log.u_bbr.epoch = rack_rto_max;
2314                 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2315                 log.u_bbr.lost = rack_rto_min;
2316                 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2317                 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2318                 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2319                 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2320                 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2321                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2322                     &rack->rc_inp->inp_socket->so_rcv,
2323                     &rack->rc_inp->inp_socket->so_snd,
2324                     TCP_LOG_RTT, 0,
2325                     0, &log, false, &tv);
2326         }
2327 }
2328
2329 static void
2330 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2331 {
2332         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2333                 union tcp_log_stackspecific log;
2334                 struct timeval tv;
2335
2336                 /* Convert our ms to a microsecond */
2337                 memset(&log, 0, sizeof(log));
2338                 log.u_bbr.flex1 = rtt;
2339                 log.u_bbr.flex2 = send_time;
2340                 log.u_bbr.flex3 = ack_time;
2341                 log.u_bbr.flex4 = where;
2342                 log.u_bbr.flex7 = 2;
2343                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2344                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2345                     &rack->rc_inp->inp_socket->so_rcv,
2346                     &rack->rc_inp->inp_socket->so_snd,
2347                     TCP_LOG_RTT, 0,
2348                     0, &log, false, &tv);
2349         }
2350 }
2351
2352
2353
2354 static inline void
2355 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2356 {
2357         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2358                 union tcp_log_stackspecific log;
2359                 struct timeval tv;
2360
2361                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2362                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2363                 log.u_bbr.flex1 = line;
2364                 log.u_bbr.flex2 = tick;
2365                 log.u_bbr.flex3 = tp->t_maxunacktime;
2366                 log.u_bbr.flex4 = tp->t_acktime;
2367                 log.u_bbr.flex8 = event;
2368                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2369                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2370                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2371                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2372                 log.u_bbr.pacing_gain = rack->r_must_retran;
2373                 TCP_LOG_EVENTP(tp, NULL,
2374                     &rack->rc_inp->inp_socket->so_rcv,
2375                     &rack->rc_inp->inp_socket->so_snd,
2376                     BBR_LOG_PROGRESS, 0,
2377                     0, &log, false, &tv);
2378         }
2379 }
2380
2381 static void
2382 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2383 {
2384         if (tcp_bblogging_on(rack->rc_tp)) {
2385                 union tcp_log_stackspecific log;
2386
2387                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2388                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2389                 log.u_bbr.flex1 = slot;
2390                 if (rack->rack_no_prr)
2391                         log.u_bbr.flex2 = 0;
2392                 else
2393                         log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2394                 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2395                 log.u_bbr.flex8 = rack->rc_in_persist;
2396                 log.u_bbr.timeStamp = cts;
2397                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2398                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2399                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2400                 log.u_bbr.pacing_gain = rack->r_must_retran;
2401                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2402                     &rack->rc_inp->inp_socket->so_rcv,
2403                     &rack->rc_inp->inp_socket->so_snd,
2404                     BBR_LOG_BBRSND, 0,
2405                     0, &log, false, tv);
2406         }
2407 }
2408
2409 static void
2410 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2411 {
2412         if (tcp_bblogging_on(rack->rc_tp)) {
2413                 union tcp_log_stackspecific log;
2414                 struct timeval tv;
2415
2416                 memset(&log, 0, sizeof(log));
2417                 log.u_bbr.flex1 = did_out;
2418                 log.u_bbr.flex2 = nxt_pkt;
2419                 log.u_bbr.flex3 = way_out;
2420                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2421                 if (rack->rack_no_prr)
2422                         log.u_bbr.flex5 = 0;
2423                 else
2424                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2425                 log.u_bbr.flex6 = nsegs;
2426                 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2427                 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;        /* Do we have ack-can-send set */
2428                 log.u_bbr.flex7 <<= 1;
2429                 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2430                 log.u_bbr.flex7 <<= 1;
2431                 log.u_bbr.flex7 |= rack->r_wanted_output;       /* Do we want output */
2432                 log.u_bbr.flex8 = rack->rc_in_persist;
2433                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2434                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2435                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2436                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2437                 log.u_bbr.use_lt_bw <<= 1;
2438                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2439                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2440                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2441                 log.u_bbr.pacing_gain = rack->r_must_retran;
2442                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2443                     &rack->rc_inp->inp_socket->so_rcv,
2444                     &rack->rc_inp->inp_socket->so_snd,
2445                     BBR_LOG_DOSEG_DONE, 0,
2446                     0, &log, false, &tv);
2447         }
2448 }
2449
2450 static void
2451 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2452 {
2453         if (tcp_bblogging_on(rack->rc_tp)) {
2454                 union tcp_log_stackspecific log;
2455                 struct timeval tv;
2456
2457                 memset(&log, 0, sizeof(log));
2458                 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2459                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2460                 log.u_bbr.flex4 = arg1;
2461                 log.u_bbr.flex5 = arg2;
2462                 log.u_bbr.flex6 = arg3;
2463                 log.u_bbr.flex8 = frm;
2464                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2465                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2466                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2467                 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2468                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2469                 log.u_bbr.pacing_gain = rack->r_must_retran;
2470                 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
2471                     &tptosocket(tp)->so_snd,
2472                     TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
2473         }
2474 }
2475
2476 static void
2477 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2478                           uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2479 {
2480         if (tcp_bblogging_on(rack->rc_tp)) {
2481                 union tcp_log_stackspecific log;
2482                 struct timeval tv;
2483
2484                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2485                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2486                 log.u_bbr.flex1 = slot;
2487                 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2488                 log.u_bbr.flex4 = reason;
2489                 if (rack->rack_no_prr)
2490                         log.u_bbr.flex5 = 0;
2491                 else
2492                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2493                 log.u_bbr.flex7 = hpts_calling;
2494                 log.u_bbr.flex8 = rack->rc_in_persist;
2495                 log.u_bbr.lt_epoch = cwnd_to_use;
2496                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2497                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2498                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2499                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2500                 log.u_bbr.pacing_gain = rack->r_must_retran;
2501                 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2502                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2503                     &rack->rc_inp->inp_socket->so_rcv,
2504                     &rack->rc_inp->inp_socket->so_snd,
2505                     BBR_LOG_JUSTRET, 0,
2506                     tlen, &log, false, &tv);
2507         }
2508 }
2509
2510 static void
2511 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2512                    struct timeval *tv, uint32_t flags_on_entry)
2513 {
2514         if (tcp_bblogging_on(rack->rc_tp)) {
2515                 union tcp_log_stackspecific log;
2516
2517                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2518                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2519                 log.u_bbr.flex1 = line;
2520                 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2521                 log.u_bbr.flex3 = flags_on_entry;
2522                 log.u_bbr.flex4 = us_cts;
2523                 if (rack->rack_no_prr)
2524                         log.u_bbr.flex5 = 0;
2525                 else
2526                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2527                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2528                 log.u_bbr.flex7 = hpts_removed;
2529                 log.u_bbr.flex8 = 1;
2530                 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2531                 log.u_bbr.timeStamp = us_cts;
2532                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2533                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2534                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2535                 log.u_bbr.pacing_gain = rack->r_must_retran;
2536                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2537                     &rack->rc_inp->inp_socket->so_rcv,
2538                     &rack->rc_inp->inp_socket->so_snd,
2539                     BBR_LOG_TIMERCANC, 0,
2540                     0, &log, false, tv);
2541         }
2542 }
2543
2544 static void
2545 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2546                           uint32_t flex1, uint32_t flex2,
2547                           uint32_t flex3, uint32_t flex4,
2548                           uint32_t flex5, uint32_t flex6,
2549                           uint16_t flex7, uint8_t mod)
2550 {
2551         if (tcp_bblogging_on(rack->rc_tp)) {
2552                 union tcp_log_stackspecific log;
2553                 struct timeval tv;
2554
2555                 if (mod == 1) {
2556                         /* No you can't use 1, its for the real to cancel */
2557                         return;
2558                 }
2559                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2560                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2561                 log.u_bbr.flex1 = flex1;
2562                 log.u_bbr.flex2 = flex2;
2563                 log.u_bbr.flex3 = flex3;
2564                 log.u_bbr.flex4 = flex4;
2565                 log.u_bbr.flex5 = flex5;
2566                 log.u_bbr.flex6 = flex6;
2567                 log.u_bbr.flex7 = flex7;
2568                 log.u_bbr.flex8 = mod;
2569                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2570                     &rack->rc_inp->inp_socket->so_rcv,
2571                     &rack->rc_inp->inp_socket->so_snd,
2572                     BBR_LOG_TIMERCANC, 0,
2573                     0, &log, false, &tv);
2574         }
2575 }
2576
2577 static void
2578 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2579 {
2580         if (tcp_bblogging_on(rack->rc_tp)) {
2581                 union tcp_log_stackspecific log;
2582                 struct timeval tv;
2583
2584                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2585                 log.u_bbr.flex1 = timers;
2586                 log.u_bbr.flex2 = ret;
2587                 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2588                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2589                 log.u_bbr.flex5 = cts;
2590                 if (rack->rack_no_prr)
2591                         log.u_bbr.flex6 = 0;
2592                 else
2593                         log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2594                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2595                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2596                 log.u_bbr.pacing_gain = rack->r_must_retran;
2597                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2598                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2599                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2600                     &rack->rc_inp->inp_socket->so_rcv,
2601                     &rack->rc_inp->inp_socket->so_snd,
2602                     BBR_LOG_TO_PROCESS, 0,
2603                     0, &log, false, &tv);
2604         }
2605 }
2606
2607 static void
2608 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
2609 {
2610         if (tcp_bblogging_on(rack->rc_tp)) {
2611                 union tcp_log_stackspecific log;
2612                 struct timeval tv;
2613
2614                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2615                 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2616                 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2617                 if (rack->rack_no_prr)
2618                         log.u_bbr.flex3 = 0;
2619                 else
2620                         log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2621                 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2622                 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2623                 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2624                 log.u_bbr.flex7 = line;
2625                 log.u_bbr.flex8 = frm;
2626                 log.u_bbr.pkts_out = orig_cwnd;
2627                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2628                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2629                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2630                 log.u_bbr.use_lt_bw <<= 1;
2631                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2632                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2633                     &rack->rc_inp->inp_socket->so_rcv,
2634                     &rack->rc_inp->inp_socket->so_snd,
2635                     BBR_LOG_BBRUPD, 0,
2636                     0, &log, false, &tv);
2637         }
2638 }
2639
2640 #ifdef NETFLIX_EXP_DETECTION
2641 static void
2642 rack_log_sad(struct tcp_rack *rack, int event)
2643 {
2644         if (tcp_bblogging_on(rack->rc_tp)) {
2645                 union tcp_log_stackspecific log;
2646                 struct timeval tv;
2647
2648                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2649                 log.u_bbr.flex1 = rack->r_ctl.sack_count;
2650                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2651                 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2652                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2653                 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2654                 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2655                 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2656                 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2657                 log.u_bbr.lt_epoch |= rack->do_detection;
2658                 log.u_bbr.applimited = tcp_map_minimum;
2659                 log.u_bbr.flex7 = rack->sack_attack_disable;
2660                 log.u_bbr.flex8 = event;
2661                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2662                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2663                 log.u_bbr.delivered = tcp_sad_decay_val;
2664                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2665                     &rack->rc_inp->inp_socket->so_rcv,
2666                     &rack->rc_inp->inp_socket->so_snd,
2667                     TCP_SAD_DETECTION, 0,
2668                     0, &log, false, &tv);
2669         }
2670 }
2671 #endif
2672
2673 static void
2674 rack_counter_destroy(void)
2675 {
2676         counter_u64_free(rack_fto_send);
2677         counter_u64_free(rack_fto_rsm_send);
2678         counter_u64_free(rack_nfto_resend);
2679         counter_u64_free(rack_hw_pace_init_fail);
2680         counter_u64_free(rack_hw_pace_lost);
2681         counter_u64_free(rack_non_fto_send);
2682         counter_u64_free(rack_extended_rfo);
2683         counter_u64_free(rack_ack_total);
2684         counter_u64_free(rack_express_sack);
2685         counter_u64_free(rack_sack_total);
2686         counter_u64_free(rack_move_none);
2687         counter_u64_free(rack_move_some);
2688         counter_u64_free(rack_sack_attacks_detected);
2689         counter_u64_free(rack_sack_attacks_reversed);
2690         counter_u64_free(rack_sack_used_next_merge);
2691         counter_u64_free(rack_sack_used_prev_merge);
2692         counter_u64_free(rack_tlp_tot);
2693         counter_u64_free(rack_tlp_newdata);
2694         counter_u64_free(rack_tlp_retran);
2695         counter_u64_free(rack_tlp_retran_bytes);
2696         counter_u64_free(rack_to_tot);
2697         counter_u64_free(rack_saw_enobuf);
2698         counter_u64_free(rack_saw_enobuf_hw);
2699         counter_u64_free(rack_saw_enetunreach);
2700         counter_u64_free(rack_hot_alloc);
2701         counter_u64_free(rack_to_alloc);
2702         counter_u64_free(rack_to_alloc_hard);
2703         counter_u64_free(rack_to_alloc_emerg);
2704         counter_u64_free(rack_to_alloc_limited);
2705         counter_u64_free(rack_alloc_limited_conns);
2706         counter_u64_free(rack_split_limited);
2707         counter_u64_free(rack_multi_single_eq);
2708         counter_u64_free(rack_proc_non_comp_ack);
2709         counter_u64_free(rack_sack_proc_all);
2710         counter_u64_free(rack_sack_proc_restart);
2711         counter_u64_free(rack_sack_proc_short);
2712         counter_u64_free(rack_sack_skipped_acked);
2713         counter_u64_free(rack_sack_splits);
2714         counter_u64_free(rack_input_idle_reduces);
2715         counter_u64_free(rack_collapsed_win);
2716         counter_u64_free(rack_collapsed_win_rxt);
2717         counter_u64_free(rack_collapsed_win_rxt_bytes);
2718         counter_u64_free(rack_collapsed_win_seen);
2719         counter_u64_free(rack_try_scwnd);
2720         counter_u64_free(rack_persists_sends);
2721         counter_u64_free(rack_persists_acks);
2722         counter_u64_free(rack_persists_loss);
2723         counter_u64_free(rack_persists_lost_ends);
2724 #ifdef INVARIANTS
2725         counter_u64_free(rack_adjust_map_bw);
2726 #endif
2727         COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2728         COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2729 }
2730
2731 static struct rack_sendmap *
2732 rack_alloc(struct tcp_rack *rack)
2733 {
2734         struct rack_sendmap *rsm;
2735
2736         /*
2737          * First get the top of the list it in
2738          * theory is the "hottest" rsm we have,
2739          * possibly just freed by ack processing.
2740          */
2741         if (rack->rc_free_cnt > rack_free_cache) {
2742                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2743                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2744                 counter_u64_add(rack_hot_alloc, 1);
2745                 rack->rc_free_cnt--;
2746                 return (rsm);
2747         }
2748         /*
2749          * Once we get under our free cache we probably
2750          * no longer have a "hot" one available. Lets
2751          * get one from UMA.
2752          */
2753         rsm = uma_zalloc(rack_zone, M_NOWAIT);
2754         if (rsm) {
2755                 rack->r_ctl.rc_num_maps_alloced++;
2756                 counter_u64_add(rack_to_alloc, 1);
2757                 return (rsm);
2758         }
2759         /*
2760          * Dig in to our aux rsm's (the last two) since
2761          * UMA failed to get us one.
2762          */
2763         if (rack->rc_free_cnt) {
2764                 counter_u64_add(rack_to_alloc_emerg, 1);
2765                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2766                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2767                 rack->rc_free_cnt--;
2768                 return (rsm);
2769         }
2770         return (NULL);
2771 }
2772
2773 static struct rack_sendmap *
2774 rack_alloc_full_limit(struct tcp_rack *rack)
2775 {
2776         if ((V_tcp_map_entries_limit > 0) &&
2777             (rack->do_detection == 0) &&
2778             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2779                 counter_u64_add(rack_to_alloc_limited, 1);
2780                 if (!rack->alloc_limit_reported) {
2781                         rack->alloc_limit_reported = 1;
2782                         counter_u64_add(rack_alloc_limited_conns, 1);
2783                 }
2784                 return (NULL);
2785         }
2786         return (rack_alloc(rack));
2787 }
2788
2789 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2790 static struct rack_sendmap *
2791 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2792 {
2793         struct rack_sendmap *rsm;
2794
2795         if (limit_type) {
2796                 /* currently there is only one limit type */
2797                 if (V_tcp_map_split_limit > 0 &&
2798                     (rack->do_detection == 0) &&
2799                     rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2800                         counter_u64_add(rack_split_limited, 1);
2801                         if (!rack->alloc_limit_reported) {
2802                                 rack->alloc_limit_reported = 1;
2803                                 counter_u64_add(rack_alloc_limited_conns, 1);
2804                         }
2805                         return (NULL);
2806                 }
2807         }
2808
2809         /* allocate and mark in the limit type, if set */
2810         rsm = rack_alloc(rack);
2811         if (rsm != NULL && limit_type) {
2812                 rsm->r_limit_type = limit_type;
2813                 rack->r_ctl.rc_num_split_allocs++;
2814         }
2815         return (rsm);
2816 }
2817
2818 static void
2819 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2820 {
2821         if (rsm->r_flags & RACK_APP_LIMITED) {
2822                 if (rack->r_ctl.rc_app_limited_cnt > 0) {
2823                         rack->r_ctl.rc_app_limited_cnt--;
2824                 }
2825         }
2826         if (rsm->r_limit_type) {
2827                 /* currently there is only one limit type */
2828                 rack->r_ctl.rc_num_split_allocs--;
2829         }
2830         if (rsm == rack->r_ctl.rc_first_appl) {
2831                 if (rack->r_ctl.rc_app_limited_cnt == 0)
2832                         rack->r_ctl.rc_first_appl = NULL;
2833                 else {
2834                         /* Follow the next one out */
2835                         struct rack_sendmap fe;
2836
2837                         fe.r_start = rsm->r_nseq_appl;
2838                         rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2839                 }
2840         }
2841         if (rsm == rack->r_ctl.rc_resend)
2842                 rack->r_ctl.rc_resend = NULL;
2843         if (rsm == rack->r_ctl.rc_end_appl)
2844                 rack->r_ctl.rc_end_appl = NULL;
2845         if (rack->r_ctl.rc_tlpsend == rsm)
2846                 rack->r_ctl.rc_tlpsend = NULL;
2847         if (rack->r_ctl.rc_sacklast == rsm)
2848                 rack->r_ctl.rc_sacklast = NULL;
2849         memset(rsm, 0, sizeof(struct rack_sendmap));
2850         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
2851         rack->rc_free_cnt++;
2852 }
2853
2854 static void
2855 rack_free_trim(struct tcp_rack *rack)
2856 {
2857         struct rack_sendmap *rsm;
2858
2859         /*
2860          * Free up all the tail entries until
2861          * we get our list down to the limit.
2862          */
2863         while (rack->rc_free_cnt > rack_free_cache) {
2864                 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
2865                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2866                 rack->rc_free_cnt--;
2867                 uma_zfree(rack_zone, rsm);
2868         }
2869 }
2870
2871
2872 static uint32_t
2873 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2874 {
2875         uint64_t srtt, bw, len, tim;
2876         uint32_t segsiz, def_len, minl;
2877
2878         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2879         def_len = rack_def_data_window * segsiz;
2880         if (rack->rc_gp_filled == 0) {
2881                 /*
2882                  * We have no measurement (IW is in flight?) so
2883                  * we can only guess using our data_window sysctl
2884                  * value (usually 20MSS).
2885                  */
2886                 return (def_len);
2887         }
2888         /*
2889          * Now we have a number of factors to consider.
2890          *
2891          * 1) We have a desired BDP which is usually
2892          *    at least 2.
2893          * 2) We have a minimum number of rtt's usually 1 SRTT
2894          *    but we allow it too to be more.
2895          * 3) We want to make sure a measurement last N useconds (if
2896          *    we have set rack_min_measure_usec.
2897          *
2898          * We handle the first concern here by trying to create a data
2899          * window of max(rack_def_data_window, DesiredBDP). The
2900          * second concern we handle in not letting the measurement
2901          * window end normally until at least the required SRTT's
2902          * have gone by which is done further below in
2903          * rack_enough_for_measurement(). Finally the third concern
2904          * we also handle here by calculating how long that time
2905          * would take at the current BW and then return the
2906          * max of our first calculation and that length. Note
2907          * that if rack_min_measure_usec is 0, we don't deal
2908          * with concern 3. Also for both Concern 1 and 3 an
2909          * application limited period could end the measurement
2910          * earlier.
2911          *
2912          * So lets calculate the BDP with the "known" b/w using
2913          * the SRTT has our rtt and then multiply it by the
2914          * goal.
2915          */
2916         bw = rack_get_bw(rack);
2917         srtt = (uint64_t)tp->t_srtt;
2918         len = bw * srtt;
2919         len /= (uint64_t)HPTS_USEC_IN_SEC;
2920         len *= max(1, rack_goal_bdp);
2921         /* Now we need to round up to the nearest MSS */
2922         len = roundup(len, segsiz);
2923         if (rack_min_measure_usec) {
2924                 /* Now calculate our min length for this b/w */
2925                 tim = rack_min_measure_usec;
2926                 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
2927                 if (minl == 0)
2928                         minl = 1;
2929                 minl = roundup(minl, segsiz);
2930                 if (len < minl)
2931                         len = minl;
2932         }
2933         /*
2934          * Now if we have a very small window we want
2935          * to attempt to get the window that is
2936          * as small as possible. This happens on
2937          * low b/w connections and we don't want to
2938          * span huge numbers of rtt's between measurements.
2939          *
2940          * We basically include 2 over our "MIN window" so
2941          * that the measurement can be shortened (possibly) by
2942          * an ack'ed packet.
2943          */
2944         if (len < def_len)
2945                 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
2946         else
2947                 return (max((uint32_t)len, def_len));
2948
2949 }
2950
2951 static int
2952 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
2953 {
2954         uint32_t tim, srtts, segsiz;
2955
2956         /*
2957          * Has enough time passed for the GP measurement to be valid?
2958          */
2959         if ((tp->snd_max == tp->snd_una) ||
2960             (th_ack == tp->snd_max)){
2961                 /* All is acked */
2962                 *quality = RACK_QUALITY_ALLACKED;
2963                 return (1);
2964         }
2965         if (SEQ_LT(th_ack, tp->gput_seq)) {
2966                 /* Not enough bytes yet */
2967                 return (0);
2968         }
2969         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2970         if (SEQ_LT(th_ack, tp->gput_ack) &&
2971             ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
2972                 /* Not enough bytes yet */
2973                 return (0);
2974         }
2975         if (rack->r_ctl.rc_first_appl &&
2976             (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
2977                 /*
2978                  * We are up to the app limited send point
2979                  * we have to measure irrespective of the time..
2980                  */
2981                 *quality = RACK_QUALITY_APPLIMITED;
2982                 return (1);
2983         }
2984         /* Now what about time? */
2985         srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
2986         tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
2987         if (tim >= srtts) {
2988                 *quality = RACK_QUALITY_HIGH;
2989                 return (1);
2990         }
2991         /* Nope not even a full SRTT has passed */
2992         return (0);
2993 }
2994
2995 static void
2996 rack_log_timely(struct tcp_rack *rack,
2997                 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
2998                 uint64_t up_bnd, int line, uint8_t method)
2999 {
3000         if (tcp_bblogging_on(rack->rc_tp)) {
3001                 union tcp_log_stackspecific log;
3002                 struct timeval tv;
3003
3004                 memset(&log, 0, sizeof(log));
3005                 log.u_bbr.flex1 = logged;
3006                 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3007                 log.u_bbr.flex2 <<= 4;
3008                 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3009                 log.u_bbr.flex2 <<= 4;
3010                 log.u_bbr.flex2 |= rack->rc_gp_incr;
3011                 log.u_bbr.flex2 <<= 4;
3012                 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3013                 log.u_bbr.flex3 = rack->rc_gp_incr;
3014                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3015                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3016                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3017                 log.u_bbr.flex7 = rack->rc_gp_bwred;
3018                 log.u_bbr.flex8 = method;
3019                 log.u_bbr.cur_del_rate = cur_bw;
3020                 log.u_bbr.delRate = low_bnd;
3021                 log.u_bbr.bw_inuse = up_bnd;
3022                 log.u_bbr.rttProp = rack_get_bw(rack);
3023                 log.u_bbr.pkt_epoch = line;
3024                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3025                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3026                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3027                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3028                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3029                 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3030                 log.u_bbr.cwnd_gain <<= 1;
3031                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3032                 log.u_bbr.cwnd_gain <<= 1;
3033                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3034                 log.u_bbr.cwnd_gain <<= 1;
3035                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3036                 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3037                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3038                     &rack->rc_inp->inp_socket->so_rcv,
3039                     &rack->rc_inp->inp_socket->so_snd,
3040                     TCP_TIMELY_WORK, 0,
3041                     0, &log, false, &tv);
3042         }
3043 }
3044
3045 static int
3046 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3047 {
3048         /*
3049          * Before we increase we need to know if
3050          * the estimate just made was less than
3051          * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3052          *
3053          * If we already are pacing at a fast enough
3054          * rate to push us faster there is no sense of
3055          * increasing.
3056          *
3057          * We first caculate our actual pacing rate (ss or ca multiplier
3058          * times our cur_bw).
3059          *
3060          * Then we take the last measured rate and multipy by our
3061          * maximum pacing overage to give us a max allowable rate.
3062          *
3063          * If our act_rate is smaller than our max_allowable rate
3064          * then we should increase. Else we should hold steady.
3065          *
3066          */
3067         uint64_t act_rate, max_allow_rate;
3068
3069         if (rack_timely_no_stopping)
3070                 return (1);
3071
3072         if ((cur_bw == 0) || (last_bw_est == 0)) {
3073                 /*
3074                  * Initial startup case or
3075                  * everything is acked case.
3076                  */
3077                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3078                                 __LINE__, 9);
3079                 return (1);
3080         }
3081         if (mult <= 100) {
3082                 /*
3083                  * We can always pace at or slightly above our rate.
3084                  */
3085                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3086                                 __LINE__, 9);
3087                 return (1);
3088         }
3089         act_rate = cur_bw * (uint64_t)mult;
3090         act_rate /= 100;
3091         max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3092         max_allow_rate /= 100;
3093         if (act_rate < max_allow_rate) {
3094                 /*
3095                  * Here the rate we are actually pacing at
3096                  * is smaller than 10% above our last measurement.
3097                  * This means we are pacing below what we would
3098                  * like to try to achieve (plus some wiggle room).
3099                  */
3100                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3101                                 __LINE__, 9);
3102                 return (1);
3103         } else {
3104                 /*
3105                  * Here we are already pacing at least rack_max_per_above(10%)
3106                  * what we are getting back. This indicates most likely
3107                  * that we are being limited (cwnd/rwnd/app) and can't
3108                  * get any more b/w. There is no sense of trying to
3109                  * raise up the pacing rate its not speeding us up
3110                  * and we already are pacing faster than we are getting.
3111                  */
3112                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3113                                 __LINE__, 8);
3114                 return (0);
3115         }
3116 }
3117
3118 static void
3119 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3120 {
3121         /*
3122          * When we drag bottom, we want to assure
3123          * that no multiplier is below 1.0, if so
3124          * we want to restore it to at least that.
3125          */
3126         if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3127                 /* This is unlikely we usually do not touch recovery */
3128                 rack->r_ctl.rack_per_of_gp_rec = 100;
3129         }
3130         if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3131                 rack->r_ctl.rack_per_of_gp_ca = 100;
3132         }
3133         if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3134                 rack->r_ctl.rack_per_of_gp_ss = 100;
3135         }
3136 }
3137
3138 static void
3139 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3140 {
3141         if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3142                 rack->r_ctl.rack_per_of_gp_ca = 100;
3143         }
3144         if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3145                 rack->r_ctl.rack_per_of_gp_ss = 100;
3146         }
3147 }
3148
3149 static void
3150 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3151 {
3152         int32_t  calc, logged, plus;
3153
3154         logged = 0;
3155
3156         if (override) {
3157                 /*
3158                  * override is passed when we are
3159                  * loosing b/w and making one last
3160                  * gasp at trying to not loose out
3161                  * to a new-reno flow.
3162                  */
3163                 goto extra_boost;
3164         }
3165         /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3166         if (rack->rc_gp_incr &&
3167             ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3168                 /*
3169                  * Reset and get 5 strokes more before the boost. Note
3170                  * that the count is 0 based so we have to add one.
3171                  */
3172 extra_boost:
3173                 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3174                 rack->rc_gp_timely_inc_cnt = 0;
3175         } else
3176                 plus = (uint32_t)rack_gp_increase_per;
3177         /* Must be at least 1% increase for true timely increases */
3178         if ((plus < 1) &&
3179             ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3180                 plus = 1;
3181         if (rack->rc_gp_saw_rec &&
3182             (rack->rc_gp_no_rec_chg == 0) &&
3183             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3184                                   rack->r_ctl.rack_per_of_gp_rec)) {
3185                 /* We have been in recovery ding it too */
3186                 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3187                 if (calc > 0xffff)
3188                         calc = 0xffff;
3189                 logged |= 1;
3190                 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3191                 if (rack_per_upper_bound_ss &&
3192                     (rack->rc_dragged_bottom == 0) &&
3193                     (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3194                         rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3195         }
3196         if (rack->rc_gp_saw_ca &&
3197             (rack->rc_gp_saw_ss == 0) &&
3198             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3199                                   rack->r_ctl.rack_per_of_gp_ca)) {
3200                 /* In CA */
3201                 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3202                 if (calc > 0xffff)
3203                         calc = 0xffff;
3204                 logged |= 2;
3205                 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3206                 if (rack_per_upper_bound_ca &&
3207                     (rack->rc_dragged_bottom == 0) &&
3208                     (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3209                         rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3210         }
3211         if (rack->rc_gp_saw_ss &&
3212             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3213                                   rack->r_ctl.rack_per_of_gp_ss)) {
3214                 /* In SS */
3215                 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3216                 if (calc > 0xffff)
3217                         calc = 0xffff;
3218                 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3219                 if (rack_per_upper_bound_ss &&
3220                     (rack->rc_dragged_bottom == 0) &&
3221                     (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3222                         rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3223                 logged |= 4;
3224         }
3225         if (logged &&
3226             (rack->rc_gp_incr == 0)){
3227                 /* Go into increment mode */
3228                 rack->rc_gp_incr = 1;
3229                 rack->rc_gp_timely_inc_cnt = 0;
3230         }
3231         if (rack->rc_gp_incr &&
3232             logged &&
3233             (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3234                 rack->rc_gp_timely_inc_cnt++;
3235         }
3236         rack_log_timely(rack,  logged, plus, 0, 0,
3237                         __LINE__, 1);
3238 }
3239
3240 static uint32_t
3241 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3242 {
3243         /*
3244          * norm_grad = rtt_diff / minrtt;
3245          * new_per = curper * (1 - B * norm_grad)
3246          *
3247          * B = rack_gp_decrease_per (default 10%)
3248          * rtt_dif = input var current rtt-diff
3249          * curper = input var current percentage
3250          * minrtt = from rack filter
3251          *
3252          */
3253         uint64_t perf;
3254
3255         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3256                     ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3257                      (((uint64_t)rtt_diff * (uint64_t)1000000)/
3258                       (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3259                      (uint64_t)1000000)) /
3260                 (uint64_t)1000000);
3261         if (perf > curper) {
3262                 /* TSNH */
3263                 perf = curper - 1;
3264         }
3265         return ((uint32_t)perf);
3266 }
3267
3268 static uint32_t
3269 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3270 {
3271         /*
3272          *                                   highrttthresh
3273          * result = curper * (1 - (B * ( 1 -  ------          ))
3274          *                                     gp_srtt
3275          *
3276          * B = rack_gp_decrease_per (default 10%)
3277          * highrttthresh = filter_min * rack_gp_rtt_maxmul
3278          */
3279         uint64_t perf;
3280         uint32_t highrttthresh;
3281
3282         highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3283
3284         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3285                                      ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3286                                         ((uint64_t)highrttthresh * (uint64_t)1000000) /
3287                                                     (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3288         return (perf);
3289 }
3290
3291 static void
3292 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3293 {
3294         uint64_t logvar, logvar2, logvar3;
3295         uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3296
3297         if (rack->rc_gp_incr) {
3298                 /* Turn off increment counting */
3299                 rack->rc_gp_incr = 0;
3300                 rack->rc_gp_timely_inc_cnt = 0;
3301         }
3302         ss_red = ca_red = rec_red = 0;
3303         logged = 0;
3304         /* Calculate the reduction value */
3305         if (rtt_diff < 0) {
3306                 rtt_diff *= -1;
3307         }
3308         /* Must be at least 1% reduction */
3309         if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3310                 /* We have been in recovery ding it too */
3311                 if (timely_says == 2) {
3312                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3313                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3314                         if (alt < new_per)
3315                                 val = alt;
3316                         else
3317                                 val = new_per;
3318                 } else
3319                          val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3320                 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3321                         rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3322                         rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3323                 } else {
3324                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3325                         rec_red = 0;
3326                 }
3327                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3328                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3329                 logged |= 1;
3330         }
3331         if (rack->rc_gp_saw_ss) {
3332                 /* Sent in SS */
3333                 if (timely_says == 2) {
3334                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3335                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3336                         if (alt < new_per)
3337                                 val = alt;
3338                         else
3339                                 val = new_per;
3340                 } else
3341                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3342                 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3343                         ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3344                         rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3345                 } else {
3346                         ss_red = new_per;
3347                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3348                         logvar = new_per;
3349                         logvar <<= 32;
3350                         logvar |= alt;
3351                         logvar2 = (uint32_t)rtt;
3352                         logvar2 <<= 32;
3353                         logvar2 |= (uint32_t)rtt_diff;
3354                         logvar3 = rack_gp_rtt_maxmul;
3355                         logvar3 <<= 32;
3356                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3357                         rack_log_timely(rack, timely_says,
3358                                         logvar2, logvar3,
3359                                         logvar, __LINE__, 10);
3360                 }
3361                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3362                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3363                 logged |= 4;
3364         } else if (rack->rc_gp_saw_ca) {
3365                 /* Sent in CA */
3366                 if (timely_says == 2) {
3367                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3368                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3369                         if (alt < new_per)
3370                                 val = alt;
3371                         else
3372                                 val = new_per;
3373                 } else
3374                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3375                 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3376                         ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3377                         rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3378                 } else {
3379                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3380                         ca_red = 0;
3381                         logvar = new_per;
3382                         logvar <<= 32;
3383                         logvar |= alt;
3384                         logvar2 = (uint32_t)rtt;
3385                         logvar2 <<= 32;
3386                         logvar2 |= (uint32_t)rtt_diff;
3387                         logvar3 = rack_gp_rtt_maxmul;
3388                         logvar3 <<= 32;
3389                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3390                         rack_log_timely(rack, timely_says,
3391                                         logvar2, logvar3,
3392                                         logvar, __LINE__, 10);
3393                 }
3394                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3395                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3396                 logged |= 2;
3397         }
3398         if (rack->rc_gp_timely_dec_cnt < 0x7) {
3399                 rack->rc_gp_timely_dec_cnt++;
3400                 if (rack_timely_dec_clear &&
3401                     (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3402                         rack->rc_gp_timely_dec_cnt = 0;
3403         }
3404         logvar = ss_red;
3405         logvar <<= 32;
3406         logvar |= ca_red;
3407         rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3408                         __LINE__, 2);
3409 }
3410
3411 static void
3412 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3413                      uint32_t rtt, uint32_t line, uint8_t reas)
3414 {
3415         if (tcp_bblogging_on(rack->rc_tp)) {
3416                 union tcp_log_stackspecific log;
3417                 struct timeval tv;
3418
3419                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3420                 log.u_bbr.flex1 = line;
3421                 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3422                 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3423                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3424                 log.u_bbr.flex5 = rtt;
3425                 log.u_bbr.flex6 = rack->rc_highly_buffered;
3426                 log.u_bbr.flex6 <<= 1;
3427                 log.u_bbr.flex6 |= rack->forced_ack;
3428                 log.u_bbr.flex6 <<= 1;
3429                 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3430                 log.u_bbr.flex6 <<= 1;
3431                 log.u_bbr.flex6 |= rack->in_probe_rtt;
3432                 log.u_bbr.flex6 <<= 1;
3433                 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3434                 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3435                 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3436                 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3437                 log.u_bbr.flex8 = reas;
3438                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3439                 log.u_bbr.delRate = rack_get_bw(rack);
3440                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3441                 log.u_bbr.cur_del_rate <<= 32;
3442                 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3443                 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3444                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3445                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3446                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3447                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3448                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3449                 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3450                 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3451                 log.u_bbr.rttProp = us_cts;
3452                 log.u_bbr.rttProp <<= 32;
3453                 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3454                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3455                     &rack->rc_inp->inp_socket->so_rcv,
3456                     &rack->rc_inp->inp_socket->so_snd,
3457                     BBR_LOG_RTT_SHRINKS, 0,
3458                     0, &log, false, &rack->r_ctl.act_rcv_time);
3459         }
3460 }
3461
3462 static void
3463 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3464 {
3465         uint64_t bwdp;
3466
3467         bwdp = rack_get_bw(rack);
3468         bwdp *= (uint64_t)rtt;
3469         bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3470         rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3471         if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3472                 /*
3473                  * A window protocol must be able to have 4 packets
3474                  * outstanding as the floor in order to function
3475                  * (especially considering delayed ack :D).
3476                  */
3477                 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3478         }
3479 }
3480
3481 static void
3482 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3483 {
3484         /**
3485          * ProbeRTT is a bit different in rack_pacing than in
3486          * BBR. It is like BBR in that it uses the lowering of
3487          * the RTT as a signal that we saw something new and
3488          * counts from there for how long between. But it is
3489          * different in that its quite simple. It does not
3490          * play with the cwnd and wait until we get down
3491          * to N segments outstanding and hold that for
3492          * 200ms. Instead it just sets the pacing reduction
3493          * rate to a set percentage (70 by default) and hold
3494          * that for a number of recent GP Srtt's.
3495          */
3496         uint32_t segsiz;
3497
3498         if (rack->rc_gp_dyn_mul == 0)
3499                 return;
3500
3501         if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3502                 /* We are idle */
3503                 return;
3504         }
3505         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3506             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3507                 /*
3508                  * Stop the goodput now, the idea here is
3509                  * that future measurements with in_probe_rtt
3510                  * won't register if they are not greater so
3511                  * we want to get what info (if any) is available
3512                  * now.
3513                  */
3514                 rack_do_goodput_measurement(rack->rc_tp, rack,
3515                                             rack->rc_tp->snd_una, __LINE__,
3516                                             RACK_QUALITY_PROBERTT);
3517         }
3518         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3519         rack->r_ctl.rc_time_probertt_entered = us_cts;
3520         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3521                      rack->r_ctl.rc_pace_min_segs);
3522         rack->in_probe_rtt = 1;
3523         rack->measure_saw_probe_rtt = 1;
3524         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3525         rack->r_ctl.rc_time_probertt_starts = 0;
3526         rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3527         if (rack_probertt_use_min_rtt_entry)
3528                 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3529         else
3530                 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3531         rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3532                              __LINE__, RACK_RTTS_ENTERPROBE);
3533 }
3534
3535 static void
3536 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3537 {
3538         struct rack_sendmap *rsm;
3539         uint32_t segsiz;
3540
3541         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3542                      rack->r_ctl.rc_pace_min_segs);
3543         rack->in_probe_rtt = 0;
3544         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3545             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3546                 /*
3547                  * Stop the goodput now, the idea here is
3548                  * that future measurements with in_probe_rtt
3549                  * won't register if they are not greater so
3550                  * we want to get what info (if any) is available
3551                  * now.
3552                  */
3553                 rack_do_goodput_measurement(rack->rc_tp, rack,
3554                                             rack->rc_tp->snd_una, __LINE__,
3555                                             RACK_QUALITY_PROBERTT);
3556         } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3557                 /*
3558                  * We don't have enough data to make a measurement.
3559                  * So lets just stop and start here after exiting
3560                  * probe-rtt. We probably are not interested in
3561                  * the results anyway.
3562                  */
3563                 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3564         }
3565         /*
3566          * Measurements through the current snd_max are going
3567          * to be limited by the slower pacing rate.
3568          *
3569          * We need to mark these as app-limited so we
3570          * don't collapse the b/w.
3571          */
3572         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3573         if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3574                 if (rack->r_ctl.rc_app_limited_cnt == 0)
3575                         rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3576                 else {
3577                         /*
3578                          * Go out to the end app limited and mark
3579                          * this new one as next and move the end_appl up
3580                          * to this guy.
3581                          */
3582                         if (rack->r_ctl.rc_end_appl)
3583                                 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3584                         rack->r_ctl.rc_end_appl = rsm;
3585                 }
3586                 rsm->r_flags |= RACK_APP_LIMITED;
3587                 rack->r_ctl.rc_app_limited_cnt++;
3588         }
3589         /*
3590          * Now, we need to examine our pacing rate multipliers.
3591          * If its under 100%, we need to kick it back up to
3592          * 100%. We also don't let it be over our "max" above
3593          * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3594          * Note setting clamp_atexit_prtt to 0 has the effect
3595          * of setting CA/SS to 100% always at exit (which is
3596          * the default behavior).
3597          */
3598         if (rack_probertt_clear_is) {
3599                 rack->rc_gp_incr = 0;
3600                 rack->rc_gp_bwred = 0;
3601                 rack->rc_gp_timely_inc_cnt = 0;
3602                 rack->rc_gp_timely_dec_cnt = 0;
3603         }
3604         /* Do we do any clamping at exit? */
3605         if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3606                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3607                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3608         }
3609         if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3610                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3611                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3612         }
3613         /*
3614          * Lets set rtt_diff to 0, so that we will get a "boost"
3615          * after exiting.
3616          */
3617         rack->r_ctl.rc_rtt_diff = 0;
3618
3619         /* Clear all flags so we start fresh */
3620         rack->rc_tp->t_bytes_acked = 0;
3621         rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
3622         /*
3623          * If configured to, set the cwnd and ssthresh to
3624          * our targets.
3625          */
3626         if (rack_probe_rtt_sets_cwnd) {
3627                 uint64_t ebdp;
3628                 uint32_t setto;
3629
3630                 /* Set ssthresh so we get into CA once we hit our target */
3631                 if (rack_probertt_use_min_rtt_exit == 1) {
3632                         /* Set to min rtt */
3633                         rack_set_prtt_target(rack, segsiz,
3634                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3635                 } else if (rack_probertt_use_min_rtt_exit == 2) {
3636                         /* Set to current gp rtt */
3637                         rack_set_prtt_target(rack, segsiz,
3638                                              rack->r_ctl.rc_gp_srtt);
3639                 } else if (rack_probertt_use_min_rtt_exit == 3) {
3640                         /* Set to entry gp rtt */
3641                         rack_set_prtt_target(rack, segsiz,
3642                                              rack->r_ctl.rc_entry_gp_rtt);
3643                 } else {
3644                         uint64_t sum;
3645                         uint32_t setval;
3646
3647                         sum = rack->r_ctl.rc_entry_gp_rtt;
3648                         sum *= 10;
3649                         sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3650                         if (sum >= 20) {
3651                                 /*
3652                                  * A highly buffered path needs
3653                                  * cwnd space for timely to work.
3654                                  * Lets set things up as if
3655                                  * we are heading back here again.
3656                                  */
3657                                 setval = rack->r_ctl.rc_entry_gp_rtt;
3658                         } else if (sum >= 15) {
3659                                 /*
3660                                  * Lets take the smaller of the
3661                                  * two since we are just somewhat
3662                                  * buffered.
3663                                  */
3664                                 setval = rack->r_ctl.rc_gp_srtt;
3665                                 if (setval > rack->r_ctl.rc_entry_gp_rtt)
3666                                         setval = rack->r_ctl.rc_entry_gp_rtt;
3667                         } else {
3668                                 /*
3669                                  * Here we are not highly buffered
3670                                  * and should pick the min we can to
3671                                  * keep from causing loss.
3672                                  */
3673                                 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3674                         }
3675                         rack_set_prtt_target(rack, segsiz,
3676                                              setval);
3677                 }
3678                 if (rack_probe_rtt_sets_cwnd > 1) {
3679                         /* There is a percentage here to boost */
3680                         ebdp = rack->r_ctl.rc_target_probertt_flight;
3681                         ebdp *= rack_probe_rtt_sets_cwnd;
3682                         ebdp /= 100;
3683                         setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3684                 } else
3685                         setto = rack->r_ctl.rc_target_probertt_flight;
3686                 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3687                 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3688                         /* Enforce a min */
3689                         rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3690                 }
3691                 /* If we set in the cwnd also set the ssthresh point so we are in CA */
3692                 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3693         }
3694         rack_log_rtt_shrinks(rack,  us_cts,
3695                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3696                              __LINE__, RACK_RTTS_EXITPROBE);
3697         /* Clear times last so log has all the info */
3698         rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3699         rack->r_ctl.rc_time_probertt_entered = us_cts;
3700         rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3701         rack->r_ctl.rc_time_of_last_probertt = us_cts;
3702 }
3703
3704 static void
3705 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3706 {
3707         /* Check in on probe-rtt */
3708         if (rack->rc_gp_filled == 0) {
3709                 /* We do not do p-rtt unless we have gp measurements */
3710                 return;
3711         }
3712         if (rack->in_probe_rtt) {
3713                 uint64_t no_overflow;
3714                 uint32_t endtime, must_stay;
3715
3716                 if (rack->r_ctl.rc_went_idle_time &&
3717                     ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3718                         /*
3719                          * We went idle during prtt, just exit now.
3720                          */
3721                         rack_exit_probertt(rack, us_cts);
3722                 } else if (rack_probe_rtt_safety_val &&
3723                     TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3724                     ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3725                         /*
3726                          * Probe RTT safety value triggered!
3727                          */
3728                         rack_log_rtt_shrinks(rack,  us_cts,
3729                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3730                                              __LINE__, RACK_RTTS_SAFETY);
3731                         rack_exit_probertt(rack, us_cts);
3732                 }
3733                 /* Calculate the max we will wait */
3734                 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3735                 if (rack->rc_highly_buffered)
3736                         endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3737                 /* Calculate the min we must wait */
3738                 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3739                 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3740                     TSTMP_LT(us_cts, endtime)) {
3741                         uint32_t calc;
3742                         /* Do we lower more? */
3743 no_exit:
3744                         if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3745                                 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3746                         else
3747                                 calc = 0;
3748                         calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3749                         if (calc) {
3750                                 /* Maybe */
3751                                 calc *= rack_per_of_gp_probertt_reduce;
3752                                 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3753                                 /* Limit it too */
3754                                 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3755                                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3756                         }
3757                         /* We must reach target or the time set */
3758                         return;
3759                 }
3760                 if (rack->r_ctl.rc_time_probertt_starts == 0) {
3761                         if ((TSTMP_LT(us_cts, must_stay) &&
3762                              rack->rc_highly_buffered) ||
3763                              (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3764                               rack->r_ctl.rc_target_probertt_flight)) {
3765                                 /* We are not past the must_stay time */
3766                                 goto no_exit;
3767                         }
3768                         rack_log_rtt_shrinks(rack,  us_cts,
3769                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3770                                              __LINE__, RACK_RTTS_REACHTARGET);
3771                         rack->r_ctl.rc_time_probertt_starts = us_cts;
3772                         if (rack->r_ctl.rc_time_probertt_starts == 0)
3773                                 rack->r_ctl.rc_time_probertt_starts = 1;
3774                         /* Restore back to our rate we want to pace at in prtt */
3775                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3776                 }
3777                 /*
3778                  * Setup our end time, some number of gp_srtts plus 200ms.
3779                  */
3780                 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3781                                (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3782                 if (rack_probertt_gpsrtt_cnt_div)
3783                         endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3784                 else
3785                         endtime = 0;
3786                 endtime += rack_min_probertt_hold;
3787                 endtime += rack->r_ctl.rc_time_probertt_starts;
3788                 if (TSTMP_GEQ(us_cts,  endtime)) {
3789                         /* yes, exit probertt */
3790                         rack_exit_probertt(rack, us_cts);
3791                 }
3792
3793         } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3794                 /* Go into probertt, its been too long since we went lower */
3795                 rack_enter_probertt(rack, us_cts);
3796         }
3797 }
3798
3799 static void
3800 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3801                        uint32_t rtt, int32_t rtt_diff)
3802 {
3803         uint64_t cur_bw, up_bnd, low_bnd, subfr;
3804         uint32_t losses;
3805
3806         if ((rack->rc_gp_dyn_mul == 0) ||
3807             (rack->use_fixed_rate) ||
3808             (rack->in_probe_rtt) ||
3809             (rack->rc_always_pace == 0)) {
3810                 /* No dynamic GP multiplier in play */
3811                 return;
3812         }
3813         losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3814         cur_bw = rack_get_bw(rack);
3815         /* Calculate our up and down range */
3816         up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3817         up_bnd /= 100;
3818         up_bnd += rack->r_ctl.last_gp_comp_bw;
3819
3820         subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3821         subfr /= 100;
3822         low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3823         if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3824                 /*
3825                  * This is the case where our RTT is above
3826                  * the max target and we have been configured
3827                  * to just do timely no bonus up stuff in that case.
3828                  *
3829                  * There are two configurations, set to 1, and we
3830                  * just do timely if we are over our max. If its
3831                  * set above 1 then we slam the multipliers down
3832                  * to 100 and then decrement per timely.
3833                  */
3834                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3835                                 __LINE__, 3);
3836                 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3837                         rack_validate_multipliers_at_or_below_100(rack);
3838                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3839         } else if ((last_bw_est < low_bnd) && !losses) {
3840                 /*
3841                  * We are decreasing this is a bit complicated this
3842                  * means we are loosing ground. This could be
3843                  * because another flow entered and we are competing
3844                  * for b/w with it. This will push the RTT up which
3845                  * makes timely unusable unless we want to get shoved
3846                  * into a corner and just be backed off (the age
3847                  * old problem with delay based CC).
3848                  *
3849                  * On the other hand if it was a route change we
3850                  * would like to stay somewhat contained and not
3851                  * blow out the buffers.
3852                  */
3853                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3854                                 __LINE__, 3);
3855                 rack->r_ctl.last_gp_comp_bw = cur_bw;
3856                 if (rack->rc_gp_bwred == 0) {
3857                         /* Go into reduction counting */
3858                         rack->rc_gp_bwred = 1;
3859                         rack->rc_gp_timely_dec_cnt = 0;
3860                 }
3861                 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3862                     (timely_says == 0)) {
3863                         /*
3864                          * Push another time with a faster pacing
3865                          * to try to gain back (we include override to
3866                          * get a full raise factor).
3867                          */
3868                         if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3869                             (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3870                             (timely_says == 0) ||
3871                             (rack_down_raise_thresh == 0)) {
3872                                 /*
3873                                  * Do an override up in b/w if we were
3874                                  * below the threshold or if the threshold
3875                                  * is zero we always do the raise.
3876                                  */
3877                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3878                         } else {
3879                                 /* Log it stays the same */
3880                                 rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3881                                                 __LINE__, 11);
3882                         }
3883                         rack->rc_gp_timely_dec_cnt++;
3884                         /* We are not incrementing really no-count */
3885                         rack->rc_gp_incr = 0;
3886                         rack->rc_gp_timely_inc_cnt = 0;
3887                 } else {
3888                         /*
3889                          * Lets just use the RTT
3890                          * information and give up
3891                          * pushing.
3892                          */
3893                         goto use_timely;
3894                 }
3895         } else if ((timely_says != 2) &&
3896                     !losses &&
3897                     (last_bw_est > up_bnd)) {
3898                 /*
3899                  * We are increasing b/w lets keep going, updating
3900                  * our b/w and ignoring any timely input, unless
3901                  * of course we are at our max raise (if there is one).
3902                  */
3903
3904                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3905                                 __LINE__, 3);
3906                 rack->r_ctl.last_gp_comp_bw = cur_bw;
3907                 if (rack->rc_gp_saw_ss &&
3908                     rack_per_upper_bound_ss &&
3909                      (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3910                             /*
3911                              * In cases where we can't go higher
3912                              * we should just use timely.
3913                              */
3914                             goto use_timely;
3915                 }
3916                 if (rack->rc_gp_saw_ca &&
3917                     rack_per_upper_bound_ca &&
3918                     (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
3919                             /*
3920                              * In cases where we can't go higher
3921                              * we should just use timely.
3922                              */
3923                             goto use_timely;
3924                 }
3925                 rack->rc_gp_bwred = 0;
3926                 rack->rc_gp_timely_dec_cnt = 0;
3927                 /* You get a set number of pushes if timely is trying to reduce */
3928                 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
3929                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3930                 } else {
3931                         /* Log it stays the same */
3932                         rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
3933                             __LINE__, 12);
3934                 }
3935                 return;
3936         } else {
3937                 /*
3938                  * We are staying between the lower and upper range bounds
3939                  * so use timely to decide.
3940                  */
3941                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3942                                 __LINE__, 3);
3943 use_timely:
3944                 if (timely_says) {
3945                         rack->rc_gp_incr = 0;
3946                         rack->rc_gp_timely_inc_cnt = 0;
3947                         if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
3948                             !losses &&
3949                             (last_bw_est < low_bnd)) {
3950                                 /* We are loosing ground */
3951                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3952                                 rack->rc_gp_timely_dec_cnt++;
3953                                 /* We are not incrementing really no-count */
3954                                 rack->rc_gp_incr = 0;
3955                                 rack->rc_gp_timely_inc_cnt = 0;
3956                         } else
3957                                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3958                 } else {
3959                         rack->rc_gp_bwred = 0;
3960                         rack->rc_gp_timely_dec_cnt = 0;
3961                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3962                 }
3963         }
3964 }
3965
3966 static int32_t
3967 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
3968 {
3969         int32_t timely_says;
3970         uint64_t log_mult, log_rtt_a_diff;
3971
3972         log_rtt_a_diff = rtt;
3973         log_rtt_a_diff <<= 32;
3974         log_rtt_a_diff |= (uint32_t)rtt_diff;
3975         if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
3976                     rack_gp_rtt_maxmul)) {
3977                 /* Reduce the b/w multiplier */
3978                 timely_says = 2;
3979                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3980                 log_mult <<= 32;
3981                 log_mult |= prev_rtt;
3982                 rack_log_timely(rack,  timely_says, log_mult,
3983                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3984                                 log_rtt_a_diff, __LINE__, 4);
3985         } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
3986                            ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
3987                             max(rack_gp_rtt_mindiv , 1)))) {
3988                 /* Increase the b/w multiplier */
3989                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
3990                         ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
3991                          max(rack_gp_rtt_mindiv , 1));
3992                 log_mult <<= 32;
3993                 log_mult |= prev_rtt;
3994                 timely_says = 0;
3995                 rack_log_timely(rack,  timely_says, log_mult ,
3996                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3997                                 log_rtt_a_diff, __LINE__, 5);
3998         } else {
3999                 /*
4000                  * Use a gradient to find it the timely gradient
4001                  * is:
4002                  * grad = rc_rtt_diff / min_rtt;
4003                  *
4004                  * anything below or equal to 0 will be
4005                  * a increase indication. Anything above
4006                  * zero is a decrease. Note we take care
4007                  * of the actual gradient calculation
4008                  * in the reduction (its not needed for
4009                  * increase).
4010                  */
4011                 log_mult = prev_rtt;
4012                 if (rtt_diff <= 0) {
4013                         /*
4014                          * Rttdiff is less than zero, increase the
4015                          * b/w multiplier (its 0 or negative)
4016                          */
4017                         timely_says = 0;
4018                         rack_log_timely(rack,  timely_says, log_mult,
4019                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4020                 } else {
4021                         /* Reduce the b/w multiplier */
4022                         timely_says = 1;
4023                         rack_log_timely(rack,  timely_says, log_mult,
4024                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4025                 }
4026         }
4027         return (timely_says);
4028 }
4029
4030 static void
4031 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4032                             tcp_seq th_ack, int line, uint8_t quality)
4033 {
4034         uint64_t tim, bytes_ps, ltim, stim, utim;
4035         uint32_t segsiz, bytes, reqbytes, us_cts;
4036         int32_t gput, new_rtt_diff, timely_says;
4037         uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4038         int did_add = 0;
4039
4040         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4041         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4042         if (TSTMP_GEQ(us_cts, tp->gput_ts))
4043                 tim = us_cts - tp->gput_ts;
4044         else
4045                 tim = 0;
4046         if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4047                 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4048         else
4049                 stim = 0;
4050         /*
4051          * Use the larger of the send time or ack time. This prevents us
4052          * from being influenced by ack artifacts to come up with too
4053          * high of measurement. Note that since we are spanning over many more
4054          * bytes in most of our measurements hopefully that is less likely to
4055          * occur.
4056          */
4057         if (tim > stim)
4058                 utim = max(tim, 1);
4059         else
4060                 utim = max(stim, 1);
4061         /* Lets get a msec time ltim too for the old stuff */
4062         ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4063         gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4064         reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4065         if ((tim == 0) && (stim == 0)) {
4066                 /*
4067                  * Invalid measurement time, maybe
4068                  * all on one ack/one send?
4069                  */
4070                 bytes = 0;
4071                 bytes_ps = 0;
4072                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4073                                            0, 0, 0, 10, __LINE__, NULL, quality);
4074                 goto skip_measurement;
4075         }
4076         if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4077                 /* We never made a us_rtt measurement? */
4078                 bytes = 0;
4079                 bytes_ps = 0;
4080                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4081                                            0, 0, 0, 10, __LINE__, NULL, quality);
4082                 goto skip_measurement;
4083         }
4084         /*
4085          * Calculate the maximum possible b/w this connection
4086          * could have. We base our calculation on the lowest
4087          * rtt we have seen during the measurement and the
4088          * largest rwnd the client has given us in that time. This
4089          * forms a BDP that is the maximum that we could ever
4090          * get to the client. Anything larger is not valid.
4091          *
4092          * I originally had code here that rejected measurements
4093          * where the time was less than 1/2 the latest us_rtt.
4094          * But after thinking on that I realized its wrong since
4095          * say you had a 150Mbps or even 1Gbps link, and you
4096          * were a long way away.. example I am in Europe (100ms rtt)
4097          * talking to my 1Gbps link in S.C. Now measuring say 150,000
4098          * bytes my time would be 1.2ms, and yet my rtt would say
4099          * the measurement was invalid the time was < 50ms. The
4100          * same thing is true for 150Mb (8ms of time).
4101          *
4102          * A better way I realized is to look at what the maximum
4103          * the connection could possibly do. This is gated on
4104          * the lowest RTT we have seen and the highest rwnd.
4105          * We should in theory never exceed that, if we are
4106          * then something on the path is storing up packets
4107          * and then feeding them all at once to our endpoint
4108          * messing up our measurement.
4109          */
4110         rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4111         rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4112         rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4113         if (SEQ_LT(th_ack, tp->gput_seq)) {
4114                 /* No measurement can be made */
4115                 bytes = 0;
4116                 bytes_ps = 0;
4117                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4118                                            0, 0, 0, 10, __LINE__, NULL, quality);
4119                 goto skip_measurement;
4120         } else
4121                 bytes = (th_ack - tp->gput_seq);
4122         bytes_ps = (uint64_t)bytes;
4123         /*
4124          * Don't measure a b/w for pacing unless we have gotten at least
4125          * an initial windows worth of data in this measurement interval.
4126          *
4127          * Small numbers of bytes get badly influenced by delayed ack and
4128          * other artifacts. Note we take the initial window or our
4129          * defined minimum GP (defaulting to 10 which hopefully is the
4130          * IW).
4131          */
4132         if (rack->rc_gp_filled == 0) {
4133                 /*
4134                  * The initial estimate is special. We
4135                  * have blasted out an IW worth of packets
4136                  * without a real valid ack ts results. We
4137                  * then setup the app_limited_needs_set flag,
4138                  * this should get the first ack in (probably 2
4139                  * MSS worth) to be recorded as the timestamp.
4140                  * We thus allow a smaller number of bytes i.e.
4141                  * IW - 2MSS.
4142                  */
4143                 reqbytes -= (2 * segsiz);
4144                 /* Also lets fill previous for our first measurement to be neutral */
4145                 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4146         }
4147         if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4148                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4149                                            rack->r_ctl.rc_app_limited_cnt,
4150                                            0, 0, 10, __LINE__, NULL, quality);
4151                 goto skip_measurement;
4152         }
4153         /*
4154          * We now need to calculate the Timely like status so
4155          * we can update (possibly) the b/w multipliers.
4156          */
4157         new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4158         if (rack->rc_gp_filled == 0) {
4159                 /* No previous reading */
4160                 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4161         } else {
4162                 if (rack->measure_saw_probe_rtt == 0) {
4163                         /*
4164                          * We don't want a probertt to be counted
4165                          * since it will be negative incorrectly. We
4166                          * expect to be reducing the RTT when we
4167                          * pace at a slower rate.
4168                          */
4169                         rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4170                         rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4171                 }
4172         }
4173         timely_says = rack_make_timely_judgement(rack,
4174                 rack->r_ctl.rc_gp_srtt,
4175                 rack->r_ctl.rc_rtt_diff,
4176                 rack->r_ctl.rc_prev_gp_srtt
4177                 );
4178         bytes_ps *= HPTS_USEC_IN_SEC;
4179         bytes_ps /= utim;
4180         if (bytes_ps > rack->r_ctl.last_max_bw) {
4181                 /*
4182                  * Something is on path playing
4183                  * since this b/w is not possible based
4184                  * on our BDP (highest rwnd and lowest rtt
4185                  * we saw in the measurement window).
4186                  *
4187                  * Another option here would be to
4188                  * instead skip the measurement.
4189                  */
4190                 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4191                                            bytes_ps, rack->r_ctl.last_max_bw, 0,
4192                                            11, __LINE__, NULL, quality);
4193                 bytes_ps = rack->r_ctl.last_max_bw;
4194         }
4195         /* We store gp for b/w in bytes per second */
4196         if (rack->rc_gp_filled == 0) {
4197                 /* Initial measurement */
4198                 if (bytes_ps) {
4199                         rack->r_ctl.gp_bw = bytes_ps;
4200                         rack->rc_gp_filled = 1;
4201                         rack->r_ctl.num_measurements = 1;
4202                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4203                 } else {
4204                         rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4205                                                    rack->r_ctl.rc_app_limited_cnt,
4206                                                    0, 0, 10, __LINE__, NULL, quality);
4207                 }
4208                 if (tcp_in_hpts(rack->rc_inp) &&
4209                     (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4210                         /*
4211                          * Ok we can't trust the pacer in this case
4212                          * where we transition from un-paced to paced.
4213                          * Or for that matter when the burst mitigation
4214                          * was making a wild guess and got it wrong.
4215                          * Stop the pacer and clear up all the aggregate
4216                          * delays etc.
4217                          */
4218                         tcp_hpts_remove(rack->rc_inp);
4219                         rack->r_ctl.rc_hpts_flags = 0;
4220                         rack->r_ctl.rc_last_output_to = 0;
4221                 }
4222                 did_add = 2;
4223         } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4224                 /* Still a small number run an average */
4225                 rack->r_ctl.gp_bw += bytes_ps;
4226                 addpart = rack->r_ctl.num_measurements;
4227                 rack->r_ctl.num_measurements++;
4228                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4229                         /* We have collected enough to move forward */
4230                         rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4231                 }
4232                 did_add = 3;
4233         } else {
4234                 /*
4235                  * We want to take 1/wma of the goodput and add in to 7/8th
4236                  * of the old value weighted by the srtt. So if your measurement
4237                  * period is say 2 SRTT's long you would get 1/4 as the
4238                  * value, if it was like 1/2 SRTT then you would get 1/16th.
4239                  *
4240                  * But we must be careful not to take too much i.e. if the
4241                  * srtt is say 20ms and the measurement is taken over
4242                  * 400ms our weight would be 400/20 i.e. 20. On the
4243                  * other hand if we get a measurement over 1ms with a
4244                  * 10ms rtt we only want to take a much smaller portion.
4245                  */
4246                 if (rack->r_ctl.num_measurements < 0xff) {
4247                         rack->r_ctl.num_measurements++;
4248                 }
4249                 srtt = (uint64_t)tp->t_srtt;
4250                 if (srtt == 0) {
4251                         /*
4252                          * Strange why did t_srtt go back to zero?
4253                          */
4254                         if (rack->r_ctl.rc_rack_min_rtt)
4255                                 srtt = rack->r_ctl.rc_rack_min_rtt;
4256                         else
4257                                 srtt = HPTS_USEC_IN_MSEC;
4258                 }
4259                 /*
4260                  * XXXrrs: Note for reviewers, in playing with
4261                  * dynamic pacing I discovered this GP calculation
4262                  * as done originally leads to some undesired results.
4263                  * Basically you can get longer measurements contributing
4264                  * too much to the WMA. Thus I changed it if you are doing
4265                  * dynamic adjustments to only do the aportioned adjustment
4266                  * if we have a very small (time wise) measurement. Longer
4267                  * measurements just get there weight (defaulting to 1/8)
4268                  * add to the WMA. We may want to think about changing
4269                  * this to always do that for both sides i.e. dynamic
4270                  * and non-dynamic... but considering lots of folks
4271                  * were playing with this I did not want to change the
4272                  * calculation per.se. without your thoughts.. Lawerence?
4273                  * Peter??
4274                  */
4275                 if (rack->rc_gp_dyn_mul == 0) {
4276                         subpart = rack->r_ctl.gp_bw * utim;
4277                         subpart /= (srtt * 8);
4278                         if (subpart < (rack->r_ctl.gp_bw / 2)) {
4279                                 /*
4280                                  * The b/w update takes no more
4281                                  * away then 1/2 our running total
4282                                  * so factor it in.
4283                                  */
4284                                 addpart = bytes_ps * utim;
4285                                 addpart /= (srtt * 8);
4286                         } else {
4287                                 /*
4288                                  * Don't allow a single measurement
4289                                  * to account for more than 1/2 of the
4290                                  * WMA. This could happen on a retransmission
4291                                  * where utim becomes huge compared to
4292                                  * srtt (multiple retransmissions when using
4293                                  * the sending rate which factors in all the
4294                                  * transmissions from the first one).
4295                                  */
4296                                 subpart = rack->r_ctl.gp_bw / 2;
4297                                 addpart = bytes_ps / 2;
4298                         }
4299                         resid_bw = rack->r_ctl.gp_bw - subpart;
4300                         rack->r_ctl.gp_bw = resid_bw + addpart;
4301                         did_add = 1;
4302                 } else {
4303                         if ((utim / srtt) <= 1) {
4304                                 /*
4305                                  * The b/w update was over a small period
4306                                  * of time. The idea here is to prevent a small
4307                                  * measurement time period from counting
4308                                  * too much. So we scale it based on the
4309                                  * time so it attributes less than 1/rack_wma_divisor
4310                                  * of its measurement.
4311                                  */
4312                                 subpart = rack->r_ctl.gp_bw * utim;
4313                                 subpart /= (srtt * rack_wma_divisor);
4314                                 addpart = bytes_ps * utim;
4315                                 addpart /= (srtt * rack_wma_divisor);
4316                         } else {
4317                                 /*
4318                                  * The scaled measurement was long
4319                                  * enough so lets just add in the
4320                                  * portion of the measurement i.e. 1/rack_wma_divisor
4321                                  */
4322                                 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4323                                 addpart = bytes_ps / rack_wma_divisor;
4324                         }
4325                         if ((rack->measure_saw_probe_rtt == 0) ||
4326                             (bytes_ps > rack->r_ctl.gp_bw)) {
4327                                 /*
4328                                  * For probe-rtt we only add it in
4329                                  * if its larger, all others we just
4330                                  * add in.
4331                                  */
4332                                 did_add = 1;
4333                                 resid_bw = rack->r_ctl.gp_bw - subpart;
4334                                 rack->r_ctl.gp_bw = resid_bw + addpart;
4335                         }
4336                 }
4337         }
4338         if ((rack->gp_ready == 0) &&
4339             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4340                 /* We have enough measurements now */
4341                 rack->gp_ready = 1;
4342                 rack_set_cc_pacing(rack);
4343                 if (rack->defer_options)
4344                         rack_apply_deferred_options(rack);
4345         }
4346         rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4347                                    rack_get_bw(rack), 22, did_add, NULL, quality);
4348         /* We do not update any multipliers if we are in or have seen a probe-rtt */
4349         if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4350                 rack_update_multiplier(rack, timely_says, bytes_ps,
4351                                        rack->r_ctl.rc_gp_srtt,
4352                                        rack->r_ctl.rc_rtt_diff);
4353         rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4354                                    rack_get_bw(rack), 3, line, NULL, quality);
4355         /* reset the gp srtt and setup the new prev */
4356         rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4357         /* Record the lost count for the next measurement */
4358         rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4359         /*
4360          * We restart our diffs based on the gpsrtt in the
4361          * measurement window.
4362          */
4363         rack->rc_gp_rtt_set = 0;
4364         rack->rc_gp_saw_rec = 0;
4365         rack->rc_gp_saw_ca = 0;
4366         rack->rc_gp_saw_ss = 0;
4367         rack->rc_dragged_bottom = 0;
4368 skip_measurement:
4369
4370 #ifdef STATS
4371         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4372                                  gput);
4373         /*
4374          * XXXLAS: This is a temporary hack, and should be
4375          * chained off VOI_TCP_GPUT when stats(9) grows an
4376          * API to deal with chained VOIs.
4377          */
4378         if (tp->t_stats_gput_prev > 0)
4379                 stats_voi_update_abs_s32(tp->t_stats,
4380                                          VOI_TCP_GPUT_ND,
4381                                          ((gput - tp->t_stats_gput_prev) * 100) /
4382                                          tp->t_stats_gput_prev);
4383 #endif
4384         tp->t_flags &= ~TF_GPUTINPROG;
4385         tp->t_stats_gput_prev = gput;
4386         /*
4387          * Now are we app limited now and there is space from where we
4388          * were to where we want to go?
4389          *
4390          * We don't do the other case i.e. non-applimited here since
4391          * the next send will trigger us picking up the missing data.
4392          */
4393         if (rack->r_ctl.rc_first_appl &&
4394             TCPS_HAVEESTABLISHED(tp->t_state) &&
4395             rack->r_ctl.rc_app_limited_cnt &&
4396             (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4397             ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4398              max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4399                 /*
4400                  * Yep there is enough outstanding to make a measurement here.
4401                  */
4402                 struct rack_sendmap *rsm, fe;
4403
4404                 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4405                 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4406                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4407                 rack->app_limited_needs_set = 0;
4408                 tp->gput_seq = th_ack;
4409                 if (rack->in_probe_rtt)
4410                         rack->measure_saw_probe_rtt = 1;
4411                 else if ((rack->measure_saw_probe_rtt) &&
4412                          (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4413                         rack->measure_saw_probe_rtt = 0;
4414                 if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4415                         /* There is a full window to gain info from */
4416                         tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4417                 } else {
4418                         /* We can only measure up to the applimited point */
4419                         tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4420                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4421                                 /*
4422                                  * We don't have enough to make a measurement.
4423                                  */
4424                                 tp->t_flags &= ~TF_GPUTINPROG;
4425                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4426                                                            0, 0, 0, 6, __LINE__, NULL, quality);
4427                                 return;
4428                         }
4429                 }
4430                 if (tp->t_state >= TCPS_FIN_WAIT_1) {
4431                         /*
4432                          * We will get no more data into the SB
4433                          * this means we need to have the data available
4434                          * before we start a measurement.
4435                          */
4436                         if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4437                                 /* Nope not enough data. */
4438                                 return;
4439                         }
4440                 }
4441                 tp->t_flags |= TF_GPUTINPROG;
4442                 /*
4443                  * Now we need to find the timestamp of the send at tp->gput_seq
4444                  * for the send based measurement.
4445                  */
4446                 fe.r_start = tp->gput_seq;
4447                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4448                 if (rsm) {
4449                         /* Ok send-based limit is set */
4450                         if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4451                                 /*
4452                                  * Move back to include the earlier part
4453                                  * so our ack time lines up right (this may
4454                                  * make an overlapping measurement but thats
4455                                  * ok).
4456                                  */
4457                                 tp->gput_seq = rsm->r_start;
4458                         }
4459                         if (rsm->r_flags & RACK_ACKED)
4460                                 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4461                         else
4462                                 rack->app_limited_needs_set = 1;
4463                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4464                 } else {
4465                         /*
4466                          * If we don't find the rsm due to some
4467                          * send-limit set the current time, which
4468                          * basically disables the send-limit.
4469                          */
4470                         struct timeval tv;
4471
4472                         microuptime(&tv);
4473                         rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4474                 }
4475                 rack_log_pacing_delay_calc(rack,
4476                                            tp->gput_seq,
4477                                            tp->gput_ack,
4478                                            (uint64_t)rsm,
4479                                            tp->gput_ts,
4480                                            rack->r_ctl.rc_app_limited_cnt,
4481                                            9,
4482                                            __LINE__, NULL, quality);
4483         }
4484 }
4485
4486 /*
4487  * CC wrapper hook functions
4488  */
4489 static void
4490 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4491     uint16_t type, int32_t recovery)
4492 {
4493         uint32_t prior_cwnd, acked;
4494         struct tcp_log_buffer *lgb = NULL;
4495         uint8_t labc_to_use, quality;
4496
4497         INP_WLOCK_ASSERT(tptoinpcb(tp));
4498         tp->t_ccv.nsegs = nsegs;
4499         acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
4500         if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4501                 uint32_t max;
4502
4503                 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4504                 if (tp->t_ccv.bytes_this_ack > max) {
4505                         tp->t_ccv.bytes_this_ack = max;
4506                 }
4507         }
4508 #ifdef STATS
4509         stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4510             ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4511 #endif
4512         quality = RACK_QUALITY_NONE;
4513         if ((tp->t_flags & TF_GPUTINPROG) &&
4514             rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4515                 /* Measure the Goodput */
4516                 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4517 #ifdef NETFLIX_PEAKRATE
4518                 if ((type == CC_ACK) &&
4519                     (tp->t_maxpeakrate)) {
4520                         /*
4521                          * We update t_peakrate_thr. This gives us roughly
4522                          * one update per round trip time. Note
4523                          * it will only be used if pace_always is off i.e
4524                          * we don't do this for paced flows.
4525                          */
4526                         rack_update_peakrate_thr(tp);
4527                 }
4528 #endif
4529         }
4530         /* Which way our we limited, if not cwnd limited no advance in CA */
4531         if (tp->snd_cwnd <= tp->snd_wnd)
4532                 tp->t_ccv.flags |= CCF_CWND_LIMITED;
4533         else
4534                 tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
4535         if (tp->snd_cwnd > tp->snd_ssthresh) {
4536                 tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
4537                          nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4538                 /* For the setting of a window past use the actual scwnd we are using */
4539                 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4540                         tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4541                         tp->t_ccv.flags |= CCF_ABC_SENTAWND;
4542                 }
4543         } else {
4544                 tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4545                 tp->t_bytes_acked = 0;
4546         }
4547         prior_cwnd = tp->snd_cwnd;
4548         if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4549             (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4550                 labc_to_use = rack->rc_labc;
4551         else
4552                 labc_to_use = rack_max_abc_post_recovery;
4553         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
4554                 union tcp_log_stackspecific log;
4555                 struct timeval tv;
4556
4557                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4558                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4559                 log.u_bbr.flex1 = th_ack;
4560                 log.u_bbr.flex2 = tp->t_ccv.flags;
4561                 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
4562                 log.u_bbr.flex4 = tp->t_ccv.nsegs;
4563                 log.u_bbr.flex5 = labc_to_use;
4564                 log.u_bbr.flex6 = prior_cwnd;
4565                 log.u_bbr.flex7 = V_tcp_do_newsack;
4566                 log.u_bbr.flex8 = 1;
4567                 lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4568                                      0, &log, false, NULL, NULL, 0, &tv);
4569         }
4570         if (CC_ALGO(tp)->ack_received != NULL) {
4571                 /* XXXLAS: Find a way to live without this */
4572                 tp->t_ccv.curack = th_ack;
4573                 tp->t_ccv.labc = labc_to_use;
4574                 tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
4575                 CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
4576         }
4577         if (lgb) {
4578                 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4579         }
4580         if (rack->r_must_retran) {
4581                 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4582                         /*
4583                          * We now are beyond the rxt point so lets disable
4584                          * the flag.
4585                          */
4586                         rack->r_ctl.rc_out_at_rto = 0;
4587                         rack->r_must_retran = 0;
4588                 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4589                         /*
4590                          * Only decrement the rc_out_at_rto if the cwnd advances
4591                          * at least a whole segment. Otherwise next time the peer
4592                          * acks, we won't be able to send this generaly happens
4593                          * when we are in Congestion Avoidance.
4594                          */
4595                         if (acked <= rack->r_ctl.rc_out_at_rto){
4596                                 rack->r_ctl.rc_out_at_rto -= acked;
4597                         } else {
4598                                 rack->r_ctl.rc_out_at_rto = 0;
4599                         }
4600                 }
4601         }
4602 #ifdef STATS
4603         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4604 #endif
4605         if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4606                 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4607         }
4608 #ifdef NETFLIX_PEAKRATE
4609         /* we enforce max peak rate if it is set and we are not pacing */
4610         if ((rack->rc_always_pace == 0) &&
4611             tp->t_peakrate_thr &&
4612             (tp->snd_cwnd > tp->t_peakrate_thr)) {
4613                 tp->snd_cwnd = tp->t_peakrate_thr;
4614         }
4615 #endif
4616 }
4617
4618 static void
4619 tcp_rack_partialack(struct tcpcb *tp)
4620 {
4621         struct tcp_rack *rack;
4622
4623         rack = (struct tcp_rack *)tp->t_fb_ptr;
4624         INP_WLOCK_ASSERT(tptoinpcb(tp));
4625         /*
4626          * If we are doing PRR and have enough
4627          * room to send <or> we are pacing and prr
4628          * is disabled we will want to see if we
4629          * can send data (by setting r_wanted_output to
4630          * true).
4631          */
4632         if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4633             rack->rack_no_prr)
4634                 rack->r_wanted_output = 1;
4635 }
4636
4637 static void
4638 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4639 {
4640         struct tcp_rack *rack;
4641         uint32_t orig_cwnd;
4642
4643         orig_cwnd = tp->snd_cwnd;
4644         INP_WLOCK_ASSERT(tptoinpcb(tp));
4645         rack = (struct tcp_rack *)tp->t_fb_ptr;
4646         /* only alert CC if we alerted when we entered */
4647         if (CC_ALGO(tp)->post_recovery != NULL) {
4648                 tp->t_ccv.curack = th_ack;
4649                 CC_ALGO(tp)->post_recovery(&tp->t_ccv);
4650                 if (tp->snd_cwnd < tp->snd_ssthresh) {
4651                         /*
4652                          * Rack has burst control and pacing
4653                          * so lets not set this any lower than
4654                          * snd_ssthresh per RFC-6582 (option 2).
4655                          */
4656                         tp->snd_cwnd = tp->snd_ssthresh;
4657                 }
4658         }
4659         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
4660                 union tcp_log_stackspecific log;
4661                 struct timeval tv;
4662
4663                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4664                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4665                 log.u_bbr.flex1 = th_ack;
4666                 log.u_bbr.flex2 = tp->t_ccv.flags;
4667                 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
4668                 log.u_bbr.flex4 = tp->t_ccv.nsegs;
4669                 log.u_bbr.flex5 = V_tcp_abc_l_var;
4670                 log.u_bbr.flex6 = orig_cwnd;
4671                 log.u_bbr.flex7 = V_tcp_do_newsack;
4672                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4673                 log.u_bbr.flex8 = 2;
4674                 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4675                                0, &log, false, NULL, NULL, 0, &tv);
4676         }
4677         if ((rack->rack_no_prr == 0) &&
4678             (rack->no_prr_addback == 0) &&
4679             (rack->r_ctl.rc_prr_sndcnt > 0)) {
4680                 /*
4681                  * Suck the next prr cnt back into cwnd, but
4682                  * only do that if we are not application limited.
4683                  */
4684                 if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
4685                         /*
4686                          * We are allowed to add back to the cwnd the amount we did
4687                          * not get out if:
4688                          * a) no_prr_addback is off.
4689                          * b) we are not app limited
4690                          * c) we are doing prr
4691                          * <and>
4692                          * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4693                          */
4694                         tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4695                                             rack->r_ctl.rc_prr_sndcnt);
4696                 }
4697                 rack->r_ctl.rc_prr_sndcnt = 0;
4698                 rack_log_to_prr(rack, 1, 0, __LINE__);
4699         }
4700         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
4701         tp->snd_recover = tp->snd_una;
4702         if (rack->r_ctl.dsack_persist) {
4703                 rack->r_ctl.dsack_persist--;
4704                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4705                         rack->r_ctl.num_dsack = 0;
4706                 }
4707                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4708         }
4709         EXIT_RECOVERY(tp->t_flags);
4710 }
4711
4712 static void
4713 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
4714 {
4715         struct tcp_rack *rack;
4716         uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4717
4718         INP_WLOCK_ASSERT(tptoinpcb(tp));
4719 #ifdef STATS
4720         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4721 #endif
4722         if (IN_RECOVERY(tp->t_flags) == 0) {
4723                 in_rec_at_entry = 0;
4724                 ssthresh_enter = tp->snd_ssthresh;
4725                 cwnd_enter = tp->snd_cwnd;
4726         } else
4727                 in_rec_at_entry = 1;
4728         rack = (struct tcp_rack *)tp->t_fb_ptr;
4729         switch (type) {
4730         case CC_NDUPACK:
4731                 tp->t_flags &= ~TF_WASFRECOVERY;
4732                 tp->t_flags &= ~TF_WASCRECOVERY;
4733                 if (!IN_FASTRECOVERY(tp->t_flags)) {
4734                         rack->r_ctl.rc_prr_delivered = 0;
4735                         rack->r_ctl.rc_prr_out = 0;
4736                         if (rack->rack_no_prr == 0) {
4737                                 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4738                                 rack_log_to_prr(rack, 2, in_rec_at_entry, line);
4739                         }
4740                         rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4741                         tp->snd_recover = tp->snd_max;
4742                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4743                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4744                 }
4745                 break;
4746         case CC_ECN:
4747                 if (!IN_CONGRECOVERY(tp->t_flags) ||
4748                     /*
4749                      * Allow ECN reaction on ACK to CWR, if
4750                      * that data segment was also CE marked.
4751                      */
4752                     SEQ_GEQ(ack, tp->snd_recover)) {
4753                         EXIT_CONGRECOVERY(tp->t_flags);
4754                         KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4755                         tp->snd_recover = tp->snd_max + 1;
4756                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4757                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4758                 }
4759                 break;
4760         case CC_RTO:
4761                 tp->t_dupacks = 0;
4762                 tp->t_bytes_acked = 0;
4763                 EXIT_RECOVERY(tp->t_flags);
4764                 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4765                     ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4766                 orig_cwnd = tp->snd_cwnd;
4767                 tp->snd_cwnd = ctf_fixed_maxseg(tp);
4768                 rack_log_to_prr(rack, 16, orig_cwnd, line);
4769                 if (tp->t_flags2 & TF2_ECN_PERMIT)
4770                         tp->t_flags2 |= TF2_ECN_SND_CWR;
4771                 break;
4772         case CC_RTO_ERR:
4773                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4774                 /* RTO was unnecessary, so reset everything. */
4775                 tp->snd_cwnd = tp->snd_cwnd_prev;
4776                 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4777                 tp->snd_recover = tp->snd_recover_prev;
4778                 if (tp->t_flags & TF_WASFRECOVERY) {
4779                         ENTER_FASTRECOVERY(tp->t_flags);
4780                         tp->t_flags &= ~TF_WASFRECOVERY;
4781                 }
4782                 if (tp->t_flags & TF_WASCRECOVERY) {
4783                         ENTER_CONGRECOVERY(tp->t_flags);
4784                         tp->t_flags &= ~TF_WASCRECOVERY;
4785                 }
4786                 tp->snd_nxt = tp->snd_max;
4787                 tp->t_badrxtwin = 0;
4788                 break;
4789         }
4790         if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4791             (type != CC_RTO)){
4792                 tp->t_ccv.curack = ack;
4793                 CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
4794         }
4795         if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4796                 rack_log_to_prr(rack, 15, cwnd_enter, line);
4797                 rack->r_ctl.dsack_byte_cnt = 0;
4798                 rack->r_ctl.retran_during_recovery = 0;
4799                 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4800                 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4801                 rack->r_ent_rec_ns = 1;
4802         }
4803 }
4804
4805 static inline void
4806 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4807 {
4808         uint32_t i_cwnd;
4809
4810         INP_WLOCK_ASSERT(tptoinpcb(tp));
4811
4812 #ifdef NETFLIX_STATS
4813         KMOD_TCPSTAT_INC(tcps_idle_restarts);
4814         if (tp->t_state == TCPS_ESTABLISHED)
4815                 KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4816 #endif
4817         if (CC_ALGO(tp)->after_idle != NULL)
4818                 CC_ALGO(tp)->after_idle(&tp->t_ccv);
4819
4820         if (tp->snd_cwnd == 1)
4821                 i_cwnd = tp->t_maxseg;          /* SYN(-ACK) lost */
4822         else
4823                 i_cwnd = rc_init_window(rack);
4824
4825         /*
4826          * Being idle is no different than the initial window. If the cc
4827          * clamps it down below the initial window raise it to the initial
4828          * window.
4829          */
4830         if (tp->snd_cwnd < i_cwnd) {
4831                 tp->snd_cwnd = i_cwnd;
4832         }
4833 }
4834
4835 /*
4836  * Indicate whether this ack should be delayed.  We can delay the ack if
4837  * following conditions are met:
4838  *      - There is no delayed ack timer in progress.
4839  *      - Our last ack wasn't a 0-sized window. We never want to delay
4840  *        the ack that opens up a 0-sized window.
4841  *      - LRO wasn't used for this segment. We make sure by checking that the
4842  *        segment size is not larger than the MSS.
4843  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4844  *        connection.
4845  */
4846 #define DELAY_ACK(tp, tlen)                      \
4847         (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4848         ((tp->t_flags & TF_DELACK) == 0) &&      \
4849         (tlen <= tp->t_maxseg) &&                \
4850         (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4851
4852 static struct rack_sendmap *
4853 rack_find_lowest_rsm(struct tcp_rack *rack)
4854 {
4855         struct rack_sendmap *rsm;
4856
4857         /*
4858          * Walk the time-order transmitted list looking for an rsm that is
4859          * not acked. This will be the one that was sent the longest time
4860          * ago that is still outstanding.
4861          */
4862         TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4863                 if (rsm->r_flags & RACK_ACKED) {
4864                         continue;
4865                 }
4866                 goto finish;
4867         }
4868 finish:
4869         return (rsm);
4870 }
4871
4872 static struct rack_sendmap *
4873 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4874 {
4875         struct rack_sendmap *prsm;
4876
4877         /*
4878          * Walk the sequence order list backward until we hit and arrive at
4879          * the highest seq not acked. In theory when this is called it
4880          * should be the last segment (which it was not).
4881          */
4882         prsm = rsm;
4883         RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4884                 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4885                         continue;
4886                 }
4887                 return (prsm);
4888         }
4889         return (NULL);
4890 }
4891
4892 static uint32_t
4893 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4894 {
4895         int32_t lro;
4896         uint32_t thresh;
4897
4898         /*
4899          * lro is the flag we use to determine if we have seen reordering.
4900          * If it gets set we have seen reordering. The reorder logic either
4901          * works in one of two ways:
4902          *
4903          * If reorder-fade is configured, then we track the last time we saw
4904          * re-ordering occur. If we reach the point where enough time as
4905          * passed we no longer consider reordering has occuring.
4906          *
4907          * Or if reorder-face is 0, then once we see reordering we consider
4908          * the connection to alway be subject to reordering and just set lro
4909          * to 1.
4910          *
4911          * In the end if lro is non-zero we add the extra time for
4912          * reordering in.
4913          */
4914         if (srtt == 0)
4915                 srtt = 1;
4916         if (rack->r_ctl.rc_reorder_ts) {
4917                 if (rack->r_ctl.rc_reorder_fade) {
4918                         if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
4919                                 lro = cts - rack->r_ctl.rc_reorder_ts;
4920                                 if (lro == 0) {
4921                                         /*
4922                                          * No time as passed since the last
4923                                          * reorder, mark it as reordering.
4924                                          */
4925                                         lro = 1;
4926                                 }
4927                         } else {
4928                                 /* Negative time? */
4929                                 lro = 0;
4930                         }
4931                         if (lro > rack->r_ctl.rc_reorder_fade) {
4932                                 /* Turn off reordering seen too */
4933                                 rack->r_ctl.rc_reorder_ts = 0;
4934                                 lro = 0;
4935                         }
4936                 } else {
4937                         /* Reodering does not fade */
4938                         lro = 1;
4939                 }
4940         } else {
4941                 lro = 0;
4942         }
4943         if (rack->rc_rack_tmr_std_based == 0) {
4944                 thresh = srtt + rack->r_ctl.rc_pkt_delay;
4945         } else {
4946                 /* Standards based pkt-delay is 1/4 srtt */
4947                 thresh = srtt +  (srtt >> 2);
4948         }
4949         if (lro && (rack->rc_rack_tmr_std_based == 0)) {
4950                 /* It must be set, if not you get 1/4 rtt */
4951                 if (rack->r_ctl.rc_reorder_shift)
4952                         thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
4953                 else
4954                         thresh += (srtt >> 2);
4955         }
4956         if (rack->rc_rack_use_dsack &&
4957             lro &&
4958             (rack->r_ctl.num_dsack > 0)) {
4959                 /*
4960                  * We only increase the reordering window if we
4961                  * have seen reordering <and> we have a DSACK count.
4962                  */
4963                 thresh += rack->r_ctl.num_dsack * (srtt >> 2);
4964                 rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
4965         }
4966         /* SRTT * 2 is the ceiling */
4967         if (thresh > (srtt * 2)) {
4968                 thresh = srtt * 2;
4969         }
4970         /* And we don't want it above the RTO max either */
4971         if (thresh > rack_rto_max) {
4972                 thresh = rack_rto_max;
4973         }
4974         rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
4975         return (thresh);
4976 }
4977
4978 static uint32_t
4979 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
4980                      struct rack_sendmap *rsm, uint32_t srtt)
4981 {
4982         struct rack_sendmap *prsm;
4983         uint32_t thresh, len;
4984         int segsiz;
4985
4986         if (srtt == 0)
4987                 srtt = 1;
4988         if (rack->r_ctl.rc_tlp_threshold)
4989                 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
4990         else
4991                 thresh = (srtt * 2);
4992
4993         /* Get the previous sent packet, if any */
4994         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4995         len = rsm->r_end - rsm->r_start;
4996         if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
4997                 /* Exactly like the ID */
4998                 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
4999                         uint32_t alt_thresh;
5000                         /*
5001                          * Compensate for delayed-ack with the d-ack time.
5002                          */
5003                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5004                         if (alt_thresh > thresh)
5005                                 thresh = alt_thresh;
5006                 }
5007         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5008                 /* 2.1 behavior */
5009                 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5010                 if (prsm && (len <= segsiz)) {
5011                         /*
5012                          * Two packets outstanding, thresh should be (2*srtt) +
5013                          * possible inter-packet delay (if any).
5014                          */
5015                         uint32_t inter_gap = 0;
5016                         int idx, nidx;
5017
5018                         idx = rsm->r_rtr_cnt - 1;
5019                         nidx = prsm->r_rtr_cnt - 1;
5020                         if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5021                                 /* Yes it was sent later (or at the same time) */
5022                                 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5023                         }
5024                         thresh += inter_gap;
5025                 } else if (len <= segsiz) {
5026                         /*
5027                          * Possibly compensate for delayed-ack.
5028                          */
5029                         uint32_t alt_thresh;
5030
5031                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5032                         if (alt_thresh > thresh)
5033                                 thresh = alt_thresh;
5034                 }
5035         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5036                 /* 2.2 behavior */
5037                 if (len <= segsiz) {
5038                         uint32_t alt_thresh;
5039                         /*
5040                          * Compensate for delayed-ack with the d-ack time.
5041                          */
5042                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5043                         if (alt_thresh > thresh)
5044                                 thresh = alt_thresh;
5045                 }
5046         }
5047         /* Not above an RTO */
5048         if (thresh > tp->t_rxtcur) {
5049                 thresh = tp->t_rxtcur;
5050         }
5051         /* Not above a RTO max */
5052         if (thresh > rack_rto_max) {
5053                 thresh = rack_rto_max;
5054         }
5055         /* Apply user supplied min TLP */
5056         if (thresh < rack_tlp_min) {
5057                 thresh = rack_tlp_min;
5058         }
5059         return (thresh);
5060 }
5061
5062 static uint32_t
5063 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5064 {
5065         /*
5066          * We want the rack_rtt which is the
5067          * last rtt we measured. However if that
5068          * does not exist we fallback to the srtt (which
5069          * we probably will never do) and then as a last
5070          * resort we use RACK_INITIAL_RTO if no srtt is
5071          * yet set.
5072          */
5073         if (rack->rc_rack_rtt)
5074                 return (rack->rc_rack_rtt);
5075         else if (tp->t_srtt == 0)
5076                 return (RACK_INITIAL_RTO);
5077         return (tp->t_srtt);
5078 }
5079
5080 static struct rack_sendmap *
5081 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5082 {
5083         /*
5084          * Check to see that we don't need to fall into recovery. We will
5085          * need to do so if our oldest transmit is past the time we should
5086          * have had an ack.
5087          */
5088         struct tcp_rack *rack;
5089         struct rack_sendmap *rsm;
5090         int32_t idx;
5091         uint32_t srtt, thresh;
5092
5093         rack = (struct tcp_rack *)tp->t_fb_ptr;
5094         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5095                 return (NULL);
5096         }
5097         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5098         if (rsm == NULL)
5099                 return (NULL);
5100
5101
5102         if (rsm->r_flags & RACK_ACKED) {
5103                 rsm = rack_find_lowest_rsm(rack);
5104                 if (rsm == NULL)
5105                         return (NULL);
5106         }
5107         idx = rsm->r_rtr_cnt - 1;
5108         srtt = rack_grab_rtt(tp, rack);
5109         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5110         if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5111                 return (NULL);
5112         }
5113         if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5114                 return (NULL);
5115         }
5116         /* Ok if we reach here we are over-due and this guy can be sent */
5117         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
5118         return (rsm);
5119 }
5120
5121 static uint32_t
5122 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5123 {
5124         int32_t t;
5125         int32_t tt;
5126         uint32_t ret_val;
5127
5128         t = (tp->t_srtt + (tp->t_rttvar << 2));
5129         RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5130             rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5131         rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5132         ret_val = (uint32_t)tt;
5133         return (ret_val);
5134 }
5135
5136 static uint32_t
5137 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5138 {
5139         /*
5140          * Start the FR timer, we do this based on getting the first one in
5141          * the rc_tmap. Note that if its NULL we must stop the timer. in all
5142          * events we need to stop the running timer (if its running) before
5143          * starting the new one.
5144          */
5145         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5146         uint32_t srtt_cur;
5147         int32_t idx;
5148         int32_t is_tlp_timer = 0;
5149         struct rack_sendmap *rsm;
5150
5151         if (rack->t_timers_stopped) {
5152                 /* All timers have been stopped none are to run */
5153                 return (0);
5154         }
5155         if (rack->rc_in_persist) {
5156                 /* We can't start any timer in persists */
5157                 return (rack_get_persists_timer_val(tp, rack));
5158         }
5159         rack->rc_on_min_to = 0;
5160         if ((tp->t_state < TCPS_ESTABLISHED) ||
5161             ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5162                 goto activate_rxt;
5163         }
5164         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5165         if ((rsm == NULL) || sup_rack) {
5166                 /* Nothing on the send map or no rack */
5167 activate_rxt:
5168                 time_since_sent = 0;
5169                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5170                 if (rsm) {
5171                         /*
5172                          * Should we discount the RTX timer any?
5173                          *
5174                          * We want to discount it the smallest amount.
5175                          * If a timer (Rack/TLP or RXT) has gone off more
5176                          * recently thats the discount we want to use (now - timer time).
5177                          * If the retransmit of the oldest packet was more recent then
5178                          * we want to use that (now - oldest-packet-last_transmit_time).
5179                          *
5180                          */
5181                         idx = rsm->r_rtr_cnt - 1;
5182                         if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5183                                 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5184                         else
5185                                 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5186                         if (TSTMP_GT(cts, tstmp_touse))
5187                             time_since_sent = cts - tstmp_touse;
5188                 }
5189                 if (SEQ_LT(tp->snd_una, tp->snd_max) ||
5190                     sbavail(&tptosocket(tp)->so_snd)) {
5191                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5192                         to = tp->t_rxtcur;
5193                         if (to > time_since_sent)
5194                                 to -= time_since_sent;
5195                         else
5196                                 to = rack->r_ctl.rc_min_to;
5197                         if (to == 0)
5198                                 to = 1;
5199                         /* Special case for KEEPINIT */
5200                         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5201                             (TP_KEEPINIT(tp) != 0) &&
5202                             rsm) {
5203                                 /*
5204                                  * We have to put a ceiling on the rxt timer
5205                                  * of the keep-init timeout.
5206                                  */
5207                                 uint32_t max_time, red;
5208
5209                                 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5210                                 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5211                                         red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5212                                         if (red < max_time)
5213                                                 max_time -= red;
5214                                         else
5215                                                 max_time = 1;
5216                                 }
5217                                 /* Reduce timeout to the keep value if needed */
5218                                 if (max_time < to)
5219                                         to = max_time;
5220                         }
5221                         return (to);
5222                 }
5223                 return (0);
5224         }
5225         if (rsm->r_flags & RACK_ACKED) {
5226                 rsm = rack_find_lowest_rsm(rack);
5227                 if (rsm == NULL) {
5228                         /* No lowest? */
5229                         goto activate_rxt;
5230                 }
5231         }
5232         if (rack->sack_attack_disable) {
5233                 /*
5234                  * We don't want to do
5235                  * any TLP's if you are an attacker.
5236                  * Though if you are doing what
5237                  * is expected you may still have
5238                  * SACK-PASSED marks.
5239                  */
5240                 goto activate_rxt;
5241         }
5242         /* Convert from ms to usecs */
5243         if ((rsm->r_flags & RACK_SACK_PASSED) ||
5244             (rsm->r_flags & RACK_RWND_COLLAPSED) ||
5245             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5246                 if ((tp->t_flags & TF_SENTFIN) &&
5247                     ((tp->snd_max - tp->snd_una) == 1) &&
5248                     (rsm->r_flags & RACK_HAS_FIN)) {
5249                         /*
5250                          * We don't start a rack timer if all we have is a
5251                          * FIN outstanding.
5252                          */
5253                         goto activate_rxt;
5254                 }
5255                 if ((rack->use_rack_rr == 0) &&
5256                     (IN_FASTRECOVERY(tp->t_flags)) &&
5257                     (rack->rack_no_prr == 0) &&
5258                      (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5259                         /*
5260                          * We are not cheating, in recovery  and
5261                          * not enough ack's to yet get our next
5262                          * retransmission out.
5263                          *
5264                          * Note that classified attackers do not
5265                          * get to use the rack-cheat.
5266                          */
5267                         goto activate_tlp;
5268                 }
5269                 srtt = rack_grab_rtt(tp, rack);
5270                 thresh = rack_calc_thresh_rack(rack, srtt, cts);
5271                 idx = rsm->r_rtr_cnt - 1;
5272                 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5273                 if (SEQ_GEQ(exp, cts)) {
5274                         to = exp - cts;
5275                         if (to < rack->r_ctl.rc_min_to) {
5276                                 to = rack->r_ctl.rc_min_to;
5277                                 if (rack->r_rr_config == 3)
5278                                         rack->rc_on_min_to = 1;
5279                         }
5280                 } else {
5281                         to = rack->r_ctl.rc_min_to;
5282                         if (rack->r_rr_config == 3)
5283                                 rack->rc_on_min_to = 1;
5284                 }
5285         } else {
5286                 /* Ok we need to do a TLP not RACK */
5287 activate_tlp:
5288                 if ((rack->rc_tlp_in_progress != 0) &&
5289                     (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5290                         /*
5291                          * The previous send was a TLP and we have sent
5292                          * N TLP's without sending new data.
5293                          */
5294                         goto activate_rxt;
5295                 }
5296                 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5297                 if (rsm == NULL) {
5298                         /* We found no rsm to TLP with. */
5299                         goto activate_rxt;
5300                 }
5301                 if (rsm->r_flags & RACK_HAS_FIN) {
5302                         /* If its a FIN we dont do TLP */
5303                         rsm = NULL;
5304                         goto activate_rxt;
5305                 }
5306                 idx = rsm->r_rtr_cnt - 1;
5307                 time_since_sent = 0;
5308                 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5309                         tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5310                 else
5311                         tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5312                 if (TSTMP_GT(cts, tstmp_touse))
5313                     time_since_sent = cts - tstmp_touse;
5314                 is_tlp_timer = 1;
5315                 if (tp->t_srtt) {
5316                         if ((rack->rc_srtt_measure_made == 0) &&
5317                             (tp->t_srtt == 1)) {
5318                                 /*
5319                                  * If another stack as run and set srtt to 1,
5320                                  * then the srtt was 0, so lets use the initial.
5321                                  */
5322                                 srtt = RACK_INITIAL_RTO;
5323                         } else {
5324                                 srtt_cur = tp->t_srtt;
5325                                 srtt = srtt_cur;
5326                         }
5327                 } else
5328                         srtt = RACK_INITIAL_RTO;
5329                 /*
5330                  * If the SRTT is not keeping up and the
5331                  * rack RTT has spiked we want to use
5332                  * the last RTT not the smoothed one.
5333                  */
5334                 if (rack_tlp_use_greater &&
5335                     tp->t_srtt &&
5336                     (srtt < rack_grab_rtt(tp, rack))) {
5337                         srtt = rack_grab_rtt(tp, rack);
5338                 }
5339                 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5340                 if (thresh > time_since_sent) {
5341                         to = thresh - time_since_sent;
5342                 } else {
5343                         to = rack->r_ctl.rc_min_to;
5344                         rack_log_alt_to_to_cancel(rack,
5345                                                   thresh,               /* flex1 */
5346                                                   time_since_sent,      /* flex2 */
5347                                                   tstmp_touse,          /* flex3 */
5348                                                   rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5349                                                   (uint32_t)rsm->r_tim_lastsent[idx],
5350                                                   srtt,
5351                                                   idx, 99);
5352                 }
5353                 if (to < rack_tlp_min) {
5354                         to = rack_tlp_min;
5355                 }
5356                 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5357                         /*
5358                          * If the TLP time works out to larger than the max
5359                          * RTO lets not do TLP.. just RTO.
5360                          */
5361                         goto activate_rxt;
5362                 }
5363         }
5364         if (is_tlp_timer == 0) {
5365                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5366         } else {
5367                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5368         }
5369         if (to == 0)
5370                 to = 1;
5371         return (to);
5372 }
5373
5374 static void
5375 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5376 {
5377         if (rack->rc_in_persist == 0) {
5378                 if (tp->t_flags & TF_GPUTINPROG) {
5379                         /*
5380                          * Stop the goodput now, the calling of the
5381                          * measurement function clears the flag.
5382                          */
5383                         rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5384                                                     RACK_QUALITY_PERSIST);
5385                 }
5386 #ifdef NETFLIX_SHARED_CWND
5387                 if (rack->r_ctl.rc_scw) {
5388                         tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5389                         rack->rack_scwnd_is_idle = 1;
5390                 }
5391 #endif
5392                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5393                 if (rack->r_ctl.rc_went_idle_time == 0)
5394                         rack->r_ctl.rc_went_idle_time = 1;
5395                 rack_timer_cancel(tp, rack, cts, __LINE__);
5396                 rack->r_ctl.persist_lost_ends = 0;
5397                 rack->probe_not_answered = 0;
5398                 rack->forced_ack = 0;
5399                 tp->t_rxtshift = 0;
5400                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5401                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5402                 rack->rc_in_persist = 1;
5403         }
5404 }
5405
5406 static void
5407 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5408 {
5409         if (tcp_in_hpts(rack->rc_inp)) {
5410                 tcp_hpts_remove(rack->rc_inp);
5411                 rack->r_ctl.rc_hpts_flags = 0;
5412         }
5413 #ifdef NETFLIX_SHARED_CWND
5414         if (rack->r_ctl.rc_scw) {
5415                 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5416                 rack->rack_scwnd_is_idle = 0;
5417         }
5418 #endif
5419         if (rack->rc_gp_dyn_mul &&
5420             (rack->use_fixed_rate == 0) &&
5421             (rack->rc_always_pace)) {
5422                 /*
5423                  * Do we count this as if a probe-rtt just
5424                  * finished?
5425                  */
5426                 uint32_t time_idle, idle_min;
5427
5428                 time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5429                 idle_min = rack_min_probertt_hold;
5430                 if (rack_probertt_gpsrtt_cnt_div) {
5431                         uint64_t extra;
5432                         extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5433                                 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
5434                         extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5435                         idle_min += (uint32_t)extra;
5436                 }
5437                 if (time_idle >= idle_min) {
5438                         /* Yes, we count it as a probe-rtt. */
5439                         uint32_t us_cts;
5440
5441                         us_cts = tcp_get_usecs(NULL);
5442                         if (rack->in_probe_rtt == 0) {
5443                                 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5444                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5445                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5446                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5447                         } else {
5448                                 rack_exit_probertt(rack, us_cts);
5449                         }
5450                 }
5451         }
5452         rack->rc_in_persist = 0;
5453         rack->r_ctl.rc_went_idle_time = 0;
5454         tp->t_rxtshift = 0;
5455         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5456            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5457         rack->r_ctl.rc_agg_delayed = 0;
5458         rack->r_early = 0;
5459         rack->r_late = 0;
5460         rack->r_ctl.rc_agg_early = 0;
5461 }
5462
5463 static void
5464 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5465                    struct hpts_diag *diag, struct timeval *tv)
5466 {
5467         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5468                 union tcp_log_stackspecific log;
5469
5470                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5471                 log.u_bbr.flex1 = diag->p_nxt_slot;
5472                 log.u_bbr.flex2 = diag->p_cur_slot;
5473                 log.u_bbr.flex3 = diag->slot_req;
5474                 log.u_bbr.flex4 = diag->inp_hptsslot;
5475                 log.u_bbr.flex5 = diag->slot_remaining;
5476                 log.u_bbr.flex6 = diag->need_new_to;
5477                 log.u_bbr.flex7 = diag->p_hpts_active;
5478                 log.u_bbr.flex8 = diag->p_on_min_sleep;
5479                 /* Hijack other fields as needed */
5480                 log.u_bbr.epoch = diag->have_slept;
5481                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
5482                 log.u_bbr.pkts_out = diag->co_ret;
5483                 log.u_bbr.applimited = diag->hpts_sleep_time;
5484                 log.u_bbr.delivered = diag->p_prev_slot;
5485                 log.u_bbr.inflight = diag->p_runningslot;
5486                 log.u_bbr.bw_inuse = diag->wheel_slot;
5487                 log.u_bbr.rttProp = diag->wheel_cts;
5488                 log.u_bbr.timeStamp = cts;
5489                 log.u_bbr.delRate = diag->maxslots;
5490                 log.u_bbr.cur_del_rate = diag->p_curtick;
5491                 log.u_bbr.cur_del_rate <<= 32;
5492                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
5493                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5494                     &rack->rc_inp->inp_socket->so_rcv,
5495                     &rack->rc_inp->inp_socket->so_snd,
5496                     BBR_LOG_HPTSDIAG, 0,
5497                     0, &log, false, tv);
5498         }
5499
5500 }
5501
5502 static void
5503 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5504 {
5505         if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5506                 union tcp_log_stackspecific log;
5507                 struct timeval tv;
5508
5509                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5510                 log.u_bbr.flex1 = sb->sb_flags;
5511                 log.u_bbr.flex2 = len;
5512                 log.u_bbr.flex3 = sb->sb_state;
5513                 log.u_bbr.flex8 = type;
5514                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5515                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5516                     &rack->rc_inp->inp_socket->so_rcv,
5517                     &rack->rc_inp->inp_socket->so_snd,
5518                     TCP_LOG_SB_WAKE, 0,
5519                     len, &log, false, &tv);
5520         }
5521 }
5522
5523 static void
5524 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5525       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5526 {
5527         struct hpts_diag diag;
5528         struct inpcb *inp = tptoinpcb(tp);
5529         struct timeval tv;
5530         uint32_t delayed_ack = 0;
5531         uint32_t hpts_timeout;
5532         uint32_t entry_slot = slot;
5533         uint8_t stopped;
5534         uint32_t left = 0;
5535         uint32_t us_cts;
5536
5537         if ((tp->t_state == TCPS_CLOSED) ||
5538             (tp->t_state == TCPS_LISTEN)) {
5539                 return;
5540         }
5541         if (tcp_in_hpts(inp)) {
5542                 /* Already on the pacer */
5543                 return;
5544         }
5545         stopped = rack->rc_tmr_stopped;
5546         if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5547                 left = rack->r_ctl.rc_timer_exp - cts;
5548         }
5549         rack->r_ctl.rc_timer_exp = 0;
5550         rack->r_ctl.rc_hpts_flags = 0;
5551         us_cts = tcp_get_usecs(&tv);
5552         /* Now early/late accounting */
5553         rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5554         if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5555                 /*
5556                  * We have a early carry over set,
5557                  * we can always add more time so we
5558                  * can always make this compensation.
5559                  *
5560                  * Note if ack's are allowed to wake us do not
5561                  * penalize the next timer for being awoke
5562                  * by an ack aka the rc_agg_early (non-paced mode).
5563                  */
5564                 slot += rack->r_ctl.rc_agg_early;
5565                 rack->r_early = 0;
5566                 rack->r_ctl.rc_agg_early = 0;
5567         }
5568         if (rack->r_late) {
5569                 /*
5570                  * This is harder, we can
5571                  * compensate some but it
5572                  * really depends on what
5573                  * the current pacing time is.
5574                  */
5575                 if (rack->r_ctl.rc_agg_delayed >= slot) {
5576                         /*
5577                          * We can't compensate for it all.
5578                          * And we have to have some time
5579                          * on the clock. We always have a min
5580                          * 10 slots (10 x 10 i.e. 100 usecs).
5581                          */
5582                         if (slot <= HPTS_TICKS_PER_SLOT) {
5583                                 /* We gain delay */
5584                                 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5585                                 slot = HPTS_TICKS_PER_SLOT;
5586                         } else {
5587                                 /* We take off some */
5588                                 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5589                                 slot = HPTS_TICKS_PER_SLOT;
5590                         }
5591                 } else {
5592                         slot -= rack->r_ctl.rc_agg_delayed;
5593                         rack->r_ctl.rc_agg_delayed = 0;
5594                         /* Make sure we have 100 useconds at minimum */
5595                         if (slot < HPTS_TICKS_PER_SLOT) {
5596                                 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5597                                 slot = HPTS_TICKS_PER_SLOT;
5598                         }
5599                         if (rack->r_ctl.rc_agg_delayed == 0)
5600                                 rack->r_late = 0;
5601                 }
5602         }
5603         if (slot) {
5604                 /* We are pacing too */
5605                 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5606         }
5607         hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5608 #ifdef NETFLIX_EXP_DETECTION
5609         if (rack->sack_attack_disable &&
5610             (slot < tcp_sad_pacing_interval)) {
5611                 /*
5612                  * We have a potential attacker on
5613                  * the line. We have possibly some
5614                  * (or now) pacing time set. We want to
5615                  * slow down the processing of sacks by some
5616                  * amount (if it is an attacker). Set the default
5617                  * slot for attackers in place (unless the original
5618                  * interval is longer). Its stored in
5619                  * micro-seconds, so lets convert to msecs.
5620                  */
5621                 slot = tcp_sad_pacing_interval;
5622         }
5623 #endif
5624         if (tp->t_flags & TF_DELACK) {
5625                 delayed_ack = TICKS_2_USEC(tcp_delacktime);
5626                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5627         }
5628         if (delayed_ack && ((hpts_timeout == 0) ||
5629                             (delayed_ack < hpts_timeout)))
5630                 hpts_timeout = delayed_ack;
5631         else
5632                 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5633         /*
5634          * If no timers are going to run and we will fall off the hptsi
5635          * wheel, we resort to a keep-alive timer if its configured.
5636          */
5637         if ((hpts_timeout == 0) &&
5638             (slot == 0)) {
5639                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5640                     (tp->t_state <= TCPS_CLOSING)) {
5641                         /*
5642                          * Ok we have no timer (persists, rack, tlp, rxt  or
5643                          * del-ack), we don't have segments being paced. So
5644                          * all that is left is the keepalive timer.
5645                          */
5646                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5647                                 /* Get the established keep-alive time */
5648                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5649                         } else {
5650                                 /*
5651                                  * Get the initial setup keep-alive time,
5652                                  * note that this is probably not going to
5653                                  * happen, since rack will be running a rxt timer
5654                                  * if a SYN of some sort is outstanding. It is
5655                                  * actually handled in rack_timeout_rxt().
5656                                  */
5657                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5658                         }
5659                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5660                         if (rack->in_probe_rtt) {
5661                                 /*
5662                                  * We want to instead not wake up a long time from
5663                                  * now but to wake up about the time we would
5664                                  * exit probe-rtt and initiate a keep-alive ack.
5665                                  * This will get us out of probe-rtt and update
5666                                  * our min-rtt.
5667                                  */
5668                                 hpts_timeout = rack_min_probertt_hold;
5669                         }
5670                 }
5671         }
5672         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5673             (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5674                 /*
5675                  * RACK, TLP, persists and RXT timers all are restartable
5676                  * based on actions input .. i.e we received a packet (ack
5677                  * or sack) and that changes things (rw, or snd_una etc).
5678                  * Thus we can restart them with a new value. For
5679                  * keep-alive, delayed_ack we keep track of what was left
5680                  * and restart the timer with a smaller value.
5681                  */
5682                 if (left < hpts_timeout)
5683                         hpts_timeout = left;
5684         }
5685         if (hpts_timeout) {
5686                 /*
5687                  * Hack alert for now we can't time-out over 2,147,483
5688                  * seconds (a bit more than 596 hours), which is probably ok
5689                  * :).
5690                  */
5691                 if (hpts_timeout > 0x7ffffffe)
5692                         hpts_timeout = 0x7ffffffe;
5693                 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5694         }
5695         rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5696         if ((rack->gp_ready == 0) &&
5697             (rack->use_fixed_rate == 0) &&
5698             (hpts_timeout < slot) &&
5699             (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5700                 /*
5701                  * We have no good estimate yet for the
5702                  * old clunky burst mitigation or the
5703                  * real pacing. And the tlp or rxt is smaller
5704                  * than the pacing calculation. Lets not
5705                  * pace that long since we know the calculation
5706                  * so far is not accurate.
5707                  */
5708                 slot = hpts_timeout;
5709         }
5710         /**
5711          * Turn off all the flags for queuing by default. The
5712          * flags have important meanings to what happens when
5713          * LRO interacts with the transport. Most likely (by default now)
5714          * mbuf_queueing and ack compression are on. So the transport
5715          * has a couple of flags that control what happens (if those
5716          * are not on then these flags won't have any effect since it
5717          * won't go through the queuing LRO path).
5718          *
5719          * INP_MBUF_QUEUE_READY - This flags says that I am busy
5720          *                        pacing output, so don't disturb. But
5721          *                        it also means LRO can wake me if there
5722          *                        is a SACK arrival.
5723          *
5724          * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5725          *                       with the above flag (QUEUE_READY) and
5726          *                       when present it says don't even wake me
5727          *                       if a SACK arrives.
5728          *
5729          * The idea behind these flags is that if we are pacing we
5730          * set the MBUF_QUEUE_READY and only get woken up if
5731          * a SACK arrives (which could change things) or if
5732          * our pacing timer expires. If, however, we have a rack
5733          * timer running, then we don't even want a sack to wake
5734          * us since the rack timer has to expire before we can send.
5735          *
5736          * Other cases should usually have none of the flags set
5737          * so LRO can call into us.
5738          */
5739         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5740         if (slot) {
5741                 rack->r_ctl.rc_last_output_to = us_cts + slot;
5742                 /*
5743                  * A pacing timer (slot) is being set, in
5744                  * such a case we cannot send (we are blocked by
5745                  * the timer). So lets tell LRO that it should not
5746                  * wake us unless there is a SACK. Note this only
5747                  * will be effective if mbuf queueing is on or
5748                  * compressed acks are being processed.
5749                  */
5750                 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5751                 /*
5752                  * But wait if we have a Rack timer running
5753                  * even a SACK should not disturb us (with
5754                  * the exception of r_rr_config 3).
5755                  */
5756                 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5757                     (rack->r_rr_config != 3))
5758                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5759                 if (rack->rc_ack_can_sendout_data) {
5760                         /*
5761                          * Ahh but wait, this is that special case
5762                          * where the pacing timer can be disturbed
5763                          * backout the changes (used for non-paced
5764                          * burst limiting).
5765                          */
5766                         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5767                 }
5768                 if ((rack->use_rack_rr) &&
5769                     (rack->r_rr_config < 2) &&
5770                     ((hpts_timeout) && (hpts_timeout < slot))) {
5771                         /*
5772                          * Arrange for the hpts to kick back in after the
5773                          * t-o if the t-o does not cause a send.
5774                          */
5775                         (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5776                                                    __LINE__, &diag);
5777                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5778                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5779                 } else {
5780                         (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(slot),
5781                                                    __LINE__, &diag);
5782                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5783                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5784                 }
5785         } else if (hpts_timeout) {
5786                 /*
5787                  * With respect to inp_flags2 here, lets let any new acks wake
5788                  * us up here. Since we are not pacing (no pacing timer), output
5789                  * can happen so we should let it. If its a Rack timer, then any inbound
5790                  * packet probably won't change the sending (we will be blocked)
5791                  * but it may change the prr stats so letting it in (the set defaults
5792                  * at the start of this block) are good enough.
5793                  */
5794                 (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5795                                            __LINE__, &diag);
5796                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5797                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5798         } else {
5799                 /* No timer starting */
5800 #ifdef INVARIANTS
5801                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5802                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5803                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5804                 }
5805 #endif
5806         }
5807         rack->rc_tmr_stopped = 0;
5808         if (slot)
5809                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5810 }
5811
5812 /*
5813  * RACK Timer, here we simply do logging and house keeping.
5814  * the normal rack_output() function will call the
5815  * appropriate thing to check if we need to do a RACK retransmit.
5816  * We return 1, saying don't proceed with rack_output only
5817  * when all timers have been stopped (destroyed PCB?).
5818  */
5819 static int
5820 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5821 {
5822         /*
5823          * This timer simply provides an internal trigger to send out data.
5824          * The check_recovery_mode call will see if there are needed
5825          * retransmissions, if so we will enter fast-recovery. The output
5826          * call may or may not do the same thing depending on sysctl
5827          * settings.
5828          */
5829         struct rack_sendmap *rsm;
5830
5831         counter_u64_add(rack_to_tot, 1);
5832         if (rack->r_state && (rack->r_state != tp->t_state))
5833                 rack_set_state(tp, rack);
5834         rack->rc_on_min_to = 0;
5835         rsm = rack_check_recovery_mode(tp, cts);
5836         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5837         if (rsm) {
5838                 rack->r_ctl.rc_resend = rsm;
5839                 rack->r_timer_override = 1;
5840                 if (rack->use_rack_rr) {
5841                         /*
5842                          * Don't accumulate extra pacing delay
5843                          * we are allowing the rack timer to
5844                          * over-ride pacing i.e. rrr takes precedence
5845                          * if the pacing interval is longer than the rrr
5846                          * time (in other words we get the min pacing
5847                          * time versus rrr pacing time).
5848                          */
5849                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5850                 }
5851         }
5852         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5853         if (rsm == NULL) {
5854                 /* restart a timer and return 1 */
5855                 rack_start_hpts_timer(rack, tp, cts,
5856                                       0, 0, 0);
5857                 return (1);
5858         }
5859         return (0);
5860 }
5861
5862 static void
5863 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5864 {
5865         if (rsm->m->m_len > rsm->orig_m_len) {
5866                 /*
5867                  * Mbuf grew, caused by sbcompress, our offset does
5868                  * not change.
5869                  */
5870                 rsm->orig_m_len = rsm->m->m_len;
5871         } else if (rsm->m->m_len < rsm->orig_m_len) {
5872                 /*
5873                  * Mbuf shrank, trimmed off the top by an ack, our
5874                  * offset changes.
5875                  */
5876                 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
5877                 rsm->orig_m_len = rsm->m->m_len;
5878         }
5879 }
5880
5881 static void
5882 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
5883 {
5884         struct mbuf *m;
5885         uint32_t soff;
5886
5887         if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
5888                 /* Fix up the orig_m_len and possibly the mbuf offset */
5889                 rack_adjust_orig_mlen(src_rsm);
5890         }
5891         m = src_rsm->m;
5892         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
5893         while (soff >= m->m_len) {
5894                 /* Move out past this mbuf */
5895                 soff -= m->m_len;
5896                 m = m->m_next;
5897                 KASSERT((m != NULL),
5898                         ("rsm:%p nrsm:%p hit at soff:%u null m",
5899                          src_rsm, rsm, soff));
5900         }
5901         rsm->m = m;
5902         rsm->soff = soff;
5903         rsm->orig_m_len = m->m_len;
5904 }
5905
5906 static __inline void
5907 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5908                struct rack_sendmap *rsm, uint32_t start)
5909 {
5910         int idx;
5911
5912         nrsm->r_start = start;
5913         nrsm->r_end = rsm->r_end;
5914         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
5915         nrsm->r_flags = rsm->r_flags;
5916         nrsm->r_dupack = rsm->r_dupack;
5917         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
5918         nrsm->r_rtr_bytes = 0;
5919         nrsm->r_fas = rsm->r_fas;
5920         rsm->r_end = nrsm->r_start;
5921         nrsm->r_just_ret = rsm->r_just_ret;
5922         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
5923                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
5924         }
5925         /* Now if we have SYN flag we keep it on the left edge */
5926         if (nrsm->r_flags & RACK_HAS_SYN)
5927                 nrsm->r_flags &= ~RACK_HAS_SYN;
5928         /* Now if we have a FIN flag we keep it on the right edge */
5929         if (rsm->r_flags & RACK_HAS_FIN)
5930                 rsm->r_flags &= ~RACK_HAS_FIN;
5931         /* Push bit must go to the right edge as well */
5932         if (rsm->r_flags & RACK_HAD_PUSH)
5933                 rsm->r_flags &= ~RACK_HAD_PUSH;
5934         /* Clone over the state of the hw_tls flag */
5935         nrsm->r_hw_tls = rsm->r_hw_tls;
5936         /*
5937          * Now we need to find nrsm's new location in the mbuf chain
5938          * we basically calculate a new offset, which is soff +
5939          * how much is left in original rsm. Then we walk out the mbuf
5940          * chain to find the righ position, it may be the same mbuf
5941          * or maybe not.
5942          */
5943         KASSERT(((rsm->m != NULL) ||
5944                  (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
5945                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
5946         if (rsm->m)
5947                 rack_setup_offset_for_rsm(rsm, nrsm);
5948 }
5949
5950 static struct rack_sendmap *
5951 rack_merge_rsm(struct tcp_rack *rack,
5952                struct rack_sendmap *l_rsm,
5953                struct rack_sendmap *r_rsm)
5954 {
5955         /*
5956          * We are merging two ack'd RSM's,
5957          * the l_rsm is on the left (lower seq
5958          * values) and the r_rsm is on the right
5959          * (higher seq value). The simplest way
5960          * to merge these is to move the right
5961          * one into the left. I don't think there
5962          * is any reason we need to try to find
5963          * the oldest (or last oldest retransmitted).
5964          */
5965 #ifdef INVARIANTS
5966         struct rack_sendmap *rm;
5967 #endif
5968         rack_log_map_chg(rack->rc_tp, rack, NULL,
5969                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
5970         l_rsm->r_end = r_rsm->r_end;
5971         if (l_rsm->r_dupack < r_rsm->r_dupack)
5972                 l_rsm->r_dupack = r_rsm->r_dupack;
5973         if (r_rsm->r_rtr_bytes)
5974                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
5975         if (r_rsm->r_in_tmap) {
5976                 /* This really should not happen */
5977                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
5978                 r_rsm->r_in_tmap = 0;
5979         }
5980
5981         /* Now the flags */
5982         if (r_rsm->r_flags & RACK_HAS_FIN)
5983                 l_rsm->r_flags |= RACK_HAS_FIN;
5984         if (r_rsm->r_flags & RACK_TLP)
5985                 l_rsm->r_flags |= RACK_TLP;
5986         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
5987                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
5988         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
5989             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
5990                 /*
5991                  * If both are app-limited then let the
5992                  * free lower the count. If right is app
5993                  * limited and left is not, transfer.
5994                  */
5995                 l_rsm->r_flags |= RACK_APP_LIMITED;
5996                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
5997                 if (r_rsm == rack->r_ctl.rc_first_appl)
5998                         rack->r_ctl.rc_first_appl = l_rsm;
5999         }
6000 #ifndef INVARIANTS
6001         (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6002 #else
6003         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6004         if (rm != r_rsm) {
6005                 panic("removing head in rack:%p rsm:%p rm:%p",
6006                       rack, r_rsm, rm);
6007         }
6008 #endif
6009         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6010                 /* Transfer the split limit to the map we free */
6011                 r_rsm->r_limit_type = l_rsm->r_limit_type;
6012                 l_rsm->r_limit_type = 0;
6013         }
6014         rack_free(rack, r_rsm);
6015         return (l_rsm);
6016 }
6017
6018 /*
6019  * TLP Timer, here we simply setup what segment we want to
6020  * have the TLP expire on, the normal rack_output() will then
6021  * send it out.
6022  *
6023  * We return 1, saying don't proceed with rack_output only
6024  * when all timers have been stopped (destroyed PCB?).
6025  */
6026 static int
6027 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6028 {
6029         /*
6030          * Tail Loss Probe.
6031          */
6032         struct rack_sendmap *rsm = NULL;
6033 #ifdef INVARIANTS
6034         struct rack_sendmap *insret;
6035 #endif
6036         struct socket *so = tptosocket(tp);
6037         uint32_t amm;
6038         uint32_t out, avail;
6039         int collapsed_win = 0;
6040
6041         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6042                 /* Its not time yet */
6043                 return (0);
6044         }
6045         if (ctf_progress_timeout_check(tp, true)) {
6046                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6047                 return (-ETIMEDOUT);    /* tcp_drop() */
6048         }
6049         /*
6050          * A TLP timer has expired. We have been idle for 2 rtts. So we now
6051          * need to figure out how to force a full MSS segment out.
6052          */
6053         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6054         rack->r_ctl.retran_during_recovery = 0;
6055         rack->r_ctl.dsack_byte_cnt = 0;
6056         counter_u64_add(rack_tlp_tot, 1);
6057         if (rack->r_state && (rack->r_state != tp->t_state))
6058                 rack_set_state(tp, rack);
6059         avail = sbavail(&so->so_snd);
6060         out = tp->snd_max - tp->snd_una;
6061         if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
6062                 /* special case, we need a retransmission */
6063                 collapsed_win = 1;
6064                 goto need_retran;
6065         }
6066         if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6067                 rack->r_ctl.dsack_persist--;
6068                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6069                         rack->r_ctl.num_dsack = 0;
6070                 }
6071                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6072         }
6073         if ((tp->t_flags & TF_GPUTINPROG) &&
6074             (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6075                 /*
6076                  * If this is the second in a row
6077                  * TLP and we are doing a measurement
6078                  * its time to abandon the measurement.
6079                  * Something is likely broken on
6080                  * the clients network and measuring a
6081                  * broken network does us no good.
6082                  */
6083                 tp->t_flags &= ~TF_GPUTINPROG;
6084                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6085                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
6086                                            tp->gput_seq,
6087                                            0, 0, 18, __LINE__, NULL, 0);
6088         }
6089         /*
6090          * Check our send oldest always settings, and if
6091          * there is an oldest to send jump to the need_retran.
6092          */
6093         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6094                 goto need_retran;
6095
6096         if (avail > out) {
6097                 /* New data is available */
6098                 amm = avail - out;
6099                 if (amm > ctf_fixed_maxseg(tp)) {
6100                         amm = ctf_fixed_maxseg(tp);
6101                         if ((amm + out) > tp->snd_wnd) {
6102                                 /* We are rwnd limited */
6103                                 goto need_retran;
6104                         }
6105                 } else if (amm < ctf_fixed_maxseg(tp)) {
6106                         /* not enough to fill a MTU */
6107                         goto need_retran;
6108                 }
6109                 if (IN_FASTRECOVERY(tp->t_flags)) {
6110                         /* Unlikely */
6111                         if (rack->rack_no_prr == 0) {
6112                                 if (out + amm <= tp->snd_wnd) {
6113                                         rack->r_ctl.rc_prr_sndcnt = amm;
6114                                         rack->r_ctl.rc_tlp_new_data = amm;
6115                                         rack_log_to_prr(rack, 4, 0, __LINE__);
6116                                 }
6117                         } else
6118                                 goto need_retran;
6119                 } else {
6120                         /* Set the send-new override */
6121                         if (out + amm <= tp->snd_wnd)
6122                                 rack->r_ctl.rc_tlp_new_data = amm;
6123                         else
6124                                 goto need_retran;
6125                 }
6126                 rack->r_ctl.rc_tlpsend = NULL;
6127                 counter_u64_add(rack_tlp_newdata, 1);
6128                 goto send;
6129         }
6130 need_retran:
6131         /*
6132          * Ok we need to arrange the last un-acked segment to be re-sent, or
6133          * optionally the first un-acked segment.
6134          */
6135         if (collapsed_win == 0) {
6136                 if (rack_always_send_oldest)
6137                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6138                 else {
6139                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6140                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6141                                 rsm = rack_find_high_nonack(rack, rsm);
6142                         }
6143                 }
6144                 if (rsm == NULL) {
6145 #ifdef TCP_BLACKBOX
6146                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6147 #endif
6148                         goto out;
6149                 }
6150         } else {
6151                 /*
6152                  * We must find the last segment
6153                  * that was acceptable by the client.
6154                  */
6155                 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6156                         if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6157                                 /* Found one */
6158                                 break;
6159                         }
6160                 }
6161                 if (rsm == NULL) {
6162                         /* None? if so send the first */
6163                         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6164                         if (rsm == NULL) {
6165 #ifdef TCP_BLACKBOX
6166                                 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6167 #endif
6168                                 goto out;
6169                         }
6170                 }
6171         }
6172         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6173                 /*
6174                  * We need to split this the last segment in two.
6175                  */
6176                 struct rack_sendmap *nrsm;
6177
6178                 nrsm = rack_alloc_full_limit(rack);
6179                 if (nrsm == NULL) {
6180                         /*
6181                          * No memory to split, we will just exit and punt
6182                          * off to the RXT timer.
6183                          */
6184                         goto out;
6185                 }
6186                 rack_clone_rsm(rack, nrsm, rsm,
6187                                (rsm->r_end - ctf_fixed_maxseg(tp)));
6188                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6189 #ifndef INVARIANTS
6190                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6191 #else
6192                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6193                 if (insret != NULL) {
6194                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6195                               nrsm, insret, rack, rsm);
6196                 }
6197 #endif
6198                 if (rsm->r_in_tmap) {
6199                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6200                         nrsm->r_in_tmap = 1;
6201                 }
6202                 rsm = nrsm;
6203         }
6204         rack->r_ctl.rc_tlpsend = rsm;
6205 send:
6206         /* Make sure output path knows we are doing a TLP */
6207         *doing_tlp = 1;
6208         rack->r_timer_override = 1;
6209         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6210         return (0);
6211 out:
6212         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6213         return (0);
6214 }
6215
6216 /*
6217  * Delayed ack Timer, here we simply need to setup the
6218  * ACK_NOW flag and remove the DELACK flag. From there
6219  * the output routine will send the ack out.
6220  *
6221  * We only return 1, saying don't proceed, if all timers
6222  * are stopped (destroyed PCB?).
6223  */
6224 static int
6225 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6226 {
6227
6228         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6229         tp->t_flags &= ~TF_DELACK;
6230         tp->t_flags |= TF_ACKNOW;
6231         KMOD_TCPSTAT_INC(tcps_delack);
6232         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6233         return (0);
6234 }
6235
6236 /*
6237  * Persists timer, here we simply send the
6238  * same thing as a keepalive will.
6239  * the one byte send.
6240  *
6241  * We only return 1, saying don't proceed, if all timers
6242  * are stopped (destroyed PCB?).
6243  */
6244 static int
6245 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6246 {
6247         struct tcptemp *t_template;
6248         int32_t retval = 1;
6249
6250         if (rack->rc_in_persist == 0)
6251                 return (0);
6252         if (ctf_progress_timeout_check(tp, false)) {
6253                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6254                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6255                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6256                 return (-ETIMEDOUT);    /* tcp_drop() */
6257         }
6258         /*
6259          * Persistence timer into zero window. Force a byte to be output, if
6260          * possible.
6261          */
6262         KMOD_TCPSTAT_INC(tcps_persisttimeo);
6263         /*
6264          * Hack: if the peer is dead/unreachable, we do not time out if the
6265          * window is closed.  After a full backoff, drop the connection if
6266          * the idle time (no responses to probes) reaches the maximum
6267          * backoff that we would use if retransmitting.
6268          */
6269         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6270             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6271              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6272                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6273                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6274                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6275                 retval = -ETIMEDOUT;    /* tcp_drop() */
6276                 goto out;
6277         }
6278         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6279             tp->snd_una == tp->snd_max)
6280                 rack_exit_persist(tp, rack, cts);
6281         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6282         /*
6283          * If the user has closed the socket then drop a persisting
6284          * connection after a much reduced timeout.
6285          */
6286         if (tp->t_state > TCPS_CLOSE_WAIT &&
6287             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6288                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6289                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6290                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6291                 retval = -ETIMEDOUT;    /* tcp_drop() */
6292                 goto out;
6293         }
6294         t_template = tcpip_maketemplate(rack->rc_inp);
6295         if (t_template) {
6296                 /* only set it if we were answered */
6297                 if (rack->forced_ack == 0) {
6298                         rack->forced_ack = 1;
6299                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6300                 } else {
6301                         rack->probe_not_answered = 1;
6302                         counter_u64_add(rack_persists_loss, 1);
6303                         rack->r_ctl.persist_lost_ends++;
6304                 }
6305                 counter_u64_add(rack_persists_sends, 1);
6306                 tcp_respond(tp, t_template->tt_ipgen,
6307                             &t_template->tt_t, (struct mbuf *)NULL,
6308                             tp->rcv_nxt, tp->snd_una - 1, 0);
6309                 /* This sends an ack */
6310                 if (tp->t_flags & TF_DELACK)
6311                         tp->t_flags &= ~TF_DELACK;
6312                 free(t_template, M_TEMP);
6313         }
6314         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6315                 tp->t_rxtshift++;
6316 out:
6317         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6318         rack_start_hpts_timer(rack, tp, cts,
6319                               0, 0, 0);
6320         return (retval);
6321 }
6322
6323 /*
6324  * If a keepalive goes off, we had no other timers
6325  * happening. We always return 1 here since this
6326  * routine either drops the connection or sends
6327  * out a segment with respond.
6328  */
6329 static int
6330 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6331 {
6332         struct tcptemp *t_template;
6333         struct inpcb *inp = tptoinpcb(tp);
6334
6335         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6336         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6337         /*
6338          * Keep-alive timer went off; send something or drop connection if
6339          * idle for too long.
6340          */
6341         KMOD_TCPSTAT_INC(tcps_keeptimeo);
6342         if (tp->t_state < TCPS_ESTABLISHED)
6343                 goto dropit;
6344         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6345             tp->t_state <= TCPS_CLOSING) {
6346                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6347                         goto dropit;
6348                 /*
6349                  * Send a packet designed to force a response if the peer is
6350                  * up and reachable: either an ACK if the connection is
6351                  * still alive, or an RST if the peer has closed the
6352                  * connection due to timeout or reboot. Using sequence
6353                  * number tp->snd_una-1 causes the transmitted zero-length
6354                  * segment to lie outside the receive window; by the
6355                  * protocol spec, this requires the correspondent TCP to
6356                  * respond.
6357                  */
6358                 KMOD_TCPSTAT_INC(tcps_keepprobe);
6359                 t_template = tcpip_maketemplate(inp);
6360                 if (t_template) {
6361                         if (rack->forced_ack == 0) {
6362                                 rack->forced_ack = 1;
6363                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6364                         } else {
6365                                 rack->probe_not_answered = 1;
6366                         }
6367                         tcp_respond(tp, t_template->tt_ipgen,
6368                             &t_template->tt_t, (struct mbuf *)NULL,
6369                             tp->rcv_nxt, tp->snd_una - 1, 0);
6370                         free(t_template, M_TEMP);
6371                 }
6372         }
6373         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6374         return (1);
6375 dropit:
6376         KMOD_TCPSTAT_INC(tcps_keepdrops);
6377         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6378         return (-ETIMEDOUT);    /* tcp_drop() */
6379 }
6380
6381 /*
6382  * Retransmit helper function, clear up all the ack
6383  * flags and take care of important book keeping.
6384  */
6385 static void
6386 rack_remxt_tmr(struct tcpcb *tp)
6387 {
6388         /*
6389          * The retransmit timer went off, all sack'd blocks must be
6390          * un-acked.
6391          */
6392         struct rack_sendmap *rsm, *trsm = NULL;
6393         struct tcp_rack *rack;
6394
6395         rack = (struct tcp_rack *)tp->t_fb_ptr;
6396         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6397         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6398         if (rack->r_state && (rack->r_state != tp->t_state))
6399                 rack_set_state(tp, rack);
6400         /*
6401          * Ideally we would like to be able to
6402          * mark SACK-PASS on anything not acked here.
6403          *
6404          * However, if we do that we would burst out
6405          * all that data 1ms apart. This would be unwise,
6406          * so for now we will just let the normal rxt timer
6407          * and tlp timer take care of it.
6408          *
6409          * Also we really need to stick them back in sequence
6410          * order. This way we send in the proper order and any
6411          * sacks that come floating in will "re-ack" the data.
6412          * To do this we zap the tmap with an INIT and then
6413          * walk through and place every rsm in the RB tree
6414          * back in its seq ordered place.
6415          */
6416         TAILQ_INIT(&rack->r_ctl.rc_tmap);
6417         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6418                 rsm->r_dupack = 0;
6419                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6420                 /* We must re-add it back to the tlist */
6421                 if (trsm == NULL) {
6422                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6423                 } else {
6424                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6425                 }
6426                 rsm->r_in_tmap = 1;
6427                 trsm = rsm;
6428                 if (rsm->r_flags & RACK_ACKED)
6429                         rsm->r_flags |= RACK_WAS_ACKED;
6430                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
6431                 rsm->r_flags |= RACK_MUST_RXT;
6432         }
6433         /* Clear the count (we just un-acked them) */
6434         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6435         rack->r_ctl.rc_sacked = 0;
6436         rack->r_ctl.rc_sacklast = NULL;
6437         rack->r_ctl.rc_agg_delayed = 0;
6438         rack->r_early = 0;
6439         rack->r_ctl.rc_agg_early = 0;
6440         rack->r_late = 0;
6441         /* Clear the tlp rtx mark */
6442         rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6443         if (rack->r_ctl.rc_resend != NULL)
6444                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6445         rack->r_ctl.rc_prr_sndcnt = 0;
6446         rack_log_to_prr(rack, 6, 0, __LINE__);
6447         rack->r_timer_override = 1;
6448         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6449 #ifdef NETFLIX_EXP_DETECTION
6450             || (rack->sack_attack_disable != 0)
6451 #endif
6452                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6453                 /*
6454                  * For non-sack customers new data
6455                  * needs to go out as retransmits until
6456                  * we retransmit up to snd_max.
6457                  */
6458                 rack->r_must_retran = 1;
6459                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6460                                                 rack->r_ctl.rc_sacked);
6461         }
6462         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6463 }
6464
6465 static void
6466 rack_convert_rtts(struct tcpcb *tp)
6467 {
6468         if (tp->t_srtt > 1) {
6469                 uint32_t val, frac;
6470
6471                 val = tp->t_srtt >> TCP_RTT_SHIFT;
6472                 frac = tp->t_srtt & 0x1f;
6473                 tp->t_srtt = TICKS_2_USEC(val);
6474                 /*
6475                  * frac is the fractional part of the srtt (if any)
6476                  * but its in ticks and every bit represents
6477                  * 1/32nd of a hz.
6478                  */
6479                 if (frac) {
6480                         if (hz == 1000) {
6481                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6482                         } else {
6483                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6484                         }
6485                         tp->t_srtt += frac;
6486                 }
6487         }
6488         if (tp->t_rttvar) {
6489                 uint32_t val, frac;
6490
6491                 val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6492                 frac = tp->t_rttvar & 0x1f;
6493                 tp->t_rttvar = TICKS_2_USEC(val);
6494                 /*
6495                  * frac is the fractional part of the srtt (if any)
6496                  * but its in ticks and every bit represents
6497                  * 1/32nd of a hz.
6498                  */
6499                 if (frac) {
6500                         if (hz == 1000) {
6501                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6502                         } else {
6503                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6504                         }
6505                         tp->t_rttvar += frac;
6506                 }
6507         }
6508         tp->t_rxtcur = RACK_REXMTVAL(tp);
6509         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6510                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6511         }
6512         if (tp->t_rxtcur > rack_rto_max) {
6513                 tp->t_rxtcur = rack_rto_max;
6514         }
6515 }
6516
6517 static void
6518 rack_cc_conn_init(struct tcpcb *tp)
6519 {
6520         struct tcp_rack *rack;
6521         uint32_t srtt;
6522
6523         rack = (struct tcp_rack *)tp->t_fb_ptr;
6524         srtt = tp->t_srtt;
6525         cc_conn_init(tp);
6526         /*
6527          * Now convert to rack's internal format,
6528          * if required.
6529          */
6530         if ((srtt == 0) && (tp->t_srtt != 0))
6531                 rack_convert_rtts(tp);
6532         /*
6533          * We want a chance to stay in slowstart as
6534          * we create a connection. TCP spec says that
6535          * initially ssthresh is infinite. For our
6536          * purposes that is the snd_wnd.
6537          */
6538         if (tp->snd_ssthresh < tp->snd_wnd) {
6539                 tp->snd_ssthresh = tp->snd_wnd;
6540         }
6541         /*
6542          * We also want to assure a IW worth of
6543          * data can get inflight.
6544          */
6545         if (rc_init_window(rack) < tp->snd_cwnd)
6546                 tp->snd_cwnd = rc_init_window(rack);
6547 }
6548
6549 /*
6550  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6551  * we will setup to retransmit the lowest seq number outstanding.
6552  */
6553 static int
6554 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6555 {
6556         struct inpcb *inp = tptoinpcb(tp);
6557         int32_t rexmt;
6558         int32_t retval = 0;
6559         bool isipv6;
6560
6561         if ((tp->t_flags & TF_GPUTINPROG) &&
6562             (tp->t_rxtshift)) {
6563                 /*
6564                  * We have had a second timeout
6565                  * measurements on successive rxt's are not profitable.
6566                  * It is unlikely to be of any use (the network is
6567                  * broken or the client went away).
6568                  */
6569                 tp->t_flags &= ~TF_GPUTINPROG;
6570                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6571                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
6572                                            tp->gput_seq,
6573                                            0, 0, 18, __LINE__, NULL, 0);
6574         }
6575         if (ctf_progress_timeout_check(tp, false)) {
6576                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6577                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6578                 return (-ETIMEDOUT);    /* tcp_drop() */
6579         }
6580         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6581         rack->r_ctl.retran_during_recovery = 0;
6582         rack->rc_ack_required = 1;
6583         rack->r_ctl.dsack_byte_cnt = 0;
6584         if (IN_FASTRECOVERY(tp->t_flags))
6585                 tp->t_flags |= TF_WASFRECOVERY;
6586         else
6587                 tp->t_flags &= ~TF_WASFRECOVERY;
6588         if (IN_CONGRECOVERY(tp->t_flags))
6589                 tp->t_flags |= TF_WASCRECOVERY;
6590         else
6591                 tp->t_flags &= ~TF_WASCRECOVERY;
6592         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6593             (tp->snd_una == tp->snd_max)) {
6594                 /* Nothing outstanding .. nothing to do */
6595                 return (0);
6596         }
6597         if (rack->r_ctl.dsack_persist) {
6598                 rack->r_ctl.dsack_persist--;
6599                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6600                         rack->r_ctl.num_dsack = 0;
6601                 }
6602                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6603         }
6604         /*
6605          * Rack can only run one timer  at a time, so we cannot
6606          * run a KEEPINIT (gating SYN sending) and a retransmit
6607          * timer for the SYN. So if we are in a front state and
6608          * have a KEEPINIT timer we need to check the first transmit
6609          * against now to see if we have exceeded the KEEPINIT time
6610          * (if one is set).
6611          */
6612         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6613             (TP_KEEPINIT(tp) != 0)) {
6614                 struct rack_sendmap *rsm;
6615
6616                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6617                 if (rsm) {
6618                         /* Ok we have something outstanding to test keepinit with */
6619                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6620                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6621                                 /* We have exceeded the KEEPINIT time */
6622                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6623                                 goto drop_it;
6624                         }
6625                 }
6626         }
6627         /*
6628          * Retransmission timer went off.  Message has not been acked within
6629          * retransmit interval.  Back off to a longer retransmit interval
6630          * and retransmit one segment.
6631          */
6632         rack_remxt_tmr(tp);
6633         if ((rack->r_ctl.rc_resend == NULL) ||
6634             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6635                 /*
6636                  * If the rwnd collapsed on
6637                  * the one we are retransmitting
6638                  * it does not count against the
6639                  * rxt count.
6640                  */
6641                 tp->t_rxtshift++;
6642         }
6643         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6644                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6645 drop_it:
6646                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6647                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6648                 /* XXXGL: previously t_softerror was casted to uint16_t */
6649                 MPASS(tp->t_softerror >= 0);
6650                 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
6651                 goto out;       /* tcp_drop() */
6652         }
6653         if (tp->t_state == TCPS_SYN_SENT) {
6654                 /*
6655                  * If the SYN was retransmitted, indicate CWND to be limited
6656                  * to 1 segment in cc_conn_init().
6657                  */
6658                 tp->snd_cwnd = 1;
6659         } else if (tp->t_rxtshift == 1) {
6660                 /*
6661                  * first retransmit; record ssthresh and cwnd so they can be
6662                  * recovered if this turns out to be a "bad" retransmit. A
6663                  * retransmit is considered "bad" if an ACK for this segment
6664                  * is received within RTT/2 interval; the assumption here is
6665                  * that the ACK was already in flight.  See "On Estimating
6666                  * End-to-End Network Path Properties" by Allman and Paxson
6667                  * for more details.
6668                  */
6669                 tp->snd_cwnd_prev = tp->snd_cwnd;
6670                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6671                 tp->snd_recover_prev = tp->snd_recover;
6672                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6673                 tp->t_flags |= TF_PREVVALID;
6674         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6675                 tp->t_flags &= ~TF_PREVVALID;
6676         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6677         if ((tp->t_state == TCPS_SYN_SENT) ||
6678             (tp->t_state == TCPS_SYN_RECEIVED))
6679                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6680         else
6681                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6682
6683         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6684            max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6685         /*
6686          * We enter the path for PLMTUD if connection is established or, if
6687          * connection is FIN_WAIT_1 status, reason for the last is that if
6688          * amount of data we send is very small, we could send it in couple
6689          * of packets and process straight to FIN. In that case we won't
6690          * catch ESTABLISHED state.
6691          */
6692 #ifdef INET6
6693         isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
6694 #else
6695         isipv6 = false;
6696 #endif
6697         if (((V_tcp_pmtud_blackhole_detect == 1) ||
6698             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6699             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6700             ((tp->t_state == TCPS_ESTABLISHED) ||
6701             (tp->t_state == TCPS_FIN_WAIT_1))) {
6702                 /*
6703                  * Idea here is that at each stage of mtu probe (usually,
6704                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
6705                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6706                  * should take care of that.
6707                  */
6708                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6709                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6710                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6711                     tp->t_rxtshift % 2 == 0)) {
6712                         /*
6713                          * Enter Path MTU Black-hole Detection mechanism: -
6714                          * Disable Path MTU Discovery (IP "DF" bit). -
6715                          * Reduce MTU to lower value than what we negotiated
6716                          * with peer.
6717                          */
6718                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6719                                 /* Record that we may have found a black hole. */
6720                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6721                                 /* Keep track of previous MSS. */
6722                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6723                         }
6724
6725                         /*
6726                          * Reduce the MSS to blackhole value or to the
6727                          * default in an attempt to retransmit.
6728                          */
6729 #ifdef INET6
6730                         if (isipv6 &&
6731                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6732                                 /* Use the sysctl tuneable blackhole MSS. */
6733                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6734                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6735                         } else if (isipv6) {
6736                                 /* Use the default MSS. */
6737                                 tp->t_maxseg = V_tcp_v6mssdflt;
6738                                 /*
6739                                  * Disable Path MTU Discovery when we switch
6740                                  * to minmss.
6741                                  */
6742                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6743                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6744                         }
6745 #endif
6746 #if defined(INET6) && defined(INET)
6747                         else
6748 #endif
6749 #ifdef INET
6750                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6751                                 /* Use the sysctl tuneable blackhole MSS. */
6752                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6753                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6754                         } else {
6755                                 /* Use the default MSS. */
6756                                 tp->t_maxseg = V_tcp_mssdflt;
6757                                 /*
6758                                  * Disable Path MTU Discovery when we switch
6759                                  * to minmss.
6760                                  */
6761                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6762                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6763                         }
6764 #endif
6765                 } else {
6766                         /*
6767                          * If further retransmissions are still unsuccessful
6768                          * with a lowered MTU, maybe this isn't a blackhole
6769                          * and we restore the previous MSS and blackhole
6770                          * detection flags. The limit '6' is determined by
6771                          * giving each probe stage (1448, 1188, 524) 2
6772                          * chances to recover.
6773                          */
6774                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6775                             (tp->t_rxtshift >= 6)) {
6776                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6777                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6778                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6779                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6780                         }
6781                 }
6782         }
6783         /*
6784          * Disable RFC1323 and SACK if we haven't got any response to
6785          * our third SYN to work-around some broken terminal servers
6786          * (most of which have hopefully been retired) that have bad VJ
6787          * header compression code which trashes TCP segments containing
6788          * unknown-to-them TCP options.
6789          */
6790         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6791             (tp->t_rxtshift == 3))
6792                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6793         /*
6794          * If we backed off this far, our srtt estimate is probably bogus.
6795          * Clobber it so we'll take the next rtt measurement as our srtt;
6796          * move the current srtt into rttvar to keep the current retransmit
6797          * times until then.
6798          */
6799         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6800 #ifdef INET6
6801                 if ((inp->inp_vflag & INP_IPV6) != 0)
6802                         in6_losing(inp);
6803                 else
6804 #endif
6805                         in_losing(inp);
6806                 tp->t_rttvar += tp->t_srtt;
6807                 tp->t_srtt = 0;
6808         }
6809         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6810         tp->snd_recover = tp->snd_max;
6811         tp->t_flags |= TF_ACKNOW;
6812         tp->t_rtttime = 0;
6813         rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
6814 out:
6815         return (retval);
6816 }
6817
6818 static int
6819 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
6820 {
6821         int32_t ret = 0;
6822         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6823
6824         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
6825             (tp->t_flags & TF_GPUTINPROG)) {
6826                 /*
6827                  * We have a goodput in progress
6828                  * and we have entered a late state.
6829                  * Do we have enough data in the sb
6830                  * to handle the GPUT request?
6831                  */
6832                 uint32_t bytes;
6833
6834                 bytes = tp->gput_ack - tp->gput_seq;
6835                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
6836                         bytes += tp->gput_seq - tp->snd_una;
6837                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
6838                         /*
6839                          * There are not enough bytes in the socket
6840                          * buffer that have been sent to cover this
6841                          * measurement. Cancel it.
6842                          */
6843                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6844                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
6845                                                    tp->gput_seq,
6846                                                    0, 0, 18, __LINE__, NULL, 0);
6847                         tp->t_flags &= ~TF_GPUTINPROG;
6848                 }
6849         }
6850         if (timers == 0) {
6851                 return (0);
6852         }
6853         if (tp->t_state == TCPS_LISTEN) {
6854                 /* no timers on listen sockets */
6855                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6856                         return (0);
6857                 return (1);
6858         }
6859         if ((timers & PACE_TMR_RACK) &&
6860             rack->rc_on_min_to) {
6861                 /*
6862                  * For the rack timer when we
6863                  * are on a min-timeout (which means rrr_conf = 3)
6864                  * we don't want to check the timer. It may
6865                  * be going off for a pace and thats ok we
6866                  * want to send the retransmit (if its ready).
6867                  *
6868                  * If its on a normal rack timer (non-min) then
6869                  * we will check if its expired.
6870                  */
6871                 goto skip_time_check;
6872         }
6873         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6874                 uint32_t left;
6875
6876                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6877                         ret = -1;
6878                         rack_log_to_processing(rack, cts, ret, 0);
6879                         return (0);
6880                 }
6881                 if (hpts_calling == 0) {
6882                         /*
6883                          * A user send or queued mbuf (sack) has called us? We
6884                          * return 0 and let the pacing guards
6885                          * deal with it if they should or
6886                          * should not cause a send.
6887                          */
6888                         ret = -2;
6889                         rack_log_to_processing(rack, cts, ret, 0);
6890                         return (0);
6891                 }
6892                 /*
6893                  * Ok our timer went off early and we are not paced false
6894                  * alarm, go back to sleep.
6895                  */
6896                 ret = -3;
6897                 left = rack->r_ctl.rc_timer_exp - cts;
6898                 tcp_hpts_insert(tptoinpcb(tp), HPTS_MS_TO_SLOTS(left));
6899                 rack_log_to_processing(rack, cts, ret, left);
6900                 return (1);
6901         }
6902 skip_time_check:
6903         rack->rc_tmr_stopped = 0;
6904         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6905         if (timers & PACE_TMR_DELACK) {
6906                 ret = rack_timeout_delack(tp, rack, cts);
6907         } else if (timers & PACE_TMR_RACK) {
6908                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6909                 rack->r_fast_output = 0;
6910                 ret = rack_timeout_rack(tp, rack, cts);
6911         } else if (timers & PACE_TMR_TLP) {
6912                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6913                 ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
6914         } else if (timers & PACE_TMR_RXT) {
6915                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6916                 rack->r_fast_output = 0;
6917                 ret = rack_timeout_rxt(tp, rack, cts);
6918         } else if (timers & PACE_TMR_PERSIT) {
6919                 ret = rack_timeout_persist(tp, rack, cts);
6920         } else if (timers & PACE_TMR_KEEP) {
6921                 ret = rack_timeout_keepalive(tp, rack, cts);
6922         }
6923         rack_log_to_processing(rack, cts, ret, timers);
6924         return (ret);
6925 }
6926
6927 static void
6928 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6929 {
6930         struct timeval tv;
6931         uint32_t us_cts, flags_on_entry;
6932         uint8_t hpts_removed = 0;
6933
6934         flags_on_entry = rack->r_ctl.rc_hpts_flags;
6935         us_cts = tcp_get_usecs(&tv);
6936         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
6937             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
6938              ((tp->snd_max - tp->snd_una) == 0))) {
6939                 tcp_hpts_remove(rack->rc_inp);
6940                 hpts_removed = 1;
6941                 /* If we were not delayed cancel out the flag. */
6942                 if ((tp->snd_max - tp->snd_una) == 0)
6943                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6944                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
6945         }
6946         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
6947                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
6948                 if (tcp_in_hpts(rack->rc_inp) &&
6949                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
6950                         /*
6951                          * Canceling timer's when we have no output being
6952                          * paced. We also must remove ourselves from the
6953                          * hpts.
6954                          */
6955                         tcp_hpts_remove(rack->rc_inp);
6956                         hpts_removed = 1;
6957                 }
6958                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
6959         }
6960         if (hpts_removed == 0)
6961                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
6962 }
6963
6964 static int
6965 rack_stopall(struct tcpcb *tp)
6966 {
6967         struct tcp_rack *rack;
6968         rack = (struct tcp_rack *)tp->t_fb_ptr;
6969         rack->t_timers_stopped = 1;
6970         return (0);
6971 }
6972
6973 static void
6974 rack_stop_all_timers(struct tcpcb *tp)
6975 {
6976         struct tcp_rack *rack;
6977
6978         /*
6979          * Assure no timers are running.
6980          */
6981         if (tcp_timer_active(tp, TT_PERSIST)) {
6982                 /* We enter in persists, set the flag appropriately */
6983                 rack = (struct tcp_rack *)tp->t_fb_ptr;
6984                 rack->rc_in_persist = 1;
6985         }
6986 }
6987
6988 static void
6989 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
6990     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
6991 {
6992         int32_t idx;
6993
6994         rsm->r_rtr_cnt++;
6995         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6996         rsm->r_dupack = 0;
6997         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
6998                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
6999                 rsm->r_flags |= RACK_OVERMAX;
7000         }
7001         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7002                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7003                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7004         }
7005         idx = rsm->r_rtr_cnt - 1;
7006         rsm->r_tim_lastsent[idx] = ts;
7007         /*
7008          * Here we don't add in the len of send, since its already
7009          * in snduna <->snd_max.
7010          */
7011         rsm->r_fas = ctf_flight_size(rack->rc_tp,
7012                                      rack->r_ctl.rc_sacked);
7013         if (rsm->r_flags & RACK_ACKED) {
7014                 /* Problably MTU discovery messing with us */
7015                 rsm->r_flags &= ~RACK_ACKED;
7016                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7017         }
7018         if (rsm->r_in_tmap) {
7019                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7020                 rsm->r_in_tmap = 0;
7021         }
7022         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7023         rsm->r_in_tmap = 1;
7024         /* Take off the must retransmit flag, if its on */
7025         if (rsm->r_flags & RACK_MUST_RXT) {
7026                 if (rack->r_must_retran)
7027                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
7028                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
7029                         /*
7030                          * We have retransmitted all we need. Clear
7031                          * any must retransmit flags.
7032                          */
7033                         rack->r_must_retran = 0;
7034                         rack->r_ctl.rc_out_at_rto = 0;
7035                 }
7036                 rsm->r_flags &= ~RACK_MUST_RXT;
7037         }
7038         if (rsm->r_flags & RACK_SACK_PASSED) {
7039                 /* We have retransmitted due to the SACK pass */
7040                 rsm->r_flags &= ~RACK_SACK_PASSED;
7041                 rsm->r_flags |= RACK_WAS_SACKPASS;
7042         }
7043 }
7044
7045 static uint32_t
7046 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7047     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7048 {
7049         /*
7050          * We (re-)transmitted starting at rsm->r_start for some length
7051          * (possibly less than r_end.
7052          */
7053         struct rack_sendmap *nrsm;
7054 #ifdef INVARIANTS
7055         struct rack_sendmap *insret;
7056 #endif
7057         uint32_t c_end;
7058         int32_t len;
7059
7060         len = *lenp;
7061         c_end = rsm->r_start + len;
7062         if (SEQ_GEQ(c_end, rsm->r_end)) {
7063                 /*
7064                  * We retransmitted the whole piece or more than the whole
7065                  * slopping into the next rsm.
7066                  */
7067                 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7068                 if (c_end == rsm->r_end) {
7069                         *lenp = 0;
7070                         return (0);
7071                 } else {
7072                         int32_t act_len;
7073
7074                         /* Hangs over the end return whats left */
7075                         act_len = rsm->r_end - rsm->r_start;
7076                         *lenp = (len - act_len);
7077                         return (rsm->r_end);
7078                 }
7079                 /* We don't get out of this block. */
7080         }
7081         /*
7082          * Here we retransmitted less than the whole thing which means we
7083          * have to split this into what was transmitted and what was not.
7084          */
7085         nrsm = rack_alloc_full_limit(rack);
7086         if (nrsm == NULL) {
7087                 /*
7088                  * We can't get memory, so lets not proceed.
7089                  */
7090                 *lenp = 0;
7091                 return (0);
7092         }
7093         /*
7094          * So here we are going to take the original rsm and make it what we
7095          * retransmitted. nrsm will be the tail portion we did not
7096          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7097          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7098          * 1, 6 and the new piece will be 6, 11.
7099          */
7100         rack_clone_rsm(rack, nrsm, rsm, c_end);
7101         nrsm->r_dupack = 0;
7102         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7103 #ifndef INVARIANTS
7104         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7105 #else
7106         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7107         if (insret != NULL) {
7108                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7109                       nrsm, insret, rack, rsm);
7110         }
7111 #endif
7112         if (rsm->r_in_tmap) {
7113                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7114                 nrsm->r_in_tmap = 1;
7115         }
7116         rsm->r_flags &= (~RACK_HAS_FIN);
7117         rack_update_rsm(tp, rack, rsm, ts, add_flag);
7118         /* Log a split of rsm into rsm and nrsm */
7119         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7120         *lenp = 0;
7121         return (0);
7122 }
7123
7124 static void
7125 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7126                 uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
7127                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7128 {
7129         struct tcp_rack *rack;
7130         struct rack_sendmap *rsm, *nrsm, fe;
7131 #ifdef INVARIANTS
7132         struct rack_sendmap *insret;
7133 #endif
7134         register uint32_t snd_max, snd_una;
7135
7136         /*
7137          * Add to the RACK log of packets in flight or retransmitted. If
7138          * there is a TS option we will use the TS echoed, if not we will
7139          * grab a TS.
7140          *
7141          * Retransmissions will increment the count and move the ts to its
7142          * proper place. Note that if options do not include TS's then we
7143          * won't be able to effectively use the ACK for an RTT on a retran.
7144          *
7145          * Notes about r_start and r_end. Lets consider a send starting at
7146          * sequence 1 for 10 bytes. In such an example the r_start would be
7147          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7148          * This means that r_end is actually the first sequence for the next
7149          * slot (11).
7150          *
7151          */
7152         /*
7153          * If err is set what do we do XXXrrs? should we not add the thing?
7154          * -- i.e. return if err != 0 or should we pretend we sent it? --
7155          * i.e. proceed with add ** do this for now.
7156          */
7157         INP_WLOCK_ASSERT(tptoinpcb(tp));
7158         if (err)
7159                 /*
7160                  * We don't log errors -- we could but snd_max does not
7161                  * advance in this case either.
7162                  */
7163                 return;
7164
7165         if (th_flags & TH_RST) {
7166                 /*
7167                  * We don't log resets and we return immediately from
7168                  * sending
7169                  */
7170                 return;
7171         }
7172         rack = (struct tcp_rack *)tp->t_fb_ptr;
7173         snd_una = tp->snd_una;
7174         snd_max = tp->snd_max;
7175         if (th_flags & (TH_SYN | TH_FIN)) {
7176                 /*
7177                  * The call to rack_log_output is made before bumping
7178                  * snd_max. This means we can record one extra byte on a SYN
7179                  * or FIN if seq_out is adding more on and a FIN is present
7180                  * (and we are not resending).
7181                  */
7182                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7183                         len++;
7184                 if (th_flags & TH_FIN)
7185                         len++;
7186                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7187                         /*
7188                          * The add/update as not been done for the FIN/SYN
7189                          * yet.
7190                          */
7191                         snd_max = tp->snd_nxt;
7192                 }
7193         }
7194         if (SEQ_LEQ((seq_out + len), snd_una)) {
7195                 /* Are sending an old segment to induce an ack (keep-alive)? */
7196                 return;
7197         }
7198         if (SEQ_LT(seq_out, snd_una)) {
7199                 /* huh? should we panic? */
7200                 uint32_t end;
7201
7202                 end = seq_out + len;
7203                 seq_out = snd_una;
7204                 if (SEQ_GEQ(end, seq_out))
7205                         len = end - seq_out;
7206                 else
7207                         len = 0;
7208         }
7209         if (len == 0) {
7210                 /* We don't log zero window probes */
7211                 return;
7212         }
7213         if (IN_FASTRECOVERY(tp->t_flags)) {
7214                 rack->r_ctl.rc_prr_out += len;
7215         }
7216         /* First question is it a retransmission or new? */
7217         if (seq_out == snd_max) {
7218                 /* Its new */
7219 again:
7220                 rsm = rack_alloc(rack);
7221                 if (rsm == NULL) {
7222                         /*
7223                          * Hmm out of memory and the tcb got destroyed while
7224                          * we tried to wait.
7225                          */
7226                         return;
7227                 }
7228                 if (th_flags & TH_FIN) {
7229                         rsm->r_flags = RACK_HAS_FIN|add_flag;
7230                 } else {
7231                         rsm->r_flags = add_flag;
7232                 }
7233                 if (hw_tls)
7234                         rsm->r_hw_tls = 1;
7235                 rsm->r_tim_lastsent[0] = cts;
7236                 rsm->r_rtr_cnt = 1;
7237                 rsm->r_rtr_bytes = 0;
7238                 if (th_flags & TH_SYN) {
7239                         /* The data space is one beyond snd_una */
7240                         rsm->r_flags |= RACK_HAS_SYN;
7241                 }
7242                 rsm->r_start = seq_out;
7243                 rsm->r_end = rsm->r_start + len;
7244                 rsm->r_dupack = 0;
7245                 /*
7246                  * save off the mbuf location that
7247                  * sndmbuf_noadv returned (which is
7248                  * where we started copying from)..
7249                  */
7250                 rsm->m = s_mb;
7251                 rsm->soff = s_moff;
7252                 /*
7253                  * Here we do add in the len of send, since its not yet
7254                  * reflected in in snduna <->snd_max
7255                  */
7256                 rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7257                                               rack->r_ctl.rc_sacked) +
7258                               (rsm->r_end - rsm->r_start));
7259                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7260                 if (rsm->m) {
7261                         if (rsm->m->m_len <= rsm->soff) {
7262                                 /*
7263                                  * XXXrrs Question, will this happen?
7264                                  *
7265                                  * If sbsndptr is set at the correct place
7266                                  * then s_moff should always be somewhere
7267                                  * within rsm->m. But if the sbsndptr was
7268                                  * off then that won't be true. If it occurs
7269                                  * we need to walkout to the correct location.
7270                                  */
7271                                 struct mbuf *lm;
7272
7273                                 lm = rsm->m;
7274                                 while (lm->m_len <= rsm->soff) {
7275                                         rsm->soff -= lm->m_len;
7276                                         lm = lm->m_next;
7277                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7278                                                              __func__, rack, s_moff, s_mb, rsm->soff));
7279                                 }
7280                                 rsm->m = lm;
7281                         }
7282                         rsm->orig_m_len = rsm->m->m_len;
7283                 } else
7284                         rsm->orig_m_len = 0;
7285                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7286                 /* Log a new rsm */
7287                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7288 #ifndef INVARIANTS
7289                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7290 #else
7291                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7292                 if (insret != NULL) {
7293                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7294                               nrsm, insret, rack, rsm);
7295                 }
7296 #endif
7297                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7298                 rsm->r_in_tmap = 1;
7299                 /*
7300                  * Special case detection, is there just a single
7301                  * packet outstanding when we are not in recovery?
7302                  *
7303                  * If this is true mark it so.
7304                  */
7305                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7306                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7307                         struct rack_sendmap *prsm;
7308
7309                         prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7310                         if (prsm)
7311                                 prsm->r_one_out_nr = 1;
7312                 }
7313                 return;
7314         }
7315         /*
7316          * If we reach here its a retransmission and we need to find it.
7317          */
7318         memset(&fe, 0, sizeof(fe));
7319 more:
7320         if (hintrsm && (hintrsm->r_start == seq_out)) {
7321                 rsm = hintrsm;
7322                 hintrsm = NULL;
7323         } else {
7324                 /* No hints sorry */
7325                 rsm = NULL;
7326         }
7327         if ((rsm) && (rsm->r_start == seq_out)) {
7328                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7329                 if (len == 0) {
7330                         return;
7331                 } else {
7332                         goto more;
7333                 }
7334         }
7335         /* Ok it was not the last pointer go through it the hard way. */
7336 refind:
7337         fe.r_start = seq_out;
7338         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7339         if (rsm) {
7340                 if (rsm->r_start == seq_out) {
7341                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7342                         if (len == 0) {
7343                                 return;
7344                         } else {
7345                                 goto refind;
7346                         }
7347                 }
7348                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7349                         /* Transmitted within this piece */
7350                         /*
7351                          * Ok we must split off the front and then let the
7352                          * update do the rest
7353                          */
7354                         nrsm = rack_alloc_full_limit(rack);
7355                         if (nrsm == NULL) {
7356                                 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7357                                 return;
7358                         }
7359                         /*
7360                          * copy rsm to nrsm and then trim the front of rsm
7361                          * to not include this part.
7362                          */
7363                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
7364                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7365 #ifndef INVARIANTS
7366                         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7367 #else
7368                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7369                         if (insret != NULL) {
7370                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7371                                       nrsm, insret, rack, rsm);
7372                         }
7373 #endif
7374                         if (rsm->r_in_tmap) {
7375                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7376                                 nrsm->r_in_tmap = 1;
7377                         }
7378                         rsm->r_flags &= (~RACK_HAS_FIN);
7379                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7380                         if (len == 0) {
7381                                 return;
7382                         } else if (len > 0)
7383                                 goto refind;
7384                 }
7385         }
7386         /*
7387          * Hmm not found in map did they retransmit both old and on into the
7388          * new?
7389          */
7390         if (seq_out == tp->snd_max) {
7391                 goto again;
7392         } else if (SEQ_LT(seq_out, tp->snd_max)) {
7393 #ifdef INVARIANTS
7394                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7395                        seq_out, len, tp->snd_una, tp->snd_max);
7396                 printf("Starting Dump of all rack entries\n");
7397                 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7398                         printf("rsm:%p start:%u end:%u\n",
7399                                rsm, rsm->r_start, rsm->r_end);
7400                 }
7401                 printf("Dump complete\n");
7402                 panic("seq_out not found rack:%p tp:%p",
7403                       rack, tp);
7404 #endif
7405         } else {
7406 #ifdef INVARIANTS
7407                 /*
7408                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
7409                  * flag)
7410                  */
7411                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7412                       seq_out, len, tp->snd_max, tp);
7413 #endif
7414         }
7415 }
7416
7417 /*
7418  * Record one of the RTT updates from an ack into
7419  * our sample structure.
7420  */
7421
7422 static void
7423 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7424                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7425 {
7426         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7427             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7428                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7429         }
7430         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7431             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7432                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7433         }
7434         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7435             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7436                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7437             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7438                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7439         }
7440         if ((confidence == 1) &&
7441             ((rsm == NULL) ||
7442              (rsm->r_just_ret) ||
7443              (rsm->r_one_out_nr &&
7444               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7445                 /*
7446                  * If the rsm had a just return
7447                  * hit it then we can't trust the
7448                  * rtt measurement for buffer deterimination
7449                  * Note that a confidence of 2, indicates
7450                  * SACK'd which overrides the r_just_ret or
7451                  * the r_one_out_nr. If it was a CUM-ACK and
7452                  * we had only two outstanding, but get an
7453                  * ack for only 1. Then that also lowers our
7454                  * confidence.
7455                  */
7456                 confidence = 0;
7457         }
7458         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7459             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7460                 if (rack->r_ctl.rack_rs.confidence == 0) {
7461                         /*
7462                          * We take anything with no current confidence
7463                          * saved.
7464                          */
7465                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7466                         rack->r_ctl.rack_rs.confidence = confidence;
7467                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7468                 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7469                         /*
7470                          * Once we have a confident number,
7471                          * we can update it with a smaller
7472                          * value since this confident number
7473                          * may include the DSACK time until
7474                          * the next segment (the second one) arrived.
7475                          */
7476                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7477                         rack->r_ctl.rack_rs.confidence = confidence;
7478                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7479                 }
7480         }
7481         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7482         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7483         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7484         rack->r_ctl.rack_rs.rs_rtt_cnt++;
7485 }
7486
7487 /*
7488  * Collect new round-trip time estimate
7489  * and update averages and current timeout.
7490  */
7491 static void
7492 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7493 {
7494         int32_t delta;
7495         int32_t rtt;
7496
7497         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7498                 /* No valid sample */
7499                 return;
7500         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7501                 /* We are to use the lowest RTT seen in a single ack */
7502                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7503         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7504                 /* We are to use the highest RTT seen in a single ack */
7505                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7506         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7507                 /* We are to use the average RTT seen in a single ack */
7508                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7509                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7510         } else {
7511 #ifdef INVARIANTS
7512                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7513 #endif
7514                 return;
7515         }
7516         if (rtt == 0)
7517                 rtt = 1;
7518         if (rack->rc_gp_rtt_set == 0) {
7519                 /*
7520                  * With no RTT we have to accept
7521                  * even one we are not confident of.
7522                  */
7523                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7524                 rack->rc_gp_rtt_set = 1;
7525         } else if (rack->r_ctl.rack_rs.confidence) {
7526                 /* update the running gp srtt */
7527                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7528                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7529         }
7530         if (rack->r_ctl.rack_rs.confidence) {
7531                 /*
7532                  * record the low and high for highly buffered path computation,
7533                  * we only do this if we are confident (not a retransmission).
7534                  */
7535                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7536                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7537                 }
7538                 if (rack->rc_highly_buffered == 0) {
7539                         /*
7540                          * Currently once we declare a path has
7541                          * highly buffered there is no going
7542                          * back, which may be a problem...
7543                          */
7544                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7545                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7546                                                      rack->r_ctl.rc_highest_us_rtt,
7547                                                      rack->r_ctl.rc_lowest_us_rtt,
7548                                                      RACK_RTTS_SEEHBP);
7549                                 rack->rc_highly_buffered = 1;
7550                         }
7551                 }
7552         }
7553         if ((rack->r_ctl.rack_rs.confidence) ||
7554             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7555                 /*
7556                  * If we are highly confident of it <or> it was
7557                  * never retransmitted we accept it as the last us_rtt.
7558                  */
7559                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7560                 /* The lowest rtt can be set if its was not retransmited */
7561                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7562                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7563                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
7564                                 rack->r_ctl.rc_lowest_us_rtt = 1;
7565                 }
7566         }
7567         rack = (struct tcp_rack *)tp->t_fb_ptr;
7568         if (tp->t_srtt != 0) {
7569                 /*
7570                  * We keep a simple srtt in microseconds, like our rtt
7571                  * measurement. We don't need to do any tricks with shifting
7572                  * etc. Instead we just add in 1/8th of the new measurement
7573                  * and subtract out 1/8 of the old srtt. We do the same with
7574                  * the variance after finding the absolute value of the
7575                  * difference between this sample and the current srtt.
7576                  */
7577                 delta = tp->t_srtt - rtt;
7578                 /* Take off 1/8th of the current sRTT */
7579                 tp->t_srtt -= (tp->t_srtt >> 3);
7580                 /* Add in 1/8th of the new RTT just measured */
7581                 tp->t_srtt += (rtt >> 3);
7582                 if (tp->t_srtt <= 0)
7583                         tp->t_srtt = 1;
7584                 /* Now lets make the absolute value of the variance */
7585                 if (delta < 0)
7586                         delta = -delta;
7587                 /* Subtract out 1/8th */
7588                 tp->t_rttvar -= (tp->t_rttvar >> 3);
7589                 /* Add in 1/8th of the new variance we just saw */
7590                 tp->t_rttvar += (delta >> 3);
7591                 if (tp->t_rttvar <= 0)
7592                         tp->t_rttvar = 1;
7593         } else {
7594                 /*
7595                  * No rtt measurement yet - use the unsmoothed rtt. Set the
7596                  * variance to half the rtt (so our first retransmit happens
7597                  * at 3*rtt).
7598                  */
7599                 tp->t_srtt = rtt;
7600                 tp->t_rttvar = rtt >> 1;
7601         }
7602         rack->rc_srtt_measure_made = 1;
7603         KMOD_TCPSTAT_INC(tcps_rttupdated);
7604         if (tp->t_rttupdated < UCHAR_MAX)
7605                 tp->t_rttupdated++;
7606 #ifdef STATS
7607         if (rack_stats_gets_ms_rtt == 0) {
7608                 /* Send in the microsecond rtt used for rxt timeout purposes */
7609                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7610         } else if (rack_stats_gets_ms_rtt == 1) {
7611                 /* Send in the millisecond rtt used for rxt timeout purposes */
7612                 int32_t ms_rtt;
7613
7614                 /* Round up */
7615                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7616                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7617         } else if (rack_stats_gets_ms_rtt == 2) {
7618                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
7619                 int32_t ms_rtt;
7620
7621                 /* Round up */
7622                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7623                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7624         }  else {
7625                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
7626                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7627         }
7628
7629 #endif
7630         /*
7631          * the retransmit should happen at rtt + 4 * rttvar. Because of the
7632          * way we do the smoothing, srtt and rttvar will each average +1/2
7633          * tick of bias.  When we compute the retransmit timer, we want 1/2
7634          * tick of rounding and 1 extra tick because of +-1/2 tick
7635          * uncertainty in the firing of the timer.  The bias will give us
7636          * exactly the 1.5 tick we need.  But, because the bias is
7637          * statistical, we have to test that we don't drop below the minimum
7638          * feasible timer (which is 2 ticks).
7639          */
7640         tp->t_rxtshift = 0;
7641         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7642                       max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7643         rack_log_rtt_sample(rack, rtt);
7644         tp->t_softerror = 0;
7645 }
7646
7647
7648 static void
7649 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7650 {
7651         /*
7652          * Apply to filter the inbound us-rtt at us_cts.
7653          */
7654         uint32_t old_rtt;
7655
7656         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7657         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7658                                us_rtt, us_cts);
7659         if (old_rtt > us_rtt) {
7660                 /* We just hit a new lower rtt time */
7661                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7662                                      __LINE__, RACK_RTTS_NEWRTT);
7663                 /*
7664                  * Only count it if its lower than what we saw within our
7665                  * calculated range.
7666                  */
7667                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7668                         if (rack_probertt_lower_within &&
7669                             rack->rc_gp_dyn_mul &&
7670                             (rack->use_fixed_rate == 0) &&
7671                             (rack->rc_always_pace)) {
7672                                 /*
7673                                  * We are seeing a new lower rtt very close
7674                                  * to the time that we would have entered probe-rtt.
7675                                  * This is probably due to the fact that a peer flow
7676                                  * has entered probe-rtt. Lets go in now too.
7677                                  */
7678                                 uint32_t val;
7679
7680                                 val = rack_probertt_lower_within * rack_time_between_probertt;
7681                                 val /= 100;
7682                                 if ((rack->in_probe_rtt == 0)  &&
7683                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7684                                         rack_enter_probertt(rack, us_cts);
7685                                 }
7686                         }
7687                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7688                 }
7689         }
7690 }
7691
7692 static int
7693 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7694     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7695 {
7696         uint32_t us_rtt;
7697         int32_t i, all;
7698         uint32_t t, len_acked;
7699
7700         if ((rsm->r_flags & RACK_ACKED) ||
7701             (rsm->r_flags & RACK_WAS_ACKED))
7702                 /* Already done */
7703                 return (0);
7704         if (rsm->r_no_rtt_allowed) {
7705                 /* Not allowed */
7706                 return (0);
7707         }
7708         if (ack_type == CUM_ACKED) {
7709                 if (SEQ_GT(th_ack, rsm->r_end)) {
7710                         len_acked = rsm->r_end - rsm->r_start;
7711                         all = 1;
7712                 } else {
7713                         len_acked = th_ack - rsm->r_start;
7714                         all = 0;
7715                 }
7716         } else {
7717                 len_acked = rsm->r_end - rsm->r_start;
7718                 all = 0;
7719         }
7720         if (rsm->r_rtr_cnt == 1) {
7721
7722                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7723                 if ((int)t <= 0)
7724                         t = 1;
7725                 if (!tp->t_rttlow || tp->t_rttlow > t)
7726                         tp->t_rttlow = t;
7727                 if (!rack->r_ctl.rc_rack_min_rtt ||
7728                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7729                         rack->r_ctl.rc_rack_min_rtt = t;
7730                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7731                                 rack->r_ctl.rc_rack_min_rtt = 1;
7732                         }
7733                 }
7734                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7735                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7736                 else
7737                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7738                 if (us_rtt == 0)
7739                         us_rtt = 1;
7740                 if (CC_ALGO(tp)->rttsample != NULL) {
7741                         /* Kick the RTT to the CC */
7742                         CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
7743                 }
7744                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7745                 if (ack_type == SACKED) {
7746                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7747                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7748                 } else {
7749                         /*
7750                          * We need to setup what our confidence
7751                          * is in this ack.
7752                          *
7753                          * If the rsm was app limited and it is
7754                          * less than a mss in length (the end
7755                          * of the send) then we have a gap. If we
7756                          * were app limited but say we were sending
7757                          * multiple MSS's then we are more confident
7758                          * int it.
7759                          *
7760                          * When we are not app-limited then we see if
7761                          * the rsm is being included in the current
7762                          * measurement, we tell this by the app_limited_needs_set
7763                          * flag.
7764                          *
7765                          * Note that being cwnd blocked is not applimited
7766                          * as well as the pacing delay between packets which
7767                          * are sending only 1 or 2 MSS's also will show up
7768                          * in the RTT. We probably need to examine this algorithm
7769                          * a bit more and enhance it to account for the delay
7770                          * between rsm's. We could do that by saving off the
7771                          * pacing delay of each rsm (in an rsm) and then
7772                          * factoring that in somehow though for now I am
7773                          * not sure how :)
7774                          */
7775                         int calc_conf = 0;
7776
7777                         if (rsm->r_flags & RACK_APP_LIMITED) {
7778                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7779                                         calc_conf = 0;
7780                                 else
7781                                         calc_conf = 1;
7782                         } else if (rack->app_limited_needs_set == 0) {
7783                                 calc_conf = 1;
7784                         } else {
7785                                 calc_conf = 0;
7786                         }
7787                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7788                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7789                                             calc_conf, rsm, rsm->r_rtr_cnt);
7790                 }
7791                 if ((rsm->r_flags & RACK_TLP) &&
7792                     (!IN_FASTRECOVERY(tp->t_flags))) {
7793                         /* Segment was a TLP and our retrans matched */
7794                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7795                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
7796                         }
7797                 }
7798                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7799                         /* New more recent rack_tmit_time */
7800                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7801                         rack->rc_rack_rtt = t;
7802                 }
7803                 return (1);
7804         }
7805         /*
7806          * We clear the soft/rxtshift since we got an ack.
7807          * There is no assurance we will call the commit() function
7808          * so we need to clear these to avoid incorrect handling.
7809          */
7810         tp->t_rxtshift = 0;
7811         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7812                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7813         tp->t_softerror = 0;
7814         if (to && (to->to_flags & TOF_TS) &&
7815             (ack_type == CUM_ACKED) &&
7816             (to->to_tsecr) &&
7817             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7818                 /*
7819                  * Now which timestamp does it match? In this block the ACK
7820                  * must be coming from a previous transmission.
7821                  */
7822                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7823                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7824                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7825                                 if ((int)t <= 0)
7826                                         t = 1;
7827                                 if (CC_ALGO(tp)->rttsample != NULL) {
7828                                         /*
7829                                          * Kick the RTT to the CC, here
7830                                          * we lie a bit in that we know the
7831                                          * retransmission is correct even though
7832                                          * we retransmitted. This is because
7833                                          * we match the timestamps.
7834                                          */
7835                                         if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
7836                                                 us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
7837                                         else
7838                                                 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
7839                                         CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
7840                                 }
7841                                 if ((i + 1) < rsm->r_rtr_cnt) {
7842                                         /*
7843                                          * The peer ack'd from our previous
7844                                          * transmission. We have a spurious
7845                                          * retransmission and thus we dont
7846                                          * want to update our rack_rtt.
7847                                          *
7848                                          * Hmm should there be a CC revert here?
7849                                          *
7850                                          */
7851                                         return (0);
7852                                 }
7853                                 if (!tp->t_rttlow || tp->t_rttlow > t)
7854                                         tp->t_rttlow = t;
7855                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7856                                         rack->r_ctl.rc_rack_min_rtt = t;
7857                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7858                                                 rack->r_ctl.rc_rack_min_rtt = 1;
7859                                         }
7860                                 }
7861                                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7862                                            (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7863                                         /* New more recent rack_tmit_time */
7864                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7865                                         rack->rc_rack_rtt = t;
7866                                 }
7867                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7868                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7869                                                     rsm->r_rtr_cnt);
7870                                 return (1);
7871                         }
7872                 }
7873                 goto ts_not_found;
7874         } else {
7875                 /*
7876                  * Ok its a SACK block that we retransmitted. or a windows
7877                  * machine without timestamps. We can tell nothing from the
7878                  * time-stamp since its not there or the time the peer last
7879                  * recieved a segment that moved forward its cum-ack point.
7880                  */
7881 ts_not_found:
7882                 i = rsm->r_rtr_cnt - 1;
7883                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7884                 if ((int)t <= 0)
7885                         t = 1;
7886                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7887                         /*
7888                          * We retransmitted and the ack came back in less
7889                          * than the smallest rtt we have observed. We most
7890                          * likely did an improper retransmit as outlined in
7891                          * 6.2 Step 2 point 2 in the rack-draft so we
7892                          * don't want to update our rack_rtt. We in
7893                          * theory (in future) might want to think about reverting our
7894                          * cwnd state but we won't for now.
7895                          */
7896                         return (0);
7897                 } else if (rack->r_ctl.rc_rack_min_rtt) {
7898                         /*
7899                          * We retransmitted it and the retransmit did the
7900                          * job.
7901                          */
7902                         if (!rack->r_ctl.rc_rack_min_rtt ||
7903                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7904                                 rack->r_ctl.rc_rack_min_rtt = t;
7905                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7906                                         rack->r_ctl.rc_rack_min_rtt = 1;
7907                                 }
7908                         }
7909                         if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7910                                 /* New more recent rack_tmit_time */
7911                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7912                                 rack->rc_rack_rtt = t;
7913                         }
7914                         return (1);
7915                 }
7916         }
7917         return (0);
7918 }
7919
7920 /*
7921  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7922  */
7923 static void
7924 rack_log_sack_passed(struct tcpcb *tp,
7925     struct tcp_rack *rack, struct rack_sendmap *rsm)
7926 {
7927         struct rack_sendmap *nrsm;
7928
7929         nrsm = rsm;
7930         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7931             rack_head, r_tnext) {
7932                 if (nrsm == rsm) {
7933                         /* Skip original segment he is acked */
7934                         continue;
7935                 }
7936                 if (nrsm->r_flags & RACK_ACKED) {
7937                         /*
7938                          * Skip ack'd segments, though we
7939                          * should not see these, since tmap
7940                          * should not have ack'd segments.
7941                          */
7942                         continue;
7943                 }
7944                 if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
7945                         /*
7946                          * If the peer dropped the rwnd on
7947                          * these then we don't worry about them.
7948                          */
7949                         continue;
7950                 }
7951                 if (nrsm->r_flags & RACK_SACK_PASSED) {
7952                         /*
7953                          * We found one that is already marked
7954                          * passed, we have been here before and
7955                          * so all others below this are marked.
7956                          */
7957                         break;
7958                 }
7959                 nrsm->r_flags |= RACK_SACK_PASSED;
7960                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
7961         }
7962 }
7963
7964 static void
7965 rack_need_set_test(struct tcpcb *tp,
7966                    struct tcp_rack *rack,
7967                    struct rack_sendmap *rsm,
7968                    tcp_seq th_ack,
7969                    int line,
7970                    int use_which)
7971 {
7972
7973         if ((tp->t_flags & TF_GPUTINPROG) &&
7974             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
7975                 /*
7976                  * We were app limited, and this ack
7977                  * butts up or goes beyond the point where we want
7978                  * to start our next measurement. We need
7979                  * to record the new gput_ts as here and
7980                  * possibly update the start sequence.
7981                  */
7982                 uint32_t seq, ts;
7983
7984                 if (rsm->r_rtr_cnt > 1) {
7985                         /*
7986                          * This is a retransmit, can we
7987                          * really make any assessment at this
7988                          * point?  We are not really sure of
7989                          * the timestamp, is it this or the
7990                          * previous transmission?
7991                          *
7992                          * Lets wait for something better that
7993                          * is not retransmitted.
7994                          */
7995                         return;
7996                 }
7997                 seq = tp->gput_seq;
7998                 ts = tp->gput_ts;
7999                 rack->app_limited_needs_set = 0;
8000                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8001                 /* Do we start at a new end? */
8002                 if ((use_which == RACK_USE_BEG) &&
8003                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8004                         /*
8005                          * When we get an ACK that just eats
8006                          * up some of the rsm, we set RACK_USE_BEG
8007                          * since whats at r_start (i.e. th_ack)
8008                          * is left unacked and thats where the
8009                          * measurement not starts.
8010                          */
8011                         tp->gput_seq = rsm->r_start;
8012                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8013                 }
8014                 if ((use_which == RACK_USE_END) &&
8015                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8016                             /*
8017                              * We use the end when the cumack
8018                              * is moving forward and completely
8019                              * deleting the rsm passed so basically
8020                              * r_end holds th_ack.
8021                              *
8022                              * For SACK's we also want to use the end
8023                              * since this piece just got sacked and
8024                              * we want to target anything after that
8025                              * in our measurement.
8026                              */
8027                             tp->gput_seq = rsm->r_end;
8028                             rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8029                 }
8030                 if (use_which == RACK_USE_END_OR_THACK) {
8031                         /*
8032                          * special case for ack moving forward,
8033                          * not a sack, we need to move all the
8034                          * way up to where this ack cum-ack moves
8035                          * to.
8036                          */
8037                         if (SEQ_GT(th_ack, rsm->r_end))
8038                                 tp->gput_seq = th_ack;
8039                         else
8040                                 tp->gput_seq = rsm->r_end;
8041                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8042                 }
8043                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8044                         /*
8045                          * We moved beyond this guy's range, re-calculate
8046                          * the new end point.
8047                          */
8048                         if (rack->rc_gp_filled == 0) {
8049                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8050                         } else {
8051                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8052                         }
8053                 }
8054                 /*
8055                  * We are moving the goal post, we may be able to clear the
8056                  * measure_saw_probe_rtt flag.
8057                  */
8058                 if ((rack->in_probe_rtt == 0) &&
8059                     (rack->measure_saw_probe_rtt) &&
8060                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8061                         rack->measure_saw_probe_rtt = 0;
8062                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8063                                            seq, tp->gput_seq, 0, 5, line, NULL, 0);
8064                 if (rack->rc_gp_filled &&
8065                     ((tp->gput_ack - tp->gput_seq) <
8066                      max(rc_init_window(rack), (MIN_GP_WIN *
8067                                                 ctf_fixed_maxseg(tp))))) {
8068                         uint32_t ideal_amount;
8069
8070                         ideal_amount = rack_get_measure_window(tp, rack);
8071                         if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
8072                                 /*
8073                                  * There is no sense of continuing this measurement
8074                                  * because its too small to gain us anything we
8075                                  * trust. Skip it and that way we can start a new
8076                                  * measurement quicker.
8077                                  */
8078                                 tp->t_flags &= ~TF_GPUTINPROG;
8079                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8080                                                            0, 0, 0, 6, __LINE__, NULL, 0);
8081                         } else {
8082                                 /*
8083                                  * Reset the window further out.
8084                                  */
8085                                 tp->gput_ack = tp->gput_seq + ideal_amount;
8086                         }
8087                 }
8088         }
8089 }
8090
8091 static inline int
8092 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8093 {
8094         if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8095                 /* Behind our TLP definition or right at */
8096                 return (0);
8097         }
8098         if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8099                 /* The start is beyond or right at our end of TLP definition */
8100                 return (0);
8101         }
8102         /* It has to be a sub-part of the original TLP recorded */
8103         return (1);
8104 }
8105
8106
8107 static uint32_t
8108 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8109                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8110 {
8111         uint32_t start, end, changed = 0;
8112         struct rack_sendmap stack_map;
8113         struct rack_sendmap *rsm, *nrsm, fe, *prev, *next;
8114 #ifdef INVARIANTS
8115         struct rack_sendmap *insret;
8116 #endif
8117         int32_t used_ref = 1;
8118         int moved = 0;
8119
8120         start = sack->start;
8121         end = sack->end;
8122         rsm = *prsm;
8123         memset(&fe, 0, sizeof(fe));
8124 do_rest_ofb:
8125         if ((rsm == NULL) ||
8126             (SEQ_LT(end, rsm->r_start)) ||
8127             (SEQ_GEQ(start, rsm->r_end)) ||
8128             (SEQ_LT(start, rsm->r_start))) {
8129                 /*
8130                  * We are not in the right spot,
8131                  * find the correct spot in the tree.
8132                  */
8133                 used_ref = 0;
8134                 fe.r_start = start;
8135                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8136                 moved++;
8137         }
8138         if (rsm == NULL) {
8139                 /* TSNH */
8140                 goto out;
8141         }
8142         /* Ok we have an ACK for some piece of this rsm */
8143         if (rsm->r_start != start) {
8144                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8145                         /*
8146                          * Before any splitting or hookery is
8147                          * done is it a TLP of interest i.e. rxt?
8148                          */
8149                         if ((rsm->r_flags & RACK_TLP) &&
8150                             (rsm->r_rtr_cnt > 1)) {
8151                                 /*
8152                                  * We are splitting a rxt TLP, check
8153                                  * if we need to save off the start/end
8154                                  */
8155                                 if (rack->rc_last_tlp_acked_set &&
8156                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8157                                         /*
8158                                          * We already turned this on since we are inside
8159                                          * the previous one was a partially sack now we
8160                                          * are getting another one (maybe all of it).
8161                                          *
8162                                          */
8163                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8164                                         /*
8165                                          * Lets make sure we have all of it though.
8166                                          */
8167                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8168                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8169                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8170                                                                      rack->r_ctl.last_tlp_acked_end);
8171                                         }
8172                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8173                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8174                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8175                                                                      rack->r_ctl.last_tlp_acked_end);
8176                                         }
8177                                 } else {
8178                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8179                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8180                                         rack->rc_last_tlp_past_cumack = 0;
8181                                         rack->rc_last_tlp_acked_set = 1;
8182                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8183                                 }
8184                         }
8185                         /**
8186                          * Need to split this in two pieces the before and after,
8187                          * the before remains in the map, the after must be
8188                          * added. In other words we have:
8189                          * rsm        |--------------|
8190                          * sackblk        |------->
8191                          * rsm will become
8192                          *     rsm    |---|
8193                          * and nrsm will be  the sacked piece
8194                          *     nrsm       |----------|
8195                          *
8196                          * But before we start down that path lets
8197                          * see if the sack spans over on top of
8198                          * the next guy and it is already sacked.
8199                          *
8200                          */
8201                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8202                         if (next && (next->r_flags & RACK_ACKED) &&
8203                             SEQ_GEQ(end, next->r_start)) {
8204                                 /**
8205                                  * So the next one is already acked, and
8206                                  * we can thus by hookery use our stack_map
8207                                  * to reflect the piece being sacked and
8208                                  * then adjust the two tree entries moving
8209                                  * the start and ends around. So we start like:
8210                                  *  rsm     |------------|             (not-acked)
8211                                  *  next                 |-----------| (acked)
8212                                  *  sackblk        |-------->
8213                                  *  We want to end like so:
8214                                  *  rsm     |------|                   (not-acked)
8215                                  *  next           |-----------------| (acked)
8216                                  *  nrsm           |-----|
8217                                  * Where nrsm is a temporary stack piece we
8218                                  * use to update all the gizmos.
8219                                  */
8220                                 /* Copy up our fudge block */
8221                                 nrsm = &stack_map;
8222                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8223                                 /* Now adjust our tree blocks */
8224                                 rsm->r_end = start;
8225                                 next->r_start = start;
8226                                 /* Now we must adjust back where next->m is */
8227                                 rack_setup_offset_for_rsm(rsm, next);
8228
8229                                 /* We don't need to adjust rsm, it did not change */
8230                                 /* Clear out the dup ack count of the remainder */
8231                                 rsm->r_dupack = 0;
8232                                 rsm->r_just_ret = 0;
8233                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8234                                 /* Now lets make sure our fudge block is right */
8235                                 nrsm->r_start = start;
8236                                 /* Now lets update all the stats and such */
8237                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8238                                 if (rack->app_limited_needs_set)
8239                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8240                                 changed += (nrsm->r_end - nrsm->r_start);
8241                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8242                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8243                                         rack->r_ctl.rc_reorder_ts = cts;
8244                                 }
8245                                 /*
8246                                  * Now we want to go up from rsm (the
8247                                  * one left un-acked) to the next one
8248                                  * in the tmap. We do this so when
8249                                  * we walk backwards we include marking
8250                                  * sack-passed on rsm (The one passed in
8251                                  * is skipped since it is generally called
8252                                  * on something sacked before removing it
8253                                  * from the tmap).
8254                                  */
8255                                 if (rsm->r_in_tmap) {
8256                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
8257                                         /*
8258                                          * Now that we have the next
8259                                          * one walk backwards from there.
8260                                          */
8261                                         if (nrsm && nrsm->r_in_tmap)
8262                                                 rack_log_sack_passed(tp, rack, nrsm);
8263                                 }
8264                                 /* Now are we done? */
8265                                 if (SEQ_LT(end, next->r_end) ||
8266                                     (end == next->r_end)) {
8267                                         /* Done with block */
8268                                         goto out;
8269                                 }
8270                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8271                                 counter_u64_add(rack_sack_used_next_merge, 1);
8272                                 /* Postion for the next block */
8273                                 start = next->r_end;
8274                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8275                                 if (rsm == NULL)
8276                                         goto out;
8277                         } else {
8278                                 /**
8279                                  * We can't use any hookery here, so we
8280                                  * need to split the map. We enter like
8281                                  * so:
8282                                  *  rsm      |--------|
8283                                  *  sackblk       |----->
8284                                  * We will add the new block nrsm and
8285                                  * that will be the new portion, and then
8286                                  * fall through after reseting rsm. So we
8287                                  * split and look like this:
8288                                  *  rsm      |----|
8289                                  *  sackblk       |----->
8290                                  *  nrsm          |---|
8291                                  * We then fall through reseting
8292                                  * rsm to nrsm, so the next block
8293                                  * picks it up.
8294                                  */
8295                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8296                                 if (nrsm == NULL) {
8297                                         /*
8298                                          * failed XXXrrs what can we do but loose the sack
8299                                          * info?
8300                                          */
8301                                         goto out;
8302                                 }
8303                                 counter_u64_add(rack_sack_splits, 1);
8304                                 rack_clone_rsm(rack, nrsm, rsm, start);
8305                                 rsm->r_just_ret = 0;
8306 #ifndef INVARIANTS
8307                                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8308 #else
8309                                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8310                                 if (insret != NULL) {
8311                                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8312                                               nrsm, insret, rack, rsm);
8313                                 }
8314 #endif
8315                                 if (rsm->r_in_tmap) {
8316                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8317                                         nrsm->r_in_tmap = 1;
8318                                 }
8319                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8320                                 rsm->r_flags &= (~RACK_HAS_FIN);
8321                                 /* Position us to point to the new nrsm that starts the sack blk */
8322                                 rsm = nrsm;
8323                         }
8324                 } else {
8325                         /* Already sacked this piece */
8326                         counter_u64_add(rack_sack_skipped_acked, 1);
8327                         moved++;
8328                         if (end == rsm->r_end) {
8329                                 /* Done with block */
8330                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8331                                 goto out;
8332                         } else if (SEQ_LT(end, rsm->r_end)) {
8333                                 /* A partial sack to a already sacked block */
8334                                 moved++;
8335                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8336                                 goto out;
8337                         } else {
8338                                 /*
8339                                  * The end goes beyond this guy
8340                                  * reposition the start to the
8341                                  * next block.
8342                                  */
8343                                 start = rsm->r_end;
8344                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8345                                 if (rsm == NULL)
8346                                         goto out;
8347                         }
8348                 }
8349         }
8350         if (SEQ_GEQ(end, rsm->r_end)) {
8351                 /**
8352                  * The end of this block is either beyond this guy or right
8353                  * at this guy. I.e.:
8354                  *  rsm ---                 |-----|
8355                  *  end                     |-----|
8356                  *  <or>
8357                  *  end                     |---------|
8358                  */
8359                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8360                         /*
8361                          * Is it a TLP of interest?
8362                          */
8363                         if ((rsm->r_flags & RACK_TLP) &&
8364                             (rsm->r_rtr_cnt > 1)) {
8365                                 /*
8366                                  * We are splitting a rxt TLP, check
8367                                  * if we need to save off the start/end
8368                                  */
8369                                 if (rack->rc_last_tlp_acked_set &&
8370                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8371                                         /*
8372                                          * We already turned this on since we are inside
8373                                          * the previous one was a partially sack now we
8374                                          * are getting another one (maybe all of it).
8375                                          */
8376                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8377                                         /*
8378                                          * Lets make sure we have all of it though.
8379                                          */
8380                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8381                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8382                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8383                                                                      rack->r_ctl.last_tlp_acked_end);
8384                                         }
8385                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8386                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8387                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8388                                                                      rack->r_ctl.last_tlp_acked_end);
8389                                         }
8390                                 } else {
8391                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8392                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8393                                         rack->rc_last_tlp_past_cumack = 0;
8394                                         rack->rc_last_tlp_acked_set = 1;
8395                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8396                                 }
8397                         }
8398                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8399                         changed += (rsm->r_end - rsm->r_start);
8400                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8401                         if (rsm->r_in_tmap) /* should be true */
8402                                 rack_log_sack_passed(tp, rack, rsm);
8403                         /* Is Reordering occuring? */
8404                         if (rsm->r_flags & RACK_SACK_PASSED) {
8405                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8406                                 rack->r_ctl.rc_reorder_ts = cts;
8407                         }
8408                         if (rack->app_limited_needs_set)
8409                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8410                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8411                         rsm->r_flags |= RACK_ACKED;
8412                         if (rsm->r_in_tmap) {
8413                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8414                                 rsm->r_in_tmap = 0;
8415                         }
8416                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8417                 } else {
8418                         counter_u64_add(rack_sack_skipped_acked, 1);
8419                         moved++;
8420                 }
8421                 if (end == rsm->r_end) {
8422                         /* This block only - done, setup for next */
8423                         goto out;
8424                 }
8425                 /*
8426                  * There is more not coverend by this rsm move on
8427                  * to the next block in the RB tree.
8428                  */
8429                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8430                 start = rsm->r_end;
8431                 rsm = nrsm;
8432                 if (rsm == NULL)
8433                         goto out;
8434                 goto do_rest_ofb;
8435         }
8436         /**
8437          * The end of this sack block is smaller than
8438          * our rsm i.e.:
8439          *  rsm ---                 |-----|
8440          *  end                     |--|
8441          */
8442         if ((rsm->r_flags & RACK_ACKED) == 0) {
8443                 /*
8444                  * Is it a TLP of interest?
8445                  */
8446                 if ((rsm->r_flags & RACK_TLP) &&
8447                     (rsm->r_rtr_cnt > 1)) {
8448                         /*
8449                          * We are splitting a rxt TLP, check
8450                          * if we need to save off the start/end
8451                          */
8452                         if (rack->rc_last_tlp_acked_set &&
8453                             (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8454                                 /*
8455                                  * We already turned this on since we are inside
8456                                  * the previous one was a partially sack now we
8457                                  * are getting another one (maybe all of it).
8458                                  */
8459                                 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8460                                 /*
8461                                  * Lets make sure we have all of it though.
8462                                  */
8463                                 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8464                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8465                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8466                                                              rack->r_ctl.last_tlp_acked_end);
8467                                 }
8468                                 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8469                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8470                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8471                                                              rack->r_ctl.last_tlp_acked_end);
8472                                 }
8473                         } else {
8474                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8475                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8476                                 rack->rc_last_tlp_past_cumack = 0;
8477                                 rack->rc_last_tlp_acked_set = 1;
8478                                 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8479                         }
8480                 }
8481                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8482                 if (prev &&
8483                     (prev->r_flags & RACK_ACKED)) {
8484                         /**
8485                          * Goal, we want the right remainder of rsm to shrink
8486                          * in place and span from (rsm->r_start = end) to rsm->r_end.
8487                          * We want to expand prev to go all the way
8488                          * to prev->r_end <- end.
8489                          * so in the tree we have before:
8490                          *   prev     |--------|         (acked)
8491                          *   rsm               |-------| (non-acked)
8492                          *   sackblk           |-|
8493                          * We churn it so we end up with
8494                          *   prev     |----------|       (acked)
8495                          *   rsm                 |-----| (non-acked)
8496                          *   nrsm              |-| (temporary)
8497                          *
8498                          * Note if either prev/rsm is a TLP we don't
8499                          * do this.
8500                          */
8501                         nrsm = &stack_map;
8502                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8503                         prev->r_end = end;
8504                         rsm->r_start = end;
8505                         /* Now adjust nrsm (stack copy) to be
8506                          * the one that is the small
8507                          * piece that was "sacked".
8508                          */
8509                         nrsm->r_end = end;
8510                         rsm->r_dupack = 0;
8511                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8512                         /*
8513                          * Now that the rsm has had its start moved forward
8514                          * lets go ahead and get its new place in the world.
8515                          */
8516                         rack_setup_offset_for_rsm(prev, rsm);
8517                         /*
8518                          * Now nrsm is our new little piece
8519                          * that is acked (which was merged
8520                          * to prev). Update the rtt and changed
8521                          * based on that. Also check for reordering.
8522                          */
8523                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8524                         if (rack->app_limited_needs_set)
8525                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8526                         changed += (nrsm->r_end - nrsm->r_start);
8527                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8528                         if (nrsm->r_flags & RACK_SACK_PASSED) {
8529                                 rack->r_ctl.rc_reorder_ts = cts;
8530                         }
8531                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8532                         rsm = prev;
8533                         counter_u64_add(rack_sack_used_prev_merge, 1);
8534                 } else {
8535                         /**
8536                          * This is the case where our previous
8537                          * block is not acked either, so we must
8538                          * split the block in two.
8539                          */
8540                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8541                         if (nrsm == NULL) {
8542                                 /* failed rrs what can we do but loose the sack info? */
8543                                 goto out;
8544                         }
8545                         if ((rsm->r_flags & RACK_TLP) &&
8546                             (rsm->r_rtr_cnt > 1)) {
8547                                 /*
8548                                  * We are splitting a rxt TLP, check
8549                                  * if we need to save off the start/end
8550                                  */
8551                                 if (rack->rc_last_tlp_acked_set &&
8552                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8553                                             /*
8554                                              * We already turned this on since this block is inside
8555                                              * the previous one was a partially sack now we
8556                                              * are getting another one (maybe all of it).
8557                                              */
8558                                             rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8559                                             /*
8560                                              * Lets make sure we have all of it though.
8561                                              */
8562                                             if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8563                                                     rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8564                                                     rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8565                                                                          rack->r_ctl.last_tlp_acked_end);
8566                                             }
8567                                             if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8568                                                     rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8569                                                     rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8570                                                                          rack->r_ctl.last_tlp_acked_end);
8571                                             }
8572                                     } else {
8573                                             rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8574                                             rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8575                                             rack->rc_last_tlp_acked_set = 1;
8576                                             rack->rc_last_tlp_past_cumack = 0;
8577                                             rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8578                                     }
8579                         }
8580                         /**
8581                          * In this case nrsm becomes
8582                          * nrsm->r_start = end;
8583                          * nrsm->r_end = rsm->r_end;
8584                          * which is un-acked.
8585                          * <and>
8586                          * rsm->r_end = nrsm->r_start;
8587                          * i.e. the remaining un-acked
8588                          * piece is left on the left
8589                          * hand side.
8590                          *
8591                          * So we start like this
8592                          * rsm      |----------| (not acked)
8593                          * sackblk  |---|
8594                          * build it so we have
8595                          * rsm      |---|         (acked)
8596                          * nrsm         |------|  (not acked)
8597                          */
8598                         counter_u64_add(rack_sack_splits, 1);
8599                         rack_clone_rsm(rack, nrsm, rsm, end);
8600                         rsm->r_flags &= (~RACK_HAS_FIN);
8601                         rsm->r_just_ret = 0;
8602 #ifndef INVARIANTS
8603                         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8604 #else
8605                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8606                         if (insret != NULL) {
8607                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8608                                       nrsm, insret, rack, rsm);
8609                         }
8610 #endif
8611                         if (rsm->r_in_tmap) {
8612                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8613                                 nrsm->r_in_tmap = 1;
8614                         }
8615                         nrsm->r_dupack = 0;
8616                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8617                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8618                         changed += (rsm->r_end - rsm->r_start);
8619                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8620                         if (rsm->r_in_tmap) /* should be true */
8621                                 rack_log_sack_passed(tp, rack, rsm);
8622                         /* Is Reordering occuring? */
8623                         if (rsm->r_flags & RACK_SACK_PASSED) {
8624                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8625                                 rack->r_ctl.rc_reorder_ts = cts;
8626                         }
8627                         if (rack->app_limited_needs_set)
8628                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8629                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8630                         rsm->r_flags |= RACK_ACKED;
8631                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8632                         if (rsm->r_in_tmap) {
8633                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8634                                 rsm->r_in_tmap = 0;
8635                         }
8636                 }
8637         } else if (start != end){
8638                 /*
8639                  * The block was already acked.
8640                  */
8641                 counter_u64_add(rack_sack_skipped_acked, 1);
8642                 moved++;
8643         }
8644 out:
8645         if (rsm &&
8646             ((rsm->r_flags & RACK_TLP) == 0) &&
8647             (rsm->r_flags & RACK_ACKED)) {
8648                 /*
8649                  * Now can we merge where we worked
8650                  * with either the previous or
8651                  * next block?
8652                  */
8653                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8654                 while (next) {
8655                         if (next->r_flags & RACK_TLP)
8656                                 break;
8657                         if (next->r_flags & RACK_ACKED) {
8658                         /* yep this and next can be merged */
8659                                 rsm = rack_merge_rsm(rack, rsm, next);
8660                                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8661                         } else
8662                                 break;
8663                 }
8664                 /* Now what about the previous? */
8665                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8666                 while (prev) {
8667                         if (prev->r_flags & RACK_TLP)
8668                                 break;
8669                         if (prev->r_flags & RACK_ACKED) {
8670                                 /* yep the previous and this can be merged */
8671                                 rsm = rack_merge_rsm(rack, prev, rsm);
8672                                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8673                         } else
8674                                 break;
8675                 }
8676         }
8677         if (used_ref == 0) {
8678                 counter_u64_add(rack_sack_proc_all, 1);
8679         } else {
8680                 counter_u64_add(rack_sack_proc_short, 1);
8681         }
8682         /* Save off the next one for quick reference. */
8683         if (rsm)
8684                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8685         else
8686                 nrsm = NULL;
8687         *prsm = rack->r_ctl.rc_sacklast = nrsm;
8688         /* Pass back the moved. */
8689         *moved_two = moved;
8690         return (changed);
8691 }
8692
8693 static void inline
8694 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8695 {
8696         struct rack_sendmap *tmap;
8697
8698         tmap = NULL;
8699         while (rsm && (rsm->r_flags & RACK_ACKED)) {
8700                 /* Its no longer sacked, mark it so */
8701                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8702 #ifdef INVARIANTS
8703                 if (rsm->r_in_tmap) {
8704                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
8705                               rack, rsm, rsm->r_flags);
8706                 }
8707 #endif
8708                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8709                 /* Rebuild it into our tmap */
8710                 if (tmap == NULL) {
8711                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8712                         tmap = rsm;
8713                 } else {
8714                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8715                         tmap = rsm;
8716                 }
8717                 tmap->r_in_tmap = 1;
8718                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8719         }
8720         /*
8721          * Now lets possibly clear the sack filter so we start
8722          * recognizing sacks that cover this area.
8723          */
8724         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8725
8726 }
8727
8728 static void
8729 rack_do_decay(struct tcp_rack *rack)
8730 {
8731         struct timeval res;
8732
8733 #define timersub(tvp, uvp, vvp)                                         \
8734         do {                                                            \
8735                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
8736                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
8737                 if ((vvp)->tv_usec < 0) {                               \
8738                         (vvp)->tv_sec--;                                \
8739                         (vvp)->tv_usec += 1000000;                      \
8740                 }                                                       \
8741         } while (0)
8742
8743         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8744 #undef timersub
8745
8746         rack->r_ctl.input_pkt++;
8747         if ((rack->rc_in_persist) ||
8748             (res.tv_sec >= 1) ||
8749             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8750                 /*
8751                  * Check for decay of non-SAD,
8752                  * we want all SAD detection metrics to
8753                  * decay 1/4 per second (or more) passed.
8754                  */
8755 #ifdef NETFLIX_EXP_DETECTION
8756                 uint32_t pkt_delta;
8757
8758                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8759 #endif
8760                 /* Update our saved tracking values */
8761                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8762                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8763                 /* Now do we escape without decay? */
8764 #ifdef NETFLIX_EXP_DETECTION
8765                 if (rack->rc_in_persist ||
8766                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8767                     (pkt_delta < tcp_sad_low_pps)){
8768                         /*
8769                          * We don't decay idle connections
8770                          * or ones that have a low input pps.
8771                          */
8772                         return;
8773                 }
8774                 /* Decay the counters */
8775                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8776                                                         tcp_sad_decay_val);
8777                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8778                                                          tcp_sad_decay_val);
8779                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8780                                                                tcp_sad_decay_val);
8781                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8782                                                                 tcp_sad_decay_val);
8783 #endif
8784         }
8785 }
8786
8787 static void
8788 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8789 {
8790         struct rack_sendmap *rsm;
8791 #ifdef INVARIANTS
8792         struct rack_sendmap *rm;
8793 #endif
8794
8795         /*
8796          * The ACK point is advancing to th_ack, we must drop off
8797          * the packets in the rack log and calculate any eligble
8798          * RTT's.
8799          */
8800         rack->r_wanted_output = 1;
8801
8802         /* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
8803         if ((rack->rc_last_tlp_acked_set == 1)&&
8804             (rack->rc_last_tlp_past_cumack == 1) &&
8805             (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
8806                 /*
8807                  * We have reached the point where our last rack
8808                  * tlp retransmit sequence is ahead of the cum-ack.
8809                  * This can only happen when the cum-ack moves all
8810                  * the way around (its been a full 2^^31+1 bytes
8811                  * or more since we sent a retransmitted TLP). Lets
8812                  * turn off the valid flag since its not really valid.
8813                  *
8814                  * Note since sack's also turn on this event we have
8815                  * a complication, we have to wait to age it out until
8816                  * the cum-ack is by the TLP before checking which is
8817                  * what the next else clause does.
8818                  */
8819                 rack_log_dsack_event(rack, 9, __LINE__,
8820                                      rack->r_ctl.last_tlp_acked_start,
8821                                      rack->r_ctl.last_tlp_acked_end);
8822                 rack->rc_last_tlp_acked_set = 0;
8823                 rack->rc_last_tlp_past_cumack = 0;
8824         } else if ((rack->rc_last_tlp_acked_set == 1) &&
8825                    (rack->rc_last_tlp_past_cumack == 0) &&
8826                    (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
8827                 /*
8828                  * It is safe to start aging TLP's out.
8829                  */
8830                 rack->rc_last_tlp_past_cumack = 1;
8831         }
8832         /* We do the same for the tlp send seq as well */
8833         if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8834             (rack->rc_last_sent_tlp_past_cumack == 1) &&
8835             (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
8836                 rack_log_dsack_event(rack, 9, __LINE__,
8837                                      rack->r_ctl.last_sent_tlp_seq,
8838                                      (rack->r_ctl.last_sent_tlp_seq +
8839                                       rack->r_ctl.last_sent_tlp_len));
8840                 rack->rc_last_sent_tlp_seq_valid = 0;
8841                 rack->rc_last_sent_tlp_past_cumack = 0;
8842         } else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8843                    (rack->rc_last_sent_tlp_past_cumack == 0) &&
8844                    (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
8845                 /*
8846                  * It is safe to start aging TLP's send.
8847                  */
8848                 rack->rc_last_sent_tlp_past_cumack = 1;
8849         }
8850 more:
8851         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8852         if (rsm == NULL) {
8853                 if ((th_ack - 1) == tp->iss) {
8854                         /*
8855                          * For the SYN incoming case we will not
8856                          * have called tcp_output for the sending of
8857                          * the SYN, so there will be no map. All
8858                          * other cases should probably be a panic.
8859                          */
8860                         return;
8861                 }
8862                 if (tp->t_flags & TF_SENTFIN) {
8863                         /* if we sent a FIN we often will not have map */
8864                         return;
8865                 }
8866 #ifdef INVARIANTS
8867                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8868                       tp,
8869                       tp->t_state, th_ack, rack,
8870                       tp->snd_una, tp->snd_max, tp->snd_nxt);
8871 #endif
8872                 return;
8873         }
8874         if (SEQ_LT(th_ack, rsm->r_start)) {
8875                 /* Huh map is missing this */
8876 #ifdef INVARIANTS
8877                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8878                        rsm->r_start,
8879                        th_ack, tp->t_state, rack->r_state);
8880 #endif
8881                 return;
8882         }
8883         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8884
8885         /* Now was it a retransmitted TLP? */
8886         if ((rsm->r_flags & RACK_TLP) &&
8887             (rsm->r_rtr_cnt > 1)) {
8888                 /*
8889                  * Yes, this rsm was a TLP and retransmitted, remember that
8890                  * since if a DSACK comes back on this we don't want
8891                  * to think of it as a reordered segment. This may
8892                  * get updated again with possibly even other TLPs
8893                  * in flight, but thats ok. Only when we don't send
8894                  * a retransmitted TLP for 1/2 the sequences space
8895                  * will it get turned off (above).
8896                  */
8897                 if (rack->rc_last_tlp_acked_set &&
8898                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8899                         /*
8900                          * We already turned this on since the end matches,
8901                          * the previous one was a partially ack now we
8902                          * are getting another one (maybe all of it).
8903                          */
8904                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8905                         /*
8906                          * Lets make sure we have all of it though.
8907                          */
8908                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8909                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8910                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8911                                                      rack->r_ctl.last_tlp_acked_end);
8912                         }
8913                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8914                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8915                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8916                                                      rack->r_ctl.last_tlp_acked_end);
8917                         }
8918                 } else {
8919                         rack->rc_last_tlp_past_cumack = 1;
8920                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8921                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8922                         rack->rc_last_tlp_acked_set = 1;
8923                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8924                 }
8925         }
8926         /* Now do we consume the whole thing? */
8927         if (SEQ_GEQ(th_ack, rsm->r_end)) {
8928                 /* Its all consumed. */
8929                 uint32_t left;
8930                 uint8_t newly_acked;
8931
8932                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8933                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8934                 rsm->r_rtr_bytes = 0;
8935                 /* Record the time of highest cumack sent */
8936                 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8937 #ifndef INVARIANTS
8938                 (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8939 #else
8940                 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8941                 if (rm != rsm) {
8942                         panic("removing head in rack:%p rsm:%p rm:%p",
8943                               rack, rsm, rm);
8944                 }
8945 #endif
8946                 if (rsm->r_in_tmap) {
8947                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8948                         rsm->r_in_tmap = 0;
8949                 }
8950                 newly_acked = 1;
8951                 if (rsm->r_flags & RACK_ACKED) {
8952                         /*
8953                          * It was acked on the scoreboard -- remove
8954                          * it from total
8955                          */
8956                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8957                         newly_acked = 0;
8958                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
8959                         /*
8960                          * There are segments ACKED on the
8961                          * scoreboard further up. We are seeing
8962                          * reordering.
8963                          */
8964                         rsm->r_flags &= ~RACK_SACK_PASSED;
8965                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8966                         rsm->r_flags |= RACK_ACKED;
8967                         rack->r_ctl.rc_reorder_ts = cts;
8968                         if (rack->r_ent_rec_ns) {
8969                                 /*
8970                                  * We have sent no more, and we saw an sack
8971                                  * then ack arrive.
8972                                  */
8973                                 rack->r_might_revert = 1;
8974                         }
8975                 }
8976                 if ((rsm->r_flags & RACK_TO_REXT) &&
8977                     (tp->t_flags & TF_RCVD_TSTMP) &&
8978                     (to->to_flags & TOF_TS) &&
8979                     (to->to_tsecr != 0) &&
8980                     (tp->t_flags & TF_PREVVALID)) {
8981                         /*
8982                          * We can use the timestamp to see
8983                          * if this retransmission was from the
8984                          * first transmit. If so we made a mistake.
8985                          */
8986                         tp->t_flags &= ~TF_PREVVALID;
8987                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
8988                                 /* The first transmit is what this ack is for */
8989                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
8990                         }
8991                 }
8992                 left = th_ack - rsm->r_end;
8993                 if (rack->app_limited_needs_set && newly_acked)
8994                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
8995                 /* Free back to zone */
8996                 rack_free(rack, rsm);
8997                 if (left) {
8998                         goto more;
8999                 }
9000                 /* Check for reneging */
9001                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9002                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9003                         /*
9004                          * The peer has moved snd_una up to
9005                          * the edge of this send, i.e. one
9006                          * that it had previously acked. The only
9007                          * way that can be true if the peer threw
9008                          * away data (space issues) that it had
9009                          * previously sacked (else it would have
9010                          * given us snd_una up to (rsm->r_end).
9011                          * We need to undo the acked markings here.
9012                          *
9013                          * Note we have to look to make sure th_ack is
9014                          * our rsm->r_start in case we get an old ack
9015                          * where th_ack is behind snd_una.
9016                          */
9017                         rack_peer_reneges(rack, rsm, th_ack);
9018                 }
9019                 return;
9020         }
9021         if (rsm->r_flags & RACK_ACKED) {
9022                 /*
9023                  * It was acked on the scoreboard -- remove it from
9024                  * total for the part being cum-acked.
9025                  */
9026                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9027         }
9028         /*
9029          * Clear the dup ack count for
9030          * the piece that remains.
9031          */
9032         rsm->r_dupack = 0;
9033         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9034         if (rsm->r_rtr_bytes) {
9035                 /*
9036                  * It was retransmitted adjust the
9037                  * sack holes for what was acked.
9038                  */
9039                 int ack_am;
9040
9041                 ack_am = (th_ack - rsm->r_start);
9042                 if (ack_am >= rsm->r_rtr_bytes) {
9043                         rack->r_ctl.rc_holes_rxt -= ack_am;
9044                         rsm->r_rtr_bytes -= ack_am;
9045                 }
9046         }
9047         /*
9048          * Update where the piece starts and record
9049          * the time of send of highest cumack sent.
9050          */
9051         rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9052         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9053         /* Now we need to move our offset forward too */
9054         if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9055                 /* Fix up the orig_m_len and possibly the mbuf offset */
9056                 rack_adjust_orig_mlen(rsm);
9057         }
9058         rsm->soff += (th_ack - rsm->r_start);
9059         rsm->r_start = th_ack;
9060         /* Now do we need to move the mbuf fwd too? */
9061         if (rsm->m) {
9062                 while (rsm->soff >= rsm->m->m_len) {
9063                         rsm->soff -= rsm->m->m_len;
9064                         rsm->m = rsm->m->m_next;
9065                         KASSERT((rsm->m != NULL),
9066                                 (" nrsm:%p hit at soff:%u null m",
9067                                  rsm, rsm->soff));
9068                 }
9069                 rsm->orig_m_len = rsm->m->m_len;
9070         }
9071         if (rack->app_limited_needs_set)
9072                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9073 }
9074
9075 static void
9076 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9077 {
9078         struct rack_sendmap *rsm;
9079         int sack_pass_fnd = 0;
9080
9081         if (rack->r_might_revert) {
9082                 /*
9083                  * Ok we have reordering, have not sent anything, we
9084                  * might want to revert the congestion state if nothing
9085                  * further has SACK_PASSED on it. Lets check.
9086                  *
9087                  * We also get here when we have DSACKs come in for
9088                  * all the data that we FR'd. Note that a rxt or tlp
9089                  * timer clears this from happening.
9090                  */
9091
9092                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9093                         if (rsm->r_flags & RACK_SACK_PASSED) {
9094                                 sack_pass_fnd = 1;
9095                                 break;
9096                         }
9097                 }
9098                 if (sack_pass_fnd == 0) {
9099                         /*
9100                          * We went into recovery
9101                          * incorrectly due to reordering!
9102                          */
9103                         int orig_cwnd;
9104
9105                         rack->r_ent_rec_ns = 0;
9106                         orig_cwnd = tp->snd_cwnd;
9107                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9108                         tp->snd_recover = tp->snd_una;
9109                         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
9110                         EXIT_RECOVERY(tp->t_flags);
9111                 }
9112                 rack->r_might_revert = 0;
9113         }
9114 }
9115
9116 #ifdef NETFLIX_EXP_DETECTION
9117 static void
9118 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9119 {
9120         if ((rack->do_detection || tcp_force_detection) &&
9121             tcp_sack_to_ack_thresh &&
9122             tcp_sack_to_move_thresh &&
9123             ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9124                 /*
9125                  * We have thresholds set to find
9126                  * possible attackers and disable sack.
9127                  * Check them.
9128                  */
9129                 uint64_t ackratio, moveratio, movetotal;
9130
9131                 /* Log detecting */
9132                 rack_log_sad(rack, 1);
9133                 ackratio = (uint64_t)(rack->r_ctl.sack_count);
9134                 ackratio *= (uint64_t)(1000);
9135                 if (rack->r_ctl.ack_count)
9136                         ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9137                 else {
9138                         /* We really should not hit here */
9139                         ackratio = 1000;
9140                 }
9141                 if ((rack->sack_attack_disable == 0) &&
9142                     (ackratio > rack_highest_sack_thresh_seen))
9143                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9144                 movetotal = rack->r_ctl.sack_moved_extra;
9145                 movetotal += rack->r_ctl.sack_noextra_move;
9146                 moveratio = rack->r_ctl.sack_moved_extra;
9147                 moveratio *= (uint64_t)1000;
9148                 if (movetotal)
9149                         moveratio /= movetotal;
9150                 else {
9151                         /* No moves, thats pretty good */
9152                         moveratio = 0;
9153                 }
9154                 if ((rack->sack_attack_disable == 0) &&
9155                     (moveratio > rack_highest_move_thresh_seen))
9156                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
9157                 if (rack->sack_attack_disable == 0) {
9158                         if ((ackratio > tcp_sack_to_ack_thresh) &&
9159                             (moveratio > tcp_sack_to_move_thresh)) {
9160                                 /* Disable sack processing */
9161                                 rack->sack_attack_disable = 1;
9162                                 if (rack->r_rep_attack == 0) {
9163                                         rack->r_rep_attack = 1;
9164                                         counter_u64_add(rack_sack_attacks_detected, 1);
9165                                 }
9166                                 tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
9167                                 /* Clamp the cwnd at flight size */
9168                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9169                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9170                                 rack_log_sad(rack, 2);
9171                         }
9172                 } else {
9173                         /* We are sack-disabled check for false positives */
9174                         if ((ackratio <= tcp_restoral_thresh) ||
9175                             (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9176                                 rack->sack_attack_disable = 0;
9177                                 rack_log_sad(rack, 3);
9178                                 /* Restart counting */
9179                                 rack->r_ctl.sack_count = 0;
9180                                 rack->r_ctl.sack_moved_extra = 0;
9181                                 rack->r_ctl.sack_noextra_move = 1;
9182                                 rack->r_ctl.ack_count = max(1,
9183                                       (bytes_this_ack / segsiz));
9184
9185                                 if (rack->r_rep_reverse == 0) {
9186                                         rack->r_rep_reverse = 1;
9187                                         counter_u64_add(rack_sack_attacks_reversed, 1);
9188                                 }
9189                                 /* Restore the cwnd */
9190                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9191                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9192                         }
9193                 }
9194         }
9195 }
9196 #endif
9197
9198 static int
9199 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9200 {
9201
9202         uint32_t am, l_end;
9203         int was_tlp = 0;
9204
9205         if (SEQ_GT(end, start))
9206                 am = end - start;
9207         else
9208                 am = 0;
9209         if ((rack->rc_last_tlp_acked_set ) &&
9210             (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9211             (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9212                 /*
9213                  * The DSACK is because of a TLP which we don't
9214                  * do anything with the reordering window over since
9215                  * it was not reordering that caused the DSACK but
9216                  * our previous retransmit TLP.
9217                  */
9218                 rack_log_dsack_event(rack, 7, __LINE__, start, end);
9219                 was_tlp = 1;
9220                 goto skip_dsack_round;
9221         }
9222         if (rack->rc_last_sent_tlp_seq_valid) {
9223                 l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9224                 if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9225                     (SEQ_LEQ(end, l_end))) {
9226                         /*
9227                          * This dsack is from the last sent TLP, ignore it
9228                          * for reordering purposes.
9229                          */
9230                         rack_log_dsack_event(rack, 7, __LINE__, start, end);
9231                         was_tlp = 1;
9232                         goto skip_dsack_round;
9233                 }
9234         }
9235         if (rack->rc_dsack_round_seen == 0) {
9236                 rack->rc_dsack_round_seen = 1;
9237                 rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9238                 rack->r_ctl.num_dsack++;
9239                 rack->r_ctl.dsack_persist = 16; /* 16 is from the standard */
9240                 rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9241         }
9242 skip_dsack_round:
9243         /*
9244          * We keep track of how many DSACK blocks we get
9245          * after a recovery incident.
9246          */
9247         rack->r_ctl.dsack_byte_cnt += am;
9248         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9249             rack->r_ctl.retran_during_recovery &&
9250             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9251                 /*
9252                  * False recovery most likely culprit is reordering. If
9253                  * nothing else is missing we need to revert.
9254                  */
9255                 rack->r_might_revert = 1;
9256                 rack_handle_might_revert(rack->rc_tp, rack);
9257                 rack->r_might_revert = 0;
9258                 rack->r_ctl.retran_during_recovery = 0;
9259                 rack->r_ctl.dsack_byte_cnt = 0;
9260         }
9261         return (was_tlp);
9262 }
9263
9264 static uint32_t
9265 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
9266 {
9267         return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
9268 }
9269
9270 static int32_t
9271 rack_compute_pipe(struct tcpcb *tp)
9272 {
9273         return ((int32_t)do_rack_compute_pipe(tp,
9274                                               (struct tcp_rack *)tp->t_fb_ptr,
9275                                               tp->snd_una));
9276 }
9277
9278 static void
9279 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9280 {
9281         /* Deal with changed and PRR here (in recovery only) */
9282         uint32_t pipe, snd_una;
9283
9284         rack->r_ctl.rc_prr_delivered += changed;
9285
9286         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9287                 /*
9288                  * It is all outstanding, we are application limited
9289                  * and thus we don't need more room to send anything.
9290                  * Note we use tp->snd_una here and not th_ack because
9291                  * the data as yet not been cut from the sb.
9292                  */
9293                 rack->r_ctl.rc_prr_sndcnt = 0;
9294                 return;
9295         }
9296         /* Compute prr_sndcnt */
9297         if (SEQ_GT(tp->snd_una, th_ack)) {
9298                 snd_una = tp->snd_una;
9299         } else {
9300                 snd_una = th_ack;
9301         }
9302         pipe = do_rack_compute_pipe(tp, rack, snd_una);
9303         if (pipe > tp->snd_ssthresh) {
9304                 long sndcnt;
9305
9306                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9307                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9308                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9309                 else {
9310                         rack->r_ctl.rc_prr_sndcnt = 0;
9311                         rack_log_to_prr(rack, 9, 0, __LINE__);
9312                         sndcnt = 0;
9313                 }
9314                 sndcnt++;
9315                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9316                         sndcnt -= rack->r_ctl.rc_prr_out;
9317                 else
9318                         sndcnt = 0;
9319                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9320                 rack_log_to_prr(rack, 10, 0, __LINE__);
9321         } else {
9322                 uint32_t limit;
9323
9324                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9325                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9326                 else
9327                         limit = 0;
9328                 if (changed > limit)
9329                         limit = changed;
9330                 limit += ctf_fixed_maxseg(tp);
9331                 if (tp->snd_ssthresh > pipe) {
9332                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9333                         rack_log_to_prr(rack, 11, 0, __LINE__);
9334                 } else {
9335                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9336                         rack_log_to_prr(rack, 12, 0, __LINE__);
9337                 }
9338         }
9339 }
9340
9341 static void
9342 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9343 {
9344         uint32_t changed;
9345         struct tcp_rack *rack;
9346         struct rack_sendmap *rsm;
9347         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9348         register uint32_t th_ack;
9349         int32_t i, j, k, num_sack_blks = 0;
9350         uint32_t cts, acked, ack_point;
9351         int loop_start = 0, moved_two = 0;
9352         uint32_t tsused;
9353
9354
9355         INP_WLOCK_ASSERT(tptoinpcb(tp));
9356         if (tcp_get_flags(th) & TH_RST) {
9357                 /* We don't log resets */
9358                 return;
9359         }
9360         rack = (struct tcp_rack *)tp->t_fb_ptr;
9361         cts = tcp_get_usecs(NULL);
9362         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9363         changed = 0;
9364         th_ack = th->th_ack;
9365         if (rack->sack_attack_disable == 0)
9366                 rack_do_decay(rack);
9367         if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9368                 /*
9369                  * You only get credit for
9370                  * MSS and greater (and you get extra
9371                  * credit for larger cum-ack moves).
9372                  */
9373                 int ac;
9374
9375                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9376                 rack->r_ctl.ack_count += ac;
9377                 counter_u64_add(rack_ack_total, ac);
9378         }
9379         if (rack->r_ctl.ack_count > 0xfff00000) {
9380                 /*
9381                  * reduce the number to keep us under
9382                  * a uint32_t.
9383                  */
9384                 rack->r_ctl.ack_count /= 2;
9385                 rack->r_ctl.sack_count /= 2;
9386         }
9387         if (SEQ_GT(th_ack, tp->snd_una)) {
9388                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9389                 tp->t_acktime = ticks;
9390         }
9391         if (rsm && SEQ_GT(th_ack, rsm->r_start))
9392                 changed = th_ack - rsm->r_start;
9393         if (changed) {
9394                 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9395         }
9396         if ((to->to_flags & TOF_SACK) == 0) {
9397                 /* We are done nothing left and no sack. */
9398                 rack_handle_might_revert(tp, rack);
9399                 /*
9400                  * For cases where we struck a dup-ack
9401                  * with no SACK, add to the changes so
9402                  * PRR will work right.
9403                  */
9404                 if (dup_ack_struck && (changed == 0)) {
9405                         changed += ctf_fixed_maxseg(rack->rc_tp);
9406                 }
9407                 goto out;
9408         }
9409         /* Sack block processing */
9410         if (SEQ_GT(th_ack, tp->snd_una))
9411                 ack_point = th_ack;
9412         else
9413                 ack_point = tp->snd_una;
9414         for (i = 0; i < to->to_nsacks; i++) {
9415                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9416                       &sack, sizeof(sack));
9417                 sack.start = ntohl(sack.start);
9418                 sack.end = ntohl(sack.end);
9419                 if (SEQ_GT(sack.end, sack.start) &&
9420                     SEQ_GT(sack.start, ack_point) &&
9421                     SEQ_LT(sack.start, tp->snd_max) &&
9422                     SEQ_GT(sack.end, ack_point) &&
9423                     SEQ_LEQ(sack.end, tp->snd_max)) {
9424                         sack_blocks[num_sack_blks] = sack;
9425                         num_sack_blks++;
9426                 } else if (SEQ_LEQ(sack.start, th_ack) &&
9427                            SEQ_LEQ(sack.end, th_ack)) {
9428                         int was_tlp;
9429
9430                         was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9431                         /*
9432                          * Its a D-SACK block.
9433                          */
9434                         tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9435                 }
9436         }
9437         if (rack->rc_dsack_round_seen) {
9438                 /* Is the dsack roound over? */
9439                 if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9440                         /* Yes it is */
9441                         rack->rc_dsack_round_seen = 0;
9442                         rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9443                 }
9444         }
9445         /*
9446          * Sort the SACK blocks so we can update the rack scoreboard with
9447          * just one pass.
9448          */
9449         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9450                                          num_sack_blks, th->th_ack);
9451         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9452         if (num_sack_blks == 0) {
9453                 /* Nothing to sack (DSACKs?) */
9454                 goto out_with_totals;
9455         }
9456         if (num_sack_blks < 2) {
9457                 /* Only one, we don't need to sort */
9458                 goto do_sack_work;
9459         }
9460         /* Sort the sacks */
9461         for (i = 0; i < num_sack_blks; i++) {
9462                 for (j = i + 1; j < num_sack_blks; j++) {
9463                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9464                                 sack = sack_blocks[i];
9465                                 sack_blocks[i] = sack_blocks[j];
9466                                 sack_blocks[j] = sack;
9467                         }
9468                 }
9469         }
9470         /*
9471          * Now are any of the sack block ends the same (yes some
9472          * implementations send these)?
9473          */
9474 again:
9475         if (num_sack_blks == 0)
9476                 goto out_with_totals;
9477         if (num_sack_blks > 1) {
9478                 for (i = 0; i < num_sack_blks; i++) {
9479                         for (j = i + 1; j < num_sack_blks; j++) {
9480                                 if (sack_blocks[i].end == sack_blocks[j].end) {
9481                                         /*
9482                                          * Ok these two have the same end we
9483                                          * want the smallest end and then
9484                                          * throw away the larger and start
9485                                          * again.
9486                                          */
9487                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9488                                                 /*
9489                                                  * The second block covers
9490                                                  * more area use that
9491                                                  */
9492                                                 sack_blocks[i].start = sack_blocks[j].start;
9493                                         }
9494                                         /*
9495                                          * Now collapse out the dup-sack and
9496                                          * lower the count
9497                                          */
9498                                         for (k = (j + 1); k < num_sack_blks; k++) {
9499                                                 sack_blocks[j].start = sack_blocks[k].start;
9500                                                 sack_blocks[j].end = sack_blocks[k].end;
9501                                                 j++;
9502                                         }
9503                                         num_sack_blks--;
9504                                         goto again;
9505                                 }
9506                         }
9507                 }
9508         }
9509 do_sack_work:
9510         /*
9511          * First lets look to see if
9512          * we have retransmitted and
9513          * can use the transmit next?
9514          */
9515         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9516         if (rsm &&
9517             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9518             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9519                 /*
9520                  * We probably did the FR and the next
9521                  * SACK in continues as we would expect.
9522                  */
9523                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9524                 if (acked) {
9525                         rack->r_wanted_output = 1;
9526                         changed += acked;
9527                 }
9528                 if (num_sack_blks == 1) {
9529                         /*
9530                          * This is what we would expect from
9531                          * a normal implementation to happen
9532                          * after we have retransmitted the FR,
9533                          * i.e the sack-filter pushes down
9534                          * to 1 block and the next to be retransmitted
9535                          * is the sequence in the sack block (has more
9536                          * are acked). Count this as ACK'd data to boost
9537                          * up the chances of recovering any false positives.
9538                          */
9539                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9540                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9541                         counter_u64_add(rack_express_sack, 1);
9542                         if (rack->r_ctl.ack_count > 0xfff00000) {
9543                                 /*
9544                                  * reduce the number to keep us under
9545                                  * a uint32_t.
9546                                  */
9547                                 rack->r_ctl.ack_count /= 2;
9548                                 rack->r_ctl.sack_count /= 2;
9549                         }
9550                         goto out_with_totals;
9551                 } else {
9552                         /*
9553                          * Start the loop through the
9554                          * rest of blocks, past the first block.
9555                          */
9556                         moved_two = 0;
9557                         loop_start = 1;
9558                 }
9559         }
9560         /* Its a sack of some sort */
9561         rack->r_ctl.sack_count++;
9562         if (rack->r_ctl.sack_count > 0xfff00000) {
9563                 /*
9564                  * reduce the number to keep us under
9565                  * a uint32_t.
9566                  */
9567                 rack->r_ctl.ack_count /= 2;
9568                 rack->r_ctl.sack_count /= 2;
9569         }
9570         counter_u64_add(rack_sack_total, 1);
9571         if (rack->sack_attack_disable) {
9572                 /* An attacker disablement is in place */
9573                 if (num_sack_blks > 1) {
9574                         rack->r_ctl.sack_count += (num_sack_blks - 1);
9575                         rack->r_ctl.sack_moved_extra++;
9576                         counter_u64_add(rack_move_some, 1);
9577                         if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9578                                 rack->r_ctl.sack_moved_extra /= 2;
9579                                 rack->r_ctl.sack_noextra_move /= 2;
9580                         }
9581                 }
9582                 goto out;
9583         }
9584         rsm = rack->r_ctl.rc_sacklast;
9585         for (i = loop_start; i < num_sack_blks; i++) {
9586                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9587                 if (acked) {
9588                         rack->r_wanted_output = 1;
9589                         changed += acked;
9590                 }
9591                 if (moved_two) {
9592                         /*
9593                          * If we did not get a SACK for at least a MSS and
9594                          * had to move at all, or if we moved more than our
9595                          * threshold, it counts against the "extra" move.
9596                          */
9597                         rack->r_ctl.sack_moved_extra += moved_two;
9598                         counter_u64_add(rack_move_some, 1);
9599                 } else {
9600                         /*
9601                          * else we did not have to move
9602                          * any more than we would expect.
9603                          */
9604                         rack->r_ctl.sack_noextra_move++;
9605                         counter_u64_add(rack_move_none, 1);
9606                 }
9607                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9608                         /*
9609                          * If the SACK was not a full MSS then
9610                          * we add to sack_count the number of
9611                          * MSS's (or possibly more than
9612                          * a MSS if its a TSO send) we had to skip by.
9613                          */
9614                         rack->r_ctl.sack_count += moved_two;
9615                         counter_u64_add(rack_sack_total, moved_two);
9616                 }
9617                 /*
9618                  * Now we need to setup for the next
9619                  * round. First we make sure we won't
9620                  * exceed the size of our uint32_t on
9621                  * the various counts, and then clear out
9622                  * moved_two.
9623                  */
9624                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9625                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9626                         rack->r_ctl.sack_moved_extra /= 2;
9627                         rack->r_ctl.sack_noextra_move /= 2;
9628                 }
9629                 if (rack->r_ctl.sack_count > 0xfff00000) {
9630                         rack->r_ctl.ack_count /= 2;
9631                         rack->r_ctl.sack_count /= 2;
9632                 }
9633                 moved_two = 0;
9634         }
9635 out_with_totals:
9636         if (num_sack_blks > 1) {
9637                 /*
9638                  * You get an extra stroke if
9639                  * you have more than one sack-blk, this
9640                  * could be where we are skipping forward
9641                  * and the sack-filter is still working, or
9642                  * it could be an attacker constantly
9643                  * moving us.
9644                  */
9645                 rack->r_ctl.sack_moved_extra++;
9646                 counter_u64_add(rack_move_some, 1);
9647         }
9648 out:
9649 #ifdef NETFLIX_EXP_DETECTION
9650         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9651 #endif
9652         if (changed) {
9653                 /* Something changed cancel the rack timer */
9654                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9655         }
9656         tsused = tcp_get_usecs(NULL);
9657         rsm = tcp_rack_output(tp, rack, tsused);
9658         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9659             rsm &&
9660             ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
9661                 /* Enter recovery */
9662                 entered_recovery = 1;
9663                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9664                 /*
9665                  * When we enter recovery we need to assure we send
9666                  * one packet.
9667                  */
9668                 if (rack->rack_no_prr == 0) {
9669                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9670                         rack_log_to_prr(rack, 8, 0, __LINE__);
9671                 }
9672                 rack->r_timer_override = 1;
9673                 rack->r_early = 0;
9674                 rack->r_ctl.rc_agg_early = 0;
9675         } else if (IN_FASTRECOVERY(tp->t_flags) &&
9676                    rsm &&
9677                    (rack->r_rr_config == 3)) {
9678                 /*
9679                  * Assure we can output and we get no
9680                  * remembered pace time except the retransmit.
9681                  */
9682                 rack->r_timer_override = 1;
9683                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9684                 rack->r_ctl.rc_resend = rsm;
9685         }
9686         if (IN_FASTRECOVERY(tp->t_flags) &&
9687             (rack->rack_no_prr == 0) &&
9688             (entered_recovery == 0)) {
9689                 rack_update_prr(tp, rack, changed, th_ack);
9690                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9691                      ((tcp_in_hpts(rack->rc_inp) == 0) &&
9692                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9693                         /*
9694                          * If you are pacing output you don't want
9695                          * to override.
9696                          */
9697                         rack->r_early = 0;
9698                         rack->r_ctl.rc_agg_early = 0;
9699                         rack->r_timer_override = 1;
9700                 }
9701         }
9702 }
9703
9704 static void
9705 rack_strike_dupack(struct tcp_rack *rack)
9706 {
9707         struct rack_sendmap *rsm;
9708
9709         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9710         while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9711                 rsm = TAILQ_NEXT(rsm, r_tnext);
9712                 if (rsm->r_flags & RACK_MUST_RXT) {
9713                         /* Sendmap entries that are marked to
9714                          * be retransmitted do not need dupack's
9715                          * struck. We get these marks for a number
9716                          * of reasons (rxt timeout with no sack,
9717                          * mtu change, or rwnd collapses). When
9718                          * these events occur, we know we must retransmit
9719                          * them and mark the sendmap entries. Dupack counting
9720                          * is not needed since we are already set to retransmit
9721                          * it as soon as we can.
9722                          */
9723                         continue;
9724                 }
9725         }
9726         if (rsm && (rsm->r_dupack < 0xff)) {
9727                 rsm->r_dupack++;
9728                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9729                         struct timeval tv;
9730                         uint32_t cts;
9731                         /*
9732                          * Here we see if we need to retransmit. For
9733                          * a SACK type connection if enough time has passed
9734                          * we will get a return of the rsm. For a non-sack
9735                          * connection we will get the rsm returned if the
9736                          * dupack value is 3 or more.
9737                          */
9738                         cts = tcp_get_usecs(&tv);
9739                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9740                         if (rack->r_ctl.rc_resend != NULL) {
9741                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9742                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9743                                                          rack->rc_tp->snd_una, __LINE__);
9744                                 }
9745                                 rack->r_wanted_output = 1;
9746                                 rack->r_timer_override = 1;
9747                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9748                         }
9749                 } else {
9750                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9751                 }
9752         }
9753 }
9754
9755 static void
9756 rack_check_bottom_drag(struct tcpcb *tp,
9757                        struct tcp_rack *rack,
9758                        struct socket *so, int32_t acked)
9759 {
9760         uint32_t segsiz, minseg;
9761
9762         segsiz = ctf_fixed_maxseg(tp);
9763         minseg = segsiz;
9764
9765         if (tp->snd_max == tp->snd_una) {
9766                 /*
9767                  * We are doing dynamic pacing and we are way
9768                  * under. Basically everything got acked while
9769                  * we were still waiting on the pacer to expire.
9770                  *
9771                  * This means we need to boost the b/w in
9772                  * addition to any earlier boosting of
9773                  * the multiplier.
9774                  */
9775                 rack->rc_dragged_bottom = 1;
9776                 rack_validate_multipliers_at_or_above100(rack);
9777                 /*
9778                  * Lets use the segment bytes acked plus
9779                  * the lowest RTT seen as the basis to
9780                  * form a b/w estimate. This will be off
9781                  * due to the fact that the true estimate
9782                  * should be around 1/2 the time of the RTT
9783                  * but we can settle for that.
9784                  */
9785                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9786                     acked) {
9787                         uint64_t bw, calc_bw, rtt;
9788
9789                         rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9790                         if (rtt == 0) {
9791                                 /* no us sample is there a ms one? */
9792                                 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9793                                         rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9794                                 } else {
9795                                         goto no_measurement;
9796                                 }
9797                         }
9798                         bw = acked;
9799                         calc_bw = bw * 1000000;
9800                         calc_bw /= rtt;
9801                         if (rack->r_ctl.last_max_bw &&
9802                             (rack->r_ctl.last_max_bw < calc_bw)) {
9803                                 /*
9804                                  * If we have a last calculated max bw
9805                                  * enforce it.
9806                                  */
9807                                 calc_bw = rack->r_ctl.last_max_bw;
9808                         }
9809                         /* now plop it in */
9810                         if (rack->rc_gp_filled == 0) {
9811                                 if (calc_bw > ONE_POINT_TWO_MEG) {
9812                                         /*
9813                                          * If we have no measurement
9814                                          * don't let us set in more than
9815                                          * 1.2Mbps. If we are still too
9816                                          * low after pacing with this we
9817                                          * will hopefully have a max b/w
9818                                          * available to sanity check things.
9819                                          */
9820                                         calc_bw = ONE_POINT_TWO_MEG;
9821                                 }
9822                                 rack->r_ctl.rc_rtt_diff = 0;
9823                                 rack->r_ctl.gp_bw = calc_bw;
9824                                 rack->rc_gp_filled = 1;
9825                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9826                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9827                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9828                         } else if (calc_bw > rack->r_ctl.gp_bw) {
9829                                 rack->r_ctl.rc_rtt_diff = 0;
9830                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9831                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9832                                 rack->r_ctl.gp_bw = calc_bw;
9833                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9834                         } else
9835                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9836                         if ((rack->gp_ready == 0) &&
9837                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9838                                 /* We have enough measurements now */
9839                                 rack->gp_ready = 1;
9840                                 rack_set_cc_pacing(rack);
9841                                 if (rack->defer_options)
9842                                         rack_apply_deferred_options(rack);
9843                         }
9844                         /*
9845                          * For acks over 1mss we do a extra boost to simulate
9846                          * where we would get 2 acks (we want 110 for the mul).
9847                          */
9848                         if (acked > segsiz)
9849                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9850                 } else {
9851                         /*
9852                          * zero rtt possibly?, settle for just an old increase.
9853                          */
9854 no_measurement:
9855                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
9856                 }
9857         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9858                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9859                                                minseg)) &&
9860                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9861                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9862                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9863                     (segsiz * rack_req_segs))) {
9864                 /*
9865                  * We are doing dynamic GP pacing and
9866                  * we have everything except 1MSS or less
9867                  * bytes left out. We are still pacing away.
9868                  * And there is data that could be sent, This
9869                  * means we are inserting delayed ack time in
9870                  * our measurements because we are pacing too slow.
9871                  */
9872                 rack_validate_multipliers_at_or_above100(rack);
9873                 rack->rc_dragged_bottom = 1;
9874                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9875         }
9876 }
9877
9878
9879
9880 static void
9881 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9882 {
9883         /*
9884          * The fast output path is enabled and we
9885          * have moved the cumack forward. Lets see if
9886          * we can expand forward the fast path length by
9887          * that amount. What we would ideally like to
9888          * do is increase the number of bytes in the
9889          * fast path block (left_to_send) by the
9890          * acked amount. However we have to gate that
9891          * by two factors:
9892          * 1) The amount outstanding and the rwnd of the peer
9893          *    (i.e. we don't want to exceed the rwnd of the peer).
9894          *    <and>
9895          * 2) The amount of data left in the socket buffer (i.e.
9896          *    we can't send beyond what is in the buffer).
9897          *
9898          * Note that this does not take into account any increase
9899          * in the cwnd. We will only extend the fast path by
9900          * what was acked.
9901          */
9902         uint32_t new_total, gating_val;
9903
9904         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9905         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9906                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9907         if (new_total <= gating_val) {
9908                 /* We can increase left_to_send by the acked amount */
9909                 counter_u64_add(rack_extended_rfo, 1);
9910                 rack->r_ctl.fsb.left_to_send = new_total;
9911                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9912                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
9913                          rack, rack->r_ctl.fsb.left_to_send,
9914                          sbavail(&rack->rc_inp->inp_socket->so_snd),
9915                          (tp->snd_max - tp->snd_una)));
9916
9917         }
9918 }
9919
9920 static void
9921 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9922 {
9923         /*
9924          * Here any sendmap entry that points to the
9925          * beginning mbuf must be adjusted to the correct
9926          * offset. This must be called with:
9927          * 1) The socket buffer locked
9928          * 2) snd_una adjusted to its new position.
9929          *
9930          * Note that (2) implies rack_ack_received has also
9931          * been called.
9932          *
9933          * We grab the first mbuf in the socket buffer and
9934          * then go through the front of the sendmap, recalculating
9935          * the stored offset for any sendmap entry that has
9936          * that mbuf. We must use the sb functions to do this
9937          * since its possible an add was done has well as
9938          * the subtraction we may have just completed. This should
9939          * not be a penalty though, since we just referenced the sb
9940          * to go in and trim off the mbufs that we freed (of course
9941          * there will be a penalty for the sendmap references though).
9942          */
9943         struct mbuf *m;
9944         struct rack_sendmap *rsm;
9945
9946         SOCKBUF_LOCK_ASSERT(sb);
9947         m = sb->sb_mb;
9948         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9949         if ((rsm == NULL) || (m == NULL)) {
9950                 /* Nothing outstanding */
9951                 return;
9952         }
9953         while (rsm->m && (rsm->m == m)) {
9954                 /* one to adjust */
9955 #ifdef INVARIANTS
9956                 struct mbuf *tm;
9957                 uint32_t soff;
9958
9959                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
9960                 if (rsm->orig_m_len != m->m_len) {
9961                         rack_adjust_orig_mlen(rsm);
9962                 }
9963                 if (rsm->soff != soff) {
9964                         /*
9965                          * This is not a fatal error, we anticipate it
9966                          * might happen (the else code), so we count it here
9967                          * so that under invariant we can see that it really
9968                          * does happen.
9969                          */
9970                         counter_u64_add(rack_adjust_map_bw, 1);
9971                 }
9972                 rsm->m = tm;
9973                 rsm->soff = soff;
9974                 if (tm)
9975                         rsm->orig_m_len = rsm->m->m_len;
9976                 else
9977                         rsm->orig_m_len = 0;
9978 #else
9979                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
9980                 if (rsm->m)
9981                         rsm->orig_m_len = rsm->m->m_len;
9982                 else
9983                         rsm->orig_m_len = 0;
9984 #endif
9985                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
9986                               rsm);
9987                 if (rsm == NULL)
9988                         break;
9989         }
9990 }
9991
9992 /*
9993  * Return value of 1, we do not need to call rack_process_data().
9994  * return value of 0, rack_process_data can be called.
9995  * For ret_val if its 0 the TCP is locked, if its non-zero
9996  * its unlocked and probably unsafe to touch the TCB.
9997  */
9998 static int
9999 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10000     struct tcpcb *tp, struct tcpopt *to,
10001     uint32_t tiwin, int32_t tlen,
10002     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10003 {
10004         int32_t ourfinisacked = 0;
10005         int32_t nsegs, acked_amount;
10006         int32_t acked;
10007         struct mbuf *mfree;
10008         struct tcp_rack *rack;
10009         int32_t under_pacing = 0;
10010         int32_t recovery = 0;
10011
10012         INP_WLOCK_ASSERT(tptoinpcb(tp));
10013
10014         rack = (struct tcp_rack *)tp->t_fb_ptr;
10015         if (SEQ_GT(th->th_ack, tp->snd_max)) {
10016                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10017                                       &rack->r_ctl.challenge_ack_ts,
10018                                       &rack->r_ctl.challenge_ack_cnt);
10019                 rack->r_wanted_output = 1;
10020                 return (1);
10021         }
10022         if (rack->gp_ready &&
10023             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10024                 under_pacing = 1;
10025         }
10026         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10027                 int in_rec, dup_ack_struck = 0;
10028
10029                 in_rec = IN_FASTRECOVERY(tp->t_flags);
10030                 if (rack->rc_in_persist) {
10031                         tp->t_rxtshift = 0;
10032                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10033                                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10034                 }
10035                 if ((th->th_ack == tp->snd_una) &&
10036                     (tiwin == tp->snd_wnd) &&
10037                     ((to->to_flags & TOF_SACK) == 0)) {
10038                         rack_strike_dupack(rack);
10039                         dup_ack_struck = 1;
10040                 }
10041                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10042         }
10043         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10044                 /*
10045                  * Old ack, behind (or duplicate to) the last one rcv'd
10046                  * Note: We mark reordering is occuring if its
10047                  * less than and we have not closed our window.
10048                  */
10049                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10050                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10051                 }
10052                 return (0);
10053         }
10054         /*
10055          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10056          * something we sent.
10057          */
10058         if (tp->t_flags & TF_NEEDSYN) {
10059                 /*
10060                  * T/TCP: Connection was half-synchronized, and our SYN has
10061                  * been ACK'd (so connection is now fully synchronized).  Go
10062                  * to non-starred state, increment snd_una for ACK of SYN,
10063                  * and check if we can do window scaling.
10064                  */
10065                 tp->t_flags &= ~TF_NEEDSYN;
10066                 tp->snd_una++;
10067                 /* Do window scaling? */
10068                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10069                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10070                         tp->rcv_scale = tp->request_r_scale;
10071                         /* Send window already scaled. */
10072                 }
10073         }
10074         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10075
10076         acked = BYTES_THIS_ACK(tp, th);
10077         if (acked) {
10078                 /*
10079                  * Any time we move the cum-ack forward clear
10080                  * keep-alive tied probe-not-answered. The
10081                  * persists clears its own on entry.
10082                  */
10083                 rack->probe_not_answered = 0;
10084         }
10085         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10086         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10087         /*
10088          * If we just performed our first retransmit, and the ACK arrives
10089          * within our recovery window, then it was a mistake to do the
10090          * retransmit in the first place.  Recover our original cwnd and
10091          * ssthresh, and proceed to transmit where we left off.
10092          */
10093         if ((tp->t_flags & TF_PREVVALID) &&
10094             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10095                 tp->t_flags &= ~TF_PREVVALID;
10096                 if (tp->t_rxtshift == 1 &&
10097                     (int)(ticks - tp->t_badrxtwin) < 0)
10098                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10099         }
10100         if (acked) {
10101                 /* assure we are not backed off */
10102                 tp->t_rxtshift = 0;
10103                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10104                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10105                 rack->rc_tlp_in_progress = 0;
10106                 rack->r_ctl.rc_tlp_cnt_out = 0;
10107                 /*
10108                  * If it is the RXT timer we want to
10109                  * stop it, so we can restart a TLP.
10110                  */
10111                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10112                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10113 #ifdef NETFLIX_HTTP_LOGGING
10114                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10115 #endif
10116         }
10117         /*
10118          * If we have a timestamp reply, update smoothed round trip time. If
10119          * no timestamp is present but transmit timer is running and timed
10120          * sequence number was acked, update smoothed round trip time. Since
10121          * we now have an rtt measurement, cancel the timer backoff (cf.,
10122          * Phil Karn's retransmit alg.). Recompute the initial retransmit
10123          * timer.
10124          *
10125          * Some boxes send broken timestamp replies during the SYN+ACK
10126          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10127          * and blow up the retransmit timer.
10128          */
10129         /*
10130          * If all outstanding data is acked, stop retransmit timer and
10131          * remember to restart (more output or persist). If there is more
10132          * data to be acked, restart retransmit timer, using current
10133          * (possibly backed-off) value.
10134          */
10135         if (acked == 0) {
10136                 if (ofia)
10137                         *ofia = ourfinisacked;
10138                 return (0);
10139         }
10140         if (IN_RECOVERY(tp->t_flags)) {
10141                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10142                     (SEQ_LT(th->th_ack, tp->snd_max))) {
10143                         tcp_rack_partialack(tp);
10144                 } else {
10145                         rack_post_recovery(tp, th->th_ack);
10146                         recovery = 1;
10147                 }
10148         }
10149         /*
10150          * Let the congestion control algorithm update congestion control
10151          * related information. This typically means increasing the
10152          * congestion window.
10153          */
10154         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10155         SOCKBUF_LOCK(&so->so_snd);
10156         acked_amount = min(acked, (int)sbavail(&so->so_snd));
10157         tp->snd_wnd -= acked_amount;
10158         mfree = sbcut_locked(&so->so_snd, acked_amount);
10159         if ((sbused(&so->so_snd) == 0) &&
10160             (acked > acked_amount) &&
10161             (tp->t_state >= TCPS_FIN_WAIT_1) &&
10162             (tp->t_flags & TF_SENTFIN)) {
10163                 /*
10164                  * We must be sure our fin
10165                  * was sent and acked (we can be
10166                  * in FIN_WAIT_1 without having
10167                  * sent the fin).
10168                  */
10169                 ourfinisacked = 1;
10170         }
10171         tp->snd_una = th->th_ack;
10172         if (acked_amount && sbavail(&so->so_snd))
10173                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10174         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10175         /* NB: sowwakeup_locked() does an implicit unlock. */
10176         sowwakeup_locked(so);
10177         m_freem(mfree);
10178         if (SEQ_GT(tp->snd_una, tp->snd_recover))
10179                 tp->snd_recover = tp->snd_una;
10180
10181         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10182                 tp->snd_nxt = tp->snd_una;
10183         }
10184         if (under_pacing &&
10185             (rack->use_fixed_rate == 0) &&
10186             (rack->in_probe_rtt == 0) &&
10187             rack->rc_gp_dyn_mul &&
10188             rack->rc_always_pace) {
10189                 /* Check if we are dragging bottom */
10190                 rack_check_bottom_drag(tp, rack, so, acked);
10191         }
10192         if (tp->snd_una == tp->snd_max) {
10193                 /* Nothing left outstanding */
10194                 tp->t_flags &= ~TF_PREVVALID;
10195                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10196                 rack->r_ctl.retran_during_recovery = 0;
10197                 rack->r_ctl.dsack_byte_cnt = 0;
10198                 if (rack->r_ctl.rc_went_idle_time == 0)
10199                         rack->r_ctl.rc_went_idle_time = 1;
10200                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10201                 if (sbavail(&tptosocket(tp)->so_snd) == 0)
10202                         tp->t_acktime = 0;
10203                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10204                 /* Set need output so persist might get set */
10205                 rack->r_wanted_output = 1;
10206                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10207                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10208                     (sbavail(&so->so_snd) == 0) &&
10209                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10210                         /*
10211                          * The socket was gone and the
10212                          * peer sent data (now or in the past), time to
10213                          * reset him.
10214                          */
10215                         *ret_val = 1;
10216                         /* tcp_close will kill the inp pre-log the Reset */
10217                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10218                         tp = tcp_close(tp);
10219                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10220                         return (1);
10221                 }
10222         }
10223         if (ofia)
10224                 *ofia = ourfinisacked;
10225         return (0);
10226 }
10227
10228
10229 static void
10230 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
10231                   int dir, uint32_t flags, struct rack_sendmap *rsm)
10232 {
10233         if (tcp_bblogging_on(rack->rc_tp)) {
10234                 union tcp_log_stackspecific log;
10235                 struct timeval tv;
10236
10237                 memset(&log, 0, sizeof(log));
10238                 log.u_bbr.flex1 = cnt;
10239                 log.u_bbr.flex2 = split;
10240                 log.u_bbr.flex3 = out;
10241                 log.u_bbr.flex4 = line;
10242                 log.u_bbr.flex5 = rack->r_must_retran;
10243                 log.u_bbr.flex6 = flags;
10244                 log.u_bbr.flex7 = rack->rc_has_collapsed;
10245                 log.u_bbr.flex8 = dir;  /*
10246                                          * 1 is collapsed, 0 is uncollapsed,
10247                                          * 2 is log of a rsm being marked, 3 is a split.
10248                                          */
10249                 if (rsm == NULL)
10250                         log.u_bbr.rttProp = 0;
10251                 else
10252                         log.u_bbr.rttProp = (uint64_t)rsm;
10253                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
10254                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10255                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
10256                     &rack->rc_inp->inp_socket->so_rcv,
10257                     &rack->rc_inp->inp_socket->so_snd,
10258                     TCP_RACK_LOG_COLLAPSE, 0,
10259                     0, &log, false, &tv);
10260         }
10261 }
10262
10263 static void
10264 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, int line)
10265 {
10266         /*
10267          * Here all we do is mark the collapsed point and set the flag.
10268          * This may happen again and again, but there is no
10269          * sense splitting our map until we know where the
10270          * peer finally lands in the collapse.
10271          */
10272         tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
10273         if ((rack->rc_has_collapsed == 0) ||
10274             (rack->r_ctl.last_collapse_point != (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)))
10275                 counter_u64_add(rack_collapsed_win_seen, 1);
10276         rack->r_ctl.last_collapse_point = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10277         rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
10278         rack->rc_has_collapsed = 1;
10279         rack->r_collapse_point_valid = 1;
10280         rack_log_collapse(rack, 0, 0, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
10281 }
10282
10283 static void
10284 rack_un_collapse_window(struct tcp_rack *rack, int line)
10285 {
10286         struct rack_sendmap *nrsm, *rsm, fe;
10287         int cnt = 0, split = 0;
10288 #ifdef INVARIANTS
10289         struct rack_sendmap *insret;
10290 #endif
10291
10292         memset(&fe, 0, sizeof(fe));
10293         rack->rc_has_collapsed = 0;
10294         fe.r_start = rack->r_ctl.last_collapse_point;
10295         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10296         if (rsm == NULL) {
10297                 /* Nothing to do maybe the peer ack'ed it all */
10298                 rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10299                 return;
10300         }
10301         /* Now do we need to split this one? */
10302         if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
10303                 rack_log_collapse(rack, rsm->r_start, rsm->r_end,
10304                                   rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
10305                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10306                 if (nrsm == NULL) {
10307                         /* We can't get a rsm, mark all? */
10308                         nrsm = rsm;
10309                         goto no_split;
10310                 }
10311                 /* Clone it */
10312                 split = 1;
10313                 rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
10314 #ifndef INVARIANTS
10315                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10316 #else
10317                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10318                 if (insret != NULL) {
10319                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10320                               nrsm, insret, rack, rsm);
10321                 }
10322 #endif
10323                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
10324                                  rack->r_ctl.last_collapse_point, __LINE__);
10325                 if (rsm->r_in_tmap) {
10326                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10327                         nrsm->r_in_tmap = 1;
10328                 }
10329                 /*
10330                  * Set in the new RSM as the
10331                  * collapsed starting point
10332                  */
10333                 rsm = nrsm;
10334         }
10335 no_split:
10336         RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10337                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
10338                 rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
10339                 cnt++;
10340         }
10341         if (cnt) {
10342                 counter_u64_add(rack_collapsed_win, 1);
10343         }
10344         rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10345 }
10346
10347 static void
10348 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10349                         int32_t tlen, int32_t tfo_syn)
10350 {
10351         if (DELAY_ACK(tp, tlen) || tfo_syn) {
10352                 if (rack->rc_dack_mode &&
10353                     (tlen > 500) &&
10354                     (rack->rc_dack_toggle == 1)) {
10355                         goto no_delayed_ack;
10356                 }
10357                 rack_timer_cancel(tp, rack,
10358                                   rack->r_ctl.rc_rcvtime, __LINE__);
10359                 tp->t_flags |= TF_DELACK;
10360         } else {
10361 no_delayed_ack:
10362                 rack->r_wanted_output = 1;
10363                 tp->t_flags |= TF_ACKNOW;
10364                 if (rack->rc_dack_mode) {
10365                         if (tp->t_flags & TF_DELACK)
10366                                 rack->rc_dack_toggle = 1;
10367                         else
10368                                 rack->rc_dack_toggle = 0;
10369                 }
10370         }
10371 }
10372
10373 static void
10374 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10375 {
10376         /*
10377          * If fast output is in progress, lets validate that
10378          * the new window did not shrink on us and make it
10379          * so fast output should end.
10380          */
10381         if (rack->r_fast_output) {
10382                 uint32_t out;
10383
10384                 /*
10385                  * Calculate what we will send if left as is
10386                  * and compare that to our send window.
10387                  */
10388                 out = ctf_outstanding(tp);
10389                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10390                         /* ok we have an issue */
10391                         if (out >= tp->snd_wnd) {
10392                                 /* Turn off fast output the window is met or collapsed */
10393                                 rack->r_fast_output = 0;
10394                         } else {
10395                                 /* we have some room left */
10396                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10397                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10398                                         /* If not at least 1 full segment never mind */
10399                                         rack->r_fast_output = 0;
10400                                 }
10401                         }
10402                 }
10403         }
10404 }
10405
10406
10407 /*
10408  * Return value of 1, the TCB is unlocked and most
10409  * likely gone, return value of 0, the TCP is still
10410  * locked.
10411  */
10412 static int
10413 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10414     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10415     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10416 {
10417         /*
10418          * Update window information. Don't look at window if no ACK: TAC's
10419          * send garbage on first SYN.
10420          */
10421         int32_t nsegs;
10422         int32_t tfo_syn;
10423         struct tcp_rack *rack;
10424
10425         INP_WLOCK_ASSERT(tptoinpcb(tp));
10426
10427         rack = (struct tcp_rack *)tp->t_fb_ptr;
10428         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10429         if ((thflags & TH_ACK) &&
10430             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10431             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10432             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10433                 /* keep track of pure window updates */
10434                 if (tlen == 0 &&
10435                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10436                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10437                 tp->snd_wnd = tiwin;
10438                 rack_validate_fo_sendwin_up(tp, rack);
10439                 tp->snd_wl1 = th->th_seq;
10440                 tp->snd_wl2 = th->th_ack;
10441                 if (tp->snd_wnd > tp->max_sndwnd)
10442                         tp->max_sndwnd = tp->snd_wnd;
10443                 rack->r_wanted_output = 1;
10444         } else if (thflags & TH_ACK) {
10445                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10446                         tp->snd_wnd = tiwin;
10447                         rack_validate_fo_sendwin_up(tp, rack);
10448                         tp->snd_wl1 = th->th_seq;
10449                         tp->snd_wl2 = th->th_ack;
10450                 }
10451         }
10452         if (tp->snd_wnd < ctf_outstanding(tp))
10453                 /* The peer collapsed the window */
10454                 rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
10455         else if (rack->rc_has_collapsed)
10456                 rack_un_collapse_window(rack, __LINE__);
10457         if ((rack->r_collapse_point_valid) &&
10458             (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
10459                 rack->r_collapse_point_valid = 0;
10460         /* Was persist timer active and now we have window space? */
10461         if ((rack->rc_in_persist != 0) &&
10462             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10463                                 rack->r_ctl.rc_pace_min_segs))) {
10464                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10465                 tp->snd_nxt = tp->snd_max;
10466                 /* Make sure we output to start the timer */
10467                 rack->r_wanted_output = 1;
10468         }
10469         /* Do we enter persists? */
10470         if ((rack->rc_in_persist == 0) &&
10471             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10472             TCPS_HAVEESTABLISHED(tp->t_state) &&
10473             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10474             sbavail(&tptosocket(tp)->so_snd) &&
10475             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
10476                 /*
10477                  * Here the rwnd is less than
10478                  * the pacing size, we are established,
10479                  * nothing is outstanding, and there is
10480                  * data to send. Enter persists.
10481                  */
10482                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10483         }
10484         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10485                 m_freem(m);
10486                 return (0);
10487         }
10488         /*
10489          * don't process the URG bit, ignore them drag
10490          * along the up.
10491          */
10492         tp->rcv_up = tp->rcv_nxt;
10493
10494         /*
10495          * Process the segment text, merging it into the TCP sequencing
10496          * queue, and arranging for acknowledgment of receipt if necessary.
10497          * This process logically involves adjusting tp->rcv_wnd as data is
10498          * presented to the user (this happens in tcp_usrreq.c, case
10499          * PRU_RCVD).  If a FIN has already been received on this connection
10500          * then we just ignore the text.
10501          */
10502         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10503                    IS_FASTOPEN(tp->t_flags));
10504         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10505             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10506                 tcp_seq save_start = th->th_seq;
10507                 tcp_seq save_rnxt  = tp->rcv_nxt;
10508                 int     save_tlen  = tlen;
10509
10510                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10511                 /*
10512                  * Insert segment which includes th into TCP reassembly
10513                  * queue with control block tp.  Set thflags to whether
10514                  * reassembly now includes a segment with FIN.  This handles
10515                  * the common case inline (segment is the next to be
10516                  * received on an established connection, and the queue is
10517                  * empty), avoiding linkage into and removal from the queue
10518                  * and repetition of various conversions. Set DELACK for
10519                  * segments received in order, but ack immediately when
10520                  * segments are out of order (so fast retransmit can work).
10521                  */
10522                 if (th->th_seq == tp->rcv_nxt &&
10523                     SEGQ_EMPTY(tp) &&
10524                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
10525                     tfo_syn)) {
10526 #ifdef NETFLIX_SB_LIMITS
10527                         u_int mcnt, appended;
10528
10529                         if (so->so_rcv.sb_shlim) {
10530                                 mcnt = m_memcnt(m);
10531                                 appended = 0;
10532                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10533                                     CFO_NOSLEEP, NULL) == false) {
10534                                         counter_u64_add(tcp_sb_shlim_fails, 1);
10535                                         m_freem(m);
10536                                         return (0);
10537                                 }
10538                         }
10539 #endif
10540                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10541                         tp->rcv_nxt += tlen;
10542                         if (tlen &&
10543                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10544                             (tp->t_fbyte_in == 0)) {
10545                                 tp->t_fbyte_in = ticks;
10546                                 if (tp->t_fbyte_in == 0)
10547                                         tp->t_fbyte_in = 1;
10548                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10549                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10550                         }
10551                         thflags = tcp_get_flags(th) & TH_FIN;
10552                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10553                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10554                         SOCKBUF_LOCK(&so->so_rcv);
10555                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10556                                 m_freem(m);
10557                         } else
10558 #ifdef NETFLIX_SB_LIMITS
10559                                 appended =
10560 #endif
10561                                         sbappendstream_locked(&so->so_rcv, m, 0);
10562
10563                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10564                         /* NB: sorwakeup_locked() does an implicit unlock. */
10565                         sorwakeup_locked(so);
10566 #ifdef NETFLIX_SB_LIMITS
10567                         if (so->so_rcv.sb_shlim && appended != mcnt)
10568                                 counter_fo_release(so->so_rcv.sb_shlim,
10569                                     mcnt - appended);
10570 #endif
10571                 } else {
10572                         /*
10573                          * XXX: Due to the header drop above "th" is
10574                          * theoretically invalid by now.  Fortunately
10575                          * m_adj() doesn't actually frees any mbufs when
10576                          * trimming from the head.
10577                          */
10578                         tcp_seq temp = save_start;
10579
10580                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
10581                         tp->t_flags |= TF_ACKNOW;
10582                         if (tp->t_flags & TF_WAKESOR) {
10583                                 tp->t_flags &= ~TF_WAKESOR;
10584                                 /* NB: sorwakeup_locked() does an implicit unlock. */
10585                                 sorwakeup_locked(so);
10586                         }
10587                 }
10588                 if ((tp->t_flags & TF_SACK_PERMIT) &&
10589                     (save_tlen > 0) &&
10590                     TCPS_HAVEESTABLISHED(tp->t_state)) {
10591                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10592                                 /*
10593                                  * DSACK actually handled in the fastpath
10594                                  * above.
10595                                  */
10596                                 RACK_OPTS_INC(tcp_sack_path_1);
10597                                 tcp_update_sack_list(tp, save_start,
10598                                     save_start + save_tlen);
10599                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10600                                 if ((tp->rcv_numsacks >= 1) &&
10601                                     (tp->sackblks[0].end == save_start)) {
10602                                         /*
10603                                          * Partial overlap, recorded at todrop
10604                                          * above.
10605                                          */
10606                                         RACK_OPTS_INC(tcp_sack_path_2a);
10607                                         tcp_update_sack_list(tp,
10608                                             tp->sackblks[0].start,
10609                                             tp->sackblks[0].end);
10610                                 } else {
10611                                         RACK_OPTS_INC(tcp_sack_path_2b);
10612                                         tcp_update_dsack_list(tp, save_start,
10613                                             save_start + save_tlen);
10614                                 }
10615                         } else if (tlen >= save_tlen) {
10616                                 /* Update of sackblks. */
10617                                 RACK_OPTS_INC(tcp_sack_path_3);
10618                                 tcp_update_dsack_list(tp, save_start,
10619                                     save_start + save_tlen);
10620                         } else if (tlen > 0) {
10621                                 RACK_OPTS_INC(tcp_sack_path_4);
10622                                 tcp_update_dsack_list(tp, save_start,
10623                                     save_start + tlen);
10624                         }
10625                 }
10626         } else {
10627                 m_freem(m);
10628                 thflags &= ~TH_FIN;
10629         }
10630
10631         /*
10632          * If FIN is received ACK the FIN and let the user know that the
10633          * connection is closing.
10634          */
10635         if (thflags & TH_FIN) {
10636                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10637                         /* The socket upcall is handled by socantrcvmore. */
10638                         socantrcvmore(so);
10639                         /*
10640                          * If connection is half-synchronized (ie NEEDSYN
10641                          * flag on) then delay ACK, so it may be piggybacked
10642                          * when SYN is sent. Otherwise, since we received a
10643                          * FIN then no more input can be expected, send ACK
10644                          * now.
10645                          */
10646                         if (tp->t_flags & TF_NEEDSYN) {
10647                                 rack_timer_cancel(tp, rack,
10648                                     rack->r_ctl.rc_rcvtime, __LINE__);
10649                                 tp->t_flags |= TF_DELACK;
10650                         } else {
10651                                 tp->t_flags |= TF_ACKNOW;
10652                         }
10653                         tp->rcv_nxt++;
10654                 }
10655                 switch (tp->t_state) {
10656                         /*
10657                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
10658                          * CLOSE_WAIT state.
10659                          */
10660                 case TCPS_SYN_RECEIVED:
10661                         tp->t_starttime = ticks;
10662                         /* FALLTHROUGH */
10663                 case TCPS_ESTABLISHED:
10664                         rack_timer_cancel(tp, rack,
10665                             rack->r_ctl.rc_rcvtime, __LINE__);
10666                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
10667                         break;
10668
10669                         /*
10670                          * If still in FIN_WAIT_1 STATE FIN has not been
10671                          * acked so enter the CLOSING state.
10672                          */
10673                 case TCPS_FIN_WAIT_1:
10674                         rack_timer_cancel(tp, rack,
10675                             rack->r_ctl.rc_rcvtime, __LINE__);
10676                         tcp_state_change(tp, TCPS_CLOSING);
10677                         break;
10678
10679                         /*
10680                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
10681                          * starting the time-wait timer, turning off the
10682                          * other standard timers.
10683                          */
10684                 case TCPS_FIN_WAIT_2:
10685                         rack_timer_cancel(tp, rack,
10686                             rack->r_ctl.rc_rcvtime, __LINE__);
10687                         tcp_twstart(tp);
10688                         return (1);
10689                 }
10690         }
10691         /*
10692          * Return any desired output.
10693          */
10694         if ((tp->t_flags & TF_ACKNOW) ||
10695             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10696                 rack->r_wanted_output = 1;
10697         }
10698         return (0);
10699 }
10700
10701 /*
10702  * Here nothing is really faster, its just that we
10703  * have broken out the fast-data path also just like
10704  * the fast-ack.
10705  */
10706 static int
10707 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10708     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10709     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10710 {
10711         int32_t nsegs;
10712         int32_t newsize = 0;    /* automatic sockbuf scaling */
10713         struct tcp_rack *rack;
10714 #ifdef NETFLIX_SB_LIMITS
10715         u_int mcnt, appended;
10716 #endif
10717
10718         /*
10719          * If last ACK falls within this segment's sequence numbers, record
10720          * the timestamp. NOTE that the test is modified according to the
10721          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10722          */
10723         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10724                 return (0);
10725         }
10726         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10727                 return (0);
10728         }
10729         if (tiwin && tiwin != tp->snd_wnd) {
10730                 return (0);
10731         }
10732         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10733                 return (0);
10734         }
10735         if (__predict_false((to->to_flags & TOF_TS) &&
10736             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10737                 return (0);
10738         }
10739         if (__predict_false((th->th_ack != tp->snd_una))) {
10740                 return (0);
10741         }
10742         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10743                 return (0);
10744         }
10745         if ((to->to_flags & TOF_TS) != 0 &&
10746             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10747                 tp->ts_recent_age = tcp_ts_getticks();
10748                 tp->ts_recent = to->to_tsval;
10749         }
10750         rack = (struct tcp_rack *)tp->t_fb_ptr;
10751         /*
10752          * This is a pure, in-sequence data packet with nothing on the
10753          * reassembly queue and we have enough buffer space to take it.
10754          */
10755         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10756
10757 #ifdef NETFLIX_SB_LIMITS
10758         if (so->so_rcv.sb_shlim) {
10759                 mcnt = m_memcnt(m);
10760                 appended = 0;
10761                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10762                     CFO_NOSLEEP, NULL) == false) {
10763                         counter_u64_add(tcp_sb_shlim_fails, 1);
10764                         m_freem(m);
10765                         return (1);
10766                 }
10767         }
10768 #endif
10769         /* Clean receiver SACK report if present */
10770         if (tp->rcv_numsacks)
10771                 tcp_clean_sackreport(tp);
10772         KMOD_TCPSTAT_INC(tcps_preddat);
10773         tp->rcv_nxt += tlen;
10774         if (tlen &&
10775             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10776             (tp->t_fbyte_in == 0)) {
10777                 tp->t_fbyte_in = ticks;
10778                 if (tp->t_fbyte_in == 0)
10779                         tp->t_fbyte_in = 1;
10780                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10781                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10782         }
10783         /*
10784          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10785          */
10786         tp->snd_wl1 = th->th_seq;
10787         /*
10788          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10789          */
10790         tp->rcv_up = tp->rcv_nxt;
10791         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10792         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10793         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10794
10795         /* Add data to socket buffer. */
10796         SOCKBUF_LOCK(&so->so_rcv);
10797         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10798                 m_freem(m);
10799         } else {
10800                 /*
10801                  * Set new socket buffer size. Give up when limit is
10802                  * reached.
10803                  */
10804                 if (newsize)
10805                         if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
10806                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10807                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10808 #ifdef NETFLIX_SB_LIMITS
10809                 appended =
10810 #endif
10811                         sbappendstream_locked(&so->so_rcv, m, 0);
10812                 ctf_calc_rwin(so, tp);
10813         }
10814         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10815         /* NB: sorwakeup_locked() does an implicit unlock. */
10816         sorwakeup_locked(so);
10817 #ifdef NETFLIX_SB_LIMITS
10818         if (so->so_rcv.sb_shlim && mcnt != appended)
10819                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10820 #endif
10821         rack_handle_delayed_ack(tp, rack, tlen, 0);
10822         if (tp->snd_una == tp->snd_max)
10823                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10824         return (1);
10825 }
10826
10827 /*
10828  * This subfunction is used to try to highly optimize the
10829  * fast path. We again allow window updates that are
10830  * in sequence to remain in the fast-path. We also add
10831  * in the __predict's to attempt to help the compiler.
10832  * Note that if we return a 0, then we can *not* process
10833  * it and the caller should push the packet into the
10834  * slow-path.
10835  */
10836 static int
10837 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10838     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10839     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10840 {
10841         int32_t acked;
10842         int32_t nsegs;
10843         int32_t under_pacing = 0;
10844         struct tcp_rack *rack;
10845
10846         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10847                 /* Old ack, behind (or duplicate to) the last one rcv'd */
10848                 return (0);
10849         }
10850         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10851                 /* Above what we have sent? */
10852                 return (0);
10853         }
10854         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10855                 /* We are retransmitting */
10856                 return (0);
10857         }
10858         if (__predict_false(tiwin == 0)) {
10859                 /* zero window */
10860                 return (0);
10861         }
10862         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10863                 /* We need a SYN or a FIN, unlikely.. */
10864                 return (0);
10865         }
10866         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10867                 /* Timestamp is behind .. old ack with seq wrap? */
10868                 return (0);
10869         }
10870         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10871                 /* Still recovering */
10872                 return (0);
10873         }
10874         rack = (struct tcp_rack *)tp->t_fb_ptr;
10875         if (rack->r_ctl.rc_sacked) {
10876                 /* We have sack holes on our scoreboard */
10877                 return (0);
10878         }
10879         /* Ok if we reach here, we can process a fast-ack */
10880         if (rack->gp_ready &&
10881             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10882                 under_pacing = 1;
10883         }
10884         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10885         rack_log_ack(tp, to, th, 0, 0);
10886         /* Did the window get updated? */
10887         if (tiwin != tp->snd_wnd) {
10888                 tp->snd_wnd = tiwin;
10889                 rack_validate_fo_sendwin_up(tp, rack);
10890                 tp->snd_wl1 = th->th_seq;
10891                 if (tp->snd_wnd > tp->max_sndwnd)
10892                         tp->max_sndwnd = tp->snd_wnd;
10893         }
10894         /* Do we exit persists? */
10895         if ((rack->rc_in_persist != 0) &&
10896             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10897                                rack->r_ctl.rc_pace_min_segs))) {
10898                 rack_exit_persist(tp, rack, cts);
10899         }
10900         /* Do we enter persists? */
10901         if ((rack->rc_in_persist == 0) &&
10902             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10903             TCPS_HAVEESTABLISHED(tp->t_state) &&
10904             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10905             sbavail(&tptosocket(tp)->so_snd) &&
10906             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
10907                 /*
10908                  * Here the rwnd is less than
10909                  * the pacing size, we are established,
10910                  * nothing is outstanding, and there is
10911                  * data to send. Enter persists.
10912                  */
10913                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10914         }
10915         /*
10916          * If last ACK falls within this segment's sequence numbers, record
10917          * the timestamp. NOTE that the test is modified according to the
10918          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10919          */
10920         if ((to->to_flags & TOF_TS) != 0 &&
10921             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10922                 tp->ts_recent_age = tcp_ts_getticks();
10923                 tp->ts_recent = to->to_tsval;
10924         }
10925         /*
10926          * This is a pure ack for outstanding data.
10927          */
10928         KMOD_TCPSTAT_INC(tcps_predack);
10929
10930         /*
10931          * "bad retransmit" recovery.
10932          */
10933         if ((tp->t_flags & TF_PREVVALID) &&
10934             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10935                 tp->t_flags &= ~TF_PREVVALID;
10936                 if (tp->t_rxtshift == 1 &&
10937                     (int)(ticks - tp->t_badrxtwin) < 0)
10938                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10939         }
10940         /*
10941          * Recalculate the transmit timer / rtt.
10942          *
10943          * Some boxes send broken timestamp replies during the SYN+ACK
10944          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10945          * and blow up the retransmit timer.
10946          */
10947         acked = BYTES_THIS_ACK(tp, th);
10948
10949 #ifdef TCP_HHOOK
10950         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
10951         hhook_run_tcp_est_in(tp, th, to);
10952 #endif
10953         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10954         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10955         if (acked) {
10956                 struct mbuf *mfree;
10957
10958                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
10959                 SOCKBUF_LOCK(&so->so_snd);
10960                 mfree = sbcut_locked(&so->so_snd, acked);
10961                 tp->snd_una = th->th_ack;
10962                 /* Note we want to hold the sb lock through the sendmap adjust */
10963                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10964                 /* Wake up the socket if we have room to write more */
10965                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10966                 sowwakeup_locked(so);
10967                 m_freem(mfree);
10968                 tp->t_rxtshift = 0;
10969                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10970                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10971                 rack->rc_tlp_in_progress = 0;
10972                 rack->r_ctl.rc_tlp_cnt_out = 0;
10973                 /*
10974                  * If it is the RXT timer we want to
10975                  * stop it, so we can restart a TLP.
10976                  */
10977                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10978                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10979 #ifdef NETFLIX_HTTP_LOGGING
10980                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10981 #endif
10982         }
10983         /*
10984          * Let the congestion control algorithm update congestion control
10985          * related information. This typically means increasing the
10986          * congestion window.
10987          */
10988         if (tp->snd_wnd < ctf_outstanding(tp)) {
10989                 /* The peer collapsed the window */
10990                 rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
10991         } else if (rack->rc_has_collapsed)
10992                 rack_un_collapse_window(rack, __LINE__);
10993         if ((rack->r_collapse_point_valid) &&
10994             (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
10995                 rack->r_collapse_point_valid = 0;
10996         /*
10997          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
10998          */
10999         tp->snd_wl2 = th->th_ack;
11000         tp->t_dupacks = 0;
11001         m_freem(m);
11002         /* ND6_HINT(tp);         *//* Some progress has been made. */
11003
11004         /*
11005          * If all outstanding data are acked, stop retransmit timer,
11006          * otherwise restart timer using current (possibly backed-off)
11007          * value. If process is waiting for space, wakeup/selwakeup/signal.
11008          * If data are ready to send, let tcp_output decide between more
11009          * output or persist.
11010          */
11011         if (under_pacing &&
11012             (rack->use_fixed_rate == 0) &&
11013             (rack->in_probe_rtt == 0) &&
11014             rack->rc_gp_dyn_mul &&
11015             rack->rc_always_pace) {
11016                 /* Check if we are dragging bottom */
11017                 rack_check_bottom_drag(tp, rack, so, acked);
11018         }
11019         if (tp->snd_una == tp->snd_max) {
11020                 tp->t_flags &= ~TF_PREVVALID;
11021                 rack->r_ctl.retran_during_recovery = 0;
11022                 rack->r_ctl.dsack_byte_cnt = 0;
11023                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11024                 if (rack->r_ctl.rc_went_idle_time == 0)
11025                         rack->r_ctl.rc_went_idle_time = 1;
11026                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11027                 if (sbavail(&tptosocket(tp)->so_snd) == 0)
11028                         tp->t_acktime = 0;
11029                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11030         }
11031         if (acked && rack->r_fast_output)
11032                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11033         if (sbavail(&so->so_snd)) {
11034                 rack->r_wanted_output = 1;
11035         }
11036         return (1);
11037 }
11038
11039 /*
11040  * Return value of 1, the TCB is unlocked and most
11041  * likely gone, return value of 0, the TCP is still
11042  * locked.
11043  */
11044 static int
11045 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11046     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11047     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11048 {
11049         int32_t ret_val = 0;
11050         int32_t todrop;
11051         int32_t ourfinisacked = 0;
11052         struct tcp_rack *rack;
11053
11054         INP_WLOCK_ASSERT(tptoinpcb(tp));
11055
11056         ctf_calc_rwin(so, tp);
11057         /*
11058          * If the state is SYN_SENT: if seg contains an ACK, but not for our
11059          * SYN, drop the input. if seg contains a RST, then drop the
11060          * connection. if seg does not contain SYN, then drop it. Otherwise
11061          * this is an acceptable SYN segment initialize tp->rcv_nxt and
11062          * tp->irs if seg contains ack then advance tp->snd_una if seg
11063          * contains an ECE and ECN support is enabled, the stream is ECN
11064          * capable. if SYN has been acked change to ESTABLISHED else
11065          * SYN_RCVD state arrange for segment to be acked (eventually)
11066          * continue processing rest of data/controls.
11067          */
11068         if ((thflags & TH_ACK) &&
11069             (SEQ_LEQ(th->th_ack, tp->iss) ||
11070             SEQ_GT(th->th_ack, tp->snd_max))) {
11071                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11072                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11073                 return (1);
11074         }
11075         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11076                 TCP_PROBE5(connect__refused, NULL, tp,
11077                     mtod(m, const char *), tp, th);
11078                 tp = tcp_drop(tp, ECONNREFUSED);
11079                 ctf_do_drop(m, tp);
11080                 return (1);
11081         }
11082         if (thflags & TH_RST) {
11083                 ctf_do_drop(m, tp);
11084                 return (1);
11085         }
11086         if (!(thflags & TH_SYN)) {
11087                 ctf_do_drop(m, tp);
11088                 return (1);
11089         }
11090         tp->irs = th->th_seq;
11091         tcp_rcvseqinit(tp);
11092         rack = (struct tcp_rack *)tp->t_fb_ptr;
11093         if (thflags & TH_ACK) {
11094                 int tfo_partial = 0;
11095
11096                 KMOD_TCPSTAT_INC(tcps_connects);
11097                 soisconnected(so);
11098 #ifdef MAC
11099                 mac_socketpeer_set_from_mbuf(m, so);
11100 #endif
11101                 /* Do window scaling on this connection? */
11102                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11103                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11104                         tp->rcv_scale = tp->request_r_scale;
11105                 }
11106                 tp->rcv_adv += min(tp->rcv_wnd,
11107                     TCP_MAXWIN << tp->rcv_scale);
11108                 /*
11109                  * If not all the data that was sent in the TFO SYN
11110                  * has been acked, resend the remainder right away.
11111                  */
11112                 if (IS_FASTOPEN(tp->t_flags) &&
11113                     (tp->snd_una != tp->snd_max)) {
11114                         tp->snd_nxt = th->th_ack;
11115                         tfo_partial = 1;
11116                 }
11117                 /*
11118                  * If there's data, delay ACK; if there's also a FIN ACKNOW
11119                  * will be turned on later.
11120                  */
11121                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11122                         rack_timer_cancel(tp, rack,
11123                                           rack->r_ctl.rc_rcvtime, __LINE__);
11124                         tp->t_flags |= TF_DELACK;
11125                 } else {
11126                         rack->r_wanted_output = 1;
11127                         tp->t_flags |= TF_ACKNOW;
11128                         rack->rc_dack_toggle = 0;
11129                 }
11130
11131                 tcp_ecn_input_syn_sent(tp, thflags, iptos);
11132
11133                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
11134                         /*
11135                          * We advance snd_una for the
11136                          * fast open case. If th_ack is
11137                          * acknowledging data beyond
11138                          * snd_una we can't just call
11139                          * ack-processing since the
11140                          * data stream in our send-map
11141                          * will start at snd_una + 1 (one
11142                          * beyond the SYN). If its just
11143                          * equal we don't need to do that
11144                          * and there is no send_map.
11145                          */
11146                         tp->snd_una++;
11147                 }
11148                 /*
11149                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11150                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11151                  */
11152                 tp->t_starttime = ticks;
11153                 if (tp->t_flags & TF_NEEDFIN) {
11154                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
11155                         tp->t_flags &= ~TF_NEEDFIN;
11156                         thflags &= ~TH_SYN;
11157                 } else {
11158                         tcp_state_change(tp, TCPS_ESTABLISHED);
11159                         TCP_PROBE5(connect__established, NULL, tp,
11160                             mtod(m, const char *), tp, th);
11161                         rack_cc_conn_init(tp);
11162                 }
11163         } else {
11164                 /*
11165                  * Received initial SYN in SYN-SENT[*] state => simultaneous
11166                  * open.  If segment contains CC option and there is a
11167                  * cached CC, apply TAO test. If it succeeds, connection is *
11168                  * half-synchronized. Otherwise, do 3-way handshake:
11169                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11170                  * there was no CC option, clear cached CC value.
11171                  */
11172                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
11173                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
11174         }
11175         /*
11176          * Advance th->th_seq to correspond to first data byte. If data,
11177          * trim to stay within window, dropping FIN if necessary.
11178          */
11179         th->th_seq++;
11180         if (tlen > tp->rcv_wnd) {
11181                 todrop = tlen - tp->rcv_wnd;
11182                 m_adj(m, -todrop);
11183                 tlen = tp->rcv_wnd;
11184                 thflags &= ~TH_FIN;
11185                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11186                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11187         }
11188         tp->snd_wl1 = th->th_seq - 1;
11189         tp->rcv_up = th->th_seq;
11190         /*
11191          * Client side of transaction: already sent SYN and data. If the
11192          * remote host used T/TCP to validate the SYN, our data will be
11193          * ACK'd; if so, enter normal data segment processing in the middle
11194          * of step 5, ack processing. Otherwise, goto step 6.
11195          */
11196         if (thflags & TH_ACK) {
11197                 /* For syn-sent we need to possibly update the rtt */
11198                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11199                         uint32_t t, mcts;
11200
11201                         mcts = tcp_ts_getticks();
11202                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11203                         if (!tp->t_rttlow || tp->t_rttlow > t)
11204                                 tp->t_rttlow = t;
11205                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11206                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11207                         tcp_rack_xmit_timer_commit(rack, tp);
11208                 }
11209                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11210                         return (ret_val);
11211                 /* We may have changed to FIN_WAIT_1 above */
11212                 if (tp->t_state == TCPS_FIN_WAIT_1) {
11213                         /*
11214                          * In FIN_WAIT_1 STATE in addition to the processing
11215                          * for the ESTABLISHED state if our FIN is now
11216                          * acknowledged then enter FIN_WAIT_2.
11217                          */
11218                         if (ourfinisacked) {
11219                                 /*
11220                                  * If we can't receive any more data, then
11221                                  * closing user can proceed. Starting the
11222                                  * timer is contrary to the specification,
11223                                  * but if we don't get a FIN we'll hang
11224                                  * forever.
11225                                  *
11226                                  * XXXjl: we should release the tp also, and
11227                                  * use a compressed state.
11228                                  */
11229                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11230                                         soisdisconnected(so);
11231                                         tcp_timer_activate(tp, TT_2MSL,
11232                                             (tcp_fast_finwait2_recycle ?
11233                                             tcp_finwait2_timeout :
11234                                             TP_MAXIDLE(tp)));
11235                                 }
11236                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11237                         }
11238                 }
11239         }
11240         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11241            tiwin, thflags, nxt_pkt));
11242 }
11243
11244 /*
11245  * Return value of 1, the TCB is unlocked and most
11246  * likely gone, return value of 0, the TCP is still
11247  * locked.
11248  */
11249 static int
11250 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11251     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11252     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11253 {
11254         struct tcp_rack *rack;
11255         int32_t ret_val = 0;
11256         int32_t ourfinisacked = 0;
11257
11258         ctf_calc_rwin(so, tp);
11259         if ((thflags & TH_ACK) &&
11260             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11261             SEQ_GT(th->th_ack, tp->snd_max))) {
11262                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11263                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11264                 return (1);
11265         }
11266         rack = (struct tcp_rack *)tp->t_fb_ptr;
11267         if (IS_FASTOPEN(tp->t_flags)) {
11268                 /*
11269                  * When a TFO connection is in SYN_RECEIVED, the
11270                  * only valid packets are the initial SYN, a
11271                  * retransmit/copy of the initial SYN (possibly with
11272                  * a subset of the original data), a valid ACK, a
11273                  * FIN, or a RST.
11274                  */
11275                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11276                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11277                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11278                         return (1);
11279                 } else if (thflags & TH_SYN) {
11280                         /* non-initial SYN is ignored */
11281                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11282                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11283                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11284                                 ctf_do_drop(m, NULL);
11285                                 return (0);
11286                         }
11287                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11288                         ctf_do_drop(m, NULL);
11289                         return (0);
11290                 }
11291         }
11292
11293         if ((thflags & TH_RST) ||
11294             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11295                 return (__ctf_process_rst(m, th, so, tp,
11296                                           &rack->r_ctl.challenge_ack_ts,
11297                                           &rack->r_ctl.challenge_ack_cnt));
11298         /*
11299          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11300          * it's less than ts_recent, drop it.
11301          */
11302         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11303             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11304                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11305                         return (ret_val);
11306         }
11307         /*
11308          * In the SYN-RECEIVED state, validate that the packet belongs to
11309          * this connection before trimming the data to fit the receive
11310          * window.  Check the sequence number versus IRS since we know the
11311          * sequence numbers haven't wrapped.  This is a partial fix for the
11312          * "LAND" DoS attack.
11313          */
11314         if (SEQ_LT(th->th_seq, tp->irs)) {
11315                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11316                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11317                 return (1);
11318         }
11319         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11320                               &rack->r_ctl.challenge_ack_ts,
11321                               &rack->r_ctl.challenge_ack_cnt)) {
11322                 return (ret_val);
11323         }
11324         /*
11325          * If last ACK falls within this segment's sequence numbers, record
11326          * its timestamp. NOTE: 1) That the test incorporates suggestions
11327          * from the latest proposal of the tcplw@cray.com list (Braden
11328          * 1993/04/26). 2) That updating only on newer timestamps interferes
11329          * with our earlier PAWS tests, so this check should be solely
11330          * predicated on the sequence space of this segment. 3) That we
11331          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11332          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11333          * SEG.Len, This modified check allows us to overcome RFC1323's
11334          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11335          * p.869. In such cases, we can still calculate the RTT correctly
11336          * when RCV.NXT == Last.ACK.Sent.
11337          */
11338         if ((to->to_flags & TOF_TS) != 0 &&
11339             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11340             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11341             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11342                 tp->ts_recent_age = tcp_ts_getticks();
11343                 tp->ts_recent = to->to_tsval;
11344         }
11345         tp->snd_wnd = tiwin;
11346         rack_validate_fo_sendwin_up(tp, rack);
11347         /*
11348          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11349          * is on (half-synchronized state), then queue data for later
11350          * processing; else drop segment and return.
11351          */
11352         if ((thflags & TH_ACK) == 0) {
11353                 if (IS_FASTOPEN(tp->t_flags)) {
11354                         rack_cc_conn_init(tp);
11355                 }
11356                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11357                     tiwin, thflags, nxt_pkt));
11358         }
11359         KMOD_TCPSTAT_INC(tcps_connects);
11360         if (tp->t_flags & TF_SONOTCONN) {
11361                 tp->t_flags &= ~TF_SONOTCONN;
11362                 soisconnected(so);
11363         }
11364         /* Do window scaling? */
11365         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11366             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11367                 tp->rcv_scale = tp->request_r_scale;
11368         }
11369         /*
11370          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11371          * FIN-WAIT-1
11372          */
11373         tp->t_starttime = ticks;
11374         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11375                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11376                 tp->t_tfo_pending = NULL;
11377         }
11378         if (tp->t_flags & TF_NEEDFIN) {
11379                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11380                 tp->t_flags &= ~TF_NEEDFIN;
11381         } else {
11382                 tcp_state_change(tp, TCPS_ESTABLISHED);
11383                 TCP_PROBE5(accept__established, NULL, tp,
11384                     mtod(m, const char *), tp, th);
11385                 /*
11386                  * TFO connections call cc_conn_init() during SYN
11387                  * processing.  Calling it again here for such connections
11388                  * is not harmless as it would undo the snd_cwnd reduction
11389                  * that occurs when a TFO SYN|ACK is retransmitted.
11390                  */
11391                 if (!IS_FASTOPEN(tp->t_flags))
11392                         rack_cc_conn_init(tp);
11393         }
11394         /*
11395          * Account for the ACK of our SYN prior to
11396          * regular ACK processing below, except for
11397          * simultaneous SYN, which is handled later.
11398          */
11399         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11400                 tp->snd_una++;
11401         /*
11402          * If segment contains data or ACK, will call tcp_reass() later; if
11403          * not, do so now to pass queued data to user.
11404          */
11405         if (tlen == 0 && (thflags & TH_FIN) == 0) {
11406                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11407                     (struct mbuf *)0);
11408                 if (tp->t_flags & TF_WAKESOR) {
11409                         tp->t_flags &= ~TF_WAKESOR;
11410                         /* NB: sorwakeup_locked() does an implicit unlock. */
11411                         sorwakeup_locked(so);
11412                 }
11413         }
11414         tp->snd_wl1 = th->th_seq - 1;
11415         /* For syn-recv we need to possibly update the rtt */
11416         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11417                 uint32_t t, mcts;
11418
11419                 mcts = tcp_ts_getticks();
11420                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11421                 if (!tp->t_rttlow || tp->t_rttlow > t)
11422                         tp->t_rttlow = t;
11423                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11424                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11425                 tcp_rack_xmit_timer_commit(rack, tp);
11426         }
11427         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11428                 return (ret_val);
11429         }
11430         if (tp->t_state == TCPS_FIN_WAIT_1) {
11431                 /* We could have went to FIN_WAIT_1 (or EST) above */
11432                 /*
11433                  * In FIN_WAIT_1 STATE in addition to the processing for the
11434                  * ESTABLISHED state if our FIN is now acknowledged then
11435                  * enter FIN_WAIT_2.
11436                  */
11437                 if (ourfinisacked) {
11438                         /*
11439                          * If we can't receive any more data, then closing
11440                          * user can proceed. Starting the timer is contrary
11441                          * to the specification, but if we don't get a FIN
11442                          * we'll hang forever.
11443                          *
11444                          * XXXjl: we should release the tp also, and use a
11445                          * compressed state.
11446                          */
11447                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11448                                 soisdisconnected(so);
11449                                 tcp_timer_activate(tp, TT_2MSL,
11450                                     (tcp_fast_finwait2_recycle ?
11451                                     tcp_finwait2_timeout :
11452                                     TP_MAXIDLE(tp)));
11453                         }
11454                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
11455                 }
11456         }
11457         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11458             tiwin, thflags, nxt_pkt));
11459 }
11460
11461 /*
11462  * Return value of 1, the TCB is unlocked and most
11463  * likely gone, return value of 0, the TCP is still
11464  * locked.
11465  */
11466 static int
11467 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11468     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11469     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11470 {
11471         int32_t ret_val = 0;
11472         struct tcp_rack *rack;
11473
11474         /*
11475          * Header prediction: check for the two common cases of a
11476          * uni-directional data xfer.  If the packet has no control flags,
11477          * is in-sequence, the window didn't change and we're not
11478          * retransmitting, it's a candidate.  If the length is zero and the
11479          * ack moved forward, we're the sender side of the xfer.  Just free
11480          * the data acked & wake any higher level process that was blocked
11481          * waiting for space.  If the length is non-zero and the ack didn't
11482          * move, we're the receiver side.  If we're getting packets in-order
11483          * (the reassembly queue is empty), add the data toc The socket
11484          * buffer and note that we need a delayed ack. Make sure that the
11485          * hidden state-flags are also off. Since we check for
11486          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11487          */
11488         rack = (struct tcp_rack *)tp->t_fb_ptr;
11489         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11490             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11491             __predict_true(SEGQ_EMPTY(tp)) &&
11492             __predict_true(th->th_seq == tp->rcv_nxt)) {
11493                 if (tlen == 0) {
11494                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11495                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11496                                 return (0);
11497                         }
11498                 } else {
11499                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11500                             tiwin, nxt_pkt, iptos)) {
11501                                 return (0);
11502                         }
11503                 }
11504         }
11505         ctf_calc_rwin(so, tp);
11506
11507         if ((thflags & TH_RST) ||
11508             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11509                 return (__ctf_process_rst(m, th, so, tp,
11510                                           &rack->r_ctl.challenge_ack_ts,
11511                                           &rack->r_ctl.challenge_ack_cnt));
11512
11513         /*
11514          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11515          * synchronized state.
11516          */
11517         if (thflags & TH_SYN) {
11518                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11519                 return (ret_val);
11520         }
11521         /*
11522          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11523          * it's less than ts_recent, drop it.
11524          */
11525         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11526             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11527                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11528                         return (ret_val);
11529         }
11530         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11531                               &rack->r_ctl.challenge_ack_ts,
11532                               &rack->r_ctl.challenge_ack_cnt)) {
11533                 return (ret_val);
11534         }
11535         /*
11536          * If last ACK falls within this segment's sequence numbers, record
11537          * its timestamp. NOTE: 1) That the test incorporates suggestions
11538          * from the latest proposal of the tcplw@cray.com list (Braden
11539          * 1993/04/26). 2) That updating only on newer timestamps interferes
11540          * with our earlier PAWS tests, so this check should be solely
11541          * predicated on the sequence space of this segment. 3) That we
11542          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11543          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11544          * SEG.Len, This modified check allows us to overcome RFC1323's
11545          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11546          * p.869. In such cases, we can still calculate the RTT correctly
11547          * when RCV.NXT == Last.ACK.Sent.
11548          */
11549         if ((to->to_flags & TOF_TS) != 0 &&
11550             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11551             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11552             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11553                 tp->ts_recent_age = tcp_ts_getticks();
11554                 tp->ts_recent = to->to_tsval;
11555         }
11556         /*
11557          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11558          * is on (half-synchronized state), then queue data for later
11559          * processing; else drop segment and return.
11560          */
11561         if ((thflags & TH_ACK) == 0) {
11562                 if (tp->t_flags & TF_NEEDSYN) {
11563                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11564                             tiwin, thflags, nxt_pkt));
11565
11566                 } else if (tp->t_flags & TF_ACKNOW) {
11567                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11568                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11569                         return (ret_val);
11570                 } else {
11571                         ctf_do_drop(m, NULL);
11572                         return (0);
11573                 }
11574         }
11575         /*
11576          * Ack processing.
11577          */
11578         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11579                 return (ret_val);
11580         }
11581         if (sbavail(&so->so_snd)) {
11582                 if (ctf_progress_timeout_check(tp, true)) {
11583                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11584                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11585                         return (1);
11586                 }
11587         }
11588         /* State changes only happen in rack_process_data() */
11589         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11590             tiwin, thflags, nxt_pkt));
11591 }
11592
11593 /*
11594  * Return value of 1, the TCB is unlocked and most
11595  * likely gone, return value of 0, the TCP is still
11596  * locked.
11597  */
11598 static int
11599 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11600     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11601     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11602 {
11603         int32_t ret_val = 0;
11604         struct tcp_rack *rack;
11605
11606         rack = (struct tcp_rack *)tp->t_fb_ptr;
11607         ctf_calc_rwin(so, tp);
11608         if ((thflags & TH_RST) ||
11609             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11610                 return (__ctf_process_rst(m, th, so, tp,
11611                                           &rack->r_ctl.challenge_ack_ts,
11612                                           &rack->r_ctl.challenge_ack_cnt));
11613         /*
11614          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11615          * synchronized state.
11616          */
11617         if (thflags & TH_SYN) {
11618                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11619                 return (ret_val);
11620         }
11621         /*
11622          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11623          * it's less than ts_recent, drop it.
11624          */
11625         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11626             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11627                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11628                         return (ret_val);
11629         }
11630         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11631                               &rack->r_ctl.challenge_ack_ts,
11632                               &rack->r_ctl.challenge_ack_cnt)) {
11633                 return (ret_val);
11634         }
11635         /*
11636          * If last ACK falls within this segment's sequence numbers, record
11637          * its timestamp. NOTE: 1) That the test incorporates suggestions
11638          * from the latest proposal of the tcplw@cray.com list (Braden
11639          * 1993/04/26). 2) That updating only on newer timestamps interferes
11640          * with our earlier PAWS tests, so this check should be solely
11641          * predicated on the sequence space of this segment. 3) That we
11642          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11643          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11644          * SEG.Len, This modified check allows us to overcome RFC1323's
11645          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11646          * p.869. In such cases, we can still calculate the RTT correctly
11647          * when RCV.NXT == Last.ACK.Sent.
11648          */
11649         if ((to->to_flags & TOF_TS) != 0 &&
11650             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11651             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11652             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11653                 tp->ts_recent_age = tcp_ts_getticks();
11654                 tp->ts_recent = to->to_tsval;
11655         }
11656         /*
11657          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11658          * is on (half-synchronized state), then queue data for later
11659          * processing; else drop segment and return.
11660          */
11661         if ((thflags & TH_ACK) == 0) {
11662                 if (tp->t_flags & TF_NEEDSYN) {
11663                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11664                             tiwin, thflags, nxt_pkt));
11665
11666                 } else if (tp->t_flags & TF_ACKNOW) {
11667                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11668                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11669                         return (ret_val);
11670                 } else {
11671                         ctf_do_drop(m, NULL);
11672                         return (0);
11673                 }
11674         }
11675         /*
11676          * Ack processing.
11677          */
11678         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11679                 return (ret_val);
11680         }
11681         if (sbavail(&so->so_snd)) {
11682                 if (ctf_progress_timeout_check(tp, true)) {
11683                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11684                                                 tp, tick, PROGRESS_DROP, __LINE__);
11685                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11686                         return (1);
11687                 }
11688         }
11689         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11690             tiwin, thflags, nxt_pkt));
11691 }
11692
11693 static int
11694 rack_check_data_after_close(struct mbuf *m,
11695     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11696 {
11697         struct tcp_rack *rack;
11698
11699         rack = (struct tcp_rack *)tp->t_fb_ptr;
11700         if (rack->rc_allow_data_af_clo == 0) {
11701         close_now:
11702                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11703                 /* tcp_close will kill the inp pre-log the Reset */
11704                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11705                 tp = tcp_close(tp);
11706                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11707                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11708                 return (1);
11709         }
11710         if (sbavail(&so->so_snd) == 0)
11711                 goto close_now;
11712         /* Ok we allow data that is ignored and a followup reset */
11713         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11714         tp->rcv_nxt = th->th_seq + *tlen;
11715         tp->t_flags2 |= TF2_DROP_AF_DATA;
11716         rack->r_wanted_output = 1;
11717         *tlen = 0;
11718         return (0);
11719 }
11720
11721 /*
11722  * Return value of 1, the TCB is unlocked and most
11723  * likely gone, return value of 0, the TCP is still
11724  * locked.
11725  */
11726 static int
11727 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11728     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11729     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11730 {
11731         int32_t ret_val = 0;
11732         int32_t ourfinisacked = 0;
11733         struct tcp_rack *rack;
11734
11735         rack = (struct tcp_rack *)tp->t_fb_ptr;
11736         ctf_calc_rwin(so, tp);
11737
11738         if ((thflags & TH_RST) ||
11739             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11740                 return (__ctf_process_rst(m, th, so, tp,
11741                                           &rack->r_ctl.challenge_ack_ts,
11742                                           &rack->r_ctl.challenge_ack_cnt));
11743         /*
11744          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11745          * synchronized state.
11746          */
11747         if (thflags & TH_SYN) {
11748                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11749                 return (ret_val);
11750         }
11751         /*
11752          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11753          * it's less than ts_recent, drop it.
11754          */
11755         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11756             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11757                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11758                         return (ret_val);
11759         }
11760         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11761                               &rack->r_ctl.challenge_ack_ts,
11762                               &rack->r_ctl.challenge_ack_cnt)) {
11763                 return (ret_val);
11764         }
11765         /*
11766          * If new data are received on a connection after the user processes
11767          * are gone, then RST the other end.
11768          */
11769         if ((tp->t_flags & TF_CLOSED) && tlen &&
11770             rack_check_data_after_close(m, tp, &tlen, th, so))
11771                 return (1);
11772         /*
11773          * If last ACK falls within this segment's sequence numbers, record
11774          * its timestamp. NOTE: 1) That the test incorporates suggestions
11775          * from the latest proposal of the tcplw@cray.com list (Braden
11776          * 1993/04/26). 2) That updating only on newer timestamps interferes
11777          * with our earlier PAWS tests, so this check should be solely
11778          * predicated on the sequence space of this segment. 3) That we
11779          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11780          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11781          * SEG.Len, This modified check allows us to overcome RFC1323's
11782          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11783          * p.869. In such cases, we can still calculate the RTT correctly
11784          * when RCV.NXT == Last.ACK.Sent.
11785          */
11786         if ((to->to_flags & TOF_TS) != 0 &&
11787             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11788             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11789             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11790                 tp->ts_recent_age = tcp_ts_getticks();
11791                 tp->ts_recent = to->to_tsval;
11792         }
11793         /*
11794          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11795          * is on (half-synchronized state), then queue data for later
11796          * processing; else drop segment and return.
11797          */
11798         if ((thflags & TH_ACK) == 0) {
11799                 if (tp->t_flags & TF_NEEDSYN) {
11800                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11801                             tiwin, thflags, nxt_pkt));
11802                 } else if (tp->t_flags & TF_ACKNOW) {
11803                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11804                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11805                         return (ret_val);
11806                 } else {
11807                         ctf_do_drop(m, NULL);
11808                         return (0);
11809                 }
11810         }
11811         /*
11812          * Ack processing.
11813          */
11814         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11815                 return (ret_val);
11816         }
11817         if (ourfinisacked) {
11818                 /*
11819                  * If we can't receive any more data, then closing user can
11820                  * proceed. Starting the timer is contrary to the
11821                  * specification, but if we don't get a FIN we'll hang
11822                  * forever.
11823                  *
11824                  * XXXjl: we should release the tp also, and use a
11825                  * compressed state.
11826                  */
11827                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11828                         soisdisconnected(so);
11829                         tcp_timer_activate(tp, TT_2MSL,
11830                             (tcp_fast_finwait2_recycle ?
11831                             tcp_finwait2_timeout :
11832                             TP_MAXIDLE(tp)));
11833                 }
11834                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11835         }
11836         if (sbavail(&so->so_snd)) {
11837                 if (ctf_progress_timeout_check(tp, true)) {
11838                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11839                                                 tp, tick, PROGRESS_DROP, __LINE__);
11840                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11841                         return (1);
11842                 }
11843         }
11844         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11845             tiwin, thflags, nxt_pkt));
11846 }
11847
11848 /*
11849  * Return value of 1, the TCB is unlocked and most
11850  * likely gone, return value of 0, the TCP is still
11851  * locked.
11852  */
11853 static int
11854 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11855     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11856     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11857 {
11858         int32_t ret_val = 0;
11859         int32_t ourfinisacked = 0;
11860         struct tcp_rack *rack;
11861
11862         rack = (struct tcp_rack *)tp->t_fb_ptr;
11863         ctf_calc_rwin(so, tp);
11864
11865         if ((thflags & TH_RST) ||
11866             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11867                 return (__ctf_process_rst(m, th, so, tp,
11868                                           &rack->r_ctl.challenge_ack_ts,
11869                                           &rack->r_ctl.challenge_ack_cnt));
11870         /*
11871          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11872          * synchronized state.
11873          */
11874         if (thflags & TH_SYN) {
11875                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11876                 return (ret_val);
11877         }
11878         /*
11879          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11880          * it's less than ts_recent, drop it.
11881          */
11882         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11883             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11884                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11885                         return (ret_val);
11886         }
11887         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11888                               &rack->r_ctl.challenge_ack_ts,
11889                               &rack->r_ctl.challenge_ack_cnt)) {
11890                 return (ret_val);
11891         }
11892         /*
11893          * If new data are received on a connection after the user processes
11894          * are gone, then RST the other end.
11895          */
11896         if ((tp->t_flags & TF_CLOSED) && tlen &&
11897             rack_check_data_after_close(m, tp, &tlen, th, so))
11898                 return (1);
11899         /*
11900          * If last ACK falls within this segment's sequence numbers, record
11901          * its timestamp. NOTE: 1) That the test incorporates suggestions
11902          * from the latest proposal of the tcplw@cray.com list (Braden
11903          * 1993/04/26). 2) That updating only on newer timestamps interferes
11904          * with our earlier PAWS tests, so this check should be solely
11905          * predicated on the sequence space of this segment. 3) That we
11906          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11907          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11908          * SEG.Len, This modified check allows us to overcome RFC1323's
11909          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11910          * p.869. In such cases, we can still calculate the RTT correctly
11911          * when RCV.NXT == Last.ACK.Sent.
11912          */
11913         if ((to->to_flags & TOF_TS) != 0 &&
11914             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11915             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11916             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11917                 tp->ts_recent_age = tcp_ts_getticks();
11918                 tp->ts_recent = to->to_tsval;
11919         }
11920         /*
11921          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11922          * is on (half-synchronized state), then queue data for later
11923          * processing; else drop segment and return.
11924          */
11925         if ((thflags & TH_ACK) == 0) {
11926                 if (tp->t_flags & TF_NEEDSYN) {
11927                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11928                             tiwin, thflags, nxt_pkt));
11929                 } else if (tp->t_flags & TF_ACKNOW) {
11930                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11931                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11932                         return (ret_val);
11933                 } else {
11934                         ctf_do_drop(m, NULL);
11935                         return (0);
11936                 }
11937         }
11938         /*
11939          * Ack processing.
11940          */
11941         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11942                 return (ret_val);
11943         }
11944         if (ourfinisacked) {
11945                 tcp_twstart(tp);
11946                 m_freem(m);
11947                 return (1);
11948         }
11949         if (sbavail(&so->so_snd)) {
11950                 if (ctf_progress_timeout_check(tp, true)) {
11951                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11952                                                 tp, tick, PROGRESS_DROP, __LINE__);
11953                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11954                         return (1);
11955                 }
11956         }
11957         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11958             tiwin, thflags, nxt_pkt));
11959 }
11960
11961 /*
11962  * Return value of 1, the TCB is unlocked and most
11963  * likely gone, return value of 0, the TCP is still
11964  * locked.
11965  */
11966 static int
11967 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11968     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11969     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11970 {
11971         int32_t ret_val = 0;
11972         int32_t ourfinisacked = 0;
11973         struct tcp_rack *rack;
11974
11975         rack = (struct tcp_rack *)tp->t_fb_ptr;
11976         ctf_calc_rwin(so, tp);
11977
11978         if ((thflags & TH_RST) ||
11979             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11980                 return (__ctf_process_rst(m, th, so, tp,
11981                                           &rack->r_ctl.challenge_ack_ts,
11982                                           &rack->r_ctl.challenge_ack_cnt));
11983         /*
11984          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11985          * synchronized state.
11986          */
11987         if (thflags & TH_SYN) {
11988                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11989                 return (ret_val);
11990         }
11991         /*
11992          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11993          * it's less than ts_recent, drop it.
11994          */
11995         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11996             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11997                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11998                         return (ret_val);
11999         }
12000         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12001                               &rack->r_ctl.challenge_ack_ts,
12002                               &rack->r_ctl.challenge_ack_cnt)) {
12003                 return (ret_val);
12004         }
12005         /*
12006          * If new data are received on a connection after the user processes
12007          * are gone, then RST the other end.
12008          */
12009         if ((tp->t_flags & TF_CLOSED) && tlen &&
12010             rack_check_data_after_close(m, tp, &tlen, th, so))
12011                 return (1);
12012         /*
12013          * If last ACK falls within this segment's sequence numbers, record
12014          * its timestamp. NOTE: 1) That the test incorporates suggestions
12015          * from the latest proposal of the tcplw@cray.com list (Braden
12016          * 1993/04/26). 2) That updating only on newer timestamps interferes
12017          * with our earlier PAWS tests, so this check should be solely
12018          * predicated on the sequence space of this segment. 3) That we
12019          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12020          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12021          * SEG.Len, This modified check allows us to overcome RFC1323's
12022          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12023          * p.869. In such cases, we can still calculate the RTT correctly
12024          * when RCV.NXT == Last.ACK.Sent.
12025          */
12026         if ((to->to_flags & TOF_TS) != 0 &&
12027             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12028             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12029             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12030                 tp->ts_recent_age = tcp_ts_getticks();
12031                 tp->ts_recent = to->to_tsval;
12032         }
12033         /*
12034          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12035          * is on (half-synchronized state), then queue data for later
12036          * processing; else drop segment and return.
12037          */
12038         if ((thflags & TH_ACK) == 0) {
12039                 if (tp->t_flags & TF_NEEDSYN) {
12040                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12041                             tiwin, thflags, nxt_pkt));
12042                 } else if (tp->t_flags & TF_ACKNOW) {
12043                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12044                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12045                         return (ret_val);
12046                 } else {
12047                         ctf_do_drop(m, NULL);
12048                         return (0);
12049                 }
12050         }
12051         /*
12052          * case TCPS_LAST_ACK: Ack processing.
12053          */
12054         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12055                 return (ret_val);
12056         }
12057         if (ourfinisacked) {
12058                 tp = tcp_close(tp);
12059                 ctf_do_drop(m, tp);
12060                 return (1);
12061         }
12062         if (sbavail(&so->so_snd)) {
12063                 if (ctf_progress_timeout_check(tp, true)) {
12064                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12065                                                 tp, tick, PROGRESS_DROP, __LINE__);
12066                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12067                         return (1);
12068                 }
12069         }
12070         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12071             tiwin, thflags, nxt_pkt));
12072 }
12073
12074 /*
12075  * Return value of 1, the TCB is unlocked and most
12076  * likely gone, return value of 0, the TCP is still
12077  * locked.
12078  */
12079 static int
12080 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12081     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12082     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12083 {
12084         int32_t ret_val = 0;
12085         int32_t ourfinisacked = 0;
12086         struct tcp_rack *rack;
12087
12088         rack = (struct tcp_rack *)tp->t_fb_ptr;
12089         ctf_calc_rwin(so, tp);
12090
12091         /* Reset receive buffer auto scaling when not in bulk receive mode. */
12092         if ((thflags & TH_RST) ||
12093             (tp->t_fin_is_rst && (thflags & TH_FIN)))
12094                 return (__ctf_process_rst(m, th, so, tp,
12095                                           &rack->r_ctl.challenge_ack_ts,
12096                                           &rack->r_ctl.challenge_ack_cnt));
12097         /*
12098          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12099          * synchronized state.
12100          */
12101         if (thflags & TH_SYN) {
12102                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12103                 return (ret_val);
12104         }
12105         /*
12106          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12107          * it's less than ts_recent, drop it.
12108          */
12109         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12110             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12111                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12112                         return (ret_val);
12113         }
12114         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12115                               &rack->r_ctl.challenge_ack_ts,
12116                               &rack->r_ctl.challenge_ack_cnt)) {
12117                 return (ret_val);
12118         }
12119         /*
12120          * If new data are received on a connection after the user processes
12121          * are gone, then RST the other end.
12122          */
12123         if ((tp->t_flags & TF_CLOSED) && tlen &&
12124             rack_check_data_after_close(m, tp, &tlen, th, so))
12125                 return (1);
12126         /*
12127          * If last ACK falls within this segment's sequence numbers, record
12128          * its timestamp. NOTE: 1) That the test incorporates suggestions
12129          * from the latest proposal of the tcplw@cray.com list (Braden
12130          * 1993/04/26). 2) That updating only on newer timestamps interferes
12131          * with our earlier PAWS tests, so this check should be solely
12132          * predicated on the sequence space of this segment. 3) That we
12133          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12134          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12135          * SEG.Len, This modified check allows us to overcome RFC1323's
12136          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12137          * p.869. In such cases, we can still calculate the RTT correctly
12138          * when RCV.NXT == Last.ACK.Sent.
12139          */
12140         if ((to->to_flags & TOF_TS) != 0 &&
12141             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12142             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12143             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12144                 tp->ts_recent_age = tcp_ts_getticks();
12145                 tp->ts_recent = to->to_tsval;
12146         }
12147         /*
12148          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12149          * is on (half-synchronized state), then queue data for later
12150          * processing; else drop segment and return.
12151          */
12152         if ((thflags & TH_ACK) == 0) {
12153                 if (tp->t_flags & TF_NEEDSYN) {
12154                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12155                             tiwin, thflags, nxt_pkt));
12156                 } else if (tp->t_flags & TF_ACKNOW) {
12157                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12158                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12159                         return (ret_val);
12160                 } else {
12161                         ctf_do_drop(m, NULL);
12162                         return (0);
12163                 }
12164         }
12165         /*
12166          * Ack processing.
12167          */
12168         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12169                 return (ret_val);
12170         }
12171         if (sbavail(&so->so_snd)) {
12172                 if (ctf_progress_timeout_check(tp, true)) {
12173                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12174                                                 tp, tick, PROGRESS_DROP, __LINE__);
12175                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12176                         return (1);
12177                 }
12178         }
12179         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12180             tiwin, thflags, nxt_pkt));
12181 }
12182
12183 static void inline
12184 rack_clear_rate_sample(struct tcp_rack *rack)
12185 {
12186         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12187         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12188         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12189 }
12190
12191 static void
12192 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12193 {
12194         uint64_t bw_est, rate_wanted;
12195         int chged = 0;
12196         uint32_t user_max, orig_min, orig_max;
12197
12198         orig_min = rack->r_ctl.rc_pace_min_segs;
12199         orig_max = rack->r_ctl.rc_pace_max_segs;
12200         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12201         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12202                 chged = 1;
12203         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12204         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12205                 if (user_max != rack->r_ctl.rc_pace_max_segs)
12206                         chged = 1;
12207         }
12208         if (rack->rc_force_max_seg) {
12209                 rack->r_ctl.rc_pace_max_segs = user_max;
12210         } else if (rack->use_fixed_rate) {
12211                 bw_est = rack_get_bw(rack);
12212                 if ((rack->r_ctl.crte == NULL) ||
12213                     (bw_est != rack->r_ctl.crte->rate)) {
12214                         rack->r_ctl.rc_pace_max_segs = user_max;
12215                 } else {
12216                         /* We are pacing right at the hardware rate */
12217                         uint32_t segsiz;
12218
12219                         segsiz = min(ctf_fixed_maxseg(tp),
12220                                      rack->r_ctl.rc_pace_min_segs);
12221                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12222                                                            tp, bw_est, segsiz, 0,
12223                                                            rack->r_ctl.crte, NULL);
12224                 }
12225         } else if (rack->rc_always_pace) {
12226                 if (rack->r_ctl.gp_bw ||
12227 #ifdef NETFLIX_PEAKRATE
12228                     rack->rc_tp->t_maxpeakrate ||
12229 #endif
12230                     rack->r_ctl.init_rate) {
12231                         /* We have a rate of some sort set */
12232                         uint32_t  orig;
12233
12234                         bw_est = rack_get_bw(rack);
12235                         orig = rack->r_ctl.rc_pace_max_segs;
12236                         if (fill_override)
12237                                 rate_wanted = *fill_override;
12238                         else
12239                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12240                         if (rate_wanted) {
12241                                 /* We have something */
12242                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12243                                                                                    rate_wanted,
12244                                                                                    ctf_fixed_maxseg(rack->rc_tp));
12245                         } else
12246                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12247                         if (orig != rack->r_ctl.rc_pace_max_segs)
12248                                 chged = 1;
12249                 } else if ((rack->r_ctl.gp_bw == 0) &&
12250                            (rack->r_ctl.rc_pace_max_segs == 0)) {
12251                         /*
12252                          * If we have nothing limit us to bursting
12253                          * out IW sized pieces.
12254                          */
12255                         chged = 1;
12256                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12257                 }
12258         }
12259         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12260                 chged = 1;
12261                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12262         }
12263         if (chged)
12264                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12265 }
12266
12267
12268 static void
12269 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12270 {
12271 #ifdef INET6
12272         struct ip6_hdr *ip6 = NULL;
12273 #endif
12274 #ifdef INET
12275         struct ip *ip = NULL;
12276 #endif
12277         struct udphdr *udp = NULL;
12278
12279         /* Ok lets fill in the fast block, it can only be used with no IP options! */
12280 #ifdef INET6
12281         if (rack->r_is_v6) {
12282                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12283                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12284                 if (tp->t_port) {
12285                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12286                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12287                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12288                         udp->uh_dport = tp->t_port;
12289                         rack->r_ctl.fsb.udp = udp;
12290                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12291                 } else
12292                 {
12293                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12294                         rack->r_ctl.fsb.udp = NULL;
12295                 }
12296                 tcpip_fillheaders(rack->rc_inp,
12297                                   tp->t_port,
12298                                   ip6, rack->r_ctl.fsb.th);
12299         } else
12300 #endif                          /* INET6 */
12301 #ifdef INET
12302         {
12303                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12304                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12305                 if (tp->t_port) {
12306                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12307                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12308                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12309                         udp->uh_dport = tp->t_port;
12310                         rack->r_ctl.fsb.udp = udp;
12311                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12312                 } else
12313                 {
12314                         rack->r_ctl.fsb.udp = NULL;
12315                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12316                 }
12317                 tcpip_fillheaders(rack->rc_inp,
12318                                   tp->t_port,
12319                                   ip, rack->r_ctl.fsb.th);
12320         }
12321 #endif
12322         rack->r_fsb_inited = 1;
12323 }
12324
12325 static int
12326 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12327 {
12328         /*
12329          * Allocate the larger of spaces V6 if available else just
12330          * V4 and include udphdr (overbook)
12331          */
12332 #ifdef INET6
12333         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12334 #else
12335         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12336 #endif
12337         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12338                                             M_TCPFSB, M_NOWAIT|M_ZERO);
12339         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12340                 return (ENOMEM);
12341         }
12342         rack->r_fsb_inited = 0;
12343         return (0);
12344 }
12345
12346 static int
12347 rack_init(struct tcpcb *tp)
12348 {
12349         struct inpcb *inp = tptoinpcb(tp);
12350         struct tcp_rack *rack = NULL;
12351 #ifdef INVARIANTS
12352         struct rack_sendmap *insret;
12353 #endif
12354         uint32_t iwin, snt, us_cts;
12355         int err;
12356
12357         tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12358         if (tp->t_fb_ptr == NULL) {
12359                 /*
12360                  * We need to allocate memory but cant. The INP and INP_INFO
12361                  * locks and they are recursive (happens during setup. So a
12362                  * scheme to drop the locks fails :(
12363                  *
12364                  */
12365                 return (ENOMEM);
12366         }
12367         memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12368
12369         rack = (struct tcp_rack *)tp->t_fb_ptr;
12370         RB_INIT(&rack->r_ctl.rc_mtree);
12371         TAILQ_INIT(&rack->r_ctl.rc_free);
12372         TAILQ_INIT(&rack->r_ctl.rc_tmap);
12373         rack->rc_tp = tp;
12374         rack->rc_inp = inp;
12375         /* Set the flag */
12376         rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
12377         /* Probably not needed but lets be sure */
12378         rack_clear_rate_sample(rack);
12379         /*
12380          * Save off the default values, socket options will poke
12381          * at these if pacing is not on or we have not yet
12382          * reached where pacing is on (gp_ready/fixed enabled).
12383          * When they get set into the CC module (when gp_ready
12384          * is enabled or we enable fixed) then we will set these
12385          * values into the CC and place in here the old values
12386          * so we have a restoral. Then we will set the flag
12387          * rc_pacing_cc_set. That way whenever we turn off pacing
12388          * or switch off this stack, we will know to go restore
12389          * the saved values.
12390          */
12391         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12392         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12393         /* We want abe like behavior as well */
12394         rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12395         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12396         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12397         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12398         rack->r_ctl.roundends = tp->snd_max;
12399         if (use_rack_rr)
12400                 rack->use_rack_rr = 1;
12401         if (V_tcp_delack_enabled)
12402                 tp->t_delayed_ack = 1;
12403         else
12404                 tp->t_delayed_ack = 0;
12405 #ifdef TCP_ACCOUNTING
12406         if (rack_tcp_accounting) {
12407                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12408         }
12409 #endif
12410         if (rack_enable_shared_cwnd)
12411                 rack->rack_enable_scwnd = 1;
12412         rack->rc_user_set_max_segs = rack_hptsi_segments;
12413         rack->rc_force_max_seg = 0;
12414         if (rack_use_imac_dack)
12415                 rack->rc_dack_mode = 1;
12416         TAILQ_INIT(&rack->r_ctl.opt_list);
12417         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12418         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12419         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12420         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12421         rack->r_ctl.rc_highest_us_rtt = 0;
12422         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12423         rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12424         if (rack_use_cmp_acks)
12425                 rack->r_use_cmp_ack = 1;
12426         if (rack_disable_prr)
12427                 rack->rack_no_prr = 1;
12428         if (rack_gp_no_rec_chg)
12429                 rack->rc_gp_no_rec_chg = 1;
12430         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12431                 rack->rc_always_pace = 1;
12432                 if (rack->use_fixed_rate || rack->gp_ready)
12433                         rack_set_cc_pacing(rack);
12434         } else
12435                 rack->rc_always_pace = 0;
12436         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12437                 rack->r_mbuf_queue = 1;
12438         else
12439                 rack->r_mbuf_queue = 0;
12440         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12441                 inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12442         else
12443                 inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12444         rack_set_pace_segments(tp, rack, __LINE__, NULL);
12445         if (rack_limits_scwnd)
12446                 rack->r_limit_scw = 1;
12447         else
12448                 rack->r_limit_scw = 0;
12449         rack->rc_labc = V_tcp_abc_l_var;
12450         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12451         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12452         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12453         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12454         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12455         rack->r_ctl.rc_min_to = rack_min_to;
12456         microuptime(&rack->r_ctl.act_rcv_time);
12457         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12458         rack->rc_init_win = rack_default_init_window;
12459         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12460         if (rack_hw_up_only)
12461                 rack->r_up_only = 1;
12462         if (rack_do_dyn_mul) {
12463                 /* When dynamic adjustment is on CA needs to start at 100% */
12464                 rack->rc_gp_dyn_mul = 1;
12465                 if (rack_do_dyn_mul >= 100)
12466                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12467         } else
12468                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12469         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12470         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12471         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12472         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12473                                 rack_probertt_filter_life);
12474         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12475         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12476         rack->r_ctl.rc_time_of_last_probertt = us_cts;
12477         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12478         rack->r_ctl.rc_time_probertt_starts = 0;
12479         if (rack_dsack_std_based & 0x1) {
12480                 /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12481                 rack->rc_rack_tmr_std_based = 1;
12482         }
12483         if (rack_dsack_std_based & 0x2) {
12484                 /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12485                 rack->rc_rack_use_dsack = 1;
12486         }
12487         /* We require at least one measurement, even if the sysctl is 0 */
12488         if (rack_req_measurements)
12489                 rack->r_ctl.req_measurements = rack_req_measurements;
12490         else
12491                 rack->r_ctl.req_measurements = 1;
12492         if (rack_enable_hw_pacing)
12493                 rack->rack_hdw_pace_ena = 1;
12494         if (rack_hw_rate_caps)
12495                 rack->r_rack_hw_rate_caps = 1;
12496         /* Do we force on detection? */
12497 #ifdef NETFLIX_EXP_DETECTION
12498         if (tcp_force_detection)
12499                 rack->do_detection = 1;
12500         else
12501 #endif
12502                 rack->do_detection = 0;
12503         if (rack_non_rxt_use_cr)
12504                 rack->rack_rec_nonrxt_use_cr = 1;
12505         err = rack_init_fsb(tp, rack);
12506         if (err) {
12507                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12508                 tp->t_fb_ptr = NULL;
12509                 return (err);
12510         }
12511         if (tp->snd_una != tp->snd_max) {
12512                 /* Create a send map for the current outstanding data */
12513                 struct rack_sendmap *rsm;
12514
12515                 rsm = rack_alloc(rack);
12516                 if (rsm == NULL) {
12517                         uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12518                         tp->t_fb_ptr = NULL;
12519                         return (ENOMEM);
12520                 }
12521                 rsm->r_no_rtt_allowed = 1;
12522                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12523                 rsm->r_rtr_cnt = 1;
12524                 rsm->r_rtr_bytes = 0;
12525                 if (tp->t_flags & TF_SENTFIN)
12526                         rsm->r_flags |= RACK_HAS_FIN;
12527                 if ((tp->snd_una == tp->iss) &&
12528                     !TCPS_HAVEESTABLISHED(tp->t_state))
12529                         rsm->r_flags |= RACK_HAS_SYN;
12530                 rsm->r_start = tp->snd_una;
12531                 rsm->r_end = tp->snd_max;
12532                 rsm->r_dupack = 0;
12533                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12534                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12535                         if (rsm->m)
12536                                 rsm->orig_m_len = rsm->m->m_len;
12537                         else
12538                                 rsm->orig_m_len = 0;
12539                 } else {
12540                         /*
12541                          * This can happen if we have a stand-alone FIN or
12542                          *  SYN.
12543                          */
12544                         rsm->m = NULL;
12545                         rsm->orig_m_len = 0;
12546                         rsm->soff = 0;
12547                 }
12548 #ifndef INVARIANTS
12549                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12550 #else
12551                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12552                 if (insret != NULL) {
12553                         panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12554                               insret, rack, rsm);
12555                 }
12556 #endif
12557                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12558                 rsm->r_in_tmap = 1;
12559         }
12560         /*
12561          * Timers in Rack are kept in microseconds so lets
12562          * convert any initial incoming variables
12563          * from ticks into usecs. Note that we
12564          * also change the values of t_srtt and t_rttvar, if
12565          * they are non-zero. They are kept with a 5
12566          * bit decimal so we have to carefully convert
12567          * these to get the full precision.
12568          */
12569         rack_convert_rtts(tp);
12570         tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12571         if (rack_do_hystart) {
12572                 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
12573                 if (rack_do_hystart > 1)
12574                         tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
12575                 if (rack_do_hystart > 2)
12576                         tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
12577         }
12578         if (rack_def_profile)
12579                 rack_set_profile(rack, rack_def_profile);
12580         /* Cancel the GP measurement in progress */
12581         tp->t_flags &= ~TF_GPUTINPROG;
12582         if (SEQ_GT(tp->snd_max, tp->iss))
12583                 snt = tp->snd_max - tp->iss;
12584         else
12585                 snt = 0;
12586         iwin = rc_init_window(rack);
12587         if (snt < iwin) {
12588                 /* We are not past the initial window
12589                  * so we need to make sure cwnd is
12590                  * correct.
12591                  */
12592                 if (tp->snd_cwnd < iwin)
12593                         tp->snd_cwnd = iwin;
12594                 /*
12595                  * If we are within the initial window
12596                  * we want ssthresh to be unlimited. Setting
12597                  * it to the rwnd (which the default stack does
12598                  * and older racks) is not really a good idea
12599                  * since we want to be in SS and grow both the
12600                  * cwnd and the rwnd (via dynamic rwnd growth). If
12601                  * we set it to the rwnd then as the peer grows its
12602                  * rwnd we will be stuck in CA and never hit SS.
12603                  *
12604                  * Its far better to raise it up high (this takes the
12605                  * risk that there as been a loss already, probably
12606                  * we should have an indicator in all stacks of loss
12607                  * but we don't), but considering the normal use this
12608                  * is a risk worth taking. The consequences of not
12609                  * hitting SS are far worse than going one more time
12610                  * into it early on (before we have sent even a IW).
12611                  * It is highly unlikely that we will have had a loss
12612                  * before getting the IW out.
12613                  */
12614                 tp->snd_ssthresh = 0xffffffff;
12615         }
12616         rack_stop_all_timers(tp);
12617         /* Lets setup the fsb block */
12618         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12619         rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12620                              __LINE__, RACK_RTTS_INIT);
12621         return (0);
12622 }
12623
12624 static int
12625 rack_handoff_ok(struct tcpcb *tp)
12626 {
12627         if ((tp->t_state == TCPS_CLOSED) ||
12628             (tp->t_state == TCPS_LISTEN)) {
12629                 /* Sure no problem though it may not stick */
12630                 return (0);
12631         }
12632         if ((tp->t_state == TCPS_SYN_SENT) ||
12633             (tp->t_state == TCPS_SYN_RECEIVED)) {
12634                 /*
12635                  * We really don't know if you support sack,
12636                  * you have to get to ESTAB or beyond to tell.
12637                  */
12638                 return (EAGAIN);
12639         }
12640         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12641                 /*
12642                  * Rack will only send a FIN after all data is acknowledged.
12643                  * So in this case we have more data outstanding. We can't
12644                  * switch stacks until either all data and only the FIN
12645                  * is left (in which case rack_init() now knows how
12646                  * to deal with that) <or> all is acknowledged and we
12647                  * are only left with incoming data, though why you
12648                  * would want to switch to rack after all data is acknowledged
12649                  * I have no idea (rrs)!
12650                  */
12651                 return (EAGAIN);
12652         }
12653         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12654                 return (0);
12655         }
12656         /*
12657          * If we reach here we don't do SACK on this connection so we can
12658          * never do rack.
12659          */
12660         return (EINVAL);
12661 }
12662
12663
12664 static void
12665 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12666 {
12667         struct inpcb *inp = tptoinpcb(tp);
12668
12669         if (tp->t_fb_ptr) {
12670                 struct tcp_rack *rack;
12671                 struct rack_sendmap *rsm, *nrsm;
12672 #ifdef INVARIANTS
12673                 struct rack_sendmap *rm;
12674 #endif
12675
12676                 rack = (struct tcp_rack *)tp->t_fb_ptr;
12677                 if (tp->t_in_pkt) {
12678                         /*
12679                          * It is unsafe to process the packets since a
12680                          * reset may be lurking in them (its rare but it
12681                          * can occur). If we were to find a RST, then we
12682                          * would end up dropping the connection and the
12683                          * INP lock, so when we return the caller (tcp_usrreq)
12684                          * will blow up when it trys to unlock the inp.
12685                          */
12686                         struct mbuf *save, *m;
12687
12688                         m = tp->t_in_pkt;
12689                         tp->t_in_pkt = NULL;
12690                         tp->t_tail_pkt = NULL;
12691                         while (m) {
12692                                 save = m->m_nextpkt;
12693                                 m->m_nextpkt = NULL;
12694                                 m_freem(m);
12695                                 m = save;
12696                         }
12697                 }
12698                 tp->t_flags &= ~TF_FORCEDATA;
12699 #ifdef NETFLIX_SHARED_CWND
12700                 if (rack->r_ctl.rc_scw) {
12701                         uint32_t limit;
12702
12703                         if (rack->r_limit_scw)
12704                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12705                         else
12706                                 limit = 0;
12707                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12708                                                   rack->r_ctl.rc_scw_index,
12709                                                   limit);
12710                         rack->r_ctl.rc_scw = NULL;
12711                 }
12712 #endif
12713                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
12714                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12715                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12716                         rack->r_ctl.fsb.th = NULL;
12717                 }
12718                 /* Convert back to ticks, with  */
12719                 if (tp->t_srtt > 1) {
12720                         uint32_t val, frac;
12721
12722                         val = USEC_2_TICKS(tp->t_srtt);
12723                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12724                         tp->t_srtt = val << TCP_RTT_SHIFT;
12725                         /*
12726                          * frac is the fractional part here is left
12727                          * over from converting to hz and shifting.
12728                          * We need to convert this to the 5 bit
12729                          * remainder.
12730                          */
12731                         if (frac) {
12732                                 if (hz == 1000) {
12733                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12734                                 } else {
12735                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12736                                 }
12737                                 tp->t_srtt += frac;
12738                         }
12739                 }
12740                 if (tp->t_rttvar) {
12741                         uint32_t val, frac;
12742
12743                         val = USEC_2_TICKS(tp->t_rttvar);
12744                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12745                         tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12746                         /*
12747                          * frac is the fractional part here is left
12748                          * over from converting to hz and shifting.
12749                          * We need to convert this to the 5 bit
12750                          * remainder.
12751                          */
12752                         if (frac) {
12753                                 if (hz == 1000) {
12754                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12755                                 } else {
12756                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12757                                 }
12758                                 tp->t_rttvar += frac;
12759                         }
12760                 }
12761                 tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12762                 tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12763                 if (rack->rc_always_pace) {
12764                         tcp_decrement_paced_conn();
12765                         rack_undo_cc_pacing(rack);
12766                         rack->rc_always_pace = 0;
12767                 }
12768                 /* Clean up any options if they were not applied */
12769                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12770                         struct deferred_opt_list *dol;
12771
12772                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12773                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12774                         free(dol, M_TCPDO);
12775                 }
12776                 /* rack does not use force data but other stacks may clear it */
12777                 if (rack->r_ctl.crte != NULL) {
12778                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12779                         rack->rack_hdrw_pacing = 0;
12780                         rack->r_ctl.crte = NULL;
12781                 }
12782 #ifdef TCP_BLACKBOX
12783                 tcp_log_flowend(tp);
12784 #endif
12785                 RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12786 #ifndef INVARIANTS
12787                         (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12788 #else
12789                         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12790                         if (rm != rsm) {
12791                                 panic("At fini, rack:%p rsm:%p rm:%p",
12792                                       rack, rsm, rm);
12793                         }
12794 #endif
12795                         uma_zfree(rack_zone, rsm);
12796                 }
12797                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12798                 while (rsm) {
12799                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12800                         uma_zfree(rack_zone, rsm);
12801                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12802                 }
12803                 rack->rc_free_cnt = 0;
12804                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12805                 tp->t_fb_ptr = NULL;
12806         }
12807         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12808         inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12809         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12810         inp->inp_flags2 &= ~INP_MBUF_ACKCMP;
12811         /* Cancel the GP measurement in progress */
12812         tp->t_flags &= ~TF_GPUTINPROG;
12813         inp->inp_flags2 &= ~INP_MBUF_L_ACKS;
12814         /* Make sure snd_nxt is correctly set */
12815         tp->snd_nxt = tp->snd_max;
12816 }
12817
12818 static void
12819 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12820 {
12821         if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12822                 rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
12823         }
12824         switch (tp->t_state) {
12825         case TCPS_SYN_SENT:
12826                 rack->r_state = TCPS_SYN_SENT;
12827                 rack->r_substate = rack_do_syn_sent;
12828                 break;
12829         case TCPS_SYN_RECEIVED:
12830                 rack->r_state = TCPS_SYN_RECEIVED;
12831                 rack->r_substate = rack_do_syn_recv;
12832                 break;
12833         case TCPS_ESTABLISHED:
12834                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12835                 rack->r_state = TCPS_ESTABLISHED;
12836                 rack->r_substate = rack_do_established;
12837                 break;
12838         case TCPS_CLOSE_WAIT:
12839                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12840                 rack->r_state = TCPS_CLOSE_WAIT;
12841                 rack->r_substate = rack_do_close_wait;
12842                 break;
12843         case TCPS_FIN_WAIT_1:
12844                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12845                 rack->r_state = TCPS_FIN_WAIT_1;
12846                 rack->r_substate = rack_do_fin_wait_1;
12847                 break;
12848         case TCPS_CLOSING:
12849                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12850                 rack->r_state = TCPS_CLOSING;
12851                 rack->r_substate = rack_do_closing;
12852                 break;
12853         case TCPS_LAST_ACK:
12854                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12855                 rack->r_state = TCPS_LAST_ACK;
12856                 rack->r_substate = rack_do_lastack;
12857                 break;
12858         case TCPS_FIN_WAIT_2:
12859                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12860                 rack->r_state = TCPS_FIN_WAIT_2;
12861                 rack->r_substate = rack_do_fin_wait_2;
12862                 break;
12863         case TCPS_LISTEN:
12864         case TCPS_CLOSED:
12865         case TCPS_TIME_WAIT:
12866         default:
12867                 break;
12868         };
12869         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12870                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12871
12872 }
12873
12874 static void
12875 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12876 {
12877         /*
12878          * We received an ack, and then did not
12879          * call send or were bounced out due to the
12880          * hpts was running. Now a timer is up as well, is
12881          * it the right timer?
12882          */
12883         struct rack_sendmap *rsm;
12884         int tmr_up;
12885
12886         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12887         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12888                 return;
12889         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12890         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12891             (tmr_up == PACE_TMR_RXT)) {
12892                 /* Should be an RXT */
12893                 return;
12894         }
12895         if (rsm == NULL) {
12896                 /* Nothing outstanding? */
12897                 if (tp->t_flags & TF_DELACK) {
12898                         if (tmr_up == PACE_TMR_DELACK)
12899                                 /* We are supposed to have delayed ack up and we do */
12900                                 return;
12901                 } else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12902                         /*
12903                          * if we hit enobufs then we would expect the possibility
12904                          * of nothing outstanding and the RXT up (and the hptsi timer).
12905                          */
12906                         return;
12907                 } else if (((V_tcp_always_keepalive ||
12908                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12909                             (tp->t_state <= TCPS_CLOSING)) &&
12910                            (tmr_up == PACE_TMR_KEEP) &&
12911                            (tp->snd_max == tp->snd_una)) {
12912                         /* We should have keep alive up and we do */
12913                         return;
12914                 }
12915         }
12916         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12917                    ((tmr_up == PACE_TMR_TLP) ||
12918                     (tmr_up == PACE_TMR_RACK) ||
12919                     (tmr_up == PACE_TMR_RXT))) {
12920                 /*
12921                  * Either a Rack, TLP or RXT is fine if  we
12922                  * have outstanding data.
12923                  */
12924                 return;
12925         } else if (tmr_up == PACE_TMR_DELACK) {
12926                 /*
12927                  * If the delayed ack was going to go off
12928                  * before the rtx/tlp/rack timer were going to
12929                  * expire, then that would be the timer in control.
12930                  * Note we don't check the time here trusting the
12931                  * code is correct.
12932                  */
12933                 return;
12934         }
12935         /*
12936          * Ok the timer originally started is not what we want now.
12937          * We will force the hpts to be stopped if any, and restart
12938          * with the slot set to what was in the saved slot.
12939          */
12940         if (tcp_in_hpts(rack->rc_inp)) {
12941                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12942                         uint32_t us_cts;
12943
12944                         us_cts = tcp_get_usecs(NULL);
12945                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12946                                 rack->r_early = 1;
12947                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12948                         }
12949                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12950                 }
12951                 tcp_hpts_remove(rack->rc_inp);
12952         }
12953         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12954         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12955 }
12956
12957
12958 static void
12959 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)
12960 {
12961         if ((SEQ_LT(tp->snd_wl1, seq) ||
12962             (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
12963             (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
12964                 /* keep track of pure window updates */
12965                 if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
12966                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12967                 tp->snd_wnd = tiwin;
12968                 rack_validate_fo_sendwin_up(tp, rack);
12969                 tp->snd_wl1 = seq;
12970                 tp->snd_wl2 = ack;
12971                 if (tp->snd_wnd > tp->max_sndwnd)
12972                         tp->max_sndwnd = tp->snd_wnd;
12973             rack->r_wanted_output = 1;
12974         } else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
12975                 tp->snd_wnd = tiwin;
12976                 rack_validate_fo_sendwin_up(tp, rack);
12977                 tp->snd_wl1 = seq;
12978                 tp->snd_wl2 = ack;
12979         } else {
12980                 /* Not a valid win update */
12981                 return;
12982         }
12983         /* Do we exit persists? */
12984         if ((rack->rc_in_persist != 0) &&
12985             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12986                                 rack->r_ctl.rc_pace_min_segs))) {
12987                 rack_exit_persist(tp, rack, cts);
12988         }
12989         /* Do we enter persists? */
12990         if ((rack->rc_in_persist == 0) &&
12991             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12992             TCPS_HAVEESTABLISHED(tp->t_state) &&
12993             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12994             sbavail(&tptosocket(tp)->so_snd) &&
12995             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12996                 /*
12997                  * Here the rwnd is less than
12998                  * the pacing size, we are established,
12999                  * nothing is outstanding, and there is
13000                  * data to send. Enter persists.
13001                  */
13002                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13003         }
13004 }
13005
13006 static void
13007 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13008 {
13009
13010         if (tcp_bblogging_on(rack->rc_tp)) {
13011                 struct inpcb *inp = tptoinpcb(tp);
13012                 union tcp_log_stackspecific log;
13013                 struct timeval ltv;
13014                 char tcp_hdr_buf[60];
13015                 struct tcphdr *th;
13016                 struct timespec ts;
13017                 uint32_t orig_snd_una;
13018                 uint8_t xx = 0;
13019
13020 #ifdef NETFLIX_HTTP_LOGGING
13021                 struct http_sendfile_track *http_req;
13022
13023                 if (SEQ_GT(ae->ack, tp->snd_una)) {
13024                         http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13025                 } else {
13026                         http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13027                 }
13028 #endif
13029                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13030                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
13031                 if (rack->rack_no_prr == 0)
13032                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13033                 else
13034                         log.u_bbr.flex1 = 0;
13035                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13036                 log.u_bbr.use_lt_bw <<= 1;
13037                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13038                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13039                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13040                 log.u_bbr.pkts_out = tp->t_maxseg;
13041                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13042                 log.u_bbr.flex7 = 1;
13043                 log.u_bbr.lost = ae->flags;
13044                 log.u_bbr.cwnd_gain = ackval;
13045                 log.u_bbr.pacing_gain = 0x2;
13046                 if (ae->flags & TSTMP_HDWR) {
13047                         /* Record the hardware timestamp if present */
13048                         log.u_bbr.flex3 = M_TSTMP;
13049                         ts.tv_sec = ae->timestamp / 1000000000;
13050                         ts.tv_nsec = ae->timestamp % 1000000000;
13051                         ltv.tv_sec = ts.tv_sec;
13052                         ltv.tv_usec = ts.tv_nsec / 1000;
13053                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13054                 } else if (ae->flags & TSTMP_LRO) {
13055                         /* Record the LRO the arrival timestamp */
13056                         log.u_bbr.flex3 = M_TSTMP_LRO;
13057                         ts.tv_sec = ae->timestamp / 1000000000;
13058                         ts.tv_nsec = ae->timestamp % 1000000000;
13059                         ltv.tv_sec = ts.tv_sec;
13060                         ltv.tv_usec = ts.tv_nsec / 1000;
13061                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13062                 }
13063                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13064                 /* Log the rcv time */
13065                 log.u_bbr.delRate = ae->timestamp;
13066 #ifdef NETFLIX_HTTP_LOGGING
13067                 log.u_bbr.applimited = tp->t_http_closed;
13068                 log.u_bbr.applimited <<= 8;
13069                 log.u_bbr.applimited |= tp->t_http_open;
13070                 log.u_bbr.applimited <<= 8;
13071                 log.u_bbr.applimited |= tp->t_http_req;
13072                 if (http_req) {
13073                         /* Copy out any client req info */
13074                         /* seconds */
13075                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13076                         /* useconds */
13077                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13078                         log.u_bbr.rttProp = http_req->timestamp;
13079                         log.u_bbr.cur_del_rate = http_req->start;
13080                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13081                                 log.u_bbr.flex8 |= 1;
13082                         } else {
13083                                 log.u_bbr.flex8 |= 2;
13084                                 log.u_bbr.bw_inuse = http_req->end;
13085                         }
13086                         log.u_bbr.flex6 = http_req->start_seq;
13087                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13088                                 log.u_bbr.flex8 |= 4;
13089                                 log.u_bbr.epoch = http_req->end_seq;
13090                         }
13091                 }
13092 #endif
13093                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13094                 th = (struct tcphdr *)tcp_hdr_buf;
13095                 th->th_seq = ae->seq;
13096                 th->th_ack = ae->ack;
13097                 th->th_win = ae->win;
13098                 /* Now fill in the ports */
13099                 th->th_sport = inp->inp_fport;
13100                 th->th_dport = inp->inp_lport;
13101                 tcp_set_flags(th, ae->flags);
13102                 /* Now do we have a timestamp option? */
13103                 if (ae->flags & HAS_TSTMP) {
13104                         u_char *cp;
13105                         uint32_t val;
13106
13107                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13108                         cp = (u_char *)(th + 1);
13109                         *cp = TCPOPT_NOP;
13110                         cp++;
13111                         *cp = TCPOPT_NOP;
13112                         cp++;
13113                         *cp = TCPOPT_TIMESTAMP;
13114                         cp++;
13115                         *cp = TCPOLEN_TIMESTAMP;
13116                         cp++;
13117                         val = htonl(ae->ts_value);
13118                         bcopy((char *)&val,
13119                               (char *)cp, sizeof(uint32_t));
13120                         val = htonl(ae->ts_echo);
13121                         bcopy((char *)&val,
13122                               (char *)(cp + 4), sizeof(uint32_t));
13123                 } else
13124                         th->th_off = (sizeof(struct tcphdr) >> 2);
13125
13126                 /*
13127                  * For sane logging we need to play a little trick.
13128                  * If the ack were fully processed we would have moved
13129                  * snd_una to high_seq, but since compressed acks are
13130                  * processed in two phases, at this point (logging) snd_una
13131                  * won't be advanced. So we would see multiple acks showing
13132                  * the advancement. We can prevent that by "pretending" that
13133                  * snd_una was advanced and then un-advancing it so that the
13134                  * logging code has the right value for tlb_snd_una.
13135                  */
13136                 if (tp->snd_una != high_seq) {
13137                         orig_snd_una = tp->snd_una;
13138                         tp->snd_una = high_seq;
13139                         xx = 1;
13140                 } else
13141                         xx = 0;
13142                 TCP_LOG_EVENTP(tp, th,
13143                                &tptosocket(tp)->so_rcv,
13144                                &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
13145                                0, &log, true, &ltv);
13146                 if (xx) {
13147                         tp->snd_una = orig_snd_una;
13148                 }
13149         }
13150
13151 }
13152
13153 static void
13154 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
13155 {
13156         uint32_t us_rtt;
13157         /*
13158          * A persist or keep-alive was forced out, update our
13159          * min rtt time. Note now worry about lost responses.
13160          * When a subsequent keep-alive or persist times out
13161          * and forced_ack is still on, then the last probe
13162          * was not responded to. In such cases we have a
13163          * sysctl that controls the behavior. Either we apply
13164          * the rtt but with reduced confidence (0). Or we just
13165          * plain don't apply the rtt estimate. Having data flow
13166          * will clear the probe_not_answered flag i.e. cum-ack
13167          * move forward <or> exiting and reentering persists.
13168          */
13169
13170         rack->forced_ack = 0;
13171         rack->rc_tp->t_rxtshift = 0;
13172         if ((rack->rc_in_persist &&
13173              (tiwin == rack->rc_tp->snd_wnd)) ||
13174             (rack->rc_in_persist == 0)) {
13175                 /*
13176                  * In persists only apply the RTT update if this is
13177                  * a response to our window probe. And that
13178                  * means the rwnd sent must match the current
13179                  * snd_wnd. If it does not, then we got a
13180                  * window update ack instead. For keepalive
13181                  * we allow the answer no matter what the window.
13182                  *
13183                  * Note that if the probe_not_answered is set then
13184                  * the forced_ack_ts is the oldest one i.e. the first
13185                  * probe sent that might have been lost. This assures
13186                  * us that if we do calculate an RTT it is longer not
13187                  * some short thing.
13188                  */
13189                 if (rack->rc_in_persist)
13190                         counter_u64_add(rack_persists_acks, 1);
13191                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13192                 if (us_rtt == 0)
13193                         us_rtt = 1;
13194                 if (rack->probe_not_answered == 0) {
13195                         rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13196                         tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13197                 } else {
13198                         /* We have a retransmitted probe here too */
13199                         if (rack_apply_rtt_with_reduced_conf) {
13200                                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13201                                 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
13202                         }
13203                 }
13204         }
13205 }
13206
13207 static int
13208 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13209 {
13210         /*
13211          * Handle a "special" compressed ack mbuf. Each incoming
13212          * ack has only four possible dispositions:
13213          *
13214          * A) It moves the cum-ack forward
13215          * B) It is behind the cum-ack.
13216          * C) It is a window-update ack.
13217          * D) It is a dup-ack.
13218          *
13219          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13220          * in the incoming mbuf. We also need to still pay attention
13221          * to nxt_pkt since there may be another packet after this
13222          * one.
13223          */
13224 #ifdef TCP_ACCOUNTING
13225         uint64_t ts_val;
13226         uint64_t rdstc;
13227 #endif
13228         int segsiz;
13229         struct timespec ts;
13230         struct tcp_rack *rack;
13231         struct tcp_ackent *ae;
13232         uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13233         int cnt, i, did_out, ourfinisacked = 0;
13234         struct tcpopt to_holder, *to = NULL;
13235 #ifdef TCP_ACCOUNTING
13236         int win_up_req = 0;
13237 #endif
13238         int nsegs = 0;
13239         int under_pacing = 1;
13240         int recovery = 0;
13241 #ifdef TCP_ACCOUNTING
13242         sched_pin();
13243 #endif
13244         rack = (struct tcp_rack *)tp->t_fb_ptr;
13245         if (rack->gp_ready &&
13246             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13247                 under_pacing = 0;
13248         else
13249                 under_pacing = 1;
13250
13251         if (rack->r_state != tp->t_state)
13252                 rack_set_state(tp, rack);
13253         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13254             (tp->t_flags & TF_GPUTINPROG)) {
13255                 /*
13256                  * We have a goodput in progress
13257                  * and we have entered a late state.
13258                  * Do we have enough data in the sb
13259                  * to handle the GPUT request?
13260                  */
13261                 uint32_t bytes;
13262
13263                 bytes = tp->gput_ack - tp->gput_seq;
13264                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
13265                         bytes += tp->gput_seq - tp->snd_una;
13266                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
13267                         /*
13268                          * There are not enough bytes in the socket
13269                          * buffer that have been sent to cover this
13270                          * measurement. Cancel it.
13271                          */
13272                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13273                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
13274                                                    tp->gput_seq,
13275                                                    0, 0, 18, __LINE__, NULL, 0);
13276                         tp->t_flags &= ~TF_GPUTINPROG;
13277                 }
13278         }
13279         to = &to_holder;
13280         to->to_flags = 0;
13281         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13282                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13283         cnt = m->m_len / sizeof(struct tcp_ackent);
13284         counter_u64_add(rack_multi_single_eq, cnt);
13285         high_seq = tp->snd_una;
13286         the_win = tp->snd_wnd;
13287         win_seq = tp->snd_wl1;
13288         win_upd_ack = tp->snd_wl2;
13289         cts = tcp_tv_to_usectick(tv);
13290         ms_cts = tcp_tv_to_mssectick(tv);
13291         rack->r_ctl.rc_rcvtime = cts;
13292         segsiz = ctf_fixed_maxseg(tp);
13293         if ((rack->rc_gp_dyn_mul) &&
13294             (rack->use_fixed_rate == 0) &&
13295             (rack->rc_always_pace)) {
13296                 /* Check in on probertt */
13297                 rack_check_probe_rtt(rack, cts);
13298         }
13299         for (i = 0; i < cnt; i++) {
13300 #ifdef TCP_ACCOUNTING
13301                 ts_val = get_cyclecount();
13302 #endif
13303                 rack_clear_rate_sample(rack);
13304                 ae = ((mtod(m, struct tcp_ackent *)) + i);
13305                 /* Setup the window */
13306                 tiwin = ae->win << tp->snd_scale;
13307                 if (tiwin > rack->r_ctl.rc_high_rwnd)
13308                         rack->r_ctl.rc_high_rwnd = tiwin;
13309                 /* figure out the type of ack */
13310                 if (SEQ_LT(ae->ack, high_seq)) {
13311                         /* Case B*/
13312                         ae->ack_val_set = ACK_BEHIND;
13313                 } else if (SEQ_GT(ae->ack, high_seq)) {
13314                         /* Case A */
13315                         ae->ack_val_set = ACK_CUMACK;
13316                 } else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
13317                         /* Case D */
13318                         ae->ack_val_set = ACK_DUPACK;
13319                 } else {
13320                         /* Case C */
13321                         ae->ack_val_set = ACK_RWND;
13322                 }
13323                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13324                 /* Validate timestamp */
13325                 if (ae->flags & HAS_TSTMP) {
13326                         /* Setup for a timestamp */
13327                         to->to_flags = TOF_TS;
13328                         ae->ts_echo -= tp->ts_offset;
13329                         to->to_tsecr = ae->ts_echo;
13330                         to->to_tsval = ae->ts_value;
13331                         /*
13332                          * If echoed timestamp is later than the current time, fall back to
13333                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13334                          * were used when this connection was established.
13335                          */
13336                         if (TSTMP_GT(ae->ts_echo, ms_cts))
13337                                 to->to_tsecr = 0;
13338                         if (tp->ts_recent &&
13339                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13340                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13341 #ifdef TCP_ACCOUNTING
13342                                         rdstc = get_cyclecount();
13343                                         if (rdstc > ts_val) {
13344                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13345                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13346                                                 }
13347                                         }
13348 #endif
13349                                         continue;
13350                                 }
13351                         }
13352                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13353                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13354                                 tp->ts_recent_age = tcp_ts_getticks();
13355                                 tp->ts_recent = ae->ts_value;
13356                         }
13357                 } else {
13358                         /* Setup for a no options */
13359                         to->to_flags = 0;
13360                 }
13361                 /* Update the rcv time and perform idle reduction possibly */
13362                 if  (tp->t_idle_reduce &&
13363                      (tp->snd_max == tp->snd_una) &&
13364                      (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13365                         counter_u64_add(rack_input_idle_reduces, 1);
13366                         rack_cc_after_idle(rack, tp);
13367                 }
13368                 tp->t_rcvtime = ticks;
13369                 /* Now what about ECN of a chain of pure ACKs? */
13370                 if (tcp_ecn_input_segment(tp, ae->flags, 0,
13371                         tcp_packets_this_ack(tp, ae->ack),
13372                         ae->codepoint))
13373                         rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
13374 #ifdef TCP_ACCOUNTING
13375                 /* Count for the specific type of ack in */
13376                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13377                         tp->tcp_cnt_counters[ae->ack_val_set]++;
13378                 }
13379 #endif
13380                 /*
13381                  * Note how we could move up these in the determination
13382                  * above, but we don't so that way the timestamp checks (and ECN)
13383                  * is done first before we do any processing on the ACK.
13384                  * The non-compressed path through the code has this
13385                  * weakness (noted by @jtl) that it actually does some
13386                  * processing before verifying the timestamp information.
13387                  * We don't take that path here which is why we set
13388                  * the ack_val_set first, do the timestamp and ecn
13389                  * processing, and then look at what we have setup.
13390                  */
13391                 if (ae->ack_val_set == ACK_BEHIND) {
13392                         /*
13393                          * Case B flag reordering, if window is not closed
13394                          * or it could be a keep-alive or persists
13395                          */
13396                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13397                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13398                         }
13399                 } else if (ae->ack_val_set == ACK_DUPACK) {
13400                         /* Case D */
13401                         rack_strike_dupack(rack);
13402                 } else if (ae->ack_val_set == ACK_RWND) {
13403                         /* Case C */
13404                         if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13405                                 ts.tv_sec = ae->timestamp / 1000000000;
13406                                 ts.tv_nsec = ae->timestamp % 1000000000;
13407                                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13408                                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13409                         } else {
13410                                 rack->r_ctl.act_rcv_time = *tv;
13411                         }
13412                         if (rack->forced_ack) {
13413                                 rack_handle_probe_response(rack, tiwin,
13414                                                            tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
13415                         }
13416 #ifdef TCP_ACCOUNTING
13417                         win_up_req = 1;
13418 #endif
13419                         win_upd_ack = ae->ack;
13420                         win_seq = ae->seq;
13421                         the_win = tiwin;
13422                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13423                 } else {
13424                         /* Case A */
13425                         if (SEQ_GT(ae->ack, tp->snd_max)) {
13426                                 /*
13427                                  * We just send an ack since the incoming
13428                                  * ack is beyond the largest seq we sent.
13429                                  */
13430                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
13431                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13432                                         if (tp->t_flags && TF_ACKNOW)
13433                                                 rack->r_wanted_output = 1;
13434                                 }
13435                         } else {
13436                                 nsegs++;
13437                                 /* If the window changed setup to update */
13438                                 if (tiwin != tp->snd_wnd) {
13439                                         win_upd_ack = ae->ack;
13440                                         win_seq = ae->seq;
13441                                         the_win = tiwin;
13442                                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13443                                 }
13444 #ifdef TCP_ACCOUNTING
13445                                 /* Account for the acks */
13446                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13447                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13448                                 }
13449 #endif
13450                                 high_seq = ae->ack;
13451                                 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
13452                                         union tcp_log_stackspecific log;
13453                                         struct timeval tv;
13454
13455                                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13456                                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13457                                         log.u_bbr.flex1 = high_seq;
13458                                         log.u_bbr.flex2 = rack->r_ctl.roundends;
13459                                         log.u_bbr.flex3 = rack->r_ctl.current_round;
13460                                         log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
13461                                         log.u_bbr.flex8 = 8;
13462                                         tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
13463                                                        0, &log, false, NULL, NULL, 0, &tv);
13464                                 }
13465                                 /*
13466                                  * The draft (v3) calls for us to use SEQ_GEQ, but that
13467                                  * causes issues when we are just going app limited. Lets
13468                                  * instead use SEQ_GT <or> where its equal but more data
13469                                  * is outstanding.
13470                                  */
13471                                 if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) ||
13472                                     ((high_seq == rack->r_ctl.roundends) &&
13473                                      SEQ_GT(tp->snd_max, tp->snd_una))) {
13474                                         rack->r_ctl.current_round++;
13475                                         rack->r_ctl.roundends = tp->snd_max;
13476                                         if (CC_ALGO(tp)->newround != NULL) {
13477                                                 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
13478                                         }
13479                                 }
13480                                 /* Setup our act_rcv_time */
13481                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13482                                         ts.tv_sec = ae->timestamp / 1000000000;
13483                                         ts.tv_nsec = ae->timestamp % 1000000000;
13484                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13485                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13486                                 } else {
13487                                         rack->r_ctl.act_rcv_time = *tv;
13488                                 }
13489                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13490                                 if (rack->rc_dsack_round_seen) {
13491                                         /* Is the dsack round over? */
13492                                         if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13493                                                 /* Yes it is */
13494                                                 rack->rc_dsack_round_seen = 0;
13495                                                 rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13496                                         }
13497                                 }
13498                         }
13499                 }
13500                 /* And lets be sure to commit the rtt measurements for this ack */
13501                 tcp_rack_xmit_timer_commit(rack, tp);
13502 #ifdef TCP_ACCOUNTING
13503                 rdstc = get_cyclecount();
13504                 if (rdstc > ts_val) {
13505                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13506                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13507                                 if (ae->ack_val_set == ACK_CUMACK)
13508                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13509                         }
13510                 }
13511 #endif
13512         }
13513 #ifdef TCP_ACCOUNTING
13514         ts_val = get_cyclecount();
13515 #endif
13516         /* Tend to any collapsed window */
13517         if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
13518                 /* The peer collapsed the window */
13519                 rack_collapsed_window(rack, (tp->snd_max - high_seq), __LINE__);
13520         } else if (rack->rc_has_collapsed)
13521                 rack_un_collapse_window(rack, __LINE__);
13522         if ((rack->r_collapse_point_valid) &&
13523             (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
13524                 rack->r_collapse_point_valid = 0;
13525         acked_amount = acked = (high_seq - tp->snd_una);
13526         if (acked) {
13527                 /*
13528                  * Clear the probe not answered flag
13529                  * since cum-ack moved forward.
13530                  */
13531                 rack->probe_not_answered = 0;
13532                 if (rack->sack_attack_disable == 0)
13533                         rack_do_decay(rack);
13534                 if (acked >= segsiz) {
13535                         /*
13536                          * You only get credit for
13537                          * MSS and greater (and you get extra
13538                          * credit for larger cum-ack moves).
13539                          */
13540                         int ac;
13541
13542                         ac = acked / segsiz;
13543                         rack->r_ctl.ack_count += ac;
13544                         counter_u64_add(rack_ack_total, ac);
13545                 }
13546                 if (rack->r_ctl.ack_count > 0xfff00000) {
13547                         /*
13548                          * reduce the number to keep us under
13549                          * a uint32_t.
13550                          */
13551                         rack->r_ctl.ack_count /= 2;
13552                         rack->r_ctl.sack_count /= 2;
13553                 }
13554                 if (tp->t_flags & TF_NEEDSYN) {
13555                         /*
13556                          * T/TCP: Connection was half-synchronized, and our SYN has
13557                          * been ACK'd (so connection is now fully synchronized).  Go
13558                          * to non-starred state, increment snd_una for ACK of SYN,
13559                          * and check if we can do window scaling.
13560                          */
13561                         tp->t_flags &= ~TF_NEEDSYN;
13562                         tp->snd_una++;
13563                         acked_amount = acked = (high_seq - tp->snd_una);
13564                 }
13565                 if (acked > sbavail(&so->so_snd))
13566                         acked_amount = sbavail(&so->so_snd);
13567 #ifdef NETFLIX_EXP_DETECTION
13568                 /*
13569                  * We only care on a cum-ack move if we are in a sack-disabled
13570                  * state. We have already added in to the ack_count, and we never
13571                  * would disable on a cum-ack move, so we only care to do the
13572                  * detection if it may "undo" it, i.e. we were in disabled already.
13573                  */
13574                 if (rack->sack_attack_disable)
13575                         rack_do_detection(tp, rack, acked_amount, segsiz);
13576 #endif
13577                 if (IN_FASTRECOVERY(tp->t_flags) &&
13578                     (rack->rack_no_prr == 0))
13579                         rack_update_prr(tp, rack, acked_amount, high_seq);
13580                 if (IN_RECOVERY(tp->t_flags)) {
13581                         if (SEQ_LT(high_seq, tp->snd_recover) &&
13582                             (SEQ_LT(high_seq, tp->snd_max))) {
13583                                 tcp_rack_partialack(tp);
13584                         } else {
13585                                 rack_post_recovery(tp, high_seq);
13586                                 recovery = 1;
13587                         }
13588                 }
13589                 /* Handle the rack-log-ack part (sendmap) */
13590                 if ((sbused(&so->so_snd) == 0) &&
13591                     (acked > acked_amount) &&
13592                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
13593                     (tp->t_flags & TF_SENTFIN)) {
13594                         /*
13595                          * We must be sure our fin
13596                          * was sent and acked (we can be
13597                          * in FIN_WAIT_1 without having
13598                          * sent the fin).
13599                          */
13600                         ourfinisacked = 1;
13601                         /*
13602                          * Lets make sure snd_una is updated
13603                          * since most likely acked_amount = 0 (it
13604                          * should be).
13605                          */
13606                         tp->snd_una = high_seq;
13607                 }
13608                 /* Did we make a RTO error? */
13609                 if ((tp->t_flags & TF_PREVVALID) &&
13610                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13611                         tp->t_flags &= ~TF_PREVVALID;
13612                         if (tp->t_rxtshift == 1 &&
13613                             (int)(ticks - tp->t_badrxtwin) < 0)
13614                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
13615                 }
13616                 /* Handle the data in the socket buffer */
13617                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13618                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13619                 if (acked_amount > 0) {
13620                         struct mbuf *mfree;
13621
13622                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13623                         SOCKBUF_LOCK(&so->so_snd);
13624                         mfree = sbcut_locked(&so->so_snd, acked_amount);
13625                         tp->snd_una = high_seq;
13626                         /* Note we want to hold the sb lock through the sendmap adjust */
13627                         rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13628                         /* Wake up the socket if we have room to write more */
13629                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13630                         sowwakeup_locked(so);
13631                         m_freem(mfree);
13632                 }
13633                 /* update progress */
13634                 tp->t_acktime = ticks;
13635                 rack_log_progress_event(rack, tp, tp->t_acktime,
13636                                         PROGRESS_UPDATE, __LINE__);
13637                 /* Clear out shifts and such */
13638                 tp->t_rxtshift = 0;
13639                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13640                                    rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13641                 rack->rc_tlp_in_progress = 0;
13642                 rack->r_ctl.rc_tlp_cnt_out = 0;
13643                 /* Send recover and snd_nxt must be dragged along */
13644                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
13645                         tp->snd_recover = tp->snd_una;
13646                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13647                         tp->snd_nxt = tp->snd_una;
13648                 /*
13649                  * If the RXT timer is running we want to
13650                  * stop it, so we can restart a TLP (or new RXT).
13651                  */
13652                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13653                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13654 #ifdef NETFLIX_HTTP_LOGGING
13655                 tcp_http_check_for_comp(rack->rc_tp, high_seq);
13656 #endif
13657                 tp->snd_wl2 = high_seq;
13658                 tp->t_dupacks = 0;
13659                 if (under_pacing &&
13660                     (rack->use_fixed_rate == 0) &&
13661                     (rack->in_probe_rtt == 0) &&
13662                     rack->rc_gp_dyn_mul &&
13663                     rack->rc_always_pace) {
13664                         /* Check if we are dragging bottom */
13665                         rack_check_bottom_drag(tp, rack, so, acked);
13666                 }
13667                 if (tp->snd_una == tp->snd_max) {
13668                         tp->t_flags &= ~TF_PREVVALID;
13669                         rack->r_ctl.retran_during_recovery = 0;
13670                         rack->r_ctl.dsack_byte_cnt = 0;
13671                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13672                         if (rack->r_ctl.rc_went_idle_time == 0)
13673                                 rack->r_ctl.rc_went_idle_time = 1;
13674                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13675                         if (sbavail(&tptosocket(tp)->so_snd) == 0)
13676                                 tp->t_acktime = 0;
13677                         /* Set so we might enter persists... */
13678                         rack->r_wanted_output = 1;
13679                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13680                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13681                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13682                             (sbavail(&so->so_snd) == 0) &&
13683                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13684                                 /*
13685                                  * The socket was gone and the
13686                                  * peer sent data (not now in the past), time to
13687                                  * reset him.
13688                                  */
13689                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13690                                 /* tcp_close will kill the inp pre-log the Reset */
13691                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13692 #ifdef TCP_ACCOUNTING
13693                                 rdstc = get_cyclecount();
13694                                 if (rdstc > ts_val) {
13695                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13696                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13697                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13698                                         }
13699                                 }
13700 #endif
13701                                 m_freem(m);
13702                                 tp = tcp_close(tp);
13703                                 if (tp == NULL) {
13704 #ifdef TCP_ACCOUNTING
13705                                         sched_unpin();
13706 #endif
13707                                         return (1);
13708                                 }
13709                                 /*
13710                                  * We would normally do drop-with-reset which would
13711                                  * send back a reset. We can't since we don't have
13712                                  * all the needed bits. Instead lets arrange for
13713                                  * a call to tcp_output(). That way since we
13714                                  * are in the closed state we will generate a reset.
13715                                  *
13716                                  * Note if tcp_accounting is on we don't unpin since
13717                                  * we do that after the goto label.
13718                                  */
13719                                 goto send_out_a_rst;
13720                         }
13721                         if ((sbused(&so->so_snd) == 0) &&
13722                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
13723                             (tp->t_flags & TF_SENTFIN)) {
13724                                 /*
13725                                  * If we can't receive any more data, then closing user can
13726                                  * proceed. Starting the timer is contrary to the
13727                                  * specification, but if we don't get a FIN we'll hang
13728                                  * forever.
13729                                  *
13730                                  */
13731                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13732                                         soisdisconnected(so);
13733                                         tcp_timer_activate(tp, TT_2MSL,
13734                                                            (tcp_fast_finwait2_recycle ?
13735                                                             tcp_finwait2_timeout :
13736                                                             TP_MAXIDLE(tp)));
13737                                 }
13738                                 if (ourfinisacked == 0) {
13739                                         /*
13740                                          * We don't change to fin-wait-2 if we have our fin acked
13741                                          * which means we are probably in TCPS_CLOSING.
13742                                          */
13743                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13744                                 }
13745                         }
13746                 }
13747                 /* Wake up the socket if we have room to write more */
13748                 if (sbavail(&so->so_snd)) {
13749                         rack->r_wanted_output = 1;
13750                         if (ctf_progress_timeout_check(tp, true)) {
13751                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13752                                                         tp, tick, PROGRESS_DROP, __LINE__);
13753                                 /*
13754                                  * We cheat here and don't send a RST, we should send one
13755                                  * when the pacer drops the connection.
13756                                  */
13757 #ifdef TCP_ACCOUNTING
13758                                 rdstc = get_cyclecount();
13759                                 if (rdstc > ts_val) {
13760                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13761                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13762                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13763                                         }
13764                                 }
13765                                 sched_unpin();
13766 #endif
13767                                 (void)tcp_drop(tp, ETIMEDOUT);
13768                                 m_freem(m);
13769                                 return (1);
13770                         }
13771                 }
13772                 if (ourfinisacked) {
13773                         switch(tp->t_state) {
13774                         case TCPS_CLOSING:
13775 #ifdef TCP_ACCOUNTING
13776                                 rdstc = get_cyclecount();
13777                                 if (rdstc > ts_val) {
13778                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13779                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13780                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13781                                         }
13782                                 }
13783                                 sched_unpin();
13784 #endif
13785                                 tcp_twstart(tp);
13786                                 m_freem(m);
13787                                 return (1);
13788                                 break;
13789                         case TCPS_LAST_ACK:
13790 #ifdef TCP_ACCOUNTING
13791                                 rdstc = get_cyclecount();
13792                                 if (rdstc > ts_val) {
13793                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13794                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13795                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13796                                         }
13797                                 }
13798                                 sched_unpin();
13799 #endif
13800                                 tp = tcp_close(tp);
13801                                 ctf_do_drop(m, tp);
13802                                 return (1);
13803                                 break;
13804                         case TCPS_FIN_WAIT_1:
13805 #ifdef TCP_ACCOUNTING
13806                                 rdstc = get_cyclecount();
13807                                 if (rdstc > ts_val) {
13808                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13809                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13810                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13811                                         }
13812                                 }
13813 #endif
13814                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13815                                         soisdisconnected(so);
13816                                         tcp_timer_activate(tp, TT_2MSL,
13817                                                            (tcp_fast_finwait2_recycle ?
13818                                                             tcp_finwait2_timeout :
13819                                                             TP_MAXIDLE(tp)));
13820                                 }
13821                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13822                                 break;
13823                         default:
13824                                 break;
13825                         }
13826                 }
13827                 if (rack->r_fast_output) {
13828                         /*
13829                          * We re doing fast output.. can we expand that?
13830                          */
13831                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13832                 }
13833 #ifdef TCP_ACCOUNTING
13834                 rdstc = get_cyclecount();
13835                 if (rdstc > ts_val) {
13836                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13837                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13838                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13839                         }
13840                 }
13841
13842         } else if (win_up_req) {
13843                 rdstc = get_cyclecount();
13844                 if (rdstc > ts_val) {
13845                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13846                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13847                         }
13848                 }
13849 #endif
13850         }
13851         /* Now is there a next packet, if so we are done */
13852         m_freem(m);
13853         did_out = 0;
13854         if (nxt_pkt) {
13855 #ifdef TCP_ACCOUNTING
13856                 sched_unpin();
13857 #endif
13858                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13859                 return (0);
13860         }
13861         rack_handle_might_revert(tp, rack);
13862         ctf_calc_rwin(so, tp);
13863         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13864         send_out_a_rst:
13865                 if (tcp_output(tp) < 0) {
13866 #ifdef TCP_ACCOUNTING
13867                         sched_unpin();
13868 #endif
13869                         return (1);
13870                 }
13871                 did_out = 1;
13872         }
13873         rack_free_trim(rack);
13874 #ifdef TCP_ACCOUNTING
13875         sched_unpin();
13876 #endif
13877         rack_timer_audit(tp, rack, &so->so_snd);
13878         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13879         return (0);
13880 }
13881
13882
13883 static int
13884 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13885     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13886     int32_t nxt_pkt, struct timeval *tv)
13887 {
13888         struct inpcb *inp = tptoinpcb(tp);
13889 #ifdef TCP_ACCOUNTING
13890         uint64_t ts_val;
13891 #endif
13892         int32_t thflags, retval, did_out = 0;
13893         int32_t way_out = 0;
13894         /*
13895          * cts - is the current time from tv (caller gets ts) in microseconds.
13896          * ms_cts - is the current time from tv in milliseconds.
13897          * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
13898          */
13899         uint32_t cts, us_cts, ms_cts;
13900         uint32_t tiwin, high_seq;
13901         struct timespec ts;
13902         struct tcpopt to;
13903         struct tcp_rack *rack;
13904         struct rack_sendmap *rsm;
13905         int32_t prev_state = 0;
13906 #ifdef TCP_ACCOUNTING
13907         int ack_val_set = 0xf;
13908 #endif
13909         int nsegs;
13910
13911         NET_EPOCH_ASSERT();
13912         INP_WLOCK_ASSERT(inp);
13913
13914         /*
13915          * tv passed from common code is from either M_TSTMP_LRO or
13916          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13917          */
13918         rack = (struct tcp_rack *)tp->t_fb_ptr;
13919         if (m->m_flags & M_ACKCMP) {
13920                 /*
13921                  * All compressed ack's are ack's by definition so
13922                  * remove any ack required flag and then do the processing.
13923                  */
13924                 rack->rc_ack_required = 0;
13925                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13926         }
13927         if (m->m_flags & M_ACKCMP) {
13928                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13929         }
13930         cts = tcp_tv_to_usectick(tv);
13931         ms_cts =  tcp_tv_to_mssectick(tv);
13932         nsegs = m->m_pkthdr.lro_nsegs;
13933         counter_u64_add(rack_proc_non_comp_ack, 1);
13934         thflags = tcp_get_flags(th);
13935 #ifdef TCP_ACCOUNTING
13936         sched_pin();
13937         if (thflags & TH_ACK)
13938                 ts_val = get_cyclecount();
13939 #endif
13940         if ((m->m_flags & M_TSTMP) ||
13941             (m->m_flags & M_TSTMP_LRO)) {
13942                 mbuf_tstmp2timespec(m, &ts);
13943                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13944                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13945         } else
13946                 rack->r_ctl.act_rcv_time = *tv;
13947         kern_prefetch(rack, &prev_state);
13948         prev_state = 0;
13949         /*
13950          * Unscale the window into a 32-bit value. For the SYN_SENT state
13951          * the scale is zero.
13952          */
13953         tiwin = th->th_win << tp->snd_scale;
13954 #ifdef TCP_ACCOUNTING
13955         if (thflags & TH_ACK) {
13956                 /*
13957                  * We have a tradeoff here. We can either do what we are
13958                  * doing i.e. pinning to this CPU and then doing the accounting
13959                  * <or> we could do a critical enter, setup the rdtsc and cpu
13960                  * as in below, and then validate we are on the same CPU on
13961                  * exit. I have choosen to not do the critical enter since
13962                  * that often will gain you a context switch, and instead lock
13963                  * us (line above this if) to the same CPU with sched_pin(). This
13964                  * means we may be context switched out for a higher priority
13965                  * interupt but we won't be moved to another CPU.
13966                  *
13967                  * If this occurs (which it won't very often since we most likely
13968                  * are running this code in interupt context and only a higher
13969                  * priority will bump us ... clock?) we will falsely add in
13970                  * to the time the interupt processing time plus the ack processing
13971                  * time. This is ok since its a rare event.
13972                  */
13973                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
13974                                                     ctf_fixed_maxseg(tp));
13975         }
13976 #endif
13977         /*
13978          * Parse options on any incoming segment.
13979          */
13980         memset(&to, 0, sizeof(to));
13981         tcp_dooptions(&to, (u_char *)(th + 1),
13982             (th->th_off << 2) - sizeof(struct tcphdr),
13983             (thflags & TH_SYN) ? TO_SYN : 0);
13984         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
13985             __func__));
13986         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
13987             __func__));
13988
13989         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13990             (tp->t_flags & TF_GPUTINPROG)) {
13991                 /*
13992                  * We have a goodput in progress
13993                  * and we have entered a late state.
13994                  * Do we have enough data in the sb
13995                  * to handle the GPUT request?
13996                  */
13997                 uint32_t bytes;
13998
13999                 bytes = tp->gput_ack - tp->gput_seq;
14000                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
14001                         bytes += tp->gput_seq - tp->snd_una;
14002                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
14003                         /*
14004                          * There are not enough bytes in the socket
14005                          * buffer that have been sent to cover this
14006                          * measurement. Cancel it.
14007                          */
14008                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14009                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
14010                                                    tp->gput_seq,
14011                                                    0, 0, 18, __LINE__, NULL, 0);
14012                         tp->t_flags &= ~TF_GPUTINPROG;
14013                 }
14014         }
14015         high_seq = th->th_ack;
14016         if (tcp_bblogging_on(rack->rc_tp)) {
14017                 union tcp_log_stackspecific log;
14018                 struct timeval ltv;
14019 #ifdef NETFLIX_HTTP_LOGGING
14020                 struct http_sendfile_track *http_req;
14021
14022                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
14023                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14024                 } else {
14025                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14026                 }
14027 #endif
14028                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14029                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
14030                 if (rack->rack_no_prr == 0)
14031                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14032                 else
14033                         log.u_bbr.flex1 = 0;
14034                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14035                 log.u_bbr.use_lt_bw <<= 1;
14036                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
14037                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14038                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14039                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14040                 log.u_bbr.flex3 = m->m_flags;
14041                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14042                 log.u_bbr.lost = thflags;
14043                 log.u_bbr.pacing_gain = 0x1;
14044 #ifdef TCP_ACCOUNTING
14045                 log.u_bbr.cwnd_gain = ack_val_set;
14046 #endif
14047                 log.u_bbr.flex7 = 2;
14048                 if (m->m_flags & M_TSTMP) {
14049                         /* Record the hardware timestamp if present */
14050                         mbuf_tstmp2timespec(m, &ts);
14051                         ltv.tv_sec = ts.tv_sec;
14052                         ltv.tv_usec = ts.tv_nsec / 1000;
14053                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14054                 } else if (m->m_flags & M_TSTMP_LRO) {
14055                         /* Record the LRO the arrival timestamp */
14056                         mbuf_tstmp2timespec(m, &ts);
14057                         ltv.tv_sec = ts.tv_sec;
14058                         ltv.tv_usec = ts.tv_nsec / 1000;
14059                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14060                 }
14061                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14062                 /* Log the rcv time */
14063                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14064 #ifdef NETFLIX_HTTP_LOGGING
14065                 log.u_bbr.applimited = tp->t_http_closed;
14066                 log.u_bbr.applimited <<= 8;
14067                 log.u_bbr.applimited |= tp->t_http_open;
14068                 log.u_bbr.applimited <<= 8;
14069                 log.u_bbr.applimited |= tp->t_http_req;
14070                 if (http_req) {
14071                         /* Copy out any client req info */
14072                         /* seconds */
14073                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14074                         /* useconds */
14075                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14076                         log.u_bbr.rttProp = http_req->timestamp;
14077                         log.u_bbr.cur_del_rate = http_req->start;
14078                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14079                                 log.u_bbr.flex8 |= 1;
14080                         } else {
14081                                 log.u_bbr.flex8 |= 2;
14082                                 log.u_bbr.bw_inuse = http_req->end;
14083                         }
14084                         log.u_bbr.flex6 = http_req->start_seq;
14085                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14086                                 log.u_bbr.flex8 |= 4;
14087                                 log.u_bbr.epoch = http_req->end_seq;
14088                         }
14089                 }
14090 #endif
14091                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14092                     tlen, &log, true, &ltv);
14093         }
14094         /* Remove ack required flag if set, we have one  */
14095         if (thflags & TH_ACK)
14096                 rack->rc_ack_required = 0;
14097         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14098                 way_out = 4;
14099                 retval = 0;
14100                 m_freem(m);
14101                 goto done_with_input;
14102         }
14103         /*
14104          * If a segment with the ACK-bit set arrives in the SYN-SENT state
14105          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14106          */
14107         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14108             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14109                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14110                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14111 #ifdef TCP_ACCOUNTING
14112                 sched_unpin();
14113 #endif
14114                 return (1);
14115         }
14116         /*
14117          * If timestamps were negotiated during SYN/ACK and a
14118          * segment without a timestamp is received, silently drop
14119          * the segment, unless it is a RST segment or missing timestamps are
14120          * tolerated.
14121          * See section 3.2 of RFC 7323.
14122          */
14123         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14124             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14125                 way_out = 5;
14126                 retval = 0;
14127                 m_freem(m);
14128                 goto done_with_input;
14129         }
14130
14131         /*
14132          * Segment received on connection. Reset idle time and keep-alive
14133          * timer. XXX: This should be done after segment validation to
14134          * ignore broken/spoofed segs.
14135          */
14136         if  (tp->t_idle_reduce &&
14137              (tp->snd_max == tp->snd_una) &&
14138              (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14139                 counter_u64_add(rack_input_idle_reduces, 1);
14140                 rack_cc_after_idle(rack, tp);
14141         }
14142         tp->t_rcvtime = ticks;
14143 #ifdef STATS
14144         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14145 #endif
14146         if (tiwin > rack->r_ctl.rc_high_rwnd)
14147                 rack->r_ctl.rc_high_rwnd = tiwin;
14148         /*
14149          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14150          * this to occur after we've validated the segment.
14151          */
14152         if (tcp_ecn_input_segment(tp, thflags, tlen,
14153             tcp_packets_this_ack(tp, th->th_ack),
14154             iptos))
14155                 rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
14156
14157         /*
14158          * If echoed timestamp is later than the current time, fall back to
14159          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14160          * were used when this connection was established.
14161          */
14162         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14163                 to.to_tsecr -= tp->ts_offset;
14164                 if (TSTMP_GT(to.to_tsecr, ms_cts))
14165                         to.to_tsecr = 0;
14166         }
14167
14168         /*
14169          * If its the first time in we need to take care of options and
14170          * verify we can do SACK for rack!
14171          */
14172         if (rack->r_state == 0) {
14173                 /* Should be init'd by rack_init() */
14174                 KASSERT(rack->rc_inp != NULL,
14175                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
14176                 if (rack->rc_inp == NULL) {
14177                         rack->rc_inp = inp;
14178                 }
14179
14180                 /*
14181                  * Process options only when we get SYN/ACK back. The SYN
14182                  * case for incoming connections is handled in tcp_syncache.
14183                  * According to RFC1323 the window field in a SYN (i.e., a
14184                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14185                  * this is traditional behavior, may need to be cleaned up.
14186                  */
14187                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14188                         /* Handle parallel SYN for ECN */
14189                         tcp_ecn_input_parallel_syn(tp, thflags, iptos);
14190                         if ((to.to_flags & TOF_SCALE) &&
14191                             (tp->t_flags & TF_REQ_SCALE)) {
14192                                 tp->t_flags |= TF_RCVD_SCALE;
14193                                 tp->snd_scale = to.to_wscale;
14194                         } else
14195                                 tp->t_flags &= ~TF_REQ_SCALE;
14196                         /*
14197                          * Initial send window.  It will be updated with the
14198                          * next incoming segment to the scaled value.
14199                          */
14200                         tp->snd_wnd = th->th_win;
14201                         rack_validate_fo_sendwin_up(tp, rack);
14202                         if ((to.to_flags & TOF_TS) &&
14203                             (tp->t_flags & TF_REQ_TSTMP)) {
14204                                 tp->t_flags |= TF_RCVD_TSTMP;
14205                                 tp->ts_recent = to.to_tsval;
14206                                 tp->ts_recent_age = cts;
14207                         } else
14208                                 tp->t_flags &= ~TF_REQ_TSTMP;
14209                         if (to.to_flags & TOF_MSS) {
14210                                 tcp_mss(tp, to.to_mss);
14211                         }
14212                         if ((tp->t_flags & TF_SACK_PERMIT) &&
14213                             (to.to_flags & TOF_SACKPERM) == 0)
14214                                 tp->t_flags &= ~TF_SACK_PERMIT;
14215                         if (IS_FASTOPEN(tp->t_flags)) {
14216                                 if (to.to_flags & TOF_FASTOPEN) {
14217                                         uint16_t mss;
14218
14219                                         if (to.to_flags & TOF_MSS)
14220                                                 mss = to.to_mss;
14221                                         else
14222                                                 if ((inp->inp_vflag & INP_IPV6) != 0)
14223                                                         mss = TCP6_MSS;
14224                                                 else
14225                                                         mss = TCP_MSS;
14226                                         tcp_fastopen_update_cache(tp, mss,
14227                                             to.to_tfo_len, to.to_tfo_cookie);
14228                                 } else
14229                                         tcp_fastopen_disable_path(tp);
14230                         }
14231                 }
14232                 /*
14233                  * At this point we are at the initial call. Here we decide
14234                  * if we are doing RACK or not. We do this by seeing if
14235                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
14236                  * The code now does do dup-ack counting so if you don't
14237                  * switch back you won't get rack & TLP, but you will still
14238                  * get this stack.
14239                  */
14240
14241                 if ((rack_sack_not_required == 0) &&
14242                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14243                         tcp_switch_back_to_default(tp);
14244                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14245                             tlen, iptos);
14246 #ifdef TCP_ACCOUNTING
14247                         sched_unpin();
14248 #endif
14249                         return (1);
14250                 }
14251                 tcp_set_hpts(inp);
14252                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14253         }
14254         if (thflags & TH_FIN)
14255                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14256         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14257         if ((rack->rc_gp_dyn_mul) &&
14258             (rack->use_fixed_rate == 0) &&
14259             (rack->rc_always_pace)) {
14260                 /* Check in on probertt */
14261                 rack_check_probe_rtt(rack, us_cts);
14262         }
14263         rack_clear_rate_sample(rack);
14264         if ((rack->forced_ack) &&
14265             ((tcp_get_flags(th) & TH_RST) == 0)) {
14266                 rack_handle_probe_response(rack, tiwin, us_cts);
14267         }
14268         /*
14269          * This is the one exception case where we set the rack state
14270          * always. All other times (timers etc) we must have a rack-state
14271          * set (so we assure we have done the checks above for SACK).
14272          */
14273         rack->r_ctl.rc_rcvtime = cts;
14274         if (rack->r_state != tp->t_state)
14275                 rack_set_state(tp, rack);
14276         if (SEQ_GT(th->th_ack, tp->snd_una) &&
14277             (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14278                 kern_prefetch(rsm, &prev_state);
14279         prev_state = rack->r_state;
14280         retval = (*rack->r_substate) (m, th, so,
14281             tp, &to, drop_hdrlen,
14282             tlen, tiwin, thflags, nxt_pkt, iptos);
14283         if (retval == 0) {
14284                 /*
14285                  * If retval is 1 the tcb is unlocked and most likely the tp
14286                  * is gone.
14287                  */
14288                 INP_WLOCK_ASSERT(inp);
14289                 if ((rack->rc_gp_dyn_mul) &&
14290                     (rack->rc_always_pace) &&
14291                     (rack->use_fixed_rate == 0) &&
14292                     rack->in_probe_rtt &&
14293                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
14294                         /*
14295                          * If we are going for target, lets recheck before
14296                          * we output.
14297                          */
14298                         rack_check_probe_rtt(rack, us_cts);
14299                 }
14300                 if (rack->set_pacing_done_a_iw == 0) {
14301                         /* How much has been acked? */
14302                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14303                                 /* We have enough to set in the pacing segment size */
14304                                 rack->set_pacing_done_a_iw = 1;
14305                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
14306                         }
14307                 }
14308                 tcp_rack_xmit_timer_commit(rack, tp);
14309 #ifdef TCP_ACCOUNTING
14310                 /*
14311                  * If we set the ack_val_se to what ack processing we are doing
14312                  * we also want to track how many cycles we burned. Note
14313                  * the bits after tcp_output we let be "free". This is because
14314                  * we are also tracking the tcp_output times as well. Note the
14315                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
14316                  * 0xf cannot be returned and is what we initialize it too to
14317                  * indicate we are not doing the tabulations.
14318                  */
14319                 if (ack_val_set != 0xf) {
14320                         uint64_t crtsc;
14321
14322                         crtsc = get_cyclecount();
14323                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14324                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14325                         }
14326                 }
14327 #endif
14328                 if (nxt_pkt == 0) {
14329                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14330 do_output_now:
14331                                 if (tcp_output(tp) < 0)
14332                                         return (1);
14333                                 did_out = 1;
14334                         }
14335                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14336                         rack_free_trim(rack);
14337                 }
14338                 /* Update any rounds needed */
14339                 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
14340                         union tcp_log_stackspecific log;
14341                         struct timeval tv;
14342
14343                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14344                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14345                         log.u_bbr.flex1 = high_seq;
14346                         log.u_bbr.flex2 = rack->r_ctl.roundends;
14347                         log.u_bbr.flex3 = rack->r_ctl.current_round;
14348                         log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
14349                         log.u_bbr.flex8 = 9;
14350                         tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
14351                                        0, &log, false, NULL, NULL, 0, &tv);
14352                 }
14353                 /*
14354                  * The draft (v3) calls for us to use SEQ_GEQ, but that
14355                  * causes issues when we are just going app limited. Lets
14356                  * instead use SEQ_GT <or> where its equal but more data
14357                  * is outstanding.
14358                  */
14359                 if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) ||
14360                     ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) {
14361                         rack->r_ctl.current_round++;
14362                         rack->r_ctl.roundends = tp->snd_max;
14363                         if (CC_ALGO(tp)->newround != NULL) {
14364                                 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
14365                         }
14366                 }
14367                 if ((nxt_pkt == 0) &&
14368                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14369                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
14370                      (tp->t_flags & TF_DELACK) ||
14371                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14372                       (tp->t_state <= TCPS_CLOSING)))) {
14373                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
14374                         if ((tp->snd_max == tp->snd_una) &&
14375                             ((tp->t_flags & TF_DELACK) == 0) &&
14376                             (tcp_in_hpts(rack->rc_inp)) &&
14377                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14378                                 /* keep alive not needed if we are hptsi output yet */
14379                                 ;
14380                         } else {
14381                                 int late = 0;
14382                                 if (tcp_in_hpts(inp)) {
14383                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14384                                                 us_cts = tcp_get_usecs(NULL);
14385                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14386                                                         rack->r_early = 1;
14387                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14388                                                 } else
14389                                                         late = 1;
14390                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14391                                         }
14392                                         tcp_hpts_remove(inp);
14393                                 }
14394                                 if (late && (did_out == 0)) {
14395                                         /*
14396                                          * We are late in the sending
14397                                          * and we did not call the output
14398                                          * (this probably should not happen).
14399                                          */
14400                                         goto do_output_now;
14401                                 }
14402                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14403                         }
14404                         way_out = 1;
14405                 } else if (nxt_pkt == 0) {
14406                         /* Do we have the correct timer running? */
14407                         rack_timer_audit(tp, rack, &so->so_snd);
14408                         way_out = 2;
14409                 }
14410         done_with_input:
14411                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14412                 if (did_out)
14413                         rack->r_wanted_output = 0;
14414         }
14415 #ifdef TCP_ACCOUNTING
14416         sched_unpin();
14417 #endif
14418         return (retval);
14419 }
14420
14421 void
14422 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14423     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14424 {
14425         struct timeval tv;
14426
14427         /* First lets see if we have old packets */
14428         if (tp->t_in_pkt) {
14429                 if (ctf_do_queued_segments(so, tp, 1)) {
14430                         m_freem(m);
14431                         return;
14432                 }
14433         }
14434         if (m->m_flags & M_TSTMP_LRO) {
14435                 mbuf_tstmp2timeval(m, &tv);
14436         } else {
14437                 /* Should not be should we kassert instead? */
14438                 tcp_get_usecs(&tv);
14439         }
14440         if (rack_do_segment_nounlock(m, th, so, tp,
14441                                      drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14442                 INP_WUNLOCK(tptoinpcb(tp));
14443         }
14444 }
14445
14446 struct rack_sendmap *
14447 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14448 {
14449         struct rack_sendmap *rsm = NULL;
14450         int32_t idx;
14451         uint32_t srtt = 0, thresh = 0, ts_low = 0;
14452
14453         /* Return the next guy to be re-transmitted */
14454         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14455                 return (NULL);
14456         }
14457         if (tp->t_flags & TF_SENTFIN) {
14458                 /* retran the end FIN? */
14459                 return (NULL);
14460         }
14461         /* ok lets look at this one */
14462         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14463         if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
14464                 return (rsm);
14465         }
14466         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14467                 goto check_it;
14468         }
14469         rsm = rack_find_lowest_rsm(rack);
14470         if (rsm == NULL) {
14471                 return (NULL);
14472         }
14473 check_it:
14474         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14475             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14476                 /*
14477                  * No sack so we automatically do the 3 strikes and
14478                  * retransmit (no rack timer would be started).
14479                  */
14480
14481                 return (rsm);
14482         }
14483         if (rsm->r_flags & RACK_ACKED) {
14484                 return (NULL);
14485         }
14486         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14487             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14488                 /* Its not yet ready */
14489                 return (NULL);
14490         }
14491         srtt = rack_grab_rtt(tp, rack);
14492         idx = rsm->r_rtr_cnt - 1;
14493         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14494         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14495         if ((tsused == ts_low) ||
14496             (TSTMP_LT(tsused, ts_low))) {
14497                 /* No time since sending */
14498                 return (NULL);
14499         }
14500         if ((tsused - ts_low) < thresh) {
14501                 /* It has not been long enough yet */
14502                 return (NULL);
14503         }
14504         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14505             ((rsm->r_flags & RACK_SACK_PASSED) &&
14506              (rack->sack_attack_disable == 0))) {
14507                 /*
14508                  * We have passed the dup-ack threshold <or>
14509                  * a SACK has indicated this is missing.
14510                  * Note that if you are a declared attacker
14511                  * it is only the dup-ack threshold that
14512                  * will cause retransmits.
14513                  */
14514                 /* log retransmit reason */
14515                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14516                 rack->r_fast_output = 0;
14517                 return (rsm);
14518         }
14519         return (NULL);
14520 }
14521
14522 static void
14523 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14524                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14525                            int line, struct rack_sendmap *rsm, uint8_t quality)
14526 {
14527         if (tcp_bblogging_on(rack->rc_tp)) {
14528                 union tcp_log_stackspecific log;
14529                 struct timeval tv;
14530
14531                 memset(&log, 0, sizeof(log));
14532                 log.u_bbr.flex1 = slot;
14533                 log.u_bbr.flex2 = len;
14534                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14535                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14536                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14537                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14538                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14539                 log.u_bbr.use_lt_bw <<= 1;
14540                 log.u_bbr.use_lt_bw |= rack->r_late;
14541                 log.u_bbr.use_lt_bw <<= 1;
14542                 log.u_bbr.use_lt_bw |= rack->r_early;
14543                 log.u_bbr.use_lt_bw <<= 1;
14544                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14545                 log.u_bbr.use_lt_bw <<= 1;
14546                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14547                 log.u_bbr.use_lt_bw <<= 1;
14548                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14549                 log.u_bbr.use_lt_bw <<= 1;
14550                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14551                 log.u_bbr.use_lt_bw <<= 1;
14552                 log.u_bbr.use_lt_bw |= rack->gp_ready;
14553                 log.u_bbr.pkt_epoch = line;
14554                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14555                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14556                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14557                 log.u_bbr.bw_inuse = bw_est;
14558                 log.u_bbr.delRate = bw;
14559                 if (rack->r_ctl.gp_bw == 0)
14560                         log.u_bbr.cur_del_rate = 0;
14561                 else
14562                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
14563                 log.u_bbr.rttProp = len_time;
14564                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14565                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14566                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14567                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14568                         /* We are in slow start */
14569                         log.u_bbr.flex7 = 1;
14570                 } else {
14571                         /* we are on congestion avoidance */
14572                         log.u_bbr.flex7 = 0;
14573                 }
14574                 log.u_bbr.flex8 = method;
14575                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14576                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14577                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14578                 log.u_bbr.cwnd_gain <<= 1;
14579                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14580                 log.u_bbr.cwnd_gain <<= 1;
14581                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14582                 log.u_bbr.bbr_substate = quality;
14583                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14584                     &rack->rc_inp->inp_socket->so_rcv,
14585                     &rack->rc_inp->inp_socket->so_snd,
14586                     BBR_LOG_HPTSI_CALC, 0,
14587                     0, &log, false, &tv);
14588         }
14589 }
14590
14591 static uint32_t
14592 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14593 {
14594         uint32_t new_tso, user_max;
14595
14596         user_max = rack->rc_user_set_max_segs * mss;
14597         if (rack->rc_force_max_seg) {
14598                 return (user_max);
14599         }
14600         if (rack->use_fixed_rate &&
14601             ((rack->r_ctl.crte == NULL) ||
14602              (bw != rack->r_ctl.crte->rate))) {
14603                 /* Use the user mss since we are not exactly matched */
14604                 return (user_max);
14605         }
14606         new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14607         if (new_tso > user_max)
14608                 new_tso = user_max;
14609         return (new_tso);
14610 }
14611
14612 static int32_t
14613 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)
14614 {
14615         uint64_t lentim, fill_bw;
14616
14617         /* Lets first see if we are full, if so continue with normal rate */
14618         rack->r_via_fill_cw = 0;
14619         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14620                 return (slot);
14621         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14622                 return (slot);
14623         if (rack->r_ctl.rc_last_us_rtt == 0)
14624                 return (slot);
14625         if (rack->rc_pace_fill_if_rttin_range &&
14626             (rack->r_ctl.rc_last_us_rtt >=
14627              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14628                 /* The rtt is huge, N * smallest, lets not fill */
14629                 return (slot);
14630         }
14631         /*
14632          * first lets calculate the b/w based on the last us-rtt
14633          * and the sndwnd.
14634          */
14635         fill_bw = rack->r_ctl.cwnd_to_use;
14636         /* Take the rwnd if its smaller */
14637         if (fill_bw > rack->rc_tp->snd_wnd)
14638                 fill_bw = rack->rc_tp->snd_wnd;
14639         if (rack->r_fill_less_agg) {
14640                 /*
14641                  * Now take away the inflight (this will reduce our
14642                  * aggressiveness and yeah, if we get that much out in 1RTT
14643                  * we will have had acks come back and still be behind).
14644                  */
14645                 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14646         }
14647         /* Now lets make it into a b/w */
14648         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14649         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14650         /* We are below the min b/w */
14651         if (non_paced)
14652                 *rate_wanted = fill_bw;
14653         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14654                 return (slot);
14655         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14656                 fill_bw = rack->r_ctl.bw_rate_cap;
14657         rack->r_via_fill_cw = 1;
14658         if (rack->r_rack_hw_rate_caps &&
14659             (rack->r_ctl.crte != NULL)) {
14660                 uint64_t high_rate;
14661
14662                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14663                 if (fill_bw > high_rate) {
14664                         /* We are capping bw at the highest rate table entry */
14665                         if (*rate_wanted > high_rate) {
14666                                 /* The original rate was also capped */
14667                                 rack->r_via_fill_cw = 0;
14668                         }
14669                         rack_log_hdwr_pacing(rack,
14670                                              fill_bw, high_rate, __LINE__,
14671                                              0, 3);
14672                         fill_bw = high_rate;
14673                         if (capped)
14674                                 *capped = 1;
14675                 }
14676         } else if ((rack->r_ctl.crte == NULL) &&
14677                    (rack->rack_hdrw_pacing == 0) &&
14678                    (rack->rack_hdw_pace_ena) &&
14679                    rack->r_rack_hw_rate_caps &&
14680                    (rack->rack_attempt_hdwr_pace == 0) &&
14681                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
14682                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14683                 /*
14684                  * Ok we may have a first attempt that is greater than our top rate
14685                  * lets check.
14686                  */
14687                 uint64_t high_rate;
14688
14689                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14690                 if (high_rate) {
14691                         if (fill_bw > high_rate) {
14692                                 fill_bw = high_rate;
14693                                 if (capped)
14694                                         *capped = 1;
14695                         }
14696                 }
14697         }
14698         /*
14699          * Ok fill_bw holds our mythical b/w to fill the cwnd
14700          * in a rtt, what does that time wise equate too?
14701          */
14702         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14703         lentim /= fill_bw;
14704         *rate_wanted = fill_bw;
14705         if (non_paced || (lentim < slot)) {
14706                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14707                                            0, lentim, 12, __LINE__, NULL, 0);
14708                 return ((int32_t)lentim);
14709         } else
14710                 return (slot);
14711 }
14712
14713 static int32_t
14714 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14715 {
14716         uint64_t srtt;
14717         int32_t slot = 0;
14718         int can_start_hw_pacing = 1;
14719         int err;
14720
14721         if (rack->rc_always_pace == 0) {
14722                 /*
14723                  * We use the most optimistic possible cwnd/srtt for
14724                  * sending calculations. This will make our
14725                  * calculation anticipate getting more through
14726                  * quicker then possible. But thats ok we don't want
14727                  * the peer to have a gap in data sending.
14728                  */
14729                 uint64_t cwnd, tr_perms = 0;
14730                 int32_t reduce = 0;
14731
14732         old_method:
14733                 /*
14734                  * We keep no precise pacing with the old method
14735                  * instead we use the pacer to mitigate bursts.
14736                  */
14737                 if (rack->r_ctl.rc_rack_min_rtt)
14738                         srtt = rack->r_ctl.rc_rack_min_rtt;
14739                 else
14740                         srtt = max(tp->t_srtt, 1);
14741                 if (rack->r_ctl.rc_rack_largest_cwnd)
14742                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14743                 else
14744                         cwnd = rack->r_ctl.cwnd_to_use;
14745                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14746                 tr_perms = (cwnd * 1000) / srtt;
14747                 if (tr_perms == 0) {
14748                         tr_perms = ctf_fixed_maxseg(tp);
14749                 }
14750                 /*
14751                  * Calculate how long this will take to drain, if
14752                  * the calculation comes out to zero, thats ok we
14753                  * will use send_a_lot to possibly spin around for
14754                  * more increasing tot_len_this_send to the point
14755                  * that its going to require a pace, or we hit the
14756                  * cwnd. Which in that case we are just waiting for
14757                  * a ACK.
14758                  */
14759                 slot = len / tr_perms;
14760                 /* Now do we reduce the time so we don't run dry? */
14761                 if (slot && rack_slot_reduction) {
14762                         reduce = (slot / rack_slot_reduction);
14763                         if (reduce < slot) {
14764                                 slot -= reduce;
14765                         } else
14766                                 slot = 0;
14767                 }
14768                 slot *= HPTS_USEC_IN_MSEC;
14769                 if (rack->rc_pace_to_cwnd) {
14770                         uint64_t rate_wanted = 0;
14771
14772                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14773                         rack->rc_ack_can_sendout_data = 1;
14774                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14775                 } else
14776                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14777         } else {
14778                 uint64_t bw_est, res, lentim, rate_wanted;
14779                 uint32_t orig_val, segs, oh;
14780                 int capped = 0;
14781                 int prev_fill;
14782
14783                 if ((rack->r_rr_config == 1) && rsm) {
14784                         return (rack->r_ctl.rc_min_to);
14785                 }
14786                 if (rack->use_fixed_rate) {
14787                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14788                 } else if ((rack->r_ctl.init_rate == 0) &&
14789 #ifdef NETFLIX_PEAKRATE
14790                            (rack->rc_tp->t_maxpeakrate == 0) &&
14791 #endif
14792                            (rack->r_ctl.gp_bw == 0)) {
14793                         /* no way to yet do an estimate */
14794                         bw_est = rate_wanted = 0;
14795                 } else {
14796                         bw_est = rack_get_bw(rack);
14797                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14798                 }
14799                 if ((bw_est == 0) || (rate_wanted == 0) ||
14800                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14801                         /*
14802                          * No way yet to make a b/w estimate or
14803                          * our raise is set incorrectly.
14804                          */
14805                         goto old_method;
14806                 }
14807                 /* We need to account for all the overheads */
14808                 segs = (len + segsiz - 1) / segsiz;
14809                 /*
14810                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
14811                  * and how much data we put in each packet. Yes this
14812                  * means we may be off if we are larger than 1500 bytes
14813                  * or smaller. But this just makes us more conservative.
14814                  */
14815                 if (rack_hw_rate_min &&
14816                     (bw_est < rack_hw_rate_min))
14817                         can_start_hw_pacing = 0;
14818                 if (ETHERNET_SEGMENT_SIZE > segsiz)
14819                         oh = ETHERNET_SEGMENT_SIZE - segsiz;
14820                 else
14821                         oh = 0;
14822                 segs *= oh;
14823                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14824                 res = lentim / rate_wanted;
14825                 slot = (uint32_t)res;
14826                 orig_val = rack->r_ctl.rc_pace_max_segs;
14827                 if (rack->r_ctl.crte == NULL) {
14828                         /*
14829                          * Only do this if we are not hardware pacing
14830                          * since if we are doing hw-pacing below we will
14831                          * set make a call after setting up or changing
14832                          * the rate.
14833                          */
14834                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14835                 } else if (rack->rc_inp->inp_snd_tag == NULL) {
14836                         /*
14837                          * We lost our rate somehow, this can happen
14838                          * if the interface changed underneath us.
14839                          */
14840                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14841                         rack->r_ctl.crte = NULL;
14842                         /* Lets re-allow attempting to setup pacing */
14843                         rack->rack_hdrw_pacing = 0;
14844                         rack->rack_attempt_hdwr_pace = 0;
14845                         rack_log_hdwr_pacing(rack,
14846                                              rate_wanted, bw_est, __LINE__,
14847                                              0, 6);
14848                 }
14849                 /* Did we change the TSO size, if so log it */
14850                 if (rack->r_ctl.rc_pace_max_segs != orig_val)
14851                         rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
14852                 prev_fill = rack->r_via_fill_cw;
14853                 if ((rack->rc_pace_to_cwnd) &&
14854                     (capped == 0) &&
14855                     (rack->use_fixed_rate == 0) &&
14856                     (rack->in_probe_rtt == 0) &&
14857                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14858                         /*
14859                          * We want to pace at our rate *or* faster to
14860                          * fill the cwnd to the max if its not full.
14861                          */
14862                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14863                 }
14864                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14865                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14866                         if ((rack->rack_hdw_pace_ena) &&
14867                             (can_start_hw_pacing > 0) &&
14868                             (rack->rack_hdrw_pacing == 0) &&
14869                             (rack->rack_attempt_hdwr_pace == 0)) {
14870                                 /*
14871                                  * Lets attempt to turn on hardware pacing
14872                                  * if we can.
14873                                  */
14874                                 rack->rack_attempt_hdwr_pace = 1;
14875                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14876                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
14877                                                                        rate_wanted,
14878                                                                        RS_PACING_GEQ,
14879                                                                        &err, &rack->r_ctl.crte_prev_rate);
14880                                 if (rack->r_ctl.crte) {
14881                                         rack->rack_hdrw_pacing = 1;
14882                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14883                                                                                                  0, rack->r_ctl.crte,
14884                                                                                                  NULL);
14885                                         rack_log_hdwr_pacing(rack,
14886                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14887                                                              err, 0);
14888                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14889                                 } else {
14890                                         counter_u64_add(rack_hw_pace_init_fail, 1);
14891                                 }
14892                         } else if (rack->rack_hdrw_pacing &&
14893                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14894                                 /* Do we need to adjust our rate? */
14895                                 const struct tcp_hwrate_limit_table *nrte;
14896
14897                                 if (rack->r_up_only &&
14898                                     (rate_wanted < rack->r_ctl.crte->rate)) {
14899                                         /**
14900                                          * We have four possible states here
14901                                          * having to do with the previous time
14902                                          * and this time.
14903                                          *   previous  |  this-time
14904                                          * A)     0      |     0   -- fill_cw not in the picture
14905                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
14906                                          * C)     1      |     1   -- all rates from fill_cw
14907                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
14908                                          *
14909                                          * For case A, C and D we don't allow a drop. But for
14910                                          * case B where we now our on our steady rate we do
14911                                          * allow a drop.
14912                                          *
14913                                          */
14914                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14915                                                 goto done_w_hdwr;
14916                                 }
14917                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
14918                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14919                                         if (rack_hw_rate_to_low &&
14920                                             (bw_est < rack_hw_rate_to_low)) {
14921                                                 /*
14922                                                  * The pacing rate is too low for hardware, but
14923                                                  * do allow hardware pacing to be restarted.
14924                                                  */
14925                                                 rack_log_hdwr_pacing(rack,
14926                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
14927                                                              0, 5);
14928                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14929                                                 rack->r_ctl.crte = NULL;
14930                                                 rack->rack_attempt_hdwr_pace = 0;
14931                                                 rack->rack_hdrw_pacing = 0;
14932                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14933                                                 goto done_w_hdwr;
14934                                         }
14935                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
14936                                                                    rack->rc_tp,
14937                                                                    rack->rc_inp->inp_route.ro_nh->nh_ifp,
14938                                                                    rate_wanted,
14939                                                                    RS_PACING_GEQ,
14940                                                                    &err, &rack->r_ctl.crte_prev_rate);
14941                                         if (nrte == NULL) {
14942                                                 /* Lost the rate */
14943                                                 rack->rack_hdrw_pacing = 0;
14944                                                 rack->r_ctl.crte = NULL;
14945                                                 rack_log_hdwr_pacing(rack,
14946                                                                      rate_wanted, 0, __LINE__,
14947                                                                      err, 1);
14948                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14949                                                 counter_u64_add(rack_hw_pace_lost, 1);
14950                                         } else if (nrte != rack->r_ctl.crte) {
14951                                                 rack->r_ctl.crte = nrte;
14952                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
14953                                                                                                          segsiz, 0,
14954                                                                                                          rack->r_ctl.crte,
14955                                                                                                          NULL);
14956                                                 rack_log_hdwr_pacing(rack,
14957                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14958                                                                      err, 2);
14959                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
14960                                         }
14961                                 } else {
14962                                         /* We just need to adjust the segment size */
14963                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14964                                         rack_log_hdwr_pacing(rack,
14965                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14966                                                              0, 4);
14967                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14968                                 }
14969                         }
14970                 }
14971                 if ((rack->r_ctl.crte != NULL) &&
14972                     (rack->r_ctl.crte->rate == rate_wanted)) {
14973                         /*
14974                          * We need to add a extra if the rates
14975                          * are exactly matched. The idea is
14976                          * we want the software to make sure the
14977                          * queue is empty before adding more, this
14978                          * gives us N MSS extra pace times where
14979                          * N is our sysctl
14980                          */
14981                         slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
14982                 }
14983 done_w_hdwr:
14984                 if (rack_limit_time_with_srtt &&
14985                     (rack->use_fixed_rate == 0) &&
14986 #ifdef NETFLIX_PEAKRATE
14987                     (rack->rc_tp->t_maxpeakrate == 0) &&
14988 #endif
14989                     (rack->rack_hdrw_pacing == 0)) {
14990                         /*
14991                          * Sanity check, we do not allow the pacing delay
14992                          * to be longer than the SRTT of the path. If it is
14993                          * a slow path, then adding a packet should increase
14994                          * the RTT and compensate for this i.e. the srtt will
14995                          * be greater so the allowed pacing time will be greater.
14996                          *
14997                          * Note this restriction is not for where a peak rate
14998                          * is set, we are doing fixed pacing or hardware pacing.
14999                          */
15000                         if (rack->rc_tp->t_srtt)
15001                                 srtt = rack->rc_tp->t_srtt;
15002                         else
15003                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
15004                         if (srtt < (uint64_t)slot) {
15005                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15006                                 slot = srtt;
15007                         }
15008                 }
15009                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15010         }
15011         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15012                 /*
15013                  * If this rate is seeing enobufs when it
15014                  * goes to send then either the nic is out
15015                  * of gas or we are mis-estimating the time
15016                  * somehow and not letting the queue empty
15017                  * completely. Lets add to the pacing time.
15018                  */
15019                 int hw_boost_delay;
15020
15021                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15022                 if (hw_boost_delay > rack_enobuf_hw_max)
15023                         hw_boost_delay = rack_enobuf_hw_max;
15024                 else if (hw_boost_delay < rack_enobuf_hw_min)
15025                         hw_boost_delay = rack_enobuf_hw_min;
15026                 slot += hw_boost_delay;
15027         }
15028         return (slot);
15029 }
15030
15031 static void
15032 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15033     tcp_seq startseq, uint32_t sb_offset)
15034 {
15035         struct rack_sendmap *my_rsm = NULL;
15036         struct rack_sendmap fe;
15037
15038         if (tp->t_state < TCPS_ESTABLISHED) {
15039                 /*
15040                  * We don't start any measurements if we are
15041                  * not at least established.
15042                  */
15043                 return;
15044         }
15045         if (tp->t_state >= TCPS_FIN_WAIT_1) {
15046                 /*
15047                  * We will get no more data into the SB
15048                  * this means we need to have the data available
15049                  * before we start a measurement.
15050                  */
15051
15052                 if (sbavail(&tptosocket(tp)->so_snd) <
15053                     max(rc_init_window(rack),
15054                         (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15055                         /* Nope not enough data */
15056                         return;
15057                 }
15058         }
15059         tp->t_flags |= TF_GPUTINPROG;
15060         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15061         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15062         tp->gput_seq = startseq;
15063         rack->app_limited_needs_set = 0;
15064         if (rack->in_probe_rtt)
15065                 rack->measure_saw_probe_rtt = 1;
15066         else if ((rack->measure_saw_probe_rtt) &&
15067                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15068                 rack->measure_saw_probe_rtt = 0;
15069         if (rack->rc_gp_filled)
15070                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15071         else {
15072                 /* Special case initial measurement */
15073                 struct timeval tv;
15074
15075                 tp->gput_ts = tcp_get_usecs(&tv);
15076                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15077         }
15078         /*
15079          * We take a guess out into the future,
15080          * if we have no measurement and no
15081          * initial rate, we measure the first
15082          * initial-windows worth of data to
15083          * speed up getting some GP measurement and
15084          * thus start pacing.
15085          */
15086         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15087                 rack->app_limited_needs_set = 1;
15088                 tp->gput_ack = startseq + max(rc_init_window(rack),
15089                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15090                 rack_log_pacing_delay_calc(rack,
15091                                            tp->gput_seq,
15092                                            tp->gput_ack,
15093                                            0,
15094                                            tp->gput_ts,
15095                                            rack->r_ctl.rc_app_limited_cnt,
15096                                            9,
15097                                            __LINE__, NULL, 0);
15098                 return;
15099         }
15100         if (sb_offset) {
15101                 /*
15102                  * We are out somewhere in the sb
15103                  * can we use the already outstanding data?
15104                  */
15105                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
15106                         /*
15107                          * Yes first one is good and in this case
15108                          * the tp->gput_ts is correctly set based on
15109                          * the last ack that arrived (no need to
15110                          * set things up when an ack comes in).
15111                          */
15112                         my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15113                         if ((my_rsm == NULL) ||
15114                             (my_rsm->r_rtr_cnt != 1)) {
15115                                 /* retransmission? */
15116                                 goto use_latest;
15117                         }
15118                 } else {
15119                         if (rack->r_ctl.rc_first_appl == NULL) {
15120                                 /*
15121                                  * If rc_first_appl is NULL
15122                                  * then the cnt should be 0.
15123                                  * This is probably an error, maybe
15124                                  * a KASSERT would be approprate.
15125                                  */
15126                                 goto use_latest;
15127                         }
15128                         /*
15129                          * If we have a marker pointer to the last one that is
15130                          * app limited we can use that, but we need to set
15131                          * things up so that when it gets ack'ed we record
15132                          * the ack time (if its not already acked).
15133                          */
15134                         rack->app_limited_needs_set = 1;
15135                         /*
15136                          * We want to get to the rsm that is either
15137                          * next with space i.e. over 1 MSS or the one
15138                          * after that (after the app-limited).
15139                          */
15140                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15141                                          rack->r_ctl.rc_first_appl);
15142                         if (my_rsm) {
15143                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15144                                         /* Have to use the next one */
15145                                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15146                                                          my_rsm);
15147                                 else {
15148                                         /* Use after the first MSS of it is acked */
15149                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15150                                         goto start_set;
15151                                 }
15152                         }
15153                         if ((my_rsm == NULL) ||
15154                             (my_rsm->r_rtr_cnt != 1)) {
15155                                 /*
15156                                  * Either its a retransmit or
15157                                  * the last is the app-limited one.
15158                                  */
15159                                 goto use_latest;
15160                         }
15161                 }
15162                 tp->gput_seq = my_rsm->r_start;
15163 start_set:
15164                 if (my_rsm->r_flags & RACK_ACKED) {
15165                         /*
15166                          * This one has been acked use the arrival ack time
15167                          */
15168                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15169                         rack->app_limited_needs_set = 0;
15170                 }
15171                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15172                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15173                 rack_log_pacing_delay_calc(rack,
15174                                            tp->gput_seq,
15175                                            tp->gput_ack,
15176                                            (uint64_t)my_rsm,
15177                                            tp->gput_ts,
15178                                            rack->r_ctl.rc_app_limited_cnt,
15179                                            9,
15180                                            __LINE__, NULL, 0);
15181                 return;
15182         }
15183
15184 use_latest:
15185         /*
15186          * We don't know how long we may have been
15187          * idle or if this is the first-send. Lets
15188          * setup the flag so we will trim off
15189          * the first ack'd data so we get a true
15190          * measurement.
15191          */
15192         rack->app_limited_needs_set = 1;
15193         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15194         /* Find this guy so we can pull the send time */
15195         fe.r_start = startseq;
15196         my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15197         if (my_rsm) {
15198                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15199                 if (my_rsm->r_flags & RACK_ACKED) {
15200                         /*
15201                          * Unlikely since its probably what was
15202                          * just transmitted (but I am paranoid).
15203                          */
15204                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15205                         rack->app_limited_needs_set = 0;
15206                 }
15207                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15208                         /* This also is unlikely */
15209                         tp->gput_seq = my_rsm->r_start;
15210                 }
15211         } else {
15212                 /*
15213                  * TSNH unless we have some send-map limit,
15214                  * and even at that it should not be hitting
15215                  * that limit (we should have stopped sending).
15216                  */
15217                 struct timeval tv;
15218
15219                 microuptime(&tv);
15220                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15221         }
15222         rack_log_pacing_delay_calc(rack,
15223                                    tp->gput_seq,
15224                                    tp->gput_ack,
15225                                    (uint64_t)my_rsm,
15226                                    tp->gput_ts,
15227                                    rack->r_ctl.rc_app_limited_cnt,
15228                                    9, __LINE__, NULL, 0);
15229 }
15230
15231 static inline uint32_t
15232 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15233     uint32_t avail, int32_t sb_offset)
15234 {
15235         uint32_t len;
15236         uint32_t sendwin;
15237
15238         if (tp->snd_wnd > cwnd_to_use)
15239                 sendwin = cwnd_to_use;
15240         else
15241                 sendwin = tp->snd_wnd;
15242         if (ctf_outstanding(tp) >= tp->snd_wnd) {
15243                 /* We never want to go over our peers rcv-window */
15244                 len = 0;
15245         } else {
15246                 uint32_t flight;
15247
15248                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15249                 if (flight >= sendwin) {
15250                         /*
15251                          * We have in flight what we are allowed by cwnd (if
15252                          * it was rwnd blocking it would have hit above out
15253                          * >= tp->snd_wnd).
15254                          */
15255                         return (0);
15256                 }
15257                 len = sendwin - flight;
15258                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15259                         /* We would send too much (beyond the rwnd) */
15260                         len = tp->snd_wnd - ctf_outstanding(tp);
15261                 }
15262                 if ((len + sb_offset) > avail) {
15263                         /*
15264                          * We don't have that much in the SB, how much is
15265                          * there?
15266                          */
15267                         len = avail - sb_offset;
15268                 }
15269         }
15270         return (len);
15271 }
15272
15273 static void
15274 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15275              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15276              int rsm_is_null, int optlen, int line, uint16_t mode)
15277 {
15278         if (tcp_bblogging_on(rack->rc_tp)) {
15279                 union tcp_log_stackspecific log;
15280                 struct timeval tv;
15281
15282                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15283                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15284                 log.u_bbr.flex1 = error;
15285                 log.u_bbr.flex2 = flags;
15286                 log.u_bbr.flex3 = rsm_is_null;
15287                 log.u_bbr.flex4 = ipoptlen;
15288                 log.u_bbr.flex5 = tp->rcv_numsacks;
15289                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15290                 log.u_bbr.flex7 = optlen;
15291                 log.u_bbr.flex8 = rack->r_fsb_inited;
15292                 log.u_bbr.applimited = rack->r_fast_output;
15293                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15294                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15295                 log.u_bbr.cwnd_gain = mode;
15296                 log.u_bbr.pkts_out = orig_len;
15297                 log.u_bbr.lt_epoch = len;
15298                 log.u_bbr.delivered = line;
15299                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15300                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15301                 tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15302                                len, &log, false, NULL, NULL, 0, &tv);
15303         }
15304 }
15305
15306
15307 static struct mbuf *
15308 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15309                    struct rack_fast_send_blk *fsb,
15310                    int32_t seglimit, int32_t segsize, int hw_tls)
15311 {
15312 #ifdef KERN_TLS
15313         struct ktls_session *tls, *ntls;
15314 #ifdef INVARIANTS
15315         struct mbuf *start;
15316 #endif
15317 #endif
15318         struct mbuf *m, *n, **np, *smb;
15319         struct mbuf *top;
15320         int32_t off, soff;
15321         int32_t len = *plen;
15322         int32_t fragsize;
15323         int32_t len_cp = 0;
15324         uint32_t mlen, frags;
15325
15326         soff = off = the_off;
15327         smb = m = the_m;
15328         np = &top;
15329         top = NULL;
15330 #ifdef KERN_TLS
15331         if (hw_tls && (m->m_flags & M_EXTPG))
15332                 tls = m->m_epg_tls;
15333         else
15334                 tls = NULL;
15335 #ifdef INVARIANTS
15336         start = m;
15337 #endif
15338 #endif
15339         while (len > 0) {
15340                 if (m == NULL) {
15341                         *plen = len_cp;
15342                         break;
15343                 }
15344 #ifdef KERN_TLS
15345                 if (hw_tls) {
15346                         if (m->m_flags & M_EXTPG)
15347                                 ntls = m->m_epg_tls;
15348                         else
15349                                 ntls = NULL;
15350
15351                         /*
15352                          * Avoid mixing TLS records with handshake
15353                          * data or TLS records from different
15354                          * sessions.
15355                          */
15356                         if (tls != ntls) {
15357                                 MPASS(m != start);
15358                                 *plen = len_cp;
15359                                 break;
15360                         }
15361                 }
15362 #endif
15363                 mlen = min(len, m->m_len - off);
15364                 if (seglimit) {
15365                         /*
15366                          * For M_EXTPG mbufs, add 3 segments
15367                          * + 1 in case we are crossing page boundaries
15368                          * + 2 in case the TLS hdr/trailer are used
15369                          * It is cheaper to just add the segments
15370                          * than it is to take the cache miss to look
15371                          * at the mbuf ext_pgs state in detail.
15372                          */
15373                         if (m->m_flags & M_EXTPG) {
15374                                 fragsize = min(segsize, PAGE_SIZE);
15375                                 frags = 3;
15376                         } else {
15377                                 fragsize = segsize;
15378                                 frags = 0;
15379                         }
15380
15381                         /* Break if we really can't fit anymore. */
15382                         if ((frags + 1) >= seglimit) {
15383                                 *plen = len_cp;
15384                                 break;
15385                         }
15386
15387                         /*
15388                          * Reduce size if you can't copy the whole
15389                          * mbuf. If we can't copy the whole mbuf, also
15390                          * adjust len so the loop will end after this
15391                          * mbuf.
15392                          */
15393                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15394                                 mlen = (seglimit - frags - 1) * fragsize;
15395                                 len = mlen;
15396                                 *plen = len_cp + len;
15397                         }
15398                         frags += howmany(mlen, fragsize);
15399                         if (frags == 0)
15400                                 frags++;
15401                         seglimit -= frags;
15402                         KASSERT(seglimit > 0,
15403                             ("%s: seglimit went too low", __func__));
15404                 }
15405                 n = m_get(M_NOWAIT, m->m_type);
15406                 *np = n;
15407                 if (n == NULL)
15408                         goto nospace;
15409                 n->m_len = mlen;
15410                 soff += mlen;
15411                 len_cp += n->m_len;
15412                 if (m->m_flags & (M_EXT|M_EXTPG)) {
15413                         n->m_data = m->m_data + off;
15414                         mb_dupcl(n, m);
15415                 } else {
15416                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15417                             (u_int)n->m_len);
15418                 }
15419                 len -= n->m_len;
15420                 off = 0;
15421                 m = m->m_next;
15422                 np = &n->m_next;
15423                 if (len || (soff == smb->m_len)) {
15424                         /*
15425                          * We have more so we move forward  or
15426                          * we have consumed the entire mbuf and
15427                          * len has fell to 0.
15428                          */
15429                         soff = 0;
15430                         smb = m;
15431                 }
15432
15433         }
15434         if (fsb != NULL) {
15435                 fsb->m = smb;
15436                 fsb->off = soff;
15437                 if (smb) {
15438                         /*
15439                          * Save off the size of the mbuf. We do
15440                          * this so that we can recognize when it
15441                          * has been trimmed by sbcut() as acks
15442                          * come in.
15443                          */
15444                         fsb->o_m_len = smb->m_len;
15445                 } else {
15446                         /*
15447                          * This is the case where the next mbuf went to NULL. This
15448                          * means with this copy we have sent everything in the sb.
15449                          * In theory we could clear the fast_output flag, but lets
15450                          * not since its possible that we could get more added
15451                          * and acks that call the extend function which would let
15452                          * us send more.
15453                          */
15454                         fsb->o_m_len = 0;
15455                 }
15456         }
15457         return (top);
15458 nospace:
15459         if (top)
15460                 m_freem(top);
15461         return (NULL);
15462
15463 }
15464
15465 /*
15466  * This is a copy of m_copym(), taking the TSO segment size/limit
15467  * constraints into account, and advancing the sndptr as it goes.
15468  */
15469 static struct mbuf *
15470 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15471                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15472 {
15473         struct mbuf *m, *n;
15474         int32_t soff;
15475
15476         soff = rack->r_ctl.fsb.off;
15477         m = rack->r_ctl.fsb.m;
15478         if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15479                 /*
15480                  * The mbuf had the front of it chopped off by an ack
15481                  * we need to adjust the soff/off by that difference.
15482                  */
15483                 uint32_t delta;
15484
15485                 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15486                 soff -= delta;
15487         } else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15488                 /*
15489                  * The mbuf was expanded probably by
15490                  * a m_compress. Just update o_m_len.
15491                  */
15492                 rack->r_ctl.fsb.o_m_len = m->m_len;
15493         }
15494         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15495         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15496         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15497                                  __FUNCTION__,
15498                                  rack, *plen, m, m->m_len));
15499         /* Save off the right location before we copy and advance */
15500         *s_soff = soff;
15501         *s_mb = rack->r_ctl.fsb.m;
15502         n = rack_fo_base_copym(m, soff, plen,
15503                                &rack->r_ctl.fsb,
15504                                seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15505         return (n);
15506 }
15507
15508 static int
15509 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15510                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15511 {
15512         /*
15513          * Enter the fast retransmit path. We are given that a sched_pin is
15514          * in place (if accounting is compliled in) and the cycle count taken
15515          * at the entry is in the ts_val. The concept her is that the rsm
15516          * now holds the mbuf offsets and such so we can directly transmit
15517          * without a lot of overhead, the len field is already set for
15518          * us to prohibit us from sending too much (usually its 1MSS).
15519          */
15520         struct ip *ip = NULL;
15521         struct udphdr *udp = NULL;
15522         struct tcphdr *th = NULL;
15523         struct mbuf *m = NULL;
15524         struct inpcb *inp;
15525         uint8_t *cpto;
15526         struct tcp_log_buffer *lgb;
15527 #ifdef TCP_ACCOUNTING
15528         uint64_t crtsc;
15529         int cnt_thru = 1;
15530 #endif
15531         struct tcpopt to;
15532         u_char opt[TCP_MAXOLEN];
15533         uint32_t hdrlen, optlen;
15534         int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
15535         uint16_t flags;
15536         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15537         uint32_t if_hw_tsomaxsegsize;
15538
15539 #ifdef INET6
15540         struct ip6_hdr *ip6 = NULL;
15541
15542         if (rack->r_is_v6) {
15543                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15544                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15545         } else
15546 #endif                          /* INET6 */
15547         {
15548                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15549                 hdrlen = sizeof(struct tcpiphdr);
15550         }
15551         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15552                 goto failed;
15553         }
15554         if (doing_tlp) {
15555                 /* Its a TLP add the flag, it may already be there but be sure */
15556                 rsm->r_flags |= RACK_TLP;
15557         } else {
15558                 /* If it was a TLP it is not not on this retransmit */
15559                 rsm->r_flags &= ~RACK_TLP;
15560         }
15561         startseq = rsm->r_start;
15562         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15563         inp = rack->rc_inp;
15564         to.to_flags = 0;
15565         flags = tcp_outflags[tp->t_state];
15566         if (flags & (TH_SYN|TH_RST)) {
15567                 goto failed;
15568         }
15569         if (rsm->r_flags & RACK_HAS_FIN) {
15570                 /* We can't send a FIN here */
15571                 goto failed;
15572         }
15573         if (flags & TH_FIN) {
15574                 /* We never send a FIN */
15575                 flags &= ~TH_FIN;
15576         }
15577         if (tp->t_flags & TF_RCVD_TSTMP) {
15578                 to.to_tsval = ms_cts + tp->ts_offset;
15579                 to.to_tsecr = tp->ts_recent;
15580                 to.to_flags = TOF_TS;
15581         }
15582         optlen = tcp_addoptions(&to, opt);
15583         hdrlen += optlen;
15584         udp = rack->r_ctl.fsb.udp;
15585         if (udp)
15586                 hdrlen += sizeof(struct udphdr);
15587         if (rack->r_ctl.rc_pace_max_segs)
15588                 max_val = rack->r_ctl.rc_pace_max_segs;
15589         else if (rack->rc_user_set_max_segs)
15590                 max_val = rack->rc_user_set_max_segs * segsiz;
15591         else
15592                 max_val = len;
15593         if ((tp->t_flags & TF_TSO) &&
15594             V_tcp_do_tso &&
15595             (len > segsiz) &&
15596             (tp->t_port == 0))
15597                 tso = 1;
15598 #ifdef INET6
15599         if (MHLEN < hdrlen + max_linkhdr)
15600                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15601         else
15602 #endif
15603                 m = m_gethdr(M_NOWAIT, MT_DATA);
15604         if (m == NULL)
15605                 goto failed;
15606         m->m_data += max_linkhdr;
15607         m->m_len = hdrlen;
15608         th = rack->r_ctl.fsb.th;
15609         /* Establish the len to send */
15610         if (len > max_val)
15611                 len = max_val;
15612         if ((tso) && (len + optlen > tp->t_maxseg)) {
15613                 uint32_t if_hw_tsomax;
15614                 int32_t max_len;
15615
15616                 /* extract TSO information */
15617                 if_hw_tsomax = tp->t_tsomax;
15618                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15619                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15620                 /*
15621                  * Check if we should limit by maximum payload
15622                  * length:
15623                  */
15624                 if (if_hw_tsomax != 0) {
15625                         /* compute maximum TSO length */
15626                         max_len = (if_hw_tsomax - hdrlen -
15627                                    max_linkhdr);
15628                         if (max_len <= 0) {
15629                                 goto failed;
15630                         } else if (len > max_len) {
15631                                 len = max_len;
15632                         }
15633                 }
15634                 if (len <= segsiz) {
15635                         /*
15636                          * In case there are too many small fragments don't
15637                          * use TSO:
15638                          */
15639                         tso = 0;
15640                 }
15641         } else {
15642                 tso = 0;
15643         }
15644         if ((tso == 0) && (len > segsiz))
15645                 len = segsiz;
15646         if ((len == 0) ||
15647             (len <= MHLEN - hdrlen - max_linkhdr)) {
15648                 goto failed;
15649         }
15650         th->th_seq = htonl(rsm->r_start);
15651         th->th_ack = htonl(tp->rcv_nxt);
15652         /*
15653          * The PUSH bit should only be applied
15654          * if the full retransmission is made. If
15655          * we are sending less than this is the
15656          * left hand edge and should not have
15657          * the PUSH bit.
15658          */
15659         if ((rsm->r_flags & RACK_HAD_PUSH) &&
15660             (len == (rsm->r_end - rsm->r_start)))
15661                 flags |= TH_PUSH;
15662         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15663         if (th->th_win == 0) {
15664                 tp->t_sndzerowin++;
15665                 tp->t_flags |= TF_RXWIN0SENT;
15666         } else
15667                 tp->t_flags &= ~TF_RXWIN0SENT;
15668         if (rsm->r_flags & RACK_TLP) {
15669                 /*
15670                  * TLP should not count in retran count, but
15671                  * in its own bin
15672                  */
15673                 counter_u64_add(rack_tlp_retran, 1);
15674                 counter_u64_add(rack_tlp_retran_bytes, len);
15675         } else {
15676                 tp->t_sndrexmitpack++;
15677                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15678                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15679         }
15680 #ifdef STATS
15681         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15682                                  len);
15683 #endif
15684         if (rsm->m == NULL)
15685                 goto failed;
15686         if (rsm->orig_m_len != rsm->m->m_len) {
15687                 /* Fix up the orig_m_len and possibly the mbuf offset */
15688                 rack_adjust_orig_mlen(rsm);
15689         }
15690         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15691         if (len <= segsiz) {
15692                 /*
15693                  * Must have ran out of mbufs for the copy
15694                  * shorten it to no longer need tso. Lets
15695                  * not put on sendalot since we are low on
15696                  * mbufs.
15697                  */
15698                 tso = 0;
15699         }
15700         if ((m->m_next == NULL) || (len <= 0)){
15701                 goto failed;
15702         }
15703         if (udp) {
15704                 if (rack->r_is_v6)
15705                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15706                 else
15707                         ulen = hdrlen + len - sizeof(struct ip);
15708                 udp->uh_ulen = htons(ulen);
15709         }
15710         m->m_pkthdr.rcvif = (struct ifnet *)0;
15711         if (TCPS_HAVERCVDSYN(tp->t_state) &&
15712             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
15713                 int ect = tcp_ecn_output_established(tp, &flags, len, true);
15714                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
15715                     (tp->t_flags2 & TF2_ECN_SND_ECE))
15716                     tp->t_flags2 &= ~TF2_ECN_SND_ECE;
15717 #ifdef INET6
15718                 if (rack->r_is_v6) {
15719                     ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
15720                     ip6->ip6_flow |= htonl(ect << 20);
15721                 }
15722                 else
15723 #endif
15724                 {
15725                     ip->ip_tos &= ~IPTOS_ECN_MASK;
15726                     ip->ip_tos |= ect;
15727                 }
15728         }
15729         tcp_set_flags(th, flags);
15730         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15731 #ifdef INET6
15732         if (rack->r_is_v6) {
15733                 if (tp->t_port) {
15734                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15735                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15736                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15737                         th->th_sum = htons(0);
15738                         UDPSTAT_INC(udps_opackets);
15739                 } else {
15740                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15741                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15742                         th->th_sum = in6_cksum_pseudo(ip6,
15743                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15744                                                       0);
15745                 }
15746         }
15747 #endif
15748 #if defined(INET6) && defined(INET)
15749         else
15750 #endif
15751 #ifdef INET
15752         {
15753                 if (tp->t_port) {
15754                         m->m_pkthdr.csum_flags = CSUM_UDP;
15755                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15756                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15757                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15758                         th->th_sum = htons(0);
15759                         UDPSTAT_INC(udps_opackets);
15760                 } else {
15761                         m->m_pkthdr.csum_flags = CSUM_TCP;
15762                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15763                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15764                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15765                                                                         IPPROTO_TCP + len + optlen));
15766                 }
15767                 /* IP version must be set here for ipv4/ipv6 checking later */
15768                 KASSERT(ip->ip_v == IPVERSION,
15769                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15770         }
15771 #endif
15772         if (tso) {
15773                 KASSERT(len > tp->t_maxseg - optlen,
15774                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15775                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15776                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15777         }
15778 #ifdef INET6
15779         if (rack->r_is_v6) {
15780                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15781                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15782                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15783                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15784                 else
15785                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15786         }
15787 #endif
15788 #if defined(INET) && defined(INET6)
15789         else
15790 #endif
15791 #ifdef INET
15792         {
15793                 ip->ip_len = htons(m->m_pkthdr.len);
15794                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15795                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15796                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15797                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15798                                 ip->ip_off |= htons(IP_DF);
15799                         }
15800                 } else {
15801                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15802                 }
15803         }
15804 #endif
15805         /* Time to copy in our header */
15806         cpto = mtod(m, uint8_t *);
15807         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15808         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15809         if (optlen) {
15810                 bcopy(opt, th + 1, optlen);
15811                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15812         } else {
15813                 th->th_off = sizeof(struct tcphdr) >> 2;
15814         }
15815         if (tcp_bblogging_on(rack->rc_tp)) {
15816                 union tcp_log_stackspecific log;
15817
15818                 if (rsm->r_flags & RACK_RWND_COLLAPSED) {
15819                         rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
15820                         counter_u64_add(rack_collapsed_win_rxt, 1);
15821                         counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
15822                 }
15823                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15824                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15825                 if (rack->rack_no_prr)
15826                         log.u_bbr.flex1 = 0;
15827                 else
15828                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15829                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15830                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15831                 log.u_bbr.flex4 = max_val;
15832                 log.u_bbr.flex5 = 0;
15833                 /* Save off the early/late values */
15834                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15835                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15836                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15837                 if (doing_tlp == 0)
15838                         log.u_bbr.flex8 = 1;
15839                 else
15840                         log.u_bbr.flex8 = 2;
15841                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15842                 log.u_bbr.flex7 = 55;
15843                 log.u_bbr.pkts_out = tp->t_maxseg;
15844                 log.u_bbr.timeStamp = cts;
15845                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15846                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15847                 log.u_bbr.delivered = 0;
15848                 lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15849                                      len, &log, false, NULL, NULL, 0, tv);
15850         } else
15851                 lgb = NULL;
15852 #ifdef INET6
15853         if (rack->r_is_v6) {
15854                 error = ip6_output(m, NULL,
15855                                    &inp->inp_route6,
15856                                    0, NULL, NULL, inp);
15857         }
15858         else
15859 #endif
15860 #ifdef INET
15861         {
15862                 error = ip_output(m, NULL,
15863                                   &inp->inp_route,
15864                                   0, 0, inp);
15865         }
15866 #endif
15867         m = NULL;
15868         if (lgb) {
15869                 lgb->tlb_errno = error;
15870                 lgb = NULL;
15871         }
15872         if (error) {
15873                 goto failed;
15874         }
15875         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15876                         rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
15877         if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15878                 rack->rc_tlp_in_progress = 1;
15879                 rack->r_ctl.rc_tlp_cnt_out++;
15880         }
15881         if (error == 0) {
15882                 tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
15883                 if (doing_tlp) {
15884                         rack->rc_last_sent_tlp_past_cumack = 0;
15885                         rack->rc_last_sent_tlp_seq_valid = 1;
15886                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
15887                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
15888                 }
15889         }
15890         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15891         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15892         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15893                 rack->r_ctl.retran_during_recovery += len;
15894         {
15895                 int idx;
15896
15897                 idx = (len / segsiz) + 3;
15898                 if (idx >= TCP_MSS_ACCT_ATIMER)
15899                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15900                 else
15901                         counter_u64_add(rack_out_size[idx], 1);
15902         }
15903         if (tp->t_rtttime == 0) {
15904                 tp->t_rtttime = ticks;
15905                 tp->t_rtseq = startseq;
15906                 KMOD_TCPSTAT_INC(tcps_segstimed);
15907         }
15908         counter_u64_add(rack_fto_rsm_send, 1);
15909         if (error && (error == ENOBUFS)) {
15910                 if (rack->r_ctl.crte != NULL) {
15911                         tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
15912                 } else
15913                         tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
15914                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15915                 if (rack->rc_enobuf < 0x7f)
15916                         rack->rc_enobuf++;
15917                 if (slot < (10 * HPTS_USEC_IN_MSEC))
15918                         slot = 10 * HPTS_USEC_IN_MSEC;
15919         } else
15920                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
15921         if ((slot == 0) ||
15922             (rack->rc_always_pace == 0) ||
15923             (rack->r_rr_config == 1)) {
15924                 /*
15925                  * We have no pacing set or we
15926                  * are using old-style rack or
15927                  * we are overridden to use the old 1ms pacing.
15928                  */
15929                 slot = rack->r_ctl.rc_min_to;
15930         }
15931         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
15932 #ifdef TCP_ACCOUNTING
15933         crtsc = get_cyclecount();
15934         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15935                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15936         }
15937         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15938                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15939         }
15940         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15941                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
15942         }
15943         sched_unpin();
15944 #endif
15945         return (0);
15946 failed:
15947         if (m)
15948                 m_free(m);
15949         return (-1);
15950 }
15951
15952 static void
15953 rack_sndbuf_autoscale(struct tcp_rack *rack)
15954 {
15955         /*
15956          * Automatic sizing of send socket buffer.  Often the send buffer
15957          * size is not optimally adjusted to the actual network conditions
15958          * at hand (delay bandwidth product).  Setting the buffer size too
15959          * small limits throughput on links with high bandwidth and high
15960          * delay (eg. trans-continental/oceanic links).  Setting the
15961          * buffer size too big consumes too much real kernel memory,
15962          * especially with many connections on busy servers.
15963          *
15964          * The criteria to step up the send buffer one notch are:
15965          *  1. receive window of remote host is larger than send buffer
15966          *     (with a fudge factor of 5/4th);
15967          *  2. send buffer is filled to 7/8th with data (so we actually
15968          *     have data to make use of it);
15969          *  3. send buffer fill has not hit maximal automatic size;
15970          *  4. our send window (slow start and cogestion controlled) is
15971          *     larger than sent but unacknowledged data in send buffer.
15972          *
15973          * Note that the rack version moves things much faster since
15974          * we want to avoid hitting cache lines in the rack_fast_output()
15975          * path so this is called much less often and thus moves
15976          * the SB forward by a percentage.
15977          */
15978         struct socket *so;
15979         struct tcpcb *tp;
15980         uint32_t sendwin, scaleup;
15981
15982         tp = rack->rc_tp;
15983         so = rack->rc_inp->inp_socket;
15984         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
15985         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
15986                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
15987                     sbused(&so->so_snd) >=
15988                     (so->so_snd.sb_hiwat / 8 * 7) &&
15989                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
15990                     sendwin >= (sbused(&so->so_snd) -
15991                     (tp->snd_nxt - tp->snd_una))) {
15992                         if (rack_autosndbuf_inc)
15993                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
15994                         else
15995                                 scaleup = V_tcp_autosndbuf_inc;
15996                         if (scaleup < V_tcp_autosndbuf_inc)
15997                                 scaleup = V_tcp_autosndbuf_inc;
15998                         scaleup += so->so_snd.sb_hiwat;
15999                         if (scaleup > V_tcp_autosndbuf_max)
16000                                 scaleup = V_tcp_autosndbuf_max;
16001                         if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
16002                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16003                 }
16004         }
16005 }
16006
16007 static int
16008 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16009                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16010 {
16011         /*
16012          * Enter to do fast output. We are given that the sched_pin is
16013          * in place (if accounting is compiled in) and the cycle count taken
16014          * at entry is in place in ts_val. The idea here is that
16015          * we know how many more bytes needs to be sent (presumably either
16016          * during pacing or to fill the cwnd and that was greater than
16017          * the max-burst). We have how much to send and all the info we
16018          * need to just send.
16019          */
16020 #ifdef INET
16021         struct ip *ip = NULL;
16022 #endif
16023         struct udphdr *udp = NULL;
16024         struct tcphdr *th = NULL;
16025         struct mbuf *m, *s_mb;
16026         struct inpcb *inp;
16027         uint8_t *cpto;
16028         struct tcp_log_buffer *lgb;
16029 #ifdef TCP_ACCOUNTING
16030         uint64_t crtsc;
16031 #endif
16032         struct tcpopt to;
16033         u_char opt[TCP_MAXOLEN];
16034         uint32_t hdrlen, optlen;
16035 #ifdef TCP_ACCOUNTING
16036         int cnt_thru = 1;
16037 #endif
16038         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
16039         uint16_t flags;
16040         uint32_t s_soff;
16041         uint32_t if_hw_tsomaxsegcount = 0, startseq;
16042         uint32_t if_hw_tsomaxsegsize;
16043         uint16_t add_flag = RACK_SENT_FP;
16044 #ifdef INET6
16045         struct ip6_hdr *ip6 = NULL;
16046
16047         if (rack->r_is_v6) {
16048                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16049                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16050         } else
16051 #endif                          /* INET6 */
16052         {
16053 #ifdef INET
16054                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16055                 hdrlen = sizeof(struct tcpiphdr);
16056 #endif
16057         }
16058         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16059                 m = NULL;
16060                 goto failed;
16061         }
16062         startseq = tp->snd_max;
16063         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16064         inp = rack->rc_inp;
16065         len = rack->r_ctl.fsb.left_to_send;
16066         to.to_flags = 0;
16067         flags = rack->r_ctl.fsb.tcp_flags;
16068         if (tp->t_flags & TF_RCVD_TSTMP) {
16069                 to.to_tsval = ms_cts + tp->ts_offset;
16070                 to.to_tsecr = tp->ts_recent;
16071                 to.to_flags = TOF_TS;
16072         }
16073         optlen = tcp_addoptions(&to, opt);
16074         hdrlen += optlen;
16075         udp = rack->r_ctl.fsb.udp;
16076         if (udp)
16077                 hdrlen += sizeof(struct udphdr);
16078         if (rack->r_ctl.rc_pace_max_segs)
16079                 max_val = rack->r_ctl.rc_pace_max_segs;
16080         else if (rack->rc_user_set_max_segs)
16081                 max_val = rack->rc_user_set_max_segs * segsiz;
16082         else
16083                 max_val = len;
16084         if ((tp->t_flags & TF_TSO) &&
16085             V_tcp_do_tso &&
16086             (len > segsiz) &&
16087             (tp->t_port == 0))
16088                 tso = 1;
16089 again:
16090 #ifdef INET6
16091         if (MHLEN < hdrlen + max_linkhdr)
16092                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16093         else
16094 #endif
16095                 m = m_gethdr(M_NOWAIT, MT_DATA);
16096         if (m == NULL)
16097                 goto failed;
16098         m->m_data += max_linkhdr;
16099         m->m_len = hdrlen;
16100         th = rack->r_ctl.fsb.th;
16101         /* Establish the len to send */
16102         if (len > max_val)
16103                 len = max_val;
16104         if ((tso) && (len + optlen > tp->t_maxseg)) {
16105                 uint32_t if_hw_tsomax;
16106                 int32_t max_len;
16107
16108                 /* extract TSO information */
16109                 if_hw_tsomax = tp->t_tsomax;
16110                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16111                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16112                 /*
16113                  * Check if we should limit by maximum payload
16114                  * length:
16115                  */
16116                 if (if_hw_tsomax != 0) {
16117                         /* compute maximum TSO length */
16118                         max_len = (if_hw_tsomax - hdrlen -
16119                                    max_linkhdr);
16120                         if (max_len <= 0) {
16121                                 goto failed;
16122                         } else if (len > max_len) {
16123                                 len = max_len;
16124                         }
16125                 }
16126                 if (len <= segsiz) {
16127                         /*
16128                          * In case there are too many small fragments don't
16129                          * use TSO:
16130                          */
16131                         tso = 0;
16132                 }
16133         } else {
16134                 tso = 0;
16135         }
16136         if ((tso == 0) && (len > segsiz))
16137                 len = segsiz;
16138         if ((len == 0) ||
16139             (len <= MHLEN - hdrlen - max_linkhdr)) {
16140                 goto failed;
16141         }
16142         sb_offset = tp->snd_max - tp->snd_una;
16143         th->th_seq = htonl(tp->snd_max);
16144         th->th_ack = htonl(tp->rcv_nxt);
16145         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16146         if (th->th_win == 0) {
16147                 tp->t_sndzerowin++;
16148                 tp->t_flags |= TF_RXWIN0SENT;
16149         } else
16150                 tp->t_flags &= ~TF_RXWIN0SENT;
16151         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
16152         KMOD_TCPSTAT_INC(tcps_sndpack);
16153         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16154 #ifdef STATS
16155         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16156                                  len);
16157 #endif
16158         if (rack->r_ctl.fsb.m == NULL)
16159                 goto failed;
16160
16161         /* s_mb and s_soff are saved for rack_log_output */
16162         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16163                                     &s_mb, &s_soff);
16164         if (len <= segsiz) {
16165                 /*
16166                  * Must have ran out of mbufs for the copy
16167                  * shorten it to no longer need tso. Lets
16168                  * not put on sendalot since we are low on
16169                  * mbufs.
16170                  */
16171                 tso = 0;
16172         }
16173         if (rack->r_ctl.fsb.rfo_apply_push &&
16174             (len == rack->r_ctl.fsb.left_to_send)) {
16175                 flags |= TH_PUSH;
16176                 add_flag |= RACK_HAD_PUSH;
16177         }
16178         if ((m->m_next == NULL) || (len <= 0)){
16179                 goto failed;
16180         }
16181         if (udp) {
16182                 if (rack->r_is_v6)
16183                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
16184                 else
16185                         ulen = hdrlen + len - sizeof(struct ip);
16186                 udp->uh_ulen = htons(ulen);
16187         }
16188         m->m_pkthdr.rcvif = (struct ifnet *)0;
16189         if (TCPS_HAVERCVDSYN(tp->t_state) &&
16190             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
16191                 int ect = tcp_ecn_output_established(tp, &flags, len, false);
16192                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
16193                     (tp->t_flags2 & TF2_ECN_SND_ECE))
16194                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
16195 #ifdef INET6
16196                 if (rack->r_is_v6) {
16197                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
16198                         ip6->ip6_flow |= htonl(ect << 20);
16199                 }
16200                 else
16201 #endif
16202                 {
16203 #ifdef INET
16204                         ip->ip_tos &= ~IPTOS_ECN_MASK;
16205                         ip->ip_tos |= ect;
16206 #endif
16207                 }
16208         }
16209         tcp_set_flags(th, flags);
16210         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
16211 #ifdef INET6
16212         if (rack->r_is_v6) {
16213                 if (tp->t_port) {
16214                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16215                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16216                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16217                         th->th_sum = htons(0);
16218                         UDPSTAT_INC(udps_opackets);
16219                 } else {
16220                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16221                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16222                         th->th_sum = in6_cksum_pseudo(ip6,
16223                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16224                                                       0);
16225                 }
16226         }
16227 #endif
16228 #if defined(INET6) && defined(INET)
16229         else
16230 #endif
16231 #ifdef INET
16232         {
16233                 if (tp->t_port) {
16234                         m->m_pkthdr.csum_flags = CSUM_UDP;
16235                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16236                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16237                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16238                         th->th_sum = htons(0);
16239                         UDPSTAT_INC(udps_opackets);
16240                 } else {
16241                         m->m_pkthdr.csum_flags = CSUM_TCP;
16242                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16243                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
16244                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16245                                                                         IPPROTO_TCP + len + optlen));
16246                 }
16247                 /* IP version must be set here for ipv4/ipv6 checking later */
16248                 KASSERT(ip->ip_v == IPVERSION,
16249                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
16250         }
16251 #endif
16252         if (tso) {
16253                 KASSERT(len > tp->t_maxseg - optlen,
16254                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
16255                 m->m_pkthdr.csum_flags |= CSUM_TSO;
16256                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16257         }
16258 #ifdef INET6
16259         if (rack->r_is_v6) {
16260                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16261                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16262                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16263                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16264                 else
16265                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16266         }
16267 #endif
16268 #if defined(INET) && defined(INET6)
16269         else
16270 #endif
16271 #ifdef INET
16272         {
16273                 ip->ip_len = htons(m->m_pkthdr.len);
16274                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16275                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16276                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16277                         if (tp->t_port == 0 || len < V_tcp_minmss) {
16278                                 ip->ip_off |= htons(IP_DF);
16279                         }
16280                 } else {
16281                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16282                 }
16283         }
16284 #endif
16285         /* Time to copy in our header */
16286         cpto = mtod(m, uint8_t *);
16287         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16288         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16289         if (optlen) {
16290                 bcopy(opt, th + 1, optlen);
16291                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16292         } else {
16293                 th->th_off = sizeof(struct tcphdr) >> 2;
16294         }
16295         if (tcp_bblogging_on(rack->rc_tp)) {
16296                 union tcp_log_stackspecific log;
16297
16298                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16299                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16300                 if (rack->rack_no_prr)
16301                         log.u_bbr.flex1 = 0;
16302                 else
16303                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16304                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16305                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16306                 log.u_bbr.flex4 = max_val;
16307                 log.u_bbr.flex5 = 0;
16308                 /* Save off the early/late values */
16309                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16310                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16311                 log.u_bbr.bw_inuse = rack_get_bw(rack);
16312                 log.u_bbr.flex8 = 0;
16313                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16314                 log.u_bbr.flex7 = 44;
16315                 log.u_bbr.pkts_out = tp->t_maxseg;
16316                 log.u_bbr.timeStamp = cts;
16317                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16318                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16319                 log.u_bbr.delivered = 0;
16320                 lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16321                                      len, &log, false, NULL, NULL, 0, tv);
16322         } else
16323                 lgb = NULL;
16324 #ifdef INET6
16325         if (rack->r_is_v6) {
16326                 error = ip6_output(m, NULL,
16327                                    &inp->inp_route6,
16328                                    0, NULL, NULL, inp);
16329         }
16330 #endif
16331 #if defined(INET) && defined(INET6)
16332         else
16333 #endif
16334 #ifdef INET
16335         {
16336                 error = ip_output(m, NULL,
16337                                   &inp->inp_route,
16338                                   0, 0, inp);
16339         }
16340 #endif
16341         if (lgb) {
16342                 lgb->tlb_errno = error;
16343                 lgb = NULL;
16344         }
16345         if (error) {
16346                 *send_err = error;
16347                 m = NULL;
16348                 goto failed;
16349         }
16350         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16351                         NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16352         m = NULL;
16353         if (tp->snd_una == tp->snd_max) {
16354                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
16355                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16356                 tp->t_acktime = ticks;
16357         }
16358         if (error == 0)
16359                 tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16360
16361         rack->forced_ack = 0;   /* If we send something zap the FA flag */
16362         tot_len += len;
16363         if ((tp->t_flags & TF_GPUTINPROG) == 0)
16364                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16365         tp->snd_max += len;
16366         tp->snd_nxt = tp->snd_max;
16367         {
16368                 int idx;
16369
16370                 idx = (len / segsiz) + 3;
16371                 if (idx >= TCP_MSS_ACCT_ATIMER)
16372                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16373                 else
16374                         counter_u64_add(rack_out_size[idx], 1);
16375         }
16376         if (len <= rack->r_ctl.fsb.left_to_send)
16377                 rack->r_ctl.fsb.left_to_send -= len;
16378         else
16379                 rack->r_ctl.fsb.left_to_send = 0;
16380         if (rack->r_ctl.fsb.left_to_send < segsiz) {
16381                 rack->r_fast_output = 0;
16382                 rack->r_ctl.fsb.left_to_send = 0;
16383                 /* At the end of fast_output scale up the sb */
16384                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16385                 rack_sndbuf_autoscale(rack);
16386                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16387         }
16388         if (tp->t_rtttime == 0) {
16389                 tp->t_rtttime = ticks;
16390                 tp->t_rtseq = startseq;
16391                 KMOD_TCPSTAT_INC(tcps_segstimed);
16392         }
16393         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16394             (max_val > len) &&
16395             (tso == 0)) {
16396                 max_val -= len;
16397                 len = segsiz;
16398                 th = rack->r_ctl.fsb.th;
16399 #ifdef TCP_ACCOUNTING
16400                 cnt_thru++;
16401 #endif
16402                 goto again;
16403         }
16404         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16405         counter_u64_add(rack_fto_send, 1);
16406         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16407         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16408 #ifdef TCP_ACCOUNTING
16409         crtsc = get_cyclecount();
16410         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16411                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16412         }
16413         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16414                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16415         }
16416         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16417                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16418         }
16419         sched_unpin();
16420 #endif
16421         return (0);
16422 failed:
16423         if (m)
16424                 m_free(m);
16425         rack->r_fast_output = 0;
16426         return (-1);
16427 }
16428
16429 static struct rack_sendmap *
16430 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
16431 {
16432         struct rack_sendmap *rsm = NULL;
16433         struct rack_sendmap fe;
16434         int thresh;
16435
16436 restart:
16437         fe.r_start = rack->r_ctl.last_collapse_point;
16438         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
16439         if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
16440                 /* Nothing, strange turn off validity  */
16441                 rack->r_collapse_point_valid = 0;
16442                 return (NULL);
16443         }
16444         /* Can we send it yet? */
16445         if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
16446                 /*
16447                  * Receiver window has not grown enough for
16448                  * the segment to be put on the wire.
16449                  */
16450                 return (NULL);
16451         }
16452         if (rsm->r_flags & RACK_ACKED) {
16453                 /*
16454                  * It has been sacked, lets move to the
16455                  * next one if possible.
16456                  */
16457                 rack->r_ctl.last_collapse_point = rsm->r_end;
16458                 /* Are we done? */
16459                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16460                             rack->r_ctl.high_collapse_point)) {
16461                         rack->r_collapse_point_valid = 0;
16462                         return (NULL);
16463                 }
16464                 goto restart;
16465         }
16466         /* Now has it been long enough ? */
16467         thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
16468         if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
16469                 rack_log_collapse(rack, rsm->r_start,
16470                                   (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16471                                   thresh, __LINE__, 6, rsm->r_flags, rsm);
16472                 return (rsm);
16473         }
16474         /* Not enough time */
16475         rack_log_collapse(rack, rsm->r_start,
16476                           (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16477                           thresh, __LINE__, 7, rsm->r_flags, rsm);
16478         return (NULL);
16479 }
16480
16481 static int
16482 rack_output(struct tcpcb *tp)
16483 {
16484         struct socket *so;
16485         uint32_t recwin;
16486         uint32_t sb_offset, s_moff = 0;
16487         int32_t len, error = 0;
16488         uint16_t flags;
16489         struct mbuf *m, *s_mb = NULL;
16490         struct mbuf *mb;
16491         uint32_t if_hw_tsomaxsegcount = 0;
16492         uint32_t if_hw_tsomaxsegsize;
16493         int32_t segsiz, minseg;
16494         long tot_len_this_send = 0;
16495 #ifdef INET
16496         struct ip *ip = NULL;
16497 #endif
16498         struct udphdr *udp = NULL;
16499         struct tcp_rack *rack;
16500         struct tcphdr *th;
16501         uint8_t pass = 0;
16502         uint8_t mark = 0;
16503         uint8_t wanted_cookie = 0;
16504         u_char opt[TCP_MAXOLEN];
16505         unsigned ipoptlen, optlen, hdrlen, ulen=0;
16506         uint32_t rack_seq;
16507
16508 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16509         unsigned ipsec_optlen = 0;
16510
16511 #endif
16512         int32_t idle, sendalot;
16513         int32_t sub_from_prr = 0;
16514         volatile int32_t sack_rxmit;
16515         struct rack_sendmap *rsm = NULL;
16516         int32_t tso, mtu;
16517         struct tcpopt to;
16518         int32_t slot = 0;
16519         int32_t sup_rack = 0;
16520         uint32_t cts, ms_cts, delayed, early;
16521         uint16_t add_flag = RACK_SENT_SP;
16522         /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16523         uint8_t hpts_calling,  doing_tlp = 0;
16524         uint32_t cwnd_to_use, pace_max_seg;
16525         int32_t do_a_prefetch = 0;
16526         int32_t prefetch_rsm = 0;
16527         int32_t orig_len = 0;
16528         struct timeval tv;
16529         int32_t prefetch_so_done = 0;
16530         struct tcp_log_buffer *lgb;
16531         struct inpcb *inp = tptoinpcb(tp);
16532         struct sockbuf *sb;
16533         uint64_t ts_val = 0;
16534 #ifdef TCP_ACCOUNTING
16535         uint64_t crtsc;
16536 #endif
16537 #ifdef INET6
16538         struct ip6_hdr *ip6 = NULL;
16539         int32_t isipv6;
16540 #endif
16541         bool hw_tls = false;
16542
16543         NET_EPOCH_ASSERT();
16544         INP_WLOCK_ASSERT(inp);
16545
16546         /* setup and take the cache hits here */
16547         rack = (struct tcp_rack *)tp->t_fb_ptr;
16548 #ifdef TCP_ACCOUNTING
16549         sched_pin();
16550         ts_val = get_cyclecount();
16551 #endif
16552         hpts_calling = inp->inp_hpts_calls;
16553 #ifdef TCP_OFFLOAD
16554         if (tp->t_flags & TF_TOE) {
16555 #ifdef TCP_ACCOUNTING
16556                 sched_unpin();
16557 #endif
16558                 return (tcp_offload_output(tp));
16559         }
16560 #endif
16561         /*
16562          * For TFO connections in SYN_RECEIVED, only allow the initial
16563          * SYN|ACK and those sent by the retransmit timer.
16564          */
16565         if (IS_FASTOPEN(tp->t_flags) &&
16566             (tp->t_state == TCPS_SYN_RECEIVED) &&
16567             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16568             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16569 #ifdef TCP_ACCOUNTING
16570                 sched_unpin();
16571 #endif
16572                 return (0);
16573         }
16574 #ifdef INET6
16575         if (rack->r_state) {
16576                 /* Use the cache line loaded if possible */
16577                 isipv6 = rack->r_is_v6;
16578         } else {
16579                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16580         }
16581 #endif
16582         early = 0;
16583         cts = tcp_get_usecs(&tv);
16584         ms_cts = tcp_tv_to_mssectick(&tv);
16585         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16586             tcp_in_hpts(rack->rc_inp)) {
16587                 /*
16588                  * We are on the hpts for some timer but not hptsi output.
16589                  * Remove from the hpts unconditionally.
16590                  */
16591                 rack_timer_cancel(tp, rack, cts, __LINE__);
16592         }
16593         /* Are we pacing and late? */
16594         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16595             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16596                 /* We are delayed */
16597                 delayed = cts - rack->r_ctl.rc_last_output_to;
16598         } else {
16599                 delayed = 0;
16600         }
16601         /* Do the timers, which may override the pacer */
16602         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16603                 int retval;
16604
16605                 retval = rack_process_timers(tp, rack, cts, hpts_calling,
16606                     &doing_tlp);
16607                 if (retval != 0) {
16608                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16609 #ifdef TCP_ACCOUNTING
16610                         sched_unpin();
16611 #endif
16612                         /*
16613                          * If timers want tcp_drop(), then pass error out,
16614                          * otherwise suppress it.
16615                          */
16616                         return (retval < 0 ? retval : 0);
16617                 }
16618         }
16619         if (rack->rc_in_persist) {
16620                 if (tcp_in_hpts(rack->rc_inp) == 0) {
16621                         /* Timer is not running */
16622                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16623                 }
16624 #ifdef TCP_ACCOUNTING
16625                 sched_unpin();
16626 #endif
16627                 return (0);
16628         }
16629         if ((rack->rc_ack_required == 1) &&
16630             (rack->r_timer_override == 0)){
16631                 /* A timeout occurred and no ack has arrived */
16632                 if (tcp_in_hpts(rack->rc_inp) == 0) {
16633                         /* Timer is not running */
16634                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16635                 }
16636 #ifdef TCP_ACCOUNTING
16637                 sched_unpin();
16638 #endif
16639                 return (0);
16640         }
16641         if ((rack->r_timer_override) ||
16642             (rack->rc_ack_can_sendout_data) ||
16643             (delayed) ||
16644             (tp->t_state < TCPS_ESTABLISHED)) {
16645                 rack->rc_ack_can_sendout_data = 0;
16646                 if (tcp_in_hpts(rack->rc_inp))
16647                         tcp_hpts_remove(rack->rc_inp);
16648         } else if (tcp_in_hpts(rack->rc_inp)) {
16649                 /*
16650                  * On the hpts you can't pass even if ACKNOW is on, we will
16651                  * when the hpts fires.
16652                  */
16653 #ifdef TCP_ACCOUNTING
16654                 crtsc = get_cyclecount();
16655                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16656                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16657                 }
16658                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16659                         tp->tcp_cnt_counters[SND_BLOCKED]++;
16660                 }
16661                 sched_unpin();
16662 #endif
16663                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16664                 return (0);
16665         }
16666         rack->rc_inp->inp_hpts_calls = 0;
16667         /* Finish out both pacing early and late accounting */
16668         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16669             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16670                 early = rack->r_ctl.rc_last_output_to - cts;
16671         } else
16672                 early = 0;
16673         if (delayed) {
16674                 rack->r_ctl.rc_agg_delayed += delayed;
16675                 rack->r_late = 1;
16676         } else if (early) {
16677                 rack->r_ctl.rc_agg_early += early;
16678                 rack->r_early = 1;
16679         }
16680         /* Now that early/late accounting is done turn off the flag */
16681         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16682         rack->r_wanted_output = 0;
16683         rack->r_timer_override = 0;
16684         if ((tp->t_state != rack->r_state) &&
16685             TCPS_HAVEESTABLISHED(tp->t_state)) {
16686                 rack_set_state(tp, rack);
16687         }
16688         if ((rack->r_fast_output) &&
16689             (doing_tlp == 0) &&
16690             (tp->rcv_numsacks == 0)) {
16691                 int ret;
16692
16693                 error = 0;
16694                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16695                 if (ret >= 0)
16696                         return(ret);
16697                 else if (error) {
16698                         inp = rack->rc_inp;
16699                         so = inp->inp_socket;
16700                         sb = &so->so_snd;
16701                         goto nomore;
16702                 }
16703         }
16704         inp = rack->rc_inp;
16705         /*
16706          * For TFO connections in SYN_SENT or SYN_RECEIVED,
16707          * only allow the initial SYN or SYN|ACK and those sent
16708          * by the retransmit timer.
16709          */
16710         if (IS_FASTOPEN(tp->t_flags) &&
16711             ((tp->t_state == TCPS_SYN_RECEIVED) ||
16712              (tp->t_state == TCPS_SYN_SENT)) &&
16713             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16714             (tp->t_rxtshift == 0)) {              /* not a retransmit */
16715                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16716                 so = inp->inp_socket;
16717                 sb = &so->so_snd;
16718                 goto just_return_nolock;
16719         }
16720         /*
16721          * Determine length of data that should be transmitted, and flags
16722          * that will be used. If there is some data or critical controls
16723          * (SYN, RST) to send, then transmit; otherwise, investigate
16724          * further.
16725          */
16726         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16727         if (tp->t_idle_reduce) {
16728                 if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16729                         rack_cc_after_idle(rack, tp);
16730         }
16731         tp->t_flags &= ~TF_LASTIDLE;
16732         if (idle) {
16733                 if (tp->t_flags & TF_MORETOCOME) {
16734                         tp->t_flags |= TF_LASTIDLE;
16735                         idle = 0;
16736                 }
16737         }
16738         if ((tp->snd_una == tp->snd_max) &&
16739             rack->r_ctl.rc_went_idle_time &&
16740             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16741                 idle = cts - rack->r_ctl.rc_went_idle_time;
16742                 if (idle > rack_min_probertt_hold) {
16743                         /* Count as a probe rtt */
16744                         if (rack->in_probe_rtt == 0) {
16745                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16746                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16747                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16748                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16749                         } else {
16750                                 rack_exit_probertt(rack, cts);
16751                         }
16752                 }
16753                 idle = 0;
16754         }
16755         if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16756                 rack_init_fsb_block(tp, rack);
16757 again:
16758         /*
16759          * If we've recently taken a timeout, snd_max will be greater than
16760          * snd_nxt.  There may be SACK information that allows us to avoid
16761          * resending already delivered data.  Adjust snd_nxt accordingly.
16762          */
16763         sendalot = 0;
16764         cts = tcp_get_usecs(&tv);
16765         ms_cts = tcp_tv_to_mssectick(&tv);
16766         tso = 0;
16767         mtu = 0;
16768         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16769         minseg = segsiz;
16770         if (rack->r_ctl.rc_pace_max_segs == 0)
16771                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16772         else
16773                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16774         sb_offset = tp->snd_max - tp->snd_una;
16775         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16776         flags = tcp_outflags[tp->t_state];
16777         while (rack->rc_free_cnt < rack_free_cache) {
16778                 rsm = rack_alloc(rack);
16779                 if (rsm == NULL) {
16780                         if (inp->inp_hpts_calls)
16781                                 /* Retry in a ms */
16782                                 slot = (1 * HPTS_USEC_IN_MSEC);
16783                         so = inp->inp_socket;
16784                         sb = &so->so_snd;
16785                         goto just_return_nolock;
16786                 }
16787                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16788                 rack->rc_free_cnt++;
16789                 rsm = NULL;
16790         }
16791         if (inp->inp_hpts_calls)
16792                 inp->inp_hpts_calls = 0;
16793         sack_rxmit = 0;
16794         len = 0;
16795         rsm = NULL;
16796         if (flags & TH_RST) {
16797                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16798                 so = inp->inp_socket;
16799                 sb = &so->so_snd;
16800                 goto send;
16801         }
16802         if (rack->r_ctl.rc_resend) {
16803                 /* Retransmit timer */
16804                 rsm = rack->r_ctl.rc_resend;
16805                 rack->r_ctl.rc_resend = NULL;
16806                 len = rsm->r_end - rsm->r_start;
16807                 sack_rxmit = 1;
16808                 sendalot = 0;
16809                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16810                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16811                          __func__, __LINE__,
16812                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16813                 sb_offset = rsm->r_start - tp->snd_una;
16814                 if (len >= segsiz)
16815                         len = segsiz;
16816         } else if (rack->r_collapse_point_valid &&
16817                    ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
16818                 /*
16819                  * If an RSM is returned then enough time has passed
16820                  * for us to retransmit it. Move up the collapse point,
16821                  * since this rsm has its chance to retransmit now.
16822                  */
16823                 tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
16824                 rack->r_ctl.last_collapse_point = rsm->r_end;
16825                 /* Are we done? */
16826                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16827                             rack->r_ctl.high_collapse_point))
16828                         rack->r_collapse_point_valid = 0;
16829                 sack_rxmit = 1;
16830                 /* We are not doing a TLP */
16831                 doing_tlp = 0;
16832                 len = rsm->r_end - rsm->r_start;
16833                 sb_offset = rsm->r_start - tp->snd_una;
16834                 sendalot = 0;
16835                 if ((rack->full_size_rxt == 0) &&
16836                     (rack->shape_rxt_to_pacing_min == 0) &&
16837                     (len >= segsiz))
16838                         len = segsiz;
16839         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16840                 /* We have a retransmit that takes precedence */
16841                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16842                     ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
16843                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16844                         /* Enter recovery if not induced by a time-out */
16845                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
16846                 }
16847 #ifdef INVARIANTS
16848                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16849                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16850                               tp, rack, rsm, rsm->r_start, tp->snd_una);
16851                 }
16852 #endif
16853                 len = rsm->r_end - rsm->r_start;
16854                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16855                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16856                          __func__, __LINE__,
16857                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16858                 sb_offset = rsm->r_start - tp->snd_una;
16859                 sendalot = 0;
16860                 if (len >= segsiz)
16861                         len = segsiz;
16862                 if (len > 0) {
16863                         sack_rxmit = 1;
16864                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16865                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16866                             min(len, segsiz));
16867                 }
16868         } else if (rack->r_ctl.rc_tlpsend) {
16869                 /* Tail loss probe */
16870                 long cwin;
16871                 long tlen;
16872
16873                 /*
16874                  * Check if we can do a TLP with a RACK'd packet
16875                  * this can happen if we are not doing the rack
16876                  * cheat and we skipped to a TLP and it
16877                  * went off.
16878                  */
16879                 rsm = rack->r_ctl.rc_tlpsend;
16880                 /* We are doing a TLP make sure the flag is preent */
16881                 rsm->r_flags |= RACK_TLP;
16882                 rack->r_ctl.rc_tlpsend = NULL;
16883                 sack_rxmit = 1;
16884                 tlen = rsm->r_end - rsm->r_start;
16885                 if (tlen > segsiz)
16886                         tlen = segsiz;
16887                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16888                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16889                          __func__, __LINE__,
16890                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16891                 sb_offset = rsm->r_start - tp->snd_una;
16892                 cwin = min(tp->snd_wnd, tlen);
16893                 len = cwin;
16894         }
16895         if (rack->r_must_retran &&
16896             (doing_tlp == 0) &&
16897             (SEQ_GT(tp->snd_max, tp->snd_una)) &&
16898             (rsm == NULL)) {
16899                 /*
16900                  * There are two different ways that we
16901                  * can get into this block:
16902                  * a) This is a non-sack connection, we had a time-out
16903                  *    and thus r_must_retran was set and everything
16904                  *    left outstanding as been marked for retransmit.
16905                  * b) The MTU of the path shrank, so that everything
16906                  *    was marked to be retransmitted with the smaller
16907                  *    mtu and r_must_retran was set.
16908                  *
16909                  * This means that we expect the sendmap (outstanding)
16910                  * to all be marked must. We can use the tmap to
16911                  * look at them.
16912                  *
16913                  */
16914                 int sendwin, flight;
16915
16916                 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
16917                 flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
16918                 if (flight >= sendwin) {
16919                         /*
16920                          * We can't send yet.
16921                          */
16922                         so = inp->inp_socket;
16923                         sb = &so->so_snd;
16924                         goto just_return_nolock;
16925                 }
16926                 /*
16927                  * This is the case a/b mentioned above. All
16928                  * outstanding/not-acked should be marked.
16929                  * We can use the tmap to find them.
16930                  */
16931                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16932                 if (rsm == NULL) {
16933                         /* TSNH */
16934                         rack->r_must_retran = 0;
16935                         rack->r_ctl.rc_out_at_rto = 0;
16936                         so = inp->inp_socket;
16937                         sb = &so->so_snd;
16938                         goto just_return_nolock;
16939                 }
16940                 if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
16941                         /*
16942                          * The first one does not have the flag, did we collapse
16943                          * further up in our list?
16944                          */
16945                         rack->r_must_retran = 0;
16946                         rack->r_ctl.rc_out_at_rto = 0;
16947                         rsm = NULL;
16948                         sack_rxmit = 0;
16949                 } else {
16950                         sack_rxmit = 1;
16951                         len = rsm->r_end - rsm->r_start;
16952                         sb_offset = rsm->r_start - tp->snd_una;
16953                         sendalot = 0;
16954                         if ((rack->full_size_rxt == 0) &&
16955                             (rack->shape_rxt_to_pacing_min == 0) &&
16956                             (len >= segsiz))
16957                                 len = segsiz;
16958                         /*
16959                          * Delay removing the flag RACK_MUST_RXT so
16960                          * that the fastpath for retransmit will
16961                          * work with this rsm.
16962                          */
16963                 }
16964         }
16965         /*
16966          * Enforce a connection sendmap count limit if set
16967          * as long as we are not retransmiting.
16968          */
16969         if ((rsm == NULL) &&
16970             (rack->do_detection == 0) &&
16971             (V_tcp_map_entries_limit > 0) &&
16972             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
16973                 counter_u64_add(rack_to_alloc_limited, 1);
16974                 if (!rack->alloc_limit_reported) {
16975                         rack->alloc_limit_reported = 1;
16976                         counter_u64_add(rack_alloc_limited_conns, 1);
16977                 }
16978                 so = inp->inp_socket;
16979                 sb = &so->so_snd;
16980                 goto just_return_nolock;
16981         }
16982         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
16983                 /* we are retransmitting the fin */
16984                 len--;
16985                 if (len) {
16986                         /*
16987                          * When retransmitting data do *not* include the
16988                          * FIN. This could happen from a TLP probe.
16989                          */
16990                         flags &= ~TH_FIN;
16991                 }
16992         }
16993         if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
16994             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
16995                 int ret;
16996
16997                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
16998                 if (ret == 0)
16999                         return (0);
17000         }
17001         so = inp->inp_socket;
17002         sb = &so->so_snd;
17003         if (do_a_prefetch == 0) {
17004                 kern_prefetch(sb, &do_a_prefetch);
17005                 do_a_prefetch = 1;
17006         }
17007 #ifdef NETFLIX_SHARED_CWND
17008         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
17009             rack->rack_enable_scwnd) {
17010                 /* We are doing cwnd sharing */
17011                 if (rack->gp_ready &&
17012                     (rack->rack_attempted_scwnd == 0) &&
17013                     (rack->r_ctl.rc_scw == NULL) &&
17014                     tp->t_lib) {
17015                         /* The pcbid is in, lets make an attempt */
17016                         counter_u64_add(rack_try_scwnd, 1);
17017                         rack->rack_attempted_scwnd = 1;
17018                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
17019                                                                    &rack->r_ctl.rc_scw_index,
17020                                                                    segsiz);
17021                 }
17022                 if (rack->r_ctl.rc_scw &&
17023                     (rack->rack_scwnd_is_idle == 1) &&
17024                     sbavail(&so->so_snd)) {
17025                         /* we are no longer out of data */
17026                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17027                         rack->rack_scwnd_is_idle = 0;
17028                 }
17029                 if (rack->r_ctl.rc_scw) {
17030                         /* First lets update and get the cwnd */
17031                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
17032                                                                     rack->r_ctl.rc_scw_index,
17033                                                                     tp->snd_cwnd, tp->snd_wnd, segsiz);
17034                 }
17035         }
17036 #endif
17037         /*
17038          * Get standard flags, and add SYN or FIN if requested by 'hidden'
17039          * state flags.
17040          */
17041         if (tp->t_flags & TF_NEEDFIN)
17042                 flags |= TH_FIN;
17043         if (tp->t_flags & TF_NEEDSYN)
17044                 flags |= TH_SYN;
17045         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
17046                 void *end_rsm;
17047                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17048                 if (end_rsm)
17049                         kern_prefetch(end_rsm, &prefetch_rsm);
17050                 prefetch_rsm = 1;
17051         }
17052         SOCKBUF_LOCK(sb);
17053         /*
17054          * If snd_nxt == snd_max and we have transmitted a FIN, the
17055          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17056          * negative length.  This can also occur when TCP opens up its
17057          * congestion window while receiving additional duplicate acks after
17058          * fast-retransmit because TCP will reset snd_nxt to snd_max after
17059          * the fast-retransmit.
17060          *
17061          * In the normal retransmit-FIN-only case, however, snd_nxt will be
17062          * set to snd_una, the sb_offset will be 0, and the length may wind
17063          * up 0.
17064          *
17065          * If sack_rxmit is true we are retransmitting from the scoreboard
17066          * in which case len is already set.
17067          */
17068         if ((sack_rxmit == 0) &&
17069             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17070                 uint32_t avail;
17071
17072                 avail = sbavail(sb);
17073                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17074                         sb_offset = tp->snd_nxt - tp->snd_una;
17075                 else
17076                         sb_offset = 0;
17077                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17078                         if (rack->r_ctl.rc_tlp_new_data) {
17079                                 /* TLP is forcing out new data */
17080                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17081                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17082                                 }
17083                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17084                                         if (tp->snd_wnd > sb_offset)
17085                                                 len = tp->snd_wnd - sb_offset;
17086                                         else
17087                                                 len = 0;
17088                                 } else {
17089                                         len = rack->r_ctl.rc_tlp_new_data;
17090                                 }
17091                                 rack->r_ctl.rc_tlp_new_data = 0;
17092                         }  else {
17093                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17094                         }
17095                         if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17096                                 /*
17097                                  * For prr=off, we need to send only 1 MSS
17098                                  * at a time. We do this because another sack could
17099                                  * be arriving that causes us to send retransmits and
17100                                  * we don't want to be on a long pace due to a larger send
17101                                  * that keeps us from sending out the retransmit.
17102                                  */
17103                                 len = segsiz;
17104                         }
17105                 } else {
17106                         uint32_t outstanding;
17107                         /*
17108                          * We are inside of a Fast recovery episode, this
17109                          * is caused by a SACK or 3 dup acks. At this point
17110                          * we have sent all the retransmissions and we rely
17111                          * on PRR to dictate what we will send in the form of
17112                          * new data.
17113                          */
17114
17115                         outstanding = tp->snd_max - tp->snd_una;
17116                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17117                                 if (tp->snd_wnd > outstanding) {
17118                                         len = tp->snd_wnd - outstanding;
17119                                         /* Check to see if we have the data */
17120                                         if ((sb_offset + len) > avail) {
17121                                                 /* It does not all fit */
17122                                                 if (avail > sb_offset)
17123                                                         len = avail - sb_offset;
17124                                                 else
17125                                                         len = 0;
17126                                         }
17127                                 } else {
17128                                         len = 0;
17129                                 }
17130                         } else if (avail > sb_offset) {
17131                                 len = avail - sb_offset;
17132                         } else {
17133                                 len = 0;
17134                         }
17135                         if (len > 0) {
17136                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
17137                                         len = rack->r_ctl.rc_prr_sndcnt;
17138                                 }
17139                                 if (len > 0) {
17140                                         sub_from_prr = 1;
17141                                 }
17142                         }
17143                         if (len > segsiz) {
17144                                 /*
17145                                  * We should never send more than a MSS when
17146                                  * retransmitting or sending new data in prr
17147                                  * mode unless the override flag is on. Most
17148                                  * likely the PRR algorithm is not going to
17149                                  * let us send a lot as well :-)
17150                                  */
17151                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
17152                                         len = segsiz;
17153                                 }
17154                         } else if (len < segsiz) {
17155                                 /*
17156                                  * Do we send any? The idea here is if the
17157                                  * send empty's the socket buffer we want to
17158                                  * do it. However if not then lets just wait
17159                                  * for our prr_sndcnt to get bigger.
17160                                  */
17161                                 long leftinsb;
17162
17163                                 leftinsb = sbavail(sb) - sb_offset;
17164                                 if (leftinsb > len) {
17165                                         /* This send does not empty the sb */
17166                                         len = 0;
17167                                 }
17168                         }
17169                 }
17170         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17171                 /*
17172                  * If you have not established
17173                  * and are not doing FAST OPEN
17174                  * no data please.
17175                  */
17176                 if ((sack_rxmit == 0) &&
17177                     (!IS_FASTOPEN(tp->t_flags))){
17178                         len = 0;
17179                         sb_offset = 0;
17180                 }
17181         }
17182         if (prefetch_so_done == 0) {
17183                 kern_prefetch(so, &prefetch_so_done);
17184                 prefetch_so_done = 1;
17185         }
17186         /*
17187          * Lop off SYN bit if it has already been sent.  However, if this is
17188          * SYN-SENT state and if segment contains data and if we don't know
17189          * that foreign host supports TAO, suppress sending segment.
17190          */
17191         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17192             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17193                 /*
17194                  * When sending additional segments following a TFO SYN|ACK,
17195                  * do not include the SYN bit.
17196                  */
17197                 if (IS_FASTOPEN(tp->t_flags) &&
17198                     (tp->t_state == TCPS_SYN_RECEIVED))
17199                         flags &= ~TH_SYN;
17200         }
17201         /*
17202          * Be careful not to send data and/or FIN on SYN segments. This
17203          * measure is needed to prevent interoperability problems with not
17204          * fully conformant TCP implementations.
17205          */
17206         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17207                 len = 0;
17208                 flags &= ~TH_FIN;
17209         }
17210         /*
17211          * On TFO sockets, ensure no data is sent in the following cases:
17212          *
17213          *  - When retransmitting SYN|ACK on a passively-created socket
17214          *
17215          *  - When retransmitting SYN on an actively created socket
17216          *
17217          *  - When sending a zero-length cookie (cookie request) on an
17218          *    actively created socket
17219          *
17220          *  - When the socket is in the CLOSED state (RST is being sent)
17221          */
17222         if (IS_FASTOPEN(tp->t_flags) &&
17223             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17224              ((tp->t_state == TCPS_SYN_SENT) &&
17225               (tp->t_tfo_client_cookie_len == 0)) ||
17226              (flags & TH_RST))) {
17227                 sack_rxmit = 0;
17228                 len = 0;
17229         }
17230         /* Without fast-open there should never be data sent on a SYN */
17231         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17232                 tp->snd_nxt = tp->iss;
17233                 len = 0;
17234         }
17235         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17236                 /* We only send 1 MSS if we have a DSACK block */
17237                 add_flag |= RACK_SENT_W_DSACK;
17238                 len = segsiz;
17239         }
17240         orig_len = len;
17241         if (len <= 0) {
17242                 /*
17243                  * If FIN has been sent but not acked, but we haven't been
17244                  * called to retransmit, len will be < 0.  Otherwise, window
17245                  * shrank after we sent into it.  If window shrank to 0,
17246                  * cancel pending retransmit, pull snd_nxt back to (closed)
17247                  * window, and set the persist timer if it isn't already
17248                  * going.  If the window didn't close completely, just wait
17249                  * for an ACK.
17250                  *
17251                  * We also do a general check here to ensure that we will
17252                  * set the persist timer when we have data to send, but a
17253                  * 0-byte window. This makes sure the persist timer is set
17254                  * even if the packet hits one of the "goto send" lines
17255                  * below.
17256                  */
17257                 len = 0;
17258                 if ((tp->snd_wnd == 0) &&
17259                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17260                     (tp->snd_una == tp->snd_max) &&
17261                     (sb_offset < (int)sbavail(sb))) {
17262                         rack_enter_persist(tp, rack, cts);
17263                 }
17264         } else if ((rsm == NULL) &&
17265                    (doing_tlp == 0) &&
17266                    (len < pace_max_seg)) {
17267                 /*
17268                  * We are not sending a maximum sized segment for
17269                  * some reason. Should we not send anything (think
17270                  * sws or persists)?
17271                  */
17272                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17273                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17274                     (len < minseg) &&
17275                     (len < (int)(sbavail(sb) - sb_offset))) {
17276                         /*
17277                          * Here the rwnd is less than
17278                          * the minimum pacing size, this is not a retransmit,
17279                          * we are established and
17280                          * the send is not the last in the socket buffer
17281                          * we send nothing, and we may enter persists
17282                          * if nothing is outstanding.
17283                          */
17284                         len = 0;
17285                         if (tp->snd_max == tp->snd_una) {
17286                                 /*
17287                                  * Nothing out we can
17288                                  * go into persists.
17289                                  */
17290                                 rack_enter_persist(tp, rack, cts);
17291                         }
17292                      } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17293                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17294                            (len < (int)(sbavail(sb) - sb_offset)) &&
17295                            (len < minseg)) {
17296                         /*
17297                          * Here we are not retransmitting, and
17298                          * the cwnd is not so small that we could
17299                          * not send at least a min size (rxt timer
17300                          * not having gone off), We have 2 segments or
17301                          * more already in flight, its not the tail end
17302                          * of the socket buffer  and the cwnd is blocking
17303                          * us from sending out a minimum pacing segment size.
17304                          * Lets not send anything.
17305                          */
17306                         len = 0;
17307                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17308                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17309                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17310                            (len < (int)(sbavail(sb) - sb_offset)) &&
17311                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
17312                         /*
17313                          * Here we have a send window but we have
17314                          * filled it up and we can't send another pacing segment.
17315                          * We also have in flight more than 2 segments
17316                          * and we are not completing the sb i.e. we allow
17317                          * the last bytes of the sb to go out even if
17318                          * its not a full pacing segment.
17319                          */
17320                         len = 0;
17321                 } else if ((rack->r_ctl.crte != NULL) &&
17322                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17323                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17324                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17325                            (len < (int)(sbavail(sb) - sb_offset))) {
17326                         /*
17327                          * Here we are doing hardware pacing, this is not a TLP,
17328                          * we are not sending a pace max segment size, there is rwnd
17329                          * room to send at least N pace_max_seg, the cwnd is greater
17330                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
17331                          * more segments in flight and its not the tail of the socket buffer.
17332                          *
17333                          * We don't want to send instead we need to get more ack's in to
17334                          * allow us to send a full pacing segment. Normally, if we are pacing
17335                          * about the right speed, we should have finished our pacing
17336                          * send as most of the acks have come back if we are at the
17337                          * right rate. This is a bit fuzzy since return path delay
17338                          * can delay the acks, which is why we want to make sure we
17339                          * have cwnd space to have a bit more than a max pace segments in flight.
17340                          *
17341                          * If we have not gotten our acks back we are pacing at too high a
17342                          * rate delaying will not hurt and will bring our GP estimate down by
17343                          * injecting the delay. If we don't do this we will send
17344                          * 2 MSS out in response to the acks being clocked in which
17345                          * defeats the point of hw-pacing (i.e. to help us get
17346                          * larger TSO's out).
17347                          */
17348                         len = 0;
17349
17350                 }
17351
17352         }
17353         /* len will be >= 0 after this point. */
17354         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17355         rack_sndbuf_autoscale(rack);
17356         /*
17357          * Decide if we can use TCP Segmentation Offloading (if supported by
17358          * hardware).
17359          *
17360          * TSO may only be used if we are in a pure bulk sending state.  The
17361          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17362          * options prevent using TSO.  With TSO the TCP header is the same
17363          * (except for the sequence number) for all generated packets.  This
17364          * makes it impossible to transmit any options which vary per
17365          * generated segment or packet.
17366          *
17367          * IPv4 handling has a clear separation of ip options and ip header
17368          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17369          * the right thing below to provide length of just ip options and thus
17370          * checking for ipoptlen is enough to decide if ip options are present.
17371          */
17372         ipoptlen = 0;
17373 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17374         /*
17375          * Pre-calculate here as we save another lookup into the darknesses
17376          * of IPsec that way and can actually decide if TSO is ok.
17377          */
17378 #ifdef INET6
17379         if (isipv6 && IPSEC_ENABLED(ipv6))
17380                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
17381 #ifdef INET
17382         else
17383 #endif
17384 #endif                          /* INET6 */
17385 #ifdef INET
17386                 if (IPSEC_ENABLED(ipv4))
17387                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
17388 #endif                          /* INET */
17389 #endif
17390
17391 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17392         ipoptlen += ipsec_optlen;
17393 #endif
17394         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17395             (tp->t_port == 0) &&
17396             ((tp->t_flags & TF_SIGNATURE) == 0) &&
17397             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17398             ipoptlen == 0)
17399                 tso = 1;
17400         {
17401                 uint32_t outstanding __unused;
17402
17403                 outstanding = tp->snd_max - tp->snd_una;
17404                 if (tp->t_flags & TF_SENTFIN) {
17405                         /*
17406                          * If we sent a fin, snd_max is 1 higher than
17407                          * snd_una
17408                          */
17409                         outstanding--;
17410                 }
17411                 if (sack_rxmit) {
17412                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17413                                 flags &= ~TH_FIN;
17414                 } else {
17415                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17416                                    sbused(sb)))
17417                                 flags &= ~TH_FIN;
17418                 }
17419         }
17420         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17421             (long)TCP_MAXWIN << tp->rcv_scale);
17422
17423         /*
17424          * Sender silly window avoidance.   We transmit under the following
17425          * conditions when len is non-zero:
17426          *
17427          * - We have a full segment (or more with TSO) - This is the last
17428          * buffer in a write()/send() and we are either idle or running
17429          * NODELAY - we've timed out (e.g. persist timer) - we have more
17430          * then 1/2 the maximum send window's worth of data (receiver may be
17431          * limited the window size) - we need to retransmit
17432          */
17433         if (len) {
17434                 if (len >= segsiz) {
17435                         goto send;
17436                 }
17437                 /*
17438                  * NOTE! on localhost connections an 'ack' from the remote
17439                  * end may occur synchronously with the output and cause us
17440                  * to flush a buffer queued with moretocome.  XXX
17441                  *
17442                  */
17443                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
17444                     (idle || (tp->t_flags & TF_NODELAY)) &&
17445                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17446                     (tp->t_flags & TF_NOPUSH) == 0) {
17447                         pass = 2;
17448                         goto send;
17449                 }
17450                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
17451                         pass = 22;
17452                         goto send;
17453                 }
17454                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17455                         pass = 4;
17456                         goto send;
17457                 }
17458                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
17459                         pass = 5;
17460                         goto send;
17461                 }
17462                 if (sack_rxmit) {
17463                         pass = 6;
17464                         goto send;
17465                 }
17466                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17467                     (ctf_outstanding(tp) < (segsiz * 2))) {
17468                         /*
17469                          * We have less than two MSS outstanding (delayed ack)
17470                          * and our rwnd will not let us send a full sized
17471                          * MSS. Lets go ahead and let this small segment
17472                          * out because we want to try to have at least two
17473                          * packets inflight to not be caught by delayed ack.
17474                          */
17475                         pass = 12;
17476                         goto send;
17477                 }
17478         }
17479         /*
17480          * Sending of standalone window updates.
17481          *
17482          * Window updates are important when we close our window due to a
17483          * full socket buffer and are opening it again after the application
17484          * reads data from it.  Once the window has opened again and the
17485          * remote end starts to send again the ACK clock takes over and
17486          * provides the most current window information.
17487          *
17488          * We must avoid the silly window syndrome whereas every read from
17489          * the receive buffer, no matter how small, causes a window update
17490          * to be sent.  We also should avoid sending a flurry of window
17491          * updates when the socket buffer had queued a lot of data and the
17492          * application is doing small reads.
17493          *
17494          * Prevent a flurry of pointless window updates by only sending an
17495          * update when we can increase the advertized window by more than
17496          * 1/4th of the socket buffer capacity.  When the buffer is getting
17497          * full or is very small be more aggressive and send an update
17498          * whenever we can increase by two mss sized segments. In all other
17499          * situations the ACK's to new incoming data will carry further
17500          * window increases.
17501          *
17502          * Don't send an independent window update if a delayed ACK is
17503          * pending (it will get piggy-backed on it) or the remote side
17504          * already has done a half-close and won't send more data.  Skip
17505          * this if the connection is in T/TCP half-open state.
17506          */
17507         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17508             !(tp->t_flags & TF_DELACK) &&
17509             !TCPS_HAVERCVDFIN(tp->t_state)) {
17510                 /*
17511                  * "adv" is the amount we could increase the window, taking
17512                  * into account that we are limited by TCP_MAXWIN <<
17513                  * tp->rcv_scale.
17514                  */
17515                 int32_t adv;
17516                 int oldwin;
17517
17518                 adv = recwin;
17519                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17520                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
17521                         if (adv > oldwin)
17522                             adv -= oldwin;
17523                         else {
17524                                 /* We can't increase the window */
17525                                 adv = 0;
17526                         }
17527                 } else
17528                         oldwin = 0;
17529
17530                 /*
17531                  * If the new window size ends up being the same as or less
17532                  * than the old size when it is scaled, then don't force
17533                  * a window update.
17534                  */
17535                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17536                         goto dontupdate;
17537
17538                 if (adv >= (int32_t)(2 * segsiz) &&
17539                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17540                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17541                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17542                         pass = 7;
17543                         goto send;
17544                 }
17545                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17546                         pass = 23;
17547                         goto send;
17548                 }
17549         }
17550 dontupdate:
17551
17552         /*
17553          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17554          * is also a catch-all for the retransmit timer timeout case.
17555          */
17556         if (tp->t_flags & TF_ACKNOW) {
17557                 pass = 8;
17558                 goto send;
17559         }
17560         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17561                 pass = 9;
17562                 goto send;
17563         }
17564         /*
17565          * If our state indicates that FIN should be sent and we have not
17566          * yet done so, then we need to send.
17567          */
17568         if ((flags & TH_FIN) &&
17569             (tp->snd_nxt == tp->snd_una)) {
17570                 pass = 11;
17571                 goto send;
17572         }
17573         /*
17574          * No reason to send a segment, just return.
17575          */
17576 just_return:
17577         SOCKBUF_UNLOCK(sb);
17578 just_return_nolock:
17579         {
17580                 int app_limited = CTF_JR_SENT_DATA;
17581
17582                 if (tot_len_this_send > 0) {
17583                         /* Make sure snd_nxt is up to max */
17584                         rack->r_ctl.fsb.recwin = recwin;
17585                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17586                         if ((error == 0) &&
17587                             rack_use_rfo &&
17588                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
17589                             (ipoptlen == 0) &&
17590                             (tp->snd_nxt == tp->snd_max) &&
17591                             (tp->rcv_numsacks == 0) &&
17592                             rack->r_fsb_inited &&
17593                             TCPS_HAVEESTABLISHED(tp->t_state) &&
17594                             (rack->r_must_retran == 0) &&
17595                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
17596                             (len > 0) && (orig_len > 0) &&
17597                             (orig_len > len) &&
17598                             ((orig_len - len) >= segsiz) &&
17599                             ((optlen == 0) ||
17600                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17601                                 /* We can send at least one more MSS using our fsb */
17602
17603                                 rack->r_fast_output = 1;
17604                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17605                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17606                                 rack->r_ctl.fsb.tcp_flags = flags;
17607                                 rack->r_ctl.fsb.left_to_send = orig_len - len;
17608                                 if (hw_tls)
17609                                         rack->r_ctl.fsb.hw_tls = 1;
17610                                 else
17611                                         rack->r_ctl.fsb.hw_tls = 0;
17612                                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17613                                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
17614                                         rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17615                                          (tp->snd_max - tp->snd_una)));
17616                                 if (rack->r_ctl.fsb.left_to_send < segsiz)
17617                                         rack->r_fast_output = 0;
17618                                 else {
17619                                         if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17620                                                 rack->r_ctl.fsb.rfo_apply_push = 1;
17621                                         else
17622                                                 rack->r_ctl.fsb.rfo_apply_push = 0;
17623                                 }
17624                         } else
17625                                 rack->r_fast_output = 0;
17626
17627
17628                         rack_log_fsb(rack, tp, so, flags,
17629                                      ipoptlen, orig_len, len, 0,
17630                                      1, optlen, __LINE__, 1);
17631                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17632                                 tp->snd_nxt = tp->snd_max;
17633                 } else {
17634                         int end_window = 0;
17635                         uint32_t seq = tp->gput_ack;
17636
17637                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17638                         if (rsm) {
17639                                 /*
17640                                  * Mark the last sent that we just-returned (hinting
17641                                  * that delayed ack may play a role in any rtt measurement).
17642                                  */
17643                                 rsm->r_just_ret = 1;
17644                         }
17645                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17646                         rack->r_ctl.rc_agg_delayed = 0;
17647                         rack->r_early = 0;
17648                         rack->r_late = 0;
17649                         rack->r_ctl.rc_agg_early = 0;
17650                         if ((ctf_outstanding(tp) +
17651                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17652                                  minseg)) >= tp->snd_wnd) {
17653                                 /* We are limited by the rwnd */
17654                                 app_limited = CTF_JR_RWND_LIMITED;
17655                                 if (IN_FASTRECOVERY(tp->t_flags))
17656                                     rack->r_ctl.rc_prr_sndcnt = 0;
17657                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17658                                 /* We are limited by whats available -- app limited */
17659                                 app_limited = CTF_JR_APP_LIMITED;
17660                                 if (IN_FASTRECOVERY(tp->t_flags))
17661                                     rack->r_ctl.rc_prr_sndcnt = 0;
17662                         } else if ((idle == 0) &&
17663                                    ((tp->t_flags & TF_NODELAY) == 0) &&
17664                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17665                                    (len < segsiz)) {
17666                                 /*
17667                                  * No delay is not on and the
17668                                  * user is sending less than 1MSS. This
17669                                  * brings out SWS avoidance so we
17670                                  * don't send. Another app-limited case.
17671                                  */
17672                                 app_limited = CTF_JR_APP_LIMITED;
17673                         } else if (tp->t_flags & TF_NOPUSH) {
17674                                 /*
17675                                  * The user has requested no push of
17676                                  * the last segment and we are
17677                                  * at the last segment. Another app
17678                                  * limited case.
17679                                  */
17680                                 app_limited = CTF_JR_APP_LIMITED;
17681                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17682                                 /* Its the cwnd */
17683                                 app_limited = CTF_JR_CWND_LIMITED;
17684                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
17685                                    (rack->rack_no_prr == 0) &&
17686                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17687                                 app_limited = CTF_JR_PRR;
17688                         } else {
17689                                 /* Now why here are we not sending? */
17690 #ifdef NOW
17691 #ifdef INVARIANTS
17692                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17693 #endif
17694 #endif
17695                                 app_limited = CTF_JR_ASSESSING;
17696                         }
17697                         /*
17698                          * App limited in some fashion, for our pacing GP
17699                          * measurements we don't want any gap (even cwnd).
17700                          * Close  down the measurement window.
17701                          */
17702                         if (rack_cwnd_block_ends_measure &&
17703                             ((app_limited == CTF_JR_CWND_LIMITED) ||
17704                              (app_limited == CTF_JR_PRR))) {
17705                                 /*
17706                                  * The reason we are not sending is
17707                                  * the cwnd (or prr). We have been configured
17708                                  * to end the measurement window in
17709                                  * this case.
17710                                  */
17711                                 end_window = 1;
17712                         } else if (rack_rwnd_block_ends_measure &&
17713                                    (app_limited == CTF_JR_RWND_LIMITED)) {
17714                                 /*
17715                                  * We are rwnd limited and have been
17716                                  * configured to end the measurement
17717                                  * window in this case.
17718                                  */
17719                                 end_window = 1;
17720                         } else if (app_limited == CTF_JR_APP_LIMITED) {
17721                                 /*
17722                                  * A true application limited period, we have
17723                                  * ran out of data.
17724                                  */
17725                                 end_window = 1;
17726                         } else if (app_limited == CTF_JR_ASSESSING) {
17727                                 /*
17728                                  * In the assessing case we hit the end of
17729                                  * the if/else and had no known reason
17730                                  * This will panic us under invariants..
17731                                  *
17732                                  * If we get this out in logs we need to
17733                                  * investagate which reason we missed.
17734                                  */
17735                                 end_window = 1;
17736                         }
17737                         if (end_window) {
17738                                 uint8_t log = 0;
17739
17740                                 /* Adjust the Gput measurement */
17741                                 if ((tp->t_flags & TF_GPUTINPROG) &&
17742                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
17743                                         tp->gput_ack = tp->snd_max;
17744                                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17745                                                 /*
17746                                                  * There is not enough to measure.
17747                                                  */
17748                                                 tp->t_flags &= ~TF_GPUTINPROG;
17749                                                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17750                                                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
17751                                                                            tp->gput_seq,
17752                                                                            0, 0, 18, __LINE__, NULL, 0);
17753                                         } else
17754                                                 log = 1;
17755                                 }
17756                                 /* Mark the last packet has app limited */
17757                                 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17758                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17759                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
17760                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17761                                         else {
17762                                                 /*
17763                                                  * Go out to the end app limited and mark
17764                                                  * this new one as next and move the end_appl up
17765                                                  * to this guy.
17766                                                  */
17767                                                 if (rack->r_ctl.rc_end_appl)
17768                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17769                                                 rack->r_ctl.rc_end_appl = rsm;
17770                                         }
17771                                         rsm->r_flags |= RACK_APP_LIMITED;
17772                                         rack->r_ctl.rc_app_limited_cnt++;
17773                                 }
17774                                 if (log)
17775                                         rack_log_pacing_delay_calc(rack,
17776                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
17777                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17778                         }
17779                 }
17780                 /* Check if we need to go into persists or not */
17781                 if ((tp->snd_max == tp->snd_una) &&
17782                     TCPS_HAVEESTABLISHED(tp->t_state) &&
17783                     sbavail(sb) &&
17784                     (sbavail(sb) > tp->snd_wnd) &&
17785                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17786                         /* Yes lets make sure to move to persist before timer-start */
17787                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17788                 }
17789                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17790                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17791         }
17792 #ifdef NETFLIX_SHARED_CWND
17793         if ((sbavail(sb) == 0) &&
17794             rack->r_ctl.rc_scw) {
17795                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17796                 rack->rack_scwnd_is_idle = 1;
17797         }
17798 #endif
17799 #ifdef TCP_ACCOUNTING
17800         if (tot_len_this_send > 0) {
17801                 crtsc = get_cyclecount();
17802                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17803                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
17804                 }
17805                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17806                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17807                 }
17808                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17809                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17810                 }
17811         } else {
17812                 crtsc = get_cyclecount();
17813                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17814                         tp->tcp_cnt_counters[SND_LIMITED]++;
17815                 }
17816                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17817                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17818                 }
17819         }
17820         sched_unpin();
17821 #endif
17822         return (0);
17823
17824 send:
17825         if (rsm || sack_rxmit)
17826                 counter_u64_add(rack_nfto_resend, 1);
17827         else
17828                 counter_u64_add(rack_non_fto_send, 1);
17829         if ((flags & TH_FIN) &&
17830             sbavail(sb)) {
17831                 /*
17832                  * We do not transmit a FIN
17833                  * with data outstanding. We
17834                  * need to make it so all data
17835                  * is acked first.
17836                  */
17837                 flags &= ~TH_FIN;
17838         }
17839         /* Enforce stack imposed max seg size if we have one */
17840         if (rack->r_ctl.rc_pace_max_segs &&
17841             (len > rack->r_ctl.rc_pace_max_segs)) {
17842                 mark = 1;
17843                 len = rack->r_ctl.rc_pace_max_segs;
17844         }
17845         SOCKBUF_LOCK_ASSERT(sb);
17846         if (len > 0) {
17847                 if (len >= segsiz)
17848                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17849                 else
17850                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17851         }
17852         /*
17853          * Before ESTABLISHED, force sending of initial options unless TCP
17854          * set not to do any options. NOTE: we assume that the IP/TCP header
17855          * plus TCP options always fit in a single mbuf, leaving room for a
17856          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17857          * + optlen <= MCLBYTES
17858          */
17859         optlen = 0;
17860 #ifdef INET6
17861         if (isipv6)
17862                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17863         else
17864 #endif
17865                 hdrlen = sizeof(struct tcpiphdr);
17866
17867         /*
17868          * Compute options for segment. We only have to care about SYN and
17869          * established connection segments.  Options for SYN-ACK segments
17870          * are handled in TCP syncache.
17871          */
17872         to.to_flags = 0;
17873         if ((tp->t_flags & TF_NOOPT) == 0) {
17874                 /* Maximum segment size. */
17875                 if (flags & TH_SYN) {
17876                         tp->snd_nxt = tp->iss;
17877                         to.to_mss = tcp_mssopt(&inp->inp_inc);
17878                         if (tp->t_port)
17879                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
17880                         to.to_flags |= TOF_MSS;
17881
17882                         /*
17883                          * On SYN or SYN|ACK transmits on TFO connections,
17884                          * only include the TFO option if it is not a
17885                          * retransmit, as the presence of the TFO option may
17886                          * have caused the original SYN or SYN|ACK to have
17887                          * been dropped by a middlebox.
17888                          */
17889                         if (IS_FASTOPEN(tp->t_flags) &&
17890                             (tp->t_rxtshift == 0)) {
17891                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
17892                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17893                                         to.to_tfo_cookie =
17894                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
17895                                         to.to_flags |= TOF_FASTOPEN;
17896                                         wanted_cookie = 1;
17897                                 } else if (tp->t_state == TCPS_SYN_SENT) {
17898                                         to.to_tfo_len =
17899                                                 tp->t_tfo_client_cookie_len;
17900                                         to.to_tfo_cookie =
17901                                                 tp->t_tfo_cookie.client;
17902                                         to.to_flags |= TOF_FASTOPEN;
17903                                         wanted_cookie = 1;
17904                                         /*
17905                                          * If we wind up having more data to
17906                                          * send with the SYN than can fit in
17907                                          * one segment, don't send any more
17908                                          * until the SYN|ACK comes back from
17909                                          * the other end.
17910                                          */
17911                                         sendalot = 0;
17912                                 }
17913                         }
17914                 }
17915                 /* Window scaling. */
17916                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
17917                         to.to_wscale = tp->request_r_scale;
17918                         to.to_flags |= TOF_SCALE;
17919                 }
17920                 /* Timestamps. */
17921                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
17922                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
17923                         to.to_tsval = ms_cts + tp->ts_offset;
17924                         to.to_tsecr = tp->ts_recent;
17925                         to.to_flags |= TOF_TS;
17926                 }
17927                 /* Set receive buffer autosizing timestamp. */
17928                 if (tp->rfbuf_ts == 0 &&
17929                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
17930                         tp->rfbuf_ts = tcp_ts_getticks();
17931                 /* Selective ACK's. */
17932                 if (tp->t_flags & TF_SACK_PERMIT) {
17933                         if (flags & TH_SYN)
17934                                 to.to_flags |= TOF_SACKPERM;
17935                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17936                                  tp->rcv_numsacks > 0) {
17937                                 to.to_flags |= TOF_SACK;
17938                                 to.to_nsacks = tp->rcv_numsacks;
17939                                 to.to_sacks = (u_char *)tp->sackblks;
17940                         }
17941                 }
17942 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17943                 /* TCP-MD5 (RFC2385). */
17944                 if (tp->t_flags & TF_SIGNATURE)
17945                         to.to_flags |= TOF_SIGNATURE;
17946 #endif                          /* TCP_SIGNATURE */
17947
17948                 /* Processing the options. */
17949                 hdrlen += optlen = tcp_addoptions(&to, opt);
17950                 /*
17951                  * If we wanted a TFO option to be added, but it was unable
17952                  * to fit, ensure no data is sent.
17953                  */
17954                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
17955                     !(to.to_flags & TOF_FASTOPEN))
17956                         len = 0;
17957         }
17958         if (tp->t_port) {
17959                 if (V_tcp_udp_tunneling_port == 0) {
17960                         /* The port was removed?? */
17961                         SOCKBUF_UNLOCK(&so->so_snd);
17962 #ifdef TCP_ACCOUNTING
17963                         crtsc = get_cyclecount();
17964                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17965                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
17966                         }
17967                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17968                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
17969                         }
17970                         sched_unpin();
17971 #endif
17972                         return (EHOSTUNREACH);
17973                 }
17974                 hdrlen += sizeof(struct udphdr);
17975         }
17976 #ifdef INET6
17977         if (isipv6)
17978                 ipoptlen = ip6_optlen(inp);
17979         else
17980 #endif
17981                 if (inp->inp_options)
17982                         ipoptlen = inp->inp_options->m_len -
17983                                 offsetof(struct ipoption, ipopt_list);
17984                 else
17985                         ipoptlen = 0;
17986 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17987         ipoptlen += ipsec_optlen;
17988 #endif
17989
17990         /*
17991          * Adjust data length if insertion of options will bump the packet
17992          * length beyond the t_maxseg length. Clear the FIN bit because we
17993          * cut off the tail of the segment.
17994          */
17995         if (len + optlen + ipoptlen > tp->t_maxseg) {
17996                 if (tso) {
17997                         uint32_t if_hw_tsomax;
17998                         uint32_t moff;
17999                         int32_t max_len;
18000
18001                         /* extract TSO information */
18002                         if_hw_tsomax = tp->t_tsomax;
18003                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18004                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18005                         KASSERT(ipoptlen == 0,
18006                                 ("%s: TSO can't do IP options", __func__));
18007
18008                         /*
18009                          * Check if we should limit by maximum payload
18010                          * length:
18011                          */
18012                         if (if_hw_tsomax != 0) {
18013                                 /* compute maximum TSO length */
18014                                 max_len = (if_hw_tsomax - hdrlen -
18015                                            max_linkhdr);
18016                                 if (max_len <= 0) {
18017                                         len = 0;
18018                                 } else if (len > max_len) {
18019                                         sendalot = 1;
18020                                         len = max_len;
18021                                         mark = 2;
18022                                 }
18023                         }
18024                         /*
18025                          * Prevent the last segment from being fractional
18026                          * unless the send sockbuf can be emptied:
18027                          */
18028                         max_len = (tp->t_maxseg - optlen);
18029                         if ((sb_offset + len) < sbavail(sb)) {
18030                                 moff = len % (u_int)max_len;
18031                                 if (moff != 0) {
18032                                         mark = 3;
18033                                         len -= moff;
18034                                 }
18035                         }
18036                         /*
18037                          * In case there are too many small fragments don't
18038                          * use TSO:
18039                          */
18040                         if (len <= segsiz) {
18041                                 mark = 4;
18042                                 tso = 0;
18043                         }
18044                         /*
18045                          * Send the FIN in a separate segment after the bulk
18046                          * sending is done. We don't trust the TSO
18047                          * implementations to clear the FIN flag on all but
18048                          * the last segment.
18049                          */
18050                         if (tp->t_flags & TF_NEEDFIN) {
18051                                 sendalot = 4;
18052                         }
18053                 } else {
18054                         mark = 5;
18055                         if (optlen + ipoptlen >= tp->t_maxseg) {
18056                                 /*
18057                                  * Since we don't have enough space to put
18058                                  * the IP header chain and the TCP header in
18059                                  * one packet as required by RFC 7112, don't
18060                                  * send it. Also ensure that at least one
18061                                  * byte of the payload can be put into the
18062                                  * TCP segment.
18063                                  */
18064                                 SOCKBUF_UNLOCK(&so->so_snd);
18065                                 error = EMSGSIZE;
18066                                 sack_rxmit = 0;
18067                                 goto out;
18068                         }
18069                         len = tp->t_maxseg - optlen - ipoptlen;
18070                         sendalot = 5;
18071                 }
18072         } else {
18073                 tso = 0;
18074                 mark = 6;
18075         }
18076         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18077                 ("%s: len > IP_MAXPACKET", __func__));
18078 #ifdef DIAGNOSTIC
18079 #ifdef INET6
18080         if (max_linkhdr + hdrlen > MCLBYTES)
18081 #else
18082                 if (max_linkhdr + hdrlen > MHLEN)
18083 #endif
18084                         panic("tcphdr too big");
18085 #endif
18086
18087         /*
18088          * This KASSERT is here to catch edge cases at a well defined place.
18089          * Before, those had triggered (random) panic conditions further
18090          * down.
18091          */
18092         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18093         if ((len == 0) &&
18094             (flags & TH_FIN) &&
18095             (sbused(sb))) {
18096                 /*
18097                  * We have outstanding data, don't send a fin by itself!.
18098                  */
18099                 goto just_return;
18100         }
18101         /*
18102          * Grab a header mbuf, attaching a copy of data to be transmitted,
18103          * and initialize the header from the template for sends on this
18104          * connection.
18105          */
18106         hw_tls = tp->t_nic_ktls_xmit != 0;
18107         if (len) {
18108                 uint32_t max_val;
18109                 uint32_t moff;
18110
18111                 if (rack->r_ctl.rc_pace_max_segs)
18112                         max_val = rack->r_ctl.rc_pace_max_segs;
18113                 else if (rack->rc_user_set_max_segs)
18114                         max_val = rack->rc_user_set_max_segs * segsiz;
18115                 else
18116                         max_val = len;
18117                 /*
18118                  * We allow a limit on sending with hptsi.
18119                  */
18120                 if (len > max_val) {
18121                         mark = 7;
18122                         len = max_val;
18123                 }
18124 #ifdef INET6
18125                 if (MHLEN < hdrlen + max_linkhdr)
18126                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18127                 else
18128 #endif
18129                         m = m_gethdr(M_NOWAIT, MT_DATA);
18130
18131                 if (m == NULL) {
18132                         SOCKBUF_UNLOCK(sb);
18133                         error = ENOBUFS;
18134                         sack_rxmit = 0;
18135                         goto out;
18136                 }
18137                 m->m_data += max_linkhdr;
18138                 m->m_len = hdrlen;
18139
18140                 /*
18141                  * Start the m_copy functions from the closest mbuf to the
18142                  * sb_offset in the socket buffer chain.
18143                  */
18144                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
18145                 s_mb = mb;
18146                 s_moff = moff;
18147                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18148                         m_copydata(mb, moff, (int)len,
18149                                    mtod(m, caddr_t)+hdrlen);
18150                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18151                                 sbsndptr_adv(sb, mb, len);
18152                         m->m_len += len;
18153                 } else {
18154                         struct sockbuf *msb;
18155
18156                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18157                                 msb = NULL;
18158                         else
18159                                 msb = sb;
18160                         m->m_next = tcp_m_copym(
18161                                 mb, moff, &len,
18162                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18163                                 ((rsm == NULL) ? hw_tls : 0)
18164 #ifdef NETFLIX_COPY_ARGS
18165                                 , &s_mb, &s_moff
18166 #endif
18167                                 );
18168                         if (len <= (tp->t_maxseg - optlen)) {
18169                                 /*
18170                                  * Must have ran out of mbufs for the copy
18171                                  * shorten it to no longer need tso. Lets
18172                                  * not put on sendalot since we are low on
18173                                  * mbufs.
18174                                  */
18175                                 tso = 0;
18176                         }
18177                         if (m->m_next == NULL) {
18178                                 SOCKBUF_UNLOCK(sb);
18179                                 (void)m_free(m);
18180                                 error = ENOBUFS;
18181                                 sack_rxmit = 0;
18182                                 goto out;
18183                         }
18184                 }
18185                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18186                         if (rsm && (rsm->r_flags & RACK_TLP)) {
18187                                 /*
18188                                  * TLP should not count in retran count, but
18189                                  * in its own bin
18190                                  */
18191                                 counter_u64_add(rack_tlp_retran, 1);
18192                                 counter_u64_add(rack_tlp_retran_bytes, len);
18193                         } else {
18194                                 tp->t_sndrexmitpack++;
18195                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18196                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18197                         }
18198 #ifdef STATS
18199                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18200                                                  len);
18201 #endif
18202                 } else {
18203                         KMOD_TCPSTAT_INC(tcps_sndpack);
18204                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18205 #ifdef STATS
18206                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18207                                                  len);
18208 #endif
18209                 }
18210                 /*
18211                  * If we're sending everything we've got, set PUSH. (This
18212                  * will keep happy those implementations which only give
18213                  * data to the user when a buffer fills or a PUSH comes in.)
18214                  */
18215                 if (sb_offset + len == sbused(sb) &&
18216                     sbused(sb) &&
18217                     !(flags & TH_SYN)) {
18218                         flags |= TH_PUSH;
18219                         add_flag |= RACK_HAD_PUSH;
18220                 }
18221
18222                 SOCKBUF_UNLOCK(sb);
18223         } else {
18224                 SOCKBUF_UNLOCK(sb);
18225                 if (tp->t_flags & TF_ACKNOW)
18226                         KMOD_TCPSTAT_INC(tcps_sndacks);
18227                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
18228                         KMOD_TCPSTAT_INC(tcps_sndctrl);
18229                 else
18230                         KMOD_TCPSTAT_INC(tcps_sndwinup);
18231
18232                 m = m_gethdr(M_NOWAIT, MT_DATA);
18233                 if (m == NULL) {
18234                         error = ENOBUFS;
18235                         sack_rxmit = 0;
18236                         goto out;
18237                 }
18238 #ifdef INET6
18239                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18240                     MHLEN >= hdrlen) {
18241                         M_ALIGN(m, hdrlen);
18242                 } else
18243 #endif
18244                         m->m_data += max_linkhdr;
18245                 m->m_len = hdrlen;
18246         }
18247         SOCKBUF_UNLOCK_ASSERT(sb);
18248         m->m_pkthdr.rcvif = (struct ifnet *)0;
18249 #ifdef MAC
18250         mac_inpcb_create_mbuf(inp, m);
18251 #endif
18252         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18253 #ifdef INET6
18254                 if (isipv6)
18255                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18256                 else
18257 #endif                          /* INET6 */
18258 #ifdef INET
18259                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18260 #endif
18261                 th = rack->r_ctl.fsb.th;
18262                 udp = rack->r_ctl.fsb.udp;
18263                 if (udp) {
18264 #ifdef INET6
18265                         if (isipv6)
18266                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
18267                         else
18268 #endif                          /* INET6 */
18269                                 ulen = hdrlen + len - sizeof(struct ip);
18270                         udp->uh_ulen = htons(ulen);
18271                 }
18272         } else {
18273 #ifdef INET6
18274                 if (isipv6) {
18275                         ip6 = mtod(m, struct ip6_hdr *);
18276                         if (tp->t_port) {
18277                                 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18278                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18279                                 udp->uh_dport = tp->t_port;
18280                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
18281                                 udp->uh_ulen = htons(ulen);
18282                                 th = (struct tcphdr *)(udp + 1);
18283                         } else
18284                                 th = (struct tcphdr *)(ip6 + 1);
18285                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
18286                 } else
18287 #endif                          /* INET6 */
18288                 {
18289 #ifdef INET
18290                         ip = mtod(m, struct ip *);
18291                         if (tp->t_port) {
18292                                 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18293                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18294                                 udp->uh_dport = tp->t_port;
18295                                 ulen = hdrlen + len - sizeof(struct ip);
18296                                 udp->uh_ulen = htons(ulen);
18297                                 th = (struct tcphdr *)(udp + 1);
18298                         } else
18299                                 th = (struct tcphdr *)(ip + 1);
18300                         tcpip_fillheaders(inp, tp->t_port, ip, th);
18301 #endif
18302                 }
18303         }
18304         /*
18305          * Fill in fields, remembering maximum advertised window for use in
18306          * delaying messages about window sizes. If resending a FIN, be sure
18307          * not to use a new sequence number.
18308          */
18309         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18310             tp->snd_nxt == tp->snd_max)
18311                 tp->snd_nxt--;
18312         /*
18313          * If we are starting a connection, send ECN setup SYN packet. If we
18314          * are on a retransmit, we may resend those bits a number of times
18315          * as per RFC 3168.
18316          */
18317         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
18318                 flags |= tcp_ecn_output_syn_sent(tp);
18319         }
18320         /* Also handle parallel SYN for ECN */
18321         if (TCPS_HAVERCVDSYN(tp->t_state) &&
18322             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18323                 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
18324                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18325                     (tp->t_flags2 & TF2_ECN_SND_ECE))
18326                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18327 #ifdef INET6
18328                 if (isipv6) {
18329                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18330                         ip6->ip6_flow |= htonl(ect << 20);
18331                 }
18332                 else
18333 #endif
18334                 {
18335 #ifdef INET
18336                         ip->ip_tos &= ~IPTOS_ECN_MASK;
18337                         ip->ip_tos |= ect;
18338 #endif
18339                 }
18340         }
18341         /*
18342          * If we are doing retransmissions, then snd_nxt will not reflect
18343          * the first unsent octet.  For ACK only packets, we do not want the
18344          * sequence number of the retransmitted packet, we want the sequence
18345          * number of the next unsent octet.  So, if there is no data (and no
18346          * SYN or FIN), use snd_max instead of snd_nxt when filling in
18347          * ti_seq.  But if we are in persist state, snd_max might reflect
18348          * one byte beyond the right edge of the window, so use snd_nxt in
18349          * that case, since we know we aren't doing a retransmission.
18350          * (retransmit and persist are mutually exclusive...)
18351          */
18352         if (sack_rxmit == 0) {
18353                 if (len || (flags & (TH_SYN | TH_FIN))) {
18354                         th->th_seq = htonl(tp->snd_nxt);
18355                         rack_seq = tp->snd_nxt;
18356                 } else {
18357                         th->th_seq = htonl(tp->snd_max);
18358                         rack_seq = tp->snd_max;
18359                 }
18360         } else {
18361                 th->th_seq = htonl(rsm->r_start);
18362                 rack_seq = rsm->r_start;
18363         }
18364         th->th_ack = htonl(tp->rcv_nxt);
18365         tcp_set_flags(th, flags);
18366         /*
18367          * Calculate receive window.  Don't shrink window, but avoid silly
18368          * window syndrome.
18369          * If a RST segment is sent, advertise a window of zero.
18370          */
18371         if (flags & TH_RST) {
18372                 recwin = 0;
18373         } else {
18374                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18375                     recwin < (long)segsiz) {
18376                         recwin = 0;
18377                 }
18378                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18379                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18380                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18381         }
18382
18383         /*
18384          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18385          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18386          * handled in syncache.
18387          */
18388         if (flags & TH_SYN)
18389                 th->th_win = htons((u_short)
18390                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18391         else {
18392                 /* Avoid shrinking window with window scaling. */
18393                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
18394                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18395         }
18396         /*
18397          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18398          * window.  This may cause the remote transmitter to stall.  This
18399          * flag tells soreceive() to disable delayed acknowledgements when
18400          * draining the buffer.  This can occur if the receiver is
18401          * attempting to read more data than can be buffered prior to
18402          * transmitting on the connection.
18403          */
18404         if (th->th_win == 0) {
18405                 tp->t_sndzerowin++;
18406                 tp->t_flags |= TF_RXWIN0SENT;
18407         } else
18408                 tp->t_flags &= ~TF_RXWIN0SENT;
18409         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
18410         /* Now are we using fsb?, if so copy the template data to the mbuf */
18411         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18412                 uint8_t *cpto;
18413
18414                 cpto = mtod(m, uint8_t *);
18415                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18416                 /*
18417                  * We have just copied in:
18418                  * IP/IP6
18419                  * <optional udphdr>
18420                  * tcphdr (no options)
18421                  *
18422                  * We need to grab the correct pointers into the mbuf
18423                  * for both the tcp header, and possibly the udp header (if tunneling).
18424                  * We do this by using the offset in the copy buffer and adding it
18425                  * to the mbuf base pointer (cpto).
18426                  */
18427 #ifdef INET6
18428                 if (isipv6)
18429                         ip6 = mtod(m, struct ip6_hdr *);
18430                 else
18431 #endif                          /* INET6 */
18432 #ifdef INET
18433                         ip = mtod(m, struct ip *);
18434 #endif
18435                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18436                 /* If we have a udp header lets set it into the mbuf as well */
18437                 if (udp)
18438                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18439         }
18440 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18441         if (to.to_flags & TOF_SIGNATURE) {
18442                 /*
18443                  * Calculate MD5 signature and put it into the place
18444                  * determined before.
18445                  * NOTE: since TCP options buffer doesn't point into
18446                  * mbuf's data, calculate offset and use it.
18447                  */
18448                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18449                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18450                         /*
18451                          * Do not send segment if the calculation of MD5
18452                          * digest has failed.
18453                          */
18454                         goto out;
18455                 }
18456         }
18457 #endif
18458         if (optlen) {
18459                 bcopy(opt, th + 1, optlen);
18460                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18461         }
18462         /*
18463          * Put TCP length in extended header, and then checksum extended
18464          * header and data.
18465          */
18466         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
18467 #ifdef INET6
18468         if (isipv6) {
18469                 /*
18470                  * ip6_plen is not need to be filled now, and will be filled
18471                  * in ip6_output.
18472                  */
18473                 if (tp->t_port) {
18474                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18475                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18476                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18477                         th->th_sum = htons(0);
18478                         UDPSTAT_INC(udps_opackets);
18479                 } else {
18480                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18481                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18482                         th->th_sum = in6_cksum_pseudo(ip6,
18483                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18484                                                       0);
18485                 }
18486         }
18487 #endif
18488 #if defined(INET6) && defined(INET)
18489         else
18490 #endif
18491 #ifdef INET
18492         {
18493                 if (tp->t_port) {
18494                         m->m_pkthdr.csum_flags = CSUM_UDP;
18495                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18496                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18497                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18498                         th->th_sum = htons(0);
18499                         UDPSTAT_INC(udps_opackets);
18500                 } else {
18501                         m->m_pkthdr.csum_flags = CSUM_TCP;
18502                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18503                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
18504                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18505                                                                         IPPROTO_TCP + len + optlen));
18506                 }
18507                 /* IP version must be set here for ipv4/ipv6 checking later */
18508                 KASSERT(ip->ip_v == IPVERSION,
18509                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
18510         }
18511 #endif
18512         /*
18513          * Enable TSO and specify the size of the segments. The TCP pseudo
18514          * header checksum is always provided. XXX: Fixme: This is currently
18515          * not the case for IPv6.
18516          */
18517         if (tso) {
18518                 KASSERT(len > tp->t_maxseg - optlen,
18519                         ("%s: len <= tso_segsz", __func__));
18520                 m->m_pkthdr.csum_flags |= CSUM_TSO;
18521                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18522         }
18523         KASSERT(len + hdrlen == m_length(m, NULL),
18524                 ("%s: mbuf chain different than expected: %d + %u != %u",
18525                  __func__, len, hdrlen, m_length(m, NULL)));
18526
18527 #ifdef TCP_HHOOK
18528         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18529         hhook_run_tcp_est_out(tp, th, &to, len, tso);
18530 #endif
18531         /* We're getting ready to send; log now. */
18532         if (tcp_bblogging_on(rack->rc_tp)) {
18533                 union tcp_log_stackspecific log;
18534
18535                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18536                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18537                 if (rack->rack_no_prr)
18538                         log.u_bbr.flex1 = 0;
18539                 else
18540                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18541                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18542                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18543                 log.u_bbr.flex4 = orig_len;
18544                 /* Save off the early/late values */
18545                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18546                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18547                 log.u_bbr.bw_inuse = rack_get_bw(rack);
18548                 log.u_bbr.flex8 = 0;
18549                 if (rsm) {
18550                         if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18551                                 rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18552                                 counter_u64_add(rack_collapsed_win_rxt, 1);
18553                                 counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18554                         }
18555                         if (doing_tlp)
18556                                 log.u_bbr.flex8 = 2;
18557                         else
18558                                 log.u_bbr.flex8 = 1;
18559                 } else {
18560                         if (doing_tlp)
18561                                 log.u_bbr.flex8 = 3;
18562                         else
18563                                 log.u_bbr.flex8 = 0;
18564                 }
18565                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18566                 log.u_bbr.flex7 = mark;
18567                 log.u_bbr.flex7 <<= 8;
18568                 log.u_bbr.flex7 |= pass;
18569                 log.u_bbr.pkts_out = tp->t_maxseg;
18570                 log.u_bbr.timeStamp = cts;
18571                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18572                 log.u_bbr.lt_epoch = cwnd_to_use;
18573                 log.u_bbr.delivered = sendalot;
18574                 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18575                                      len, &log, false, NULL, NULL, 0, &tv);
18576         } else
18577                 lgb = NULL;
18578
18579         /*
18580          * Fill in IP length and desired time to live and send to IP level.
18581          * There should be a better way to handle ttl and tos; we could keep
18582          * them in the template, but need a way to checksum without them.
18583          */
18584         /*
18585          * m->m_pkthdr.len should have been set before cksum calcuration,
18586          * because in6_cksum() need it.
18587          */
18588 #ifdef INET6
18589         if (isipv6) {
18590                 /*
18591                  * we separately set hoplimit for every segment, since the
18592                  * user might want to change the value via setsockopt. Also,
18593                  * desired default hop limit might be changed via Neighbor
18594                  * Discovery.
18595                  */
18596                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18597
18598                 /*
18599                  * Set the packet size here for the benefit of DTrace
18600                  * probes. ip6_output() will set it properly; it's supposed
18601                  * to include the option header lengths as well.
18602                  */
18603                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18604
18605                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18606                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18607                 else
18608                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18609
18610                 if (tp->t_state == TCPS_SYN_SENT)
18611                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18612
18613                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18614                 /* TODO: IPv6 IP6TOS_ECT bit on */
18615                 error = ip6_output(m,
18616 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18617                                    inp->in6p_outputopts,
18618 #else
18619                                    NULL,
18620 #endif
18621                                    &inp->inp_route6,
18622                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18623                                    NULL, NULL, inp);
18624
18625                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18626                         mtu = inp->inp_route6.ro_nh->nh_mtu;
18627         }
18628 #endif                          /* INET6 */
18629 #if defined(INET) && defined(INET6)
18630         else
18631 #endif
18632 #ifdef INET
18633         {
18634                 ip->ip_len = htons(m->m_pkthdr.len);
18635 #ifdef INET6
18636                 if (inp->inp_vflag & INP_IPV6PROTO)
18637                         ip->ip_ttl = in6_selecthlim(inp, NULL);
18638 #endif                          /* INET6 */
18639                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18640                 /*
18641                  * If we do path MTU discovery, then we set DF on every
18642                  * packet. This might not be the best thing to do according
18643                  * to RFC3390 Section 2. However the tcp hostcache migitates
18644                  * the problem so it affects only the first tcp connection
18645                  * with a host.
18646                  *
18647                  * NB: Don't set DF on small MTU/MSS to have a safe
18648                  * fallback.
18649                  */
18650                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18651                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18652                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18653                                 ip->ip_off |= htons(IP_DF);
18654                         }
18655                 } else {
18656                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18657                 }
18658
18659                 if (tp->t_state == TCPS_SYN_SENT)
18660                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18661
18662                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18663
18664                 error = ip_output(m,
18665 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18666                                   inp->inp_options,
18667 #else
18668                                   NULL,
18669 #endif
18670                                   &inp->inp_route,
18671                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18672                                   inp);
18673                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18674                         mtu = inp->inp_route.ro_nh->nh_mtu;
18675         }
18676 #endif                          /* INET */
18677
18678 out:
18679         if (lgb) {
18680                 lgb->tlb_errno = error;
18681                 lgb = NULL;
18682         }
18683         /*
18684          * In transmit state, time the transmission and arrange for the
18685          * retransmit.  In persist state, just set snd_max.
18686          */
18687         if (error == 0) {
18688                 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18689                 if (rsm && doing_tlp) {
18690                         rack->rc_last_sent_tlp_past_cumack = 0;
18691                         rack->rc_last_sent_tlp_seq_valid = 1;
18692                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18693                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18694                 }
18695                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
18696                 if (rsm && (doing_tlp == 0)) {
18697                         /* Set we retransmitted */
18698                         rack->rc_gp_saw_rec = 1;
18699                 } else {
18700                         if (cwnd_to_use > tp->snd_ssthresh) {
18701                                 /* Set we sent in CA */
18702                                 rack->rc_gp_saw_ca = 1;
18703                         } else {
18704                                 /* Set we sent in SS */
18705                                 rack->rc_gp_saw_ss = 1;
18706                         }
18707                 }
18708                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18709                     (tp->t_flags & TF_SACK_PERMIT) &&
18710                     tp->rcv_numsacks > 0)
18711                         tcp_clean_dsack_blocks(tp);
18712                 tot_len_this_send += len;
18713                 if (len == 0)
18714                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18715                 else if (len == 1) {
18716                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18717                 } else if (len > 1) {
18718                         int idx;
18719
18720                         idx = (len / segsiz) + 3;
18721                         if (idx >= TCP_MSS_ACCT_ATIMER)
18722                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18723                         else
18724                                 counter_u64_add(rack_out_size[idx], 1);
18725                 }
18726         }
18727         if ((rack->rack_no_prr == 0) &&
18728             sub_from_prr &&
18729             (error == 0)) {
18730                 if (rack->r_ctl.rc_prr_sndcnt >= len)
18731                         rack->r_ctl.rc_prr_sndcnt -= len;
18732                 else
18733                         rack->r_ctl.rc_prr_sndcnt = 0;
18734         }
18735         sub_from_prr = 0;
18736         if (doing_tlp) {
18737                 /* Make sure the TLP is added */
18738                 add_flag |= RACK_TLP;
18739         } else if (rsm) {
18740                 /* If its a resend without TLP then it must not have the flag */
18741                 rsm->r_flags &= ~RACK_TLP;
18742         }
18743         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18744                         rack_to_usec_ts(&tv),
18745                         rsm, add_flag, s_mb, s_moff, hw_tls);
18746
18747
18748         if ((error == 0) &&
18749             (len > 0) &&
18750             (tp->snd_una == tp->snd_max))
18751                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18752         {
18753                 tcp_seq startseq = tp->snd_nxt;
18754
18755                 /* Track our lost count */
18756                 if (rsm && (doing_tlp == 0))
18757                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18758                 /*
18759                  * Advance snd_nxt over sequence space of this segment.
18760                  */
18761                 if (error)
18762                         /* We don't log or do anything with errors */
18763                         goto nomore;
18764                 if (doing_tlp == 0) {
18765                         if (rsm == NULL) {
18766                                 /*
18767                                  * Not a retransmission of some
18768                                  * sort, new data is going out so
18769                                  * clear our TLP count and flag.
18770                                  */
18771                                 rack->rc_tlp_in_progress = 0;
18772                                 rack->r_ctl.rc_tlp_cnt_out = 0;
18773                         }
18774                 } else {
18775                         /*
18776                          * We have just sent a TLP, mark that it is true
18777                          * and make sure our in progress is set so we
18778                          * continue to check the count.
18779                          */
18780                         rack->rc_tlp_in_progress = 1;
18781                         rack->r_ctl.rc_tlp_cnt_out++;
18782                 }
18783                 if (flags & (TH_SYN | TH_FIN)) {
18784                         if (flags & TH_SYN)
18785                                 tp->snd_nxt++;
18786                         if (flags & TH_FIN) {
18787                                 tp->snd_nxt++;
18788                                 tp->t_flags |= TF_SENTFIN;
18789                         }
18790                 }
18791                 /* In the ENOBUFS case we do *not* update snd_max */
18792                 if (sack_rxmit)
18793                         goto nomore;
18794
18795                 tp->snd_nxt += len;
18796                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18797                         if (tp->snd_una == tp->snd_max) {
18798                                 /*
18799                                  * Update the time we just added data since
18800                                  * none was outstanding.
18801                                  */
18802                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18803                                 tp->t_acktime = ticks;
18804                         }
18805                         tp->snd_max = tp->snd_nxt;
18806                         /*
18807                          * Time this transmission if not a retransmission and
18808                          * not currently timing anything.
18809                          * This is only relevant in case of switching back to
18810                          * the base stack.
18811                          */
18812                         if (tp->t_rtttime == 0) {
18813                                 tp->t_rtttime = ticks;
18814                                 tp->t_rtseq = startseq;
18815                                 KMOD_TCPSTAT_INC(tcps_segstimed);
18816                         }
18817                         if (len &&
18818                             ((tp->t_flags & TF_GPUTINPROG) == 0))
18819                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18820                 }
18821                 /*
18822                  * If we are doing FO we need to update the mbuf position and subtract
18823                  * this happens when the peer sends us duplicate information and
18824                  * we thus want to send a DSACK.
18825                  *
18826                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18827                  * turned off? If not then we are going to echo multiple DSACK blocks
18828                  * out (with the TSO), which we should not be doing.
18829                  */
18830                 if (rack->r_fast_output && len) {
18831                         if (rack->r_ctl.fsb.left_to_send > len)
18832                                 rack->r_ctl.fsb.left_to_send -= len;
18833                         else
18834                                 rack->r_ctl.fsb.left_to_send = 0;
18835                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18836                                 rack->r_fast_output = 0;
18837                         if (rack->r_fast_output) {
18838                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18839                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18840                         }
18841                 }
18842         }
18843 nomore:
18844         if (error) {
18845                 rack->r_ctl.rc_agg_delayed = 0;
18846                 rack->r_early = 0;
18847                 rack->r_late = 0;
18848                 rack->r_ctl.rc_agg_early = 0;
18849                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
18850                 /*
18851                  * Failures do not advance the seq counter above. For the
18852                  * case of ENOBUFS we will fall out and retry in 1ms with
18853                  * the hpts. Everything else will just have to retransmit
18854                  * with the timer.
18855                  *
18856                  * In any case, we do not want to loop around for another
18857                  * send without a good reason.
18858                  */
18859                 sendalot = 0;
18860                 switch (error) {
18861                 case EPERM:
18862                         tp->t_softerror = error;
18863 #ifdef TCP_ACCOUNTING
18864                         crtsc = get_cyclecount();
18865                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18866                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18867                         }
18868                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18869                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18870                         }
18871                         sched_unpin();
18872 #endif
18873                         return (error);
18874                 case ENOBUFS:
18875                         /*
18876                          * Pace us right away to retry in a some
18877                          * time
18878                          */
18879                         if (rack->r_ctl.crte != NULL) {
18880                                 tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
18881                         } else
18882                                 tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
18883                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18884                         if (rack->rc_enobuf < 0x7f)
18885                                 rack->rc_enobuf++;
18886                         if (slot < (10 * HPTS_USEC_IN_MSEC))
18887                                 slot = 10 * HPTS_USEC_IN_MSEC;
18888                         if (rack->r_ctl.crte != NULL) {
18889                                 counter_u64_add(rack_saw_enobuf_hw, 1);
18890                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
18891                         }
18892                         counter_u64_add(rack_saw_enobuf, 1);
18893                         goto enobufs;
18894                 case EMSGSIZE:
18895                         /*
18896                          * For some reason the interface we used initially
18897                          * to send segments changed to another or lowered
18898                          * its MTU. If TSO was active we either got an
18899                          * interface without TSO capabilits or TSO was
18900                          * turned off. If we obtained mtu from ip_output()
18901                          * then update it and try again.
18902                          */
18903                         if (tso)
18904                                 tp->t_flags &= ~TF_TSO;
18905                         if (mtu != 0) {
18906                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
18907                                 goto again;
18908                         }
18909                         slot = 10 * HPTS_USEC_IN_MSEC;
18910                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18911 #ifdef TCP_ACCOUNTING
18912                         crtsc = get_cyclecount();
18913                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18914                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18915                         }
18916                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18917                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18918                         }
18919                         sched_unpin();
18920 #endif
18921                         return (error);
18922                 case ENETUNREACH:
18923                         counter_u64_add(rack_saw_enetunreach, 1);
18924                 case EHOSTDOWN:
18925                 case EHOSTUNREACH:
18926                 case ENETDOWN:
18927                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
18928                                 tp->t_softerror = error;
18929                         }
18930                         /* FALLTHROUGH */
18931                 default:
18932                         slot = 10 * HPTS_USEC_IN_MSEC;
18933                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18934 #ifdef TCP_ACCOUNTING
18935                         crtsc = get_cyclecount();
18936                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18937                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18938                         }
18939                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18940                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18941                         }
18942                         sched_unpin();
18943 #endif
18944                         return (error);
18945                 }
18946         } else {
18947                 rack->rc_enobuf = 0;
18948                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18949                         rack->r_ctl.retran_during_recovery += len;
18950         }
18951         KMOD_TCPSTAT_INC(tcps_sndtotal);
18952
18953         /*
18954          * Data sent (as far as we can tell). If this advertises a larger
18955          * window than any other segment, then remember the size of the
18956          * advertised window. Any pending ACK has now been sent.
18957          */
18958         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
18959                 tp->rcv_adv = tp->rcv_nxt + recwin;
18960
18961         tp->last_ack_sent = tp->rcv_nxt;
18962         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18963 enobufs:
18964         if (sendalot) {
18965                 /* Do we need to turn off sendalot? */
18966                 if (rack->r_ctl.rc_pace_max_segs &&
18967                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
18968                         /* We hit our max. */
18969                         sendalot = 0;
18970                 } else if ((rack->rc_user_set_max_segs) &&
18971                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
18972                         /* We hit the user defined max */
18973                         sendalot = 0;
18974                 }
18975         }
18976         if ((error == 0) && (flags & TH_FIN))
18977                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
18978         if (flags & TH_RST) {
18979                 /*
18980                  * We don't send again after sending a RST.
18981                  */
18982                 slot = 0;
18983                 sendalot = 0;
18984                 if (error == 0)
18985                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
18986         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
18987                 /*
18988                  * Get our pacing rate, if an error
18989                  * occurred in sending (ENOBUF) we would
18990                  * hit the else if with slot preset. Other
18991                  * errors return.
18992                  */
18993                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
18994         }
18995         if (rsm &&
18996             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
18997             rack->use_rack_rr) {
18998                 /* Its a retransmit and we use the rack cheat? */
18999                 if ((slot == 0) ||
19000                     (rack->rc_always_pace == 0) ||
19001                     (rack->r_rr_config == 1)) {
19002                         /*
19003                          * We have no pacing set or we
19004                          * are using old-style rack or
19005                          * we are overridden to use the old 1ms pacing.
19006                          */
19007                         slot = rack->r_ctl.rc_min_to;
19008                 }
19009         }
19010         /* We have sent clear the flag */
19011         rack->r_ent_rec_ns = 0;
19012         if (rack->r_must_retran) {
19013                 if (rsm) {
19014                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
19015                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
19016                                 /*
19017                                  * We have retransmitted all.
19018                                  */
19019                                 rack->r_must_retran = 0;
19020                                 rack->r_ctl.rc_out_at_rto = 0;
19021                         }
19022                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19023                         /*
19024                          * Sending new data will also kill
19025                          * the loop.
19026                          */
19027                         rack->r_must_retran = 0;
19028                         rack->r_ctl.rc_out_at_rto = 0;
19029                 }
19030         }
19031         rack->r_ctl.fsb.recwin = recwin;
19032         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
19033             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19034                 /*
19035                  * We hit an RTO and now have past snd_max at the RTO
19036                  * clear all the WAS flags.
19037                  */
19038                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
19039         }
19040         if (slot) {
19041                 /* set the rack tcb into the slot N */
19042                 if ((error == 0) &&
19043                     rack_use_rfo &&
19044                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
19045                     (rsm == NULL) &&
19046                     (tp->snd_nxt == tp->snd_max) &&
19047                     (ipoptlen == 0) &&
19048                     (tp->rcv_numsacks == 0) &&
19049                     rack->r_fsb_inited &&
19050                     TCPS_HAVEESTABLISHED(tp->t_state) &&
19051                     (rack->r_must_retran == 0) &&
19052                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
19053                     (len > 0) && (orig_len > 0) &&
19054                     (orig_len > len) &&
19055                     ((orig_len - len) >= segsiz) &&
19056                     ((optlen == 0) ||
19057                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19058                         /* We can send at least one more MSS using our fsb */
19059
19060                         rack->r_fast_output = 1;
19061                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19062                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19063                         rack->r_ctl.fsb.tcp_flags = flags;
19064                         rack->r_ctl.fsb.left_to_send = orig_len - len;
19065                         if (hw_tls)
19066                                 rack->r_ctl.fsb.hw_tls = 1;
19067                         else
19068                                 rack->r_ctl.fsb.hw_tls = 0;
19069                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19070                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19071                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19072                                  (tp->snd_max - tp->snd_una)));
19073                         if (rack->r_ctl.fsb.left_to_send < segsiz)
19074                                 rack->r_fast_output = 0;
19075                         else {
19076                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19077                                         rack->r_ctl.fsb.rfo_apply_push = 1;
19078                                 else
19079                                         rack->r_ctl.fsb.rfo_apply_push = 0;
19080                         }
19081                 } else
19082                         rack->r_fast_output = 0;
19083                 rack_log_fsb(rack, tp, so, flags,
19084                              ipoptlen, orig_len, len, error,
19085                              (rsm == NULL), optlen, __LINE__, 2);
19086         } else if (sendalot) {
19087                 int ret;
19088
19089                 sack_rxmit = 0;
19090                 if ((error == 0) &&
19091                     rack_use_rfo &&
19092                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
19093                     (rsm == NULL) &&
19094                     (ipoptlen == 0) &&
19095                     (tp->rcv_numsacks == 0) &&
19096                     (tp->snd_nxt == tp->snd_max) &&
19097                     (rack->r_must_retran == 0) &&
19098                     rack->r_fsb_inited &&
19099                     TCPS_HAVEESTABLISHED(tp->t_state) &&
19100                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
19101                     (len > 0) && (orig_len > 0) &&
19102                     (orig_len > len) &&
19103                     ((orig_len - len) >= segsiz) &&
19104                     ((optlen == 0) ||
19105                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19106                         /* we can use fast_output for more */
19107
19108                         rack->r_fast_output = 1;
19109                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19110                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19111                         rack->r_ctl.fsb.tcp_flags = flags;
19112                         rack->r_ctl.fsb.left_to_send = orig_len - len;
19113                         if (hw_tls)
19114                                 rack->r_ctl.fsb.hw_tls = 1;
19115                         else
19116                                 rack->r_ctl.fsb.hw_tls = 0;
19117                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19118                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19119                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19120                                  (tp->snd_max - tp->snd_una)));
19121                         if (rack->r_ctl.fsb.left_to_send < segsiz) {
19122                                 rack->r_fast_output = 0;
19123                         }
19124                         if (rack->r_fast_output) {
19125                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19126                                         rack->r_ctl.fsb.rfo_apply_push = 1;
19127                                 else
19128                                         rack->r_ctl.fsb.rfo_apply_push = 0;
19129                                 rack_log_fsb(rack, tp, so, flags,
19130                                              ipoptlen, orig_len, len, error,
19131                                              (rsm == NULL), optlen, __LINE__, 3);
19132                                 error = 0;
19133                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19134                                 if (ret >= 0)
19135                                         return (ret);
19136                                 else if (error)
19137                                         goto nomore;
19138
19139                         }
19140                 }
19141                 goto again;
19142         }
19143         /* Assure when we leave that snd_nxt will point to top */
19144         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19145                 tp->snd_nxt = tp->snd_max;
19146         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19147 #ifdef TCP_ACCOUNTING
19148         crtsc = get_cyclecount() - ts_val;
19149         if (tot_len_this_send) {
19150                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19151                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
19152                 }
19153                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19154                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19155                 }
19156                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19157                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19158                 }
19159         } else {
19160                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19161                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
19162                 }
19163                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19164                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19165                 }
19166         }
19167         sched_unpin();
19168 #endif
19169         if (error == ENOBUFS)
19170                 error = 0;
19171         return (error);
19172 }
19173
19174 static void
19175 rack_update_seg(struct tcp_rack *rack)
19176 {
19177         uint32_t orig_val;
19178
19179         orig_val = rack->r_ctl.rc_pace_max_segs;
19180         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19181         if (orig_val != rack->r_ctl.rc_pace_max_segs)
19182                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19183 }
19184
19185 static void
19186 rack_mtu_change(struct tcpcb *tp)
19187 {
19188         /*
19189          * The MSS may have changed
19190          */
19191         struct tcp_rack *rack;
19192         struct rack_sendmap *rsm;
19193
19194         rack = (struct tcp_rack *)tp->t_fb_ptr;
19195         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19196                 /*
19197                  * The MTU has changed we need to resend everything
19198                  * since all we have sent is lost. We first fix
19199                  * up the mtu though.
19200                  */
19201                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19202                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
19203                 rack_remxt_tmr(tp);
19204                 rack->r_fast_output = 0;
19205                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19206                                                 rack->r_ctl.rc_sacked);
19207                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19208                 rack->r_must_retran = 1;
19209                 /* Mark all inflight to needing to be rxt'd */
19210                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19211                         rsm->r_flags |= RACK_MUST_RXT;
19212                 }
19213         }
19214         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19215         /* We don't use snd_nxt to retransmit */
19216         tp->snd_nxt = tp->snd_max;
19217 }
19218
19219 static int
19220 rack_set_profile(struct tcp_rack *rack, int prof)
19221 {
19222         int err = EINVAL;
19223         if (prof == 1) {
19224                 /* pace_always=1 */
19225                 if (rack->rc_always_pace == 0) {
19226                         if (tcp_can_enable_pacing() == 0)
19227                                 return (EBUSY);
19228                 }
19229                 rack->rc_always_pace = 1;
19230                 if (rack->use_fixed_rate || rack->gp_ready)
19231                         rack_set_cc_pacing(rack);
19232                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19233                 rack->rack_attempt_hdwr_pace = 0;
19234                 /* cmpack=1 */
19235                 if (rack_use_cmp_acks)
19236                         rack->r_use_cmp_ack = 1;
19237                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19238                     rack->r_use_cmp_ack)
19239                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19240                 /* scwnd=1 */
19241                 rack->rack_enable_scwnd = 1;
19242                 /* dynamic=100 */
19243                 rack->rc_gp_dyn_mul = 1;
19244                 /* gp_inc_ca */
19245                 rack->r_ctl.rack_per_of_gp_ca = 100;
19246                 /* rrr_conf=3 */
19247                 rack->r_rr_config = 3;
19248                 /* npush=2 */
19249                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19250                 /* fillcw=1 */
19251                 rack->rc_pace_to_cwnd = 1;
19252                 rack->rc_pace_fill_if_rttin_range = 0;
19253                 rack->rtt_limit_mul = 0;
19254                 /* noprr=1 */
19255                 rack->rack_no_prr = 1;
19256                 /* lscwnd=1 */
19257                 rack->r_limit_scw = 1;
19258                 /* gp_inc_rec */
19259                 rack->r_ctl.rack_per_of_gp_rec = 90;
19260                 err = 0;
19261
19262         } else if (prof == 3) {
19263                 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19264                 /* pace_always=1 */
19265                 if (rack->rc_always_pace == 0) {
19266                         if (tcp_can_enable_pacing() == 0)
19267                                 return (EBUSY);
19268                 }
19269                 rack->rc_always_pace = 1;
19270                 if (rack->use_fixed_rate || rack->gp_ready)
19271                         rack_set_cc_pacing(rack);
19272                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19273                 rack->rack_attempt_hdwr_pace = 0;
19274                 /* cmpack=1 */
19275                 if (rack_use_cmp_acks)
19276                         rack->r_use_cmp_ack = 1;
19277                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19278                     rack->r_use_cmp_ack)
19279                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19280                 /* scwnd=1 */
19281                 rack->rack_enable_scwnd = 1;
19282                 /* dynamic=100 */
19283                 rack->rc_gp_dyn_mul = 1;
19284                 /* gp_inc_ca */
19285                 rack->r_ctl.rack_per_of_gp_ca = 100;
19286                 /* rrr_conf=3 */
19287                 rack->r_rr_config = 3;
19288                 /* npush=2 */
19289                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19290                 /* fillcw=2 */
19291                 rack->rc_pace_to_cwnd = 1;
19292                 rack->r_fill_less_agg = 1;
19293                 rack->rc_pace_fill_if_rttin_range = 0;
19294                 rack->rtt_limit_mul = 0;
19295                 /* noprr=1 */
19296                 rack->rack_no_prr = 1;
19297                 /* lscwnd=1 */
19298                 rack->r_limit_scw = 1;
19299                 /* gp_inc_rec */
19300                 rack->r_ctl.rack_per_of_gp_rec = 90;
19301                 err = 0;
19302
19303
19304         } else if (prof == 2) {
19305                 /* cmpack=1 */
19306                 if (rack->rc_always_pace == 0) {
19307                         if (tcp_can_enable_pacing() == 0)
19308                                 return (EBUSY);
19309                 }
19310                 rack->rc_always_pace = 1;
19311                 if (rack->use_fixed_rate || rack->gp_ready)
19312                         rack_set_cc_pacing(rack);
19313                 rack->r_use_cmp_ack = 1;
19314                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19315                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19316                 /* pace_always=1 */
19317                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19318                 /* scwnd=1 */
19319                 rack->rack_enable_scwnd = 1;
19320                 /* dynamic=100 */
19321                 rack->rc_gp_dyn_mul = 1;
19322                 rack->r_ctl.rack_per_of_gp_ca = 100;
19323                 /* rrr_conf=3 */
19324                 rack->r_rr_config = 3;
19325                 /* npush=2 */
19326                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19327                 /* fillcw=1 */
19328                 rack->rc_pace_to_cwnd = 1;
19329                 rack->rc_pace_fill_if_rttin_range = 0;
19330                 rack->rtt_limit_mul = 0;
19331                 /* noprr=1 */
19332                 rack->rack_no_prr = 1;
19333                 /* lscwnd=0 */
19334                 rack->r_limit_scw = 0;
19335                 err = 0;
19336         } else if (prof == 0) {
19337                 /* This changes things back to the default settings */
19338                 err = 0;
19339                 if (rack->rc_always_pace) {
19340                         tcp_decrement_paced_conn();
19341                         rack_undo_cc_pacing(rack);
19342                         rack->rc_always_pace = 0;
19343                 }
19344                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19345                         rack->rc_always_pace = 1;
19346                         if (rack->use_fixed_rate || rack->gp_ready)
19347                                 rack_set_cc_pacing(rack);
19348                 } else
19349                         rack->rc_always_pace = 0;
19350                 if (rack_dsack_std_based & 0x1) {
19351                         /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19352                         rack->rc_rack_tmr_std_based = 1;
19353                 }
19354                 if (rack_dsack_std_based & 0x2) {
19355                         /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19356                         rack->rc_rack_use_dsack = 1;
19357                 }
19358                 if (rack_use_cmp_acks)
19359                         rack->r_use_cmp_ack = 1;
19360                 else
19361                         rack->r_use_cmp_ack = 0;
19362                 if (rack_disable_prr)
19363                         rack->rack_no_prr = 1;
19364                 else
19365                         rack->rack_no_prr = 0;
19366                 if (rack_gp_no_rec_chg)
19367                         rack->rc_gp_no_rec_chg = 1;
19368                 else
19369                         rack->rc_gp_no_rec_chg = 0;
19370                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19371                         rack->r_mbuf_queue = 1;
19372                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19373                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19374                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19375                 } else {
19376                         rack->r_mbuf_queue = 0;
19377                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19378                 }
19379                 if (rack_enable_shared_cwnd)
19380                         rack->rack_enable_scwnd = 1;
19381                 else
19382                         rack->rack_enable_scwnd = 0;
19383                 if (rack_do_dyn_mul) {
19384                         /* When dynamic adjustment is on CA needs to start at 100% */
19385                         rack->rc_gp_dyn_mul = 1;
19386                         if (rack_do_dyn_mul >= 100)
19387                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19388                 } else {
19389                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19390                         rack->rc_gp_dyn_mul = 0;
19391                 }
19392                 rack->r_rr_config = 0;
19393                 rack->r_ctl.rc_no_push_at_mrtt = 0;
19394                 rack->rc_pace_to_cwnd = 0;
19395                 rack->rc_pace_fill_if_rttin_range = 0;
19396                 rack->rtt_limit_mul = 0;
19397
19398                 if (rack_enable_hw_pacing)
19399                         rack->rack_hdw_pace_ena = 1;
19400                 else
19401                         rack->rack_hdw_pace_ena = 0;
19402                 if (rack_disable_prr)
19403                         rack->rack_no_prr = 1;
19404                 else
19405                         rack->rack_no_prr = 0;
19406                 if (rack_limits_scwnd)
19407                         rack->r_limit_scw  = 1;
19408                 else
19409                         rack->r_limit_scw  = 0;
19410                 err = 0;
19411         }
19412         return (err);
19413 }
19414
19415 static int
19416 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19417 {
19418         struct deferred_opt_list *dol;
19419
19420         dol = malloc(sizeof(struct deferred_opt_list),
19421                      M_TCPFSB, M_NOWAIT|M_ZERO);
19422         if (dol == NULL) {
19423                 /*
19424                  * No space yikes -- fail out..
19425                  */
19426                 return (0);
19427         }
19428         dol->optname = sopt_name;
19429         dol->optval = loptval;
19430         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19431         return (1);
19432 }
19433
19434 static int
19435 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19436                     uint32_t optval, uint64_t loptval)
19437 {
19438         struct epoch_tracker et;
19439         struct sockopt sopt;
19440         struct cc_newreno_opts opt;
19441         struct inpcb *inp = tptoinpcb(tp);
19442         uint64_t val;
19443         int error = 0;
19444         uint16_t ca, ss;
19445
19446         switch (sopt_name) {
19447
19448         case TCP_RACK_DSACK_OPT:
19449                 RACK_OPTS_INC(tcp_rack_dsack_opt);
19450                 if (optval & 0x1) {
19451                         rack->rc_rack_tmr_std_based = 1;
19452                 } else {
19453                         rack->rc_rack_tmr_std_based = 0;
19454                 }
19455                 if (optval & 0x2) {
19456                         rack->rc_rack_use_dsack = 1;
19457                 } else {
19458                         rack->rc_rack_use_dsack = 0;
19459                 }
19460                 rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19461                 break;
19462         case TCP_RACK_PACING_BETA:
19463                 RACK_OPTS_INC(tcp_rack_beta);
19464                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
19465                         /* This only works for newreno. */
19466                         error = EINVAL;
19467                         break;
19468                 }
19469                 if (rack->rc_pacing_cc_set) {
19470                         /*
19471                          * Set them into the real CC module
19472                          * whats in the rack pcb is the old values
19473                          * to be used on restoral/
19474                          */
19475                         sopt.sopt_dir = SOPT_SET;
19476                         opt.name = CC_NEWRENO_BETA;
19477                         opt.val = optval;
19478                         if (CC_ALGO(tp)->ctl_output != NULL)
19479                                 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
19480                         else {
19481                                 error = ENOENT;
19482                                 break;
19483                         }
19484                 } else {
19485                         /*
19486                          * Not pacing yet so set it into our local
19487                          * rack pcb storage.
19488                          */
19489                         rack->r_ctl.rc_saved_beta.beta = optval;
19490                 }
19491                 break;
19492         case TCP_RACK_TIMER_SLOP:
19493                 RACK_OPTS_INC(tcp_rack_timer_slop);
19494                 rack->r_ctl.timer_slop = optval;
19495                 if (rack->rc_tp->t_srtt) {
19496                         /*
19497                          * If we have an SRTT lets update t_rxtcur
19498                          * to have the new slop.
19499                          */
19500                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19501                                            rack_rto_min, rack_rto_max,
19502                                            rack->r_ctl.timer_slop);
19503                 }
19504                 break;
19505         case TCP_RACK_PACING_BETA_ECN:
19506                 RACK_OPTS_INC(tcp_rack_beta_ecn);
19507                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
19508                         /* This only works for newreno. */
19509                         error = EINVAL;
19510                         break;
19511                 }
19512                 if (rack->rc_pacing_cc_set) {
19513                         /*
19514                          * Set them into the real CC module
19515                          * whats in the rack pcb is the old values
19516                          * to be used on restoral/
19517                          */
19518                         sopt.sopt_dir = SOPT_SET;
19519                         opt.name = CC_NEWRENO_BETA_ECN;
19520                         opt.val = optval;
19521                         if (CC_ALGO(tp)->ctl_output != NULL)
19522                                 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
19523                         else
19524                                 error = ENOENT;
19525                 } else {
19526                         /*
19527                          * Not pacing yet so set it into our local
19528                          * rack pcb storage.
19529                          */
19530                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19531                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19532                 }
19533                 break;
19534         case TCP_DEFER_OPTIONS:
19535                 RACK_OPTS_INC(tcp_defer_opt);
19536                 if (optval) {
19537                         if (rack->gp_ready) {
19538                                 /* Too late */
19539                                 error = EINVAL;
19540                                 break;
19541                         }
19542                         rack->defer_options = 1;
19543                 } else
19544                         rack->defer_options = 0;
19545                 break;
19546         case TCP_RACK_MEASURE_CNT:
19547                 RACK_OPTS_INC(tcp_rack_measure_cnt);
19548                 if (optval && (optval <= 0xff)) {
19549                         rack->r_ctl.req_measurements = optval;
19550                 } else
19551                         error = EINVAL;
19552                 break;
19553         case TCP_REC_ABC_VAL:
19554                 RACK_OPTS_INC(tcp_rec_abc_val);
19555                 if (optval > 0)
19556                         rack->r_use_labc_for_rec = 1;
19557                 else
19558                         rack->r_use_labc_for_rec = 0;
19559                 break;
19560         case TCP_RACK_ABC_VAL:
19561                 RACK_OPTS_INC(tcp_rack_abc_val);
19562                 if ((optval > 0) && (optval < 255))
19563                         rack->rc_labc = optval;
19564                 else
19565                         error = EINVAL;
19566                 break;
19567         case TCP_HDWR_UP_ONLY:
19568                 RACK_OPTS_INC(tcp_pacing_up_only);
19569                 if (optval)
19570                         rack->r_up_only = 1;
19571                 else
19572                         rack->r_up_only = 0;
19573                 break;
19574         case TCP_PACING_RATE_CAP:
19575                 RACK_OPTS_INC(tcp_pacing_rate_cap);
19576                 rack->r_ctl.bw_rate_cap = loptval;
19577                 break;
19578         case TCP_RACK_PROFILE:
19579                 RACK_OPTS_INC(tcp_profile);
19580                 error = rack_set_profile(rack, optval);
19581                 break;
19582         case TCP_USE_CMP_ACKS:
19583                 RACK_OPTS_INC(tcp_use_cmp_acks);
19584                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19585                         /* You can't turn it off once its on! */
19586                         error = EINVAL;
19587                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19588                         rack->r_use_cmp_ack = 1;
19589                         rack->r_mbuf_queue = 1;
19590                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19591                 }
19592                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19593                         inp->inp_flags2 |= INP_MBUF_ACKCMP;
19594                 break;
19595         case TCP_SHARED_CWND_TIME_LIMIT:
19596                 RACK_OPTS_INC(tcp_lscwnd);
19597                 if (optval)
19598                         rack->r_limit_scw = 1;
19599                 else
19600                         rack->r_limit_scw = 0;
19601                 break;
19602         case TCP_RACK_PACE_TO_FILL:
19603                 RACK_OPTS_INC(tcp_fillcw);
19604                 if (optval == 0)
19605                         rack->rc_pace_to_cwnd = 0;
19606                 else {
19607                         rack->rc_pace_to_cwnd = 1;
19608                         if (optval > 1)
19609                                 rack->r_fill_less_agg = 1;
19610                 }
19611                 if ((optval >= rack_gp_rtt_maxmul) &&
19612                     rack_gp_rtt_maxmul &&
19613                     (optval < 0xf)) {
19614                         rack->rc_pace_fill_if_rttin_range = 1;
19615                         rack->rtt_limit_mul = optval;
19616                 } else {
19617                         rack->rc_pace_fill_if_rttin_range = 0;
19618                         rack->rtt_limit_mul = 0;
19619                 }
19620                 break;
19621         case TCP_RACK_NO_PUSH_AT_MAX:
19622                 RACK_OPTS_INC(tcp_npush);
19623                 if (optval == 0)
19624                         rack->r_ctl.rc_no_push_at_mrtt = 0;
19625                 else if (optval < 0xff)
19626                         rack->r_ctl.rc_no_push_at_mrtt = optval;
19627                 else
19628                         error = EINVAL;
19629                 break;
19630         case TCP_SHARED_CWND_ENABLE:
19631                 RACK_OPTS_INC(tcp_rack_scwnd);
19632                 if (optval == 0)
19633                         rack->rack_enable_scwnd = 0;
19634                 else
19635                         rack->rack_enable_scwnd = 1;
19636                 break;
19637         case TCP_RACK_MBUF_QUEUE:
19638                 /* Now do we use the LRO mbuf-queue feature */
19639                 RACK_OPTS_INC(tcp_rack_mbufq);
19640                 if (optval || rack->r_use_cmp_ack)
19641                         rack->r_mbuf_queue = 1;
19642                 else
19643                         rack->r_mbuf_queue = 0;
19644                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19645                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19646                 else
19647                         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19648                 break;
19649         case TCP_RACK_NONRXT_CFG_RATE:
19650                 RACK_OPTS_INC(tcp_rack_cfg_rate);
19651                 if (optval == 0)
19652                         rack->rack_rec_nonrxt_use_cr = 0;
19653                 else
19654                         rack->rack_rec_nonrxt_use_cr = 1;
19655                 break;
19656         case TCP_NO_PRR:
19657                 RACK_OPTS_INC(tcp_rack_noprr);
19658                 if (optval == 0)
19659                         rack->rack_no_prr = 0;
19660                 else if (optval == 1)
19661                         rack->rack_no_prr = 1;
19662                 else if (optval == 2)
19663                         rack->no_prr_addback = 1;
19664                 else
19665                         error = EINVAL;
19666                 break;
19667         case TCP_TIMELY_DYN_ADJ:
19668                 RACK_OPTS_INC(tcp_timely_dyn);
19669                 if (optval == 0)
19670                         rack->rc_gp_dyn_mul = 0;
19671                 else {
19672                         rack->rc_gp_dyn_mul = 1;
19673                         if (optval >= 100) {
19674                                 /*
19675                                  * If the user sets something 100 or more
19676                                  * its the gp_ca value.
19677                                  */
19678                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
19679                         }
19680                 }
19681                 break;
19682         case TCP_RACK_DO_DETECTION:
19683                 RACK_OPTS_INC(tcp_rack_do_detection);
19684                 if (optval == 0)
19685                         rack->do_detection = 0;
19686                 else
19687                         rack->do_detection = 1;
19688                 break;
19689         case TCP_RACK_TLP_USE:
19690                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19691                         error = EINVAL;
19692                         break;
19693                 }
19694                 RACK_OPTS_INC(tcp_tlp_use);
19695                 rack->rack_tlp_threshold_use = optval;
19696                 break;
19697         case TCP_RACK_TLP_REDUCE:
19698                 /* RACK TLP cwnd reduction (bool) */
19699                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19700                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19701                 break;
19702         /*  Pacing related ones */
19703         case TCP_RACK_PACE_ALWAYS:
19704                 /*
19705                  * zero is old rack method, 1 is new
19706                  * method using a pacing rate.
19707                  */
19708                 RACK_OPTS_INC(tcp_rack_pace_always);
19709                 if (optval > 0) {
19710                         if (rack->rc_always_pace) {
19711                                 error = EALREADY;
19712                                 break;
19713                         } else if (tcp_can_enable_pacing()) {
19714                                 rack->rc_always_pace = 1;
19715                                 if (rack->use_fixed_rate || rack->gp_ready)
19716                                         rack_set_cc_pacing(rack);
19717                         }
19718                         else {
19719                                 error = ENOSPC;
19720                                 break;
19721                         }
19722                 } else {
19723                         if (rack->rc_always_pace) {
19724                                 tcp_decrement_paced_conn();
19725                                 rack->rc_always_pace = 0;
19726                                 rack_undo_cc_pacing(rack);
19727                         }
19728                 }
19729                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19730                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19731                 else
19732                         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19733                 /* A rate may be set irate or other, if so set seg size */
19734                 rack_update_seg(rack);
19735                 break;
19736         case TCP_BBR_RACK_INIT_RATE:
19737                 RACK_OPTS_INC(tcp_initial_rate);
19738                 val = optval;
19739                 /* Change from kbits per second to bytes per second */
19740                 val *= 1000;
19741                 val /= 8;
19742                 rack->r_ctl.init_rate = val;
19743                 if (rack->rc_init_win != rack_default_init_window) {
19744                         uint32_t win, snt;
19745
19746                         /*
19747                          * Options don't always get applied
19748                          * in the order you think. So in order
19749                          * to assure we update a cwnd we need
19750                          * to check and see if we are still
19751                          * where we should raise the cwnd.
19752                          */
19753                         win = rc_init_window(rack);
19754                         if (SEQ_GT(tp->snd_max, tp->iss))
19755                                 snt = tp->snd_max - tp->iss;
19756                         else
19757                                 snt = 0;
19758                         if ((snt < win) &&
19759                             (tp->snd_cwnd < win))
19760                                 tp->snd_cwnd = win;
19761                 }
19762                 if (rack->rc_always_pace)
19763                         rack_update_seg(rack);
19764                 break;
19765         case TCP_BBR_IWINTSO:
19766                 RACK_OPTS_INC(tcp_initial_win);
19767                 if (optval && (optval <= 0xff)) {
19768                         uint32_t win, snt;
19769
19770                         rack->rc_init_win = optval;
19771                         win = rc_init_window(rack);
19772                         if (SEQ_GT(tp->snd_max, tp->iss))
19773                                 snt = tp->snd_max - tp->iss;
19774                         else
19775                                 snt = 0;
19776                         if ((snt < win) &&
19777                             (tp->t_srtt |
19778 #ifdef NETFLIX_PEAKRATE
19779                              tp->t_maxpeakrate |
19780 #endif
19781                              rack->r_ctl.init_rate)) {
19782                                 /*
19783                                  * We are not past the initial window
19784                                  * and we have some bases for pacing,
19785                                  * so we need to possibly adjust up
19786                                  * the cwnd. Note even if we don't set
19787                                  * the cwnd, its still ok to raise the rc_init_win
19788                                  * which can be used coming out of idle when we
19789                                  * would have a rate.
19790                                  */
19791                                 if (tp->snd_cwnd < win)
19792                                         tp->snd_cwnd = win;
19793                         }
19794                         if (rack->rc_always_pace)
19795                                 rack_update_seg(rack);
19796                 } else
19797                         error = EINVAL;
19798                 break;
19799         case TCP_RACK_FORCE_MSEG:
19800                 RACK_OPTS_INC(tcp_rack_force_max_seg);
19801                 if (optval)
19802                         rack->rc_force_max_seg = 1;
19803                 else
19804                         rack->rc_force_max_seg = 0;
19805                 break;
19806         case TCP_RACK_PACE_MAX_SEG:
19807                 /* Max segments size in a pace in bytes */
19808                 RACK_OPTS_INC(tcp_rack_max_seg);
19809                 rack->rc_user_set_max_segs = optval;
19810                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19811                 break;
19812         case TCP_RACK_PACE_RATE_REC:
19813                 /* Set the fixed pacing rate in Bytes per second ca */
19814                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19815                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19816                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19817                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19818                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19819                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19820                 rack->use_fixed_rate = 1;
19821                 if (rack->rc_always_pace)
19822                         rack_set_cc_pacing(rack);
19823                 rack_log_pacing_delay_calc(rack,
19824                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19825                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19826                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19827                                            __LINE__, NULL,0);
19828                 break;
19829
19830         case TCP_RACK_PACE_RATE_SS:
19831                 /* Set the fixed pacing rate in Bytes per second ca */
19832                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19833                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19834                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19835                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19836                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19837                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19838                 rack->use_fixed_rate = 1;
19839                 if (rack->rc_always_pace)
19840                         rack_set_cc_pacing(rack);
19841                 rack_log_pacing_delay_calc(rack,
19842                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19843                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19844                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19845                                            __LINE__, NULL, 0);
19846                 break;
19847
19848         case TCP_RACK_PACE_RATE_CA:
19849                 /* Set the fixed pacing rate in Bytes per second ca */
19850                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19851                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19852                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19853                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19854                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19855                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19856                 rack->use_fixed_rate = 1;
19857                 if (rack->rc_always_pace)
19858                         rack_set_cc_pacing(rack);
19859                 rack_log_pacing_delay_calc(rack,
19860                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19861                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19862                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19863                                            __LINE__, NULL, 0);
19864                 break;
19865         case TCP_RACK_GP_INCREASE_REC:
19866                 RACK_OPTS_INC(tcp_gp_inc_rec);
19867                 rack->r_ctl.rack_per_of_gp_rec = optval;
19868                 rack_log_pacing_delay_calc(rack,
19869                                            rack->r_ctl.rack_per_of_gp_ss,
19870                                            rack->r_ctl.rack_per_of_gp_ca,
19871                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19872                                            __LINE__, NULL, 0);
19873                 break;
19874         case TCP_RACK_GP_INCREASE_CA:
19875                 RACK_OPTS_INC(tcp_gp_inc_ca);
19876                 ca = optval;
19877                 if (ca < 100) {
19878                         /*
19879                          * We don't allow any reduction
19880                          * over the GP b/w.
19881                          */
19882                         error = EINVAL;
19883                         break;
19884                 }
19885                 rack->r_ctl.rack_per_of_gp_ca = ca;
19886                 rack_log_pacing_delay_calc(rack,
19887                                            rack->r_ctl.rack_per_of_gp_ss,
19888                                            rack->r_ctl.rack_per_of_gp_ca,
19889                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19890                                            __LINE__, NULL, 0);
19891                 break;
19892         case TCP_RACK_GP_INCREASE_SS:
19893                 RACK_OPTS_INC(tcp_gp_inc_ss);
19894                 ss = optval;
19895                 if (ss < 100) {
19896                         /*
19897                          * We don't allow any reduction
19898                          * over the GP b/w.
19899                          */
19900                         error = EINVAL;
19901                         break;
19902                 }
19903                 rack->r_ctl.rack_per_of_gp_ss = ss;
19904                 rack_log_pacing_delay_calc(rack,
19905                                            rack->r_ctl.rack_per_of_gp_ss,
19906                                            rack->r_ctl.rack_per_of_gp_ca,
19907                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19908                                            __LINE__, NULL, 0);
19909                 break;
19910         case TCP_RACK_RR_CONF:
19911                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
19912                 if (optval && optval <= 3)
19913                         rack->r_rr_config = optval;
19914                 else
19915                         rack->r_rr_config = 0;
19916                 break;
19917         case TCP_HDWR_RATE_CAP:
19918                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
19919                 if (optval) {
19920                         if (rack->r_rack_hw_rate_caps == 0)
19921                                 rack->r_rack_hw_rate_caps = 1;
19922                         else
19923                                 error = EALREADY;
19924                 } else {
19925                         rack->r_rack_hw_rate_caps = 0;
19926                 }
19927                 break;
19928         case TCP_BBR_HDWR_PACE:
19929                 RACK_OPTS_INC(tcp_hdwr_pacing);
19930                 if (optval){
19931                         if (rack->rack_hdrw_pacing == 0) {
19932                                 rack->rack_hdw_pace_ena = 1;
19933                                 rack->rack_attempt_hdwr_pace = 0;
19934                         } else
19935                                 error = EALREADY;
19936                 } else {
19937                         rack->rack_hdw_pace_ena = 0;
19938 #ifdef RATELIMIT
19939                         if (rack->r_ctl.crte != NULL) {
19940                                 rack->rack_hdrw_pacing = 0;
19941                                 rack->rack_attempt_hdwr_pace = 0;
19942                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
19943                                 rack->r_ctl.crte = NULL;
19944                         }
19945 #endif
19946                 }
19947                 break;
19948         /*  End Pacing related ones */
19949         case TCP_RACK_PRR_SENDALOT:
19950                 /* Allow PRR to send more than one seg */
19951                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
19952                 rack->r_ctl.rc_prr_sendalot = optval;
19953                 break;
19954         case TCP_RACK_MIN_TO:
19955                 /* Minimum time between rack t-o's in ms */
19956                 RACK_OPTS_INC(tcp_rack_min_to);
19957                 rack->r_ctl.rc_min_to = optval;
19958                 break;
19959         case TCP_RACK_EARLY_SEG:
19960                 /* If early recovery max segments */
19961                 RACK_OPTS_INC(tcp_rack_early_seg);
19962                 rack->r_ctl.rc_early_recovery_segs = optval;
19963                 break;
19964         case TCP_RACK_ENABLE_HYSTART:
19965         {
19966                 if (optval) {
19967                         tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
19968                         if (rack_do_hystart > RACK_HYSTART_ON)
19969                                 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
19970                         if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
19971                                 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
19972                 } else {
19973                         tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
19974                 }
19975         }
19976         break;
19977         case TCP_RACK_REORD_THRESH:
19978                 /* RACK reorder threshold (shift amount) */
19979                 RACK_OPTS_INC(tcp_rack_reord_thresh);
19980                 if ((optval > 0) && (optval < 31))
19981                         rack->r_ctl.rc_reorder_shift = optval;
19982                 else
19983                         error = EINVAL;
19984                 break;
19985         case TCP_RACK_REORD_FADE:
19986                 /* Does reordering fade after ms time */
19987                 RACK_OPTS_INC(tcp_rack_reord_fade);
19988                 rack->r_ctl.rc_reorder_fade = optval;
19989                 break;
19990         case TCP_RACK_TLP_THRESH:
19991                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19992                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
19993                 if (optval)
19994                         rack->r_ctl.rc_tlp_threshold = optval;
19995                 else
19996                         error = EINVAL;
19997                 break;
19998         case TCP_BBR_USE_RACK_RR:
19999                 RACK_OPTS_INC(tcp_rack_rr);
20000                 if (optval)
20001                         rack->use_rack_rr = 1;
20002                 else
20003                         rack->use_rack_rr = 0;
20004                 break;
20005         case TCP_FAST_RSM_HACK:
20006                 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
20007                 if (optval)
20008                         rack->fast_rsm_hack = 1;
20009                 else
20010                         rack->fast_rsm_hack = 0;
20011                 break;
20012         case TCP_RACK_PKT_DELAY:
20013                 /* RACK added ms i.e. rack-rtt + reord + N */
20014                 RACK_OPTS_INC(tcp_rack_pkt_delay);
20015                 rack->r_ctl.rc_pkt_delay = optval;
20016                 break;
20017         case TCP_DELACK:
20018                 RACK_OPTS_INC(tcp_rack_delayed_ack);
20019                 if (optval == 0)
20020                         tp->t_delayed_ack = 0;
20021                 else
20022                         tp->t_delayed_ack = 1;
20023                 if (tp->t_flags & TF_DELACK) {
20024                         tp->t_flags &= ~TF_DELACK;
20025                         tp->t_flags |= TF_ACKNOW;
20026                         NET_EPOCH_ENTER(et);
20027                         rack_output(tp);
20028                         NET_EPOCH_EXIT(et);
20029                 }
20030                 break;
20031
20032         case TCP_BBR_RACK_RTT_USE:
20033                 RACK_OPTS_INC(tcp_rack_rtt_use);
20034                 if ((optval != USE_RTT_HIGH) &&
20035                     (optval != USE_RTT_LOW) &&
20036                     (optval != USE_RTT_AVG))
20037                         error = EINVAL;
20038                 else
20039                         rack->r_ctl.rc_rate_sample_method = optval;
20040                 break;
20041         case TCP_DATA_AFTER_CLOSE:
20042                 RACK_OPTS_INC(tcp_data_after_close);
20043                 if (optval)
20044                         rack->rc_allow_data_af_clo = 1;
20045                 else
20046                         rack->rc_allow_data_af_clo = 0;
20047                 break;
20048         default:
20049                 break;
20050         }
20051 #ifdef NETFLIX_STATS
20052         tcp_log_socket_option(tp, sopt_name, optval, error);
20053 #endif
20054         return (error);
20055 }
20056
20057
20058 static void
20059 rack_apply_deferred_options(struct tcp_rack *rack)
20060 {
20061         struct deferred_opt_list *dol, *sdol;
20062         uint32_t s_optval;
20063
20064         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20065                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20066                 /* Disadvantage of deferal is you loose the error return */
20067                 s_optval = (uint32_t)dol->optval;
20068                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20069                 free(dol, M_TCPDO);
20070         }
20071 }
20072
20073 static void
20074 rack_hw_tls_change(struct tcpcb *tp, int chg)
20075 {
20076         /* Update HW tls state */
20077         struct tcp_rack *rack;
20078
20079         rack = (struct tcp_rack *)tp->t_fb_ptr;
20080         if (chg)
20081                 rack->r_ctl.fsb.hw_tls = 1;
20082         else
20083                 rack->r_ctl.fsb.hw_tls = 0;
20084 }
20085
20086 static int
20087 rack_pru_options(struct tcpcb *tp, int flags)
20088 {
20089         if (flags & PRUS_OOB)
20090                 return (EOPNOTSUPP);
20091         return (0);
20092 }
20093
20094 static struct tcp_function_block __tcp_rack = {
20095         .tfb_tcp_block_name = __XSTRING(STACKNAME),
20096         .tfb_tcp_output = rack_output,
20097         .tfb_do_queued_segments = ctf_do_queued_segments,
20098         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
20099         .tfb_tcp_do_segment = rack_do_segment,
20100         .tfb_tcp_ctloutput = rack_ctloutput,
20101         .tfb_tcp_fb_init = rack_init,
20102         .tfb_tcp_fb_fini = rack_fini,
20103         .tfb_tcp_timer_stop_all = rack_stopall,
20104         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20105         .tfb_tcp_handoff_ok = rack_handoff_ok,
20106         .tfb_tcp_mtu_chg = rack_mtu_change,
20107         .tfb_pru_options = rack_pru_options,
20108         .tfb_hwtls_change = rack_hw_tls_change,
20109         .tfb_compute_pipe = rack_compute_pipe,
20110         .tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
20111 };
20112
20113 /*
20114  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20115  * socket option arguments.  When it re-acquires the lock after the copy, it
20116  * has to revalidate that the connection is still valid for the socket
20117  * option.
20118  */
20119 static int
20120 rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt)
20121 {
20122 #ifdef INET6
20123         struct ip6_hdr *ip6;
20124 #endif
20125 #ifdef INET
20126         struct ip *ip;
20127 #endif
20128         struct tcpcb *tp;
20129         struct tcp_rack *rack;
20130         uint64_t loptval;
20131         int32_t error = 0, optval;
20132
20133         tp = intotcpcb(inp);
20134         rack = (struct tcp_rack *)tp->t_fb_ptr;
20135         if (rack == NULL) {
20136                 INP_WUNLOCK(inp);
20137                 return (EINVAL);
20138         }
20139 #ifdef INET6
20140         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20141 #endif
20142 #ifdef INET
20143         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20144 #endif
20145
20146         switch (sopt->sopt_level) {
20147 #ifdef INET6
20148         case IPPROTO_IPV6:
20149                 MPASS(inp->inp_vflag & INP_IPV6PROTO);
20150                 switch (sopt->sopt_name) {
20151                 case IPV6_USE_MIN_MTU:
20152                         tcp6_use_min_mtu(tp);
20153                         break;
20154                 case IPV6_TCLASS:
20155                         /*
20156                          * The DSCP codepoint has changed, update the fsb.
20157                          */
20158                         ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20159                             (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20160                         break;
20161                 }
20162                 INP_WUNLOCK(inp);
20163                 return (0);
20164 #endif
20165 #ifdef INET
20166         case IPPROTO_IP:
20167                 switch (sopt->sopt_name) {
20168                 case IP_TOS:
20169                         /*
20170                          * The DSCP codepoint has changed, update the fsb.
20171                          */
20172                         ip->ip_tos = rack->rc_inp->inp_ip_tos;
20173                         break;
20174                 case IP_TTL:
20175                         /*
20176                          * The TTL has changed, update the fsb.
20177                          */
20178                         ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20179                         break;
20180                 }
20181                 INP_WUNLOCK(inp);
20182                 return (0);
20183 #endif
20184         }
20185
20186         switch (sopt->sopt_name) {
20187         case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
20188         /*  Pacing related ones */
20189         case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
20190         case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
20191         case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
20192         case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
20193         case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
20194         case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
20195         case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
20196         case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
20197         case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
20198         case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
20199         case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
20200         case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
20201         case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
20202         case TCP_HDWR_RATE_CAP:                 /*  URL:hdwrcap boolean */
20203         case TCP_PACING_RATE_CAP:               /*  URL:cap  -- used by side-channel */
20204         case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
20205        /* End pacing related */
20206         case TCP_FAST_RSM_HACK:                 /*  URL:frsm_hack */
20207         case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20208         case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
20209         case TCP_RACK_MIN_TO:                   /*  URL:min_to */
20210         case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
20211         case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
20212         case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
20213         case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
20214         case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
20215         case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
20216         case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
20217         case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
20218         case TCP_RACK_DO_DETECTION:             /*  URL:detect */
20219         case TCP_NO_PRR:                        /*  URL:noprr */
20220         case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
20221         case TCP_DATA_AFTER_CLOSE:              /*  no URL */
20222         case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
20223         case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
20224         case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
20225         case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
20226         case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
20227         case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
20228         case TCP_RACK_PROFILE:                  /*  URL:profile */
20229         case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
20230         case TCP_RACK_ABC_VAL:                  /*  URL:labc */
20231         case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
20232         case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
20233         case TCP_DEFER_OPTIONS:                 /*  URL:defer */
20234         case TCP_RACK_DSACK_OPT:                /*  URL:dsack */
20235         case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
20236         case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
20237         case TCP_RACK_TIMER_SLOP:               /*  URL:timer_slop */
20238         case TCP_RACK_ENABLE_HYSTART:           /*  URL:hystart */
20239                 break;
20240         default:
20241                 /* Filter off all unknown options to the base stack */
20242                 return (tcp_default_ctloutput(inp, sopt));
20243                 break;
20244         }
20245         INP_WUNLOCK(inp);
20246         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20247                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20248                 /*
20249                  * We truncate it down to 32 bits for the socket-option trace this
20250                  * means rates > 34Gbps won't show right, but thats probably ok.
20251                  */
20252                 optval = (uint32_t)loptval;
20253         } else {
20254                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20255                 /* Save it in 64 bit form too */
20256                 loptval = optval;
20257         }
20258         if (error)
20259                 return (error);
20260         INP_WLOCK(inp);
20261         if (inp->inp_flags & INP_DROPPED) {
20262                 INP_WUNLOCK(inp);
20263                 return (ECONNRESET);
20264         }
20265         if (tp->t_fb != &__tcp_rack) {
20266                 INP_WUNLOCK(inp);
20267                 return (ENOPROTOOPT);
20268         }
20269         if (rack->defer_options && (rack->gp_ready == 0) &&
20270             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20271             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20272             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20273             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20274                 /* Options are beind deferred */
20275                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20276                         INP_WUNLOCK(inp);
20277                         return (0);
20278                 } else {
20279                         /* No memory to defer, fail */
20280                         INP_WUNLOCK(inp);
20281                         return (ENOMEM);
20282                 }
20283         }
20284         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20285         INP_WUNLOCK(inp);
20286         return (error);
20287 }
20288
20289 static void
20290 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20291 {
20292
20293         INP_WLOCK_ASSERT(tptoinpcb(tp));
20294         bzero(ti, sizeof(*ti));
20295
20296         ti->tcpi_state = tp->t_state;
20297         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20298                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20299         if (tp->t_flags & TF_SACK_PERMIT)
20300                 ti->tcpi_options |= TCPI_OPT_SACK;
20301         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20302                 ti->tcpi_options |= TCPI_OPT_WSCALE;
20303                 ti->tcpi_snd_wscale = tp->snd_scale;
20304                 ti->tcpi_rcv_wscale = tp->rcv_scale;
20305         }
20306         if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
20307                 ti->tcpi_options |= TCPI_OPT_ECN;
20308         if (tp->t_flags & TF_FASTOPEN)
20309                 ti->tcpi_options |= TCPI_OPT_TFO;
20310         /* still kept in ticks is t_rcvtime */
20311         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20312         /* Since we hold everything in precise useconds this is easy */
20313         ti->tcpi_rtt = tp->t_srtt;
20314         ti->tcpi_rttvar = tp->t_rttvar;
20315         ti->tcpi_rto = tp->t_rxtcur;
20316         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20317         ti->tcpi_snd_cwnd = tp->snd_cwnd;
20318         /*
20319          * FreeBSD-specific extension fields for tcp_info.
20320          */
20321         ti->tcpi_rcv_space = tp->rcv_wnd;
20322         ti->tcpi_rcv_nxt = tp->rcv_nxt;
20323         ti->tcpi_snd_wnd = tp->snd_wnd;
20324         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
20325         ti->tcpi_snd_nxt = tp->snd_nxt;
20326         ti->tcpi_snd_mss = tp->t_maxseg;
20327         ti->tcpi_rcv_mss = tp->t_maxseg;
20328         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20329         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20330         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20331 #ifdef NETFLIX_STATS
20332         ti->tcpi_total_tlp = tp->t_sndtlppack;
20333         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20334         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20335 #endif
20336 #ifdef TCP_OFFLOAD
20337         if (tp->t_flags & TF_TOE) {
20338                 ti->tcpi_options |= TCPI_OPT_TOE;
20339                 tcp_offload_tcp_info(tp, ti);
20340         }
20341 #endif
20342 }
20343
20344 static int
20345 rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt)
20346 {
20347         struct tcpcb *tp;
20348         struct tcp_rack *rack;
20349         int32_t error, optval;
20350         uint64_t val, loptval;
20351         struct  tcp_info ti;
20352         /*
20353          * Because all our options are either boolean or an int, we can just
20354          * pull everything into optval and then unlock and copy. If we ever
20355          * add a option that is not a int, then this will have quite an
20356          * impact to this routine.
20357          */
20358         error = 0;
20359         tp = intotcpcb(inp);
20360         rack = (struct tcp_rack *)tp->t_fb_ptr;
20361         if (rack == NULL) {
20362                 INP_WUNLOCK(inp);
20363                 return (EINVAL);
20364         }
20365         switch (sopt->sopt_name) {
20366         case TCP_INFO:
20367                 /* First get the info filled */
20368                 rack_fill_info(tp, &ti);
20369                 /* Fix up the rtt related fields if needed */
20370                 INP_WUNLOCK(inp);
20371                 error = sooptcopyout(sopt, &ti, sizeof ti);
20372                 return (error);
20373         /*
20374          * Beta is the congestion control value for NewReno that influences how
20375          * much of a backoff happens when loss is detected. It is normally set
20376          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20377          * when you exit recovery.
20378          */
20379         case TCP_RACK_PACING_BETA:
20380                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
20381                         error = EINVAL;
20382                 else if (rack->rc_pacing_cc_set == 0)
20383                         optval = rack->r_ctl.rc_saved_beta.beta;
20384                 else {
20385                         /*
20386                          * Reach out into the CC data and report back what
20387                          * I have previously set. Yeah it looks hackish but
20388                          * we don't want to report the saved values.
20389                          */
20390                         if (tp->t_ccv.cc_data)
20391                                 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
20392                         else
20393                                 error = EINVAL;
20394                 }
20395                 break;
20396                 /*
20397                  * Beta_ecn is the congestion control value for NewReno that influences how
20398                  * much of a backoff happens when a ECN mark is detected. It is normally set
20399                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20400                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
20401                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
20402                  */
20403
20404         case TCP_RACK_PACING_BETA_ECN:
20405                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
20406                         error = EINVAL;
20407                 else if (rack->rc_pacing_cc_set == 0)
20408                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20409                 else {
20410                         /*
20411                          * Reach out into the CC data and report back what
20412                          * I have previously set. Yeah it looks hackish but
20413                          * we don't want to report the saved values.
20414                          */
20415                         if (tp->t_ccv.cc_data)
20416                                 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
20417                         else
20418                                 error = EINVAL;
20419                 }
20420                 break;
20421         case TCP_RACK_DSACK_OPT:
20422                 optval = 0;
20423                 if (rack->rc_rack_tmr_std_based) {
20424                         optval |= 1;
20425                 }
20426                 if (rack->rc_rack_use_dsack) {
20427                         optval |= 2;
20428                 }
20429                 break;
20430         case TCP_RACK_ENABLE_HYSTART:
20431         {
20432                 if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
20433                         optval = RACK_HYSTART_ON;
20434                         if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
20435                                 optval = RACK_HYSTART_ON_W_SC;
20436                         if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
20437                                 optval = RACK_HYSTART_ON_W_SC_C;
20438                 } else {
20439                         optval = RACK_HYSTART_OFF;
20440                 }
20441         }
20442         break;
20443         case TCP_FAST_RSM_HACK:
20444                 optval = rack->fast_rsm_hack;
20445                 break;
20446         case TCP_DEFER_OPTIONS:
20447                 optval = rack->defer_options;
20448                 break;
20449         case TCP_RACK_MEASURE_CNT:
20450                 optval = rack->r_ctl.req_measurements;
20451                 break;
20452         case TCP_REC_ABC_VAL:
20453                 optval = rack->r_use_labc_for_rec;
20454                 break;
20455         case TCP_RACK_ABC_VAL:
20456                 optval = rack->rc_labc;
20457                 break;
20458         case TCP_HDWR_UP_ONLY:
20459                 optval= rack->r_up_only;
20460                 break;
20461         case TCP_PACING_RATE_CAP:
20462                 loptval = rack->r_ctl.bw_rate_cap;
20463                 break;
20464         case TCP_RACK_PROFILE:
20465                 /* You cannot retrieve a profile, its write only */
20466                 error = EINVAL;
20467                 break;
20468         case TCP_USE_CMP_ACKS:
20469                 optval = rack->r_use_cmp_ack;
20470                 break;
20471         case TCP_RACK_PACE_TO_FILL:
20472                 optval = rack->rc_pace_to_cwnd;
20473                 if (optval && rack->r_fill_less_agg)
20474                         optval++;
20475                 break;
20476         case TCP_RACK_NO_PUSH_AT_MAX:
20477                 optval = rack->r_ctl.rc_no_push_at_mrtt;
20478                 break;
20479         case TCP_SHARED_CWND_ENABLE:
20480                 optval = rack->rack_enable_scwnd;
20481                 break;
20482         case TCP_RACK_NONRXT_CFG_RATE:
20483                 optval = rack->rack_rec_nonrxt_use_cr;
20484                 break;
20485         case TCP_NO_PRR:
20486                 if (rack->rack_no_prr  == 1)
20487                         optval = 1;
20488                 else if (rack->no_prr_addback == 1)
20489                         optval = 2;
20490                 else
20491                         optval = 0;
20492                 break;
20493         case TCP_RACK_DO_DETECTION:
20494                 optval = rack->do_detection;
20495                 break;
20496         case TCP_RACK_MBUF_QUEUE:
20497                 /* Now do we use the LRO mbuf-queue feature */
20498                 optval = rack->r_mbuf_queue;
20499                 break;
20500         case TCP_TIMELY_DYN_ADJ:
20501                 optval = rack->rc_gp_dyn_mul;
20502                 break;
20503         case TCP_BBR_IWINTSO:
20504                 optval = rack->rc_init_win;
20505                 break;
20506         case TCP_RACK_TLP_REDUCE:
20507                 /* RACK TLP cwnd reduction (bool) */
20508                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20509                 break;
20510         case TCP_BBR_RACK_INIT_RATE:
20511                 val = rack->r_ctl.init_rate;
20512                 /* convert to kbits per sec */
20513                 val *= 8;
20514                 val /= 1000;
20515                 optval = (uint32_t)val;
20516                 break;
20517         case TCP_RACK_FORCE_MSEG:
20518                 optval = rack->rc_force_max_seg;
20519                 break;
20520         case TCP_RACK_PACE_MAX_SEG:
20521                 /* Max segments in a pace */
20522                 optval = rack->rc_user_set_max_segs;
20523                 break;
20524         case TCP_RACK_PACE_ALWAYS:
20525                 /* Use the always pace method */
20526                 optval = rack->rc_always_pace;
20527                 break;
20528         case TCP_RACK_PRR_SENDALOT:
20529                 /* Allow PRR to send more than one seg */
20530                 optval = rack->r_ctl.rc_prr_sendalot;
20531                 break;
20532         case TCP_RACK_MIN_TO:
20533                 /* Minimum time between rack t-o's in ms */
20534                 optval = rack->r_ctl.rc_min_to;
20535                 break;
20536         case TCP_RACK_EARLY_SEG:
20537                 /* If early recovery max segments */
20538                 optval = rack->r_ctl.rc_early_recovery_segs;
20539                 break;
20540         case TCP_RACK_REORD_THRESH:
20541                 /* RACK reorder threshold (shift amount) */
20542                 optval = rack->r_ctl.rc_reorder_shift;
20543                 break;
20544         case TCP_RACK_REORD_FADE:
20545                 /* Does reordering fade after ms time */
20546                 optval = rack->r_ctl.rc_reorder_fade;
20547                 break;
20548         case TCP_BBR_USE_RACK_RR:
20549                 /* Do we use the rack cheat for rxt */
20550                 optval = rack->use_rack_rr;
20551                 break;
20552         case TCP_RACK_RR_CONF:
20553                 optval = rack->r_rr_config;
20554                 break;
20555         case TCP_HDWR_RATE_CAP:
20556                 optval = rack->r_rack_hw_rate_caps;
20557                 break;
20558         case TCP_BBR_HDWR_PACE:
20559                 optval = rack->rack_hdw_pace_ena;
20560                 break;
20561         case TCP_RACK_TLP_THRESH:
20562                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
20563                 optval = rack->r_ctl.rc_tlp_threshold;
20564                 break;
20565         case TCP_RACK_PKT_DELAY:
20566                 /* RACK added ms i.e. rack-rtt + reord + N */
20567                 optval = rack->r_ctl.rc_pkt_delay;
20568                 break;
20569         case TCP_RACK_TLP_USE:
20570                 optval = rack->rack_tlp_threshold_use;
20571                 break;
20572         case TCP_RACK_PACE_RATE_CA:
20573                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20574                 break;
20575         case TCP_RACK_PACE_RATE_SS:
20576                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20577                 break;
20578         case TCP_RACK_PACE_RATE_REC:
20579                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20580                 break;
20581         case TCP_RACK_GP_INCREASE_SS:
20582                 optval = rack->r_ctl.rack_per_of_gp_ca;
20583                 break;
20584         case TCP_RACK_GP_INCREASE_CA:
20585                 optval = rack->r_ctl.rack_per_of_gp_ss;
20586                 break;
20587         case TCP_BBR_RACK_RTT_USE:
20588                 optval = rack->r_ctl.rc_rate_sample_method;
20589                 break;
20590         case TCP_DELACK:
20591                 optval = tp->t_delayed_ack;
20592                 break;
20593         case TCP_DATA_AFTER_CLOSE:
20594                 optval = rack->rc_allow_data_af_clo;
20595                 break;
20596         case TCP_SHARED_CWND_TIME_LIMIT:
20597                 optval = rack->r_limit_scw;
20598                 break;
20599         case TCP_RACK_TIMER_SLOP:
20600                 optval = rack->r_ctl.timer_slop;
20601                 break;
20602         default:
20603                 return (tcp_default_ctloutput(inp, sopt));
20604                 break;
20605         }
20606         INP_WUNLOCK(inp);
20607         if (error == 0) {
20608                 if (TCP_PACING_RATE_CAP)
20609                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
20610                 else
20611                         error = sooptcopyout(sopt, &optval, sizeof optval);
20612         }
20613         return (error);
20614 }
20615
20616 static int
20617 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt)
20618 {
20619         if (sopt->sopt_dir == SOPT_SET) {
20620                 return (rack_set_sockopt(inp, sopt));
20621         } else if (sopt->sopt_dir == SOPT_GET) {
20622                 return (rack_get_sockopt(inp, sopt));
20623         } else {
20624                 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
20625         }
20626 }
20627
20628 static const char *rack_stack_names[] = {
20629         __XSTRING(STACKNAME),
20630 #ifdef STACKALIAS
20631         __XSTRING(STACKALIAS),
20632 #endif
20633 };
20634
20635 static int
20636 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20637 {
20638         memset(mem, 0, size);
20639         return (0);
20640 }
20641
20642 static void
20643 rack_dtor(void *mem, int32_t size, void *arg)
20644 {
20645
20646 }
20647
20648 static bool rack_mod_inited = false;
20649
20650 static int
20651 tcp_addrack(module_t mod, int32_t type, void *data)
20652 {
20653         int32_t err = 0;
20654         int num_stacks;
20655
20656         switch (type) {
20657         case MOD_LOAD:
20658                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20659                     sizeof(struct rack_sendmap),
20660                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20661
20662                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20663                     sizeof(struct tcp_rack),
20664                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20665
20666                 sysctl_ctx_init(&rack_sysctl_ctx);
20667                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20668                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20669                     OID_AUTO,
20670 #ifdef STACKALIAS
20671                     __XSTRING(STACKALIAS),
20672 #else
20673                     __XSTRING(STACKNAME),
20674 #endif
20675                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20676                     "");
20677                 if (rack_sysctl_root == NULL) {
20678                         printf("Failed to add sysctl node\n");
20679                         err = EFAULT;
20680                         goto free_uma;
20681                 }
20682                 rack_init_sysctls();
20683                 num_stacks = nitems(rack_stack_names);
20684                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20685                     rack_stack_names, &num_stacks);
20686                 if (err) {
20687                         printf("Failed to register %s stack name for "
20688                             "%s module\n", rack_stack_names[num_stacks],
20689                             __XSTRING(MODNAME));
20690                         sysctl_ctx_free(&rack_sysctl_ctx);
20691 free_uma:
20692                         uma_zdestroy(rack_zone);
20693                         uma_zdestroy(rack_pcb_zone);
20694                         rack_counter_destroy();
20695                         printf("Failed to register rack module -- err:%d\n", err);
20696                         return (err);
20697                 }
20698                 tcp_lro_reg_mbufq();
20699                 rack_mod_inited = true;
20700                 break;
20701         case MOD_QUIESCE:
20702                 err = deregister_tcp_functions(&__tcp_rack, true, false);
20703                 break;
20704         case MOD_UNLOAD:
20705                 err = deregister_tcp_functions(&__tcp_rack, false, true);
20706                 if (err == EBUSY)
20707                         break;
20708                 if (rack_mod_inited) {
20709                         uma_zdestroy(rack_zone);
20710                         uma_zdestroy(rack_pcb_zone);
20711                         sysctl_ctx_free(&rack_sysctl_ctx);
20712                         rack_counter_destroy();
20713                         rack_mod_inited = false;
20714                 }
20715                 tcp_lro_dereg_mbufq();
20716                 err = 0;
20717                 break;
20718         default:
20719                 return (EOPNOTSUPP);
20720         }
20721         return (err);
20722 }
20723
20724 static moduledata_t tcp_rack = {
20725         .name = __XSTRING(MODNAME),
20726         .evhand = tcp_addrack,
20727         .priv = 0
20728 };
20729
20730 MODULE_VERSION(MODNAME, 1);
20731 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20732 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
20733
20734 #endif /* #if !defined(INET) && !defined(INET6) */