sys/netinet/tcp_stacks/bbr.c

   1 /*-
   2  * Copyright (c) 2016-2019
   3  *      Netflix Inc.
   4  *      All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  25  * SUCH DAMAGE.
  26  *
  27  */
  28 /**
  29  * Author: Randall Stewart <rrs@netflix.com>
  30  * This work is based on the ACM Queue paper
  31  * BBR - Congestion Based Congestion Control
  32  * and also numerous discussions with Neal, Yuchung and Van.
  33  */
  34
  35 #include <sys/cdefs.h>
  36 __FBSDID("$FreeBSD$");
  37
  38 #include "opt_inet.h"
  39 #include "opt_inet6.h"
  40 #include "opt_ipsec.h"
  41 #include "opt_tcpdebug.h"
  42 #include "opt_ratelimit.h"
  43 #include "opt_kern_tls.h"
  44 #include <sys/param.h>
  45 #include <sys/module.h>
  46 #include <sys/kernel.h>
  47 #ifdef TCP_HHOOK
  48 #include <sys/hhook.h>
  49 #endif
  50 #include <sys/malloc.h>
  51 #include <sys/mbuf.h>
  52 #include <sys/proc.h>
  53 #include <sys/socket.h>
  54 #include <sys/socketvar.h>
  55 #ifdef KERN_TLS
  56 #include <sys/ktls.h>
  57 #endif
  58 #include <sys/sysctl.h>
  59 #include <sys/systm.h>
  60 #include <sys/qmath.h>
  61 #include <sys/tree.h>
  62 #ifdef NETFLIX_STATS
  63 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
  64 #endif
  65 #include <sys/refcount.h>
  66 #include <sys/queue.h>
  67 #include <sys/eventhandler.h>
  68 #include <sys/smp.h>
  69 #include <sys/kthread.h>
  70 #include <sys/lock.h>
  71 #include <sys/mutex.h>
  72 #include <sys/tim_filter.h>
  73 #include <sys/time.h>
  74 #include <vm/uma.h>
  75 #include <sys/kern_prefetch.h>
  76
  77 #include <net/route.h>
  78 #include <net/vnet.h>
  79
  80 #define TCPSTATES               /* for logging */
  81
  82 #include <netinet/in.h>
  83 #include <netinet/in_kdtrace.h>
  84 #include <netinet/in_pcb.h>
  85 #include <netinet/ip.h>
  86 #include <netinet/ip_icmp.h>    /* required for icmp_var.h */
  87 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
  88 #include <netinet/ip_var.h>
  89 #include <netinet/ip6.h>
  90 #include <netinet6/in6_pcb.h>
  91 #include <netinet6/ip6_var.h>
  92 #define TCPOUTFLAGS
  93 #include <netinet/tcp.h>
  94 #include <netinet/tcp_fsm.h>
  95 #include <netinet/tcp_seq.h>
  96 #include <netinet/tcp_timer.h>
  97 #include <netinet/tcp_var.h>
  98 #include <netinet/tcpip.h>
  99 #include <netinet/tcp_hpts.h>
 100 #include <netinet/cc/cc.h>
 101 #include <netinet/tcp_log_buf.h>
 102 #include <netinet/tcp_ratelimit.h>
 103 #include <netinet/tcp_lro.h>
 104 #ifdef TCPDEBUG
 105 #include <netinet/tcp_debug.h>
 106 #endif                          /* TCPDEBUG */
 107 #ifdef TCP_OFFLOAD
 108 #include <netinet/tcp_offload.h>
 109 #endif
 110 #ifdef INET6
 111 #include <netinet6/tcp6_var.h>
 112 #endif
 113 #include <netinet/tcp_fastopen.h>
 114
 115 #include <netipsec/ipsec_support.h>
 116 #include <net/if.h>
 117 #include <net/if_var.h>
 118 #include <net/ethernet.h>
 119
 120 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 121 #include <netipsec/ipsec.h>
 122 #include <netipsec/ipsec6.h>
 123 #endif                          /* IPSEC */
 124
 125 #include <netinet/udp.h>
 126 #include <netinet/udp_var.h>
 127 #include <machine/in_cksum.h>
 128
 129 #ifdef MAC
 130 #include <security/mac/mac_framework.h>
 131 #endif
 132
 133 #include "sack_filter.h"
 134 #include "tcp_bbr.h"
 135 #include "rack_bbr_common.h"
 136 uma_zone_t bbr_zone;
 137 uma_zone_t bbr_pcb_zone;
 138
 139 struct sysctl_ctx_list bbr_sysctl_ctx;
 140 struct sysctl_oid *bbr_sysctl_root;
 141
 142 #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \
 143         (tv) = (value); \
 144         if ((u_long)(tv) < (u_long)(tvmin)) \
 145                 (tv) = (tvmin); \
 146         if ((u_long)(tv) > (u_long)(tvmax)) \
 147                 (tv) = (tvmax); \
 148 } while(0)
 149
 150 /*#define BBR_INVARIANT 1*/
 151
 152 /*
 153  * initial window
 154  */
 155 static uint32_t bbr_def_init_win = 10;
 156 static int32_t bbr_persist_min = 250000;        /* 250ms */
 157 static int32_t bbr_persist_max = 1000000;       /* 1 Second */
 158 static int32_t bbr_cwnd_may_shrink = 0;
 159 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP;
 160 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT;
 161 static int32_t bbr_hardware_pacing_limit = 8000;
 162 static int32_t bbr_quanta = 3;  /* How much extra quanta do we get? */
 163 static int32_t bbr_no_retran = 0;
 164 static int32_t bbr_tcp_map_entries_limit = 1500;
 165 static int32_t bbr_tcp_map_split_limit = 256;
 166
 167 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */
 168 static int32_t bbr_max_net_error_cnt = 10;
 169 /* Should the following be dynamic too -- loss wise */
 170 static int32_t bbr_rtt_gain_thresh = 0;
 171 /* Measurement controls */
 172 static int32_t bbr_use_google_algo = 1;
 173 static int32_t bbr_ts_limiting = 1;
 174 static int32_t bbr_ts_can_raise = 0;
 175 static int32_t bbr_do_red = 600;
 176 static int32_t bbr_red_scale = 20000;
 177 static int32_t bbr_red_mul = 1;
 178 static int32_t bbr_red_div = 2;
 179 static int32_t bbr_red_growth_restrict = 1;
 180 static int32_t  bbr_target_is_bbunit = 0;
 181 static int32_t bbr_drop_limit = 0;
 182 /*
 183  * How much gain do we need to see to
 184  * stay in startup?
 185  */
 186 static int32_t bbr_marks_rxt_sack_passed = 0;
 187 static int32_t bbr_start_exit = 25;
 188 static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */
 189 static int32_t bbr_startup_loss_thresh = 2000;  /* 20.00% loss */
 190 static int32_t bbr_hptsi_max_mul = 1;   /* These two mul/div assure a min pacing */
 191 static int32_t bbr_hptsi_max_div = 2;   /* time, 0 means turned off. We need this
 192                                          * if we go back ever to where the pacer
 193                                          * has priority over timers.
 194                                          */
 195 static int32_t bbr_policer_call_from_rack_to = 0;
 196 static int32_t bbr_policer_detection_enabled = 1;
 197 static int32_t bbr_min_measurements_req = 1;    /* We need at least 2
 198                                                  * measurments before we are
 199                                                  * "good" note that 2 == 1.
 200                                                  * This is because we use a >
 201                                                  * comparison. This means if
 202                                                  * min_measure was 0, it takes
 203                                                  * num-measures > min(0) and
 204                                                  * you get 1 measurement and
 205                                                  * you are good. Set to 1, you
 206                                                  * have to have two
 207                                                  * measurements (this is done
 208                                                  * to prevent it from being ok
 209                                                  * to have no measurements). */
 210 static int32_t bbr_no_pacing_until = 4;
 211
 212 static int32_t bbr_min_usec_delta = 20000;      /* 20,000 usecs */
 213 static int32_t bbr_min_peer_delta = 20;         /* 20 units */
 214 static int32_t bbr_delta_percent = 150;         /* 15.0 % */
 215
 216 static int32_t bbr_target_cwnd_mult_limit = 8;
 217 /*
 218  * bbr_cwnd_min_val is the number of
 219  * segments we hold to in the RTT probe
 220  * state typically 4.
 221  */
 222 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS;
 223
 224
 225 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS;
 226
 227 static int32_t bbr_gain_to_target = 1;
 228 static int32_t bbr_gain_gets_extra_too = 1;
 229 /*
 230  * bbr_high_gain is the 2/ln(2) value we need
 231  * to double the sending rate in startup. This
 232  * is used for both cwnd and hptsi gain's.
 233  */
 234 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1;
 235 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1;
 236 static int32_t bbr_use_lower_gain_in_startup = 1;
 237
 238 /* thresholds for reduction on drain in sub-states/drain */
 239 static int32_t bbr_drain_rtt = BBR_SRTT;
 240 static int32_t bbr_drain_floor = 88;
 241 static int32_t google_allow_early_out = 1;
 242 static int32_t google_consider_lost = 1;
 243 static int32_t bbr_drain_drop_mul = 4;
 244 static int32_t bbr_drain_drop_div = 5;
 245 static int32_t bbr_rand_ot = 50;
 246 static int32_t bbr_can_force_probertt = 0;
 247 static int32_t bbr_can_adjust_probertt = 1;
 248 static int32_t bbr_probertt_sets_rtt = 0;
 249 static int32_t bbr_can_use_ts_for_rtt = 1;
 250 static int32_t bbr_is_ratio = 0;
 251 static int32_t bbr_sub_drain_app_limit = 1;
 252 static int32_t bbr_prtt_slam_cwnd = 1;
 253 static int32_t bbr_sub_drain_slam_cwnd = 1;
 254 static int32_t bbr_slam_cwnd_in_main_drain = 1;
 255 static int32_t bbr_filter_len_sec = 6;  /* How long does the rttProp filter
 256                                          * hold */
 257 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4);
 258 /*
 259  * bbr_drain_gain is the reverse of the high_gain
 260  * designed to drain back out the standing queue
 261  * that is formed in startup by causing a larger
 262  * hptsi gain and thus drainging the packets
 263  * in flight.
 264  */
 265 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885;
 266 static int32_t bbr_rttprobe_gain = 192;
 267
 268 /*
 269  * The cwnd_gain is the default cwnd gain applied when
 270  * calculating a target cwnd. Note that the cwnd is
 271  * a secondary factor in the way BBR works (see the
 272  * paper and think about it, it will take some time).
 273  * Basically the hptsi_gain spreads the packets out
 274  * so you never get more than BDP to the peer even
 275  * if the cwnd is high. In our implemenation that
 276  * means in non-recovery/retransmission scenarios
 277  * cwnd will never be reached by the flight-size.
 278  */
 279 static int32_t bbr_cwnd_gain = BBR_UNIT * 2;
 280 static int32_t bbr_tlp_type_to_use = BBR_SRTT;
 281 static int32_t bbr_delack_time = 100000;        /* 100ms in useconds */
 282 static int32_t bbr_sack_not_required = 0;       /* set to one to allow non-sack to use bbr */
 283 static int32_t bbr_initial_bw_bps = 62500;      /* 500kbps in bytes ps */
 284 static int32_t bbr_ignore_data_after_close = 1;
 285 static int16_t bbr_hptsi_gain[] = {
 286         (BBR_UNIT *5 / 4),
 287         (BBR_UNIT * 3 / 4),
 288         BBR_UNIT,
 289         BBR_UNIT,
 290         BBR_UNIT,
 291         BBR_UNIT,
 292         BBR_UNIT,
 293         BBR_UNIT
 294 };
 295 int32_t bbr_use_rack_resend_cheat = 1;
 296 int32_t bbr_sends_full_iwnd = 1;
 297
 298 #define BBR_HPTSI_GAIN_MAX 8
 299 /*
 300  * The BBR module incorporates a number of
 301  * TCP ideas that have been put out into the IETF
 302  * over the last few years:
 303  * - Yuchung Cheng's RACK TCP (for which its named) that
 304  *    will stop us using the number of dup acks and instead
 305  *    use time as the gage of when we retransmit.
 306  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
 307  *    of Dukkipati et.al.
 308  * - Van Jacobson's et.al BBR.
 309  *
 310  * RACK depends on SACK, so if an endpoint arrives that
 311  * cannot do SACK the state machine below will shuttle the
 312  * connection back to using the "default" TCP stack that is
 313  * in FreeBSD.
 314  *
 315  * To implement BBR and RACK the original TCP stack was first decomposed
 316  * into a functional state machine with individual states
 317  * for each of the possible TCP connection states. The do_segement
 318  * functions role in life is to mandate the connection supports SACK
 319  * initially and then assure that the RACK state matches the conenction
 320  * state before calling the states do_segment function. Data processing
 321  * of inbound segments also now happens in the hpts_do_segment in general
 322  * with only one exception. This is so we can keep the connection on
 323  * a single CPU.
 324  *
 325  * Each state is simplified due to the fact that the original do_segment
 326  * has been decomposed and we *know* what state we are in (no
 327  * switches on the state) and all tests for SACK are gone. This
 328  * greatly simplifies what each state does.
 329  *
 330  * TCP output is also over-written with a new version since it
 331  * must maintain the new rack scoreboard and has had hptsi
 332  * integrated as a requirment. Still todo is to eliminate the
 333  * use of the callout_() system and use the hpts for all
 334  * timers as well.
 335  */
 336 static uint32_t bbr_rtt_probe_time = 200000;    /* 200ms in micro seconds */
 337 static uint32_t bbr_rtt_probe_cwndtarg = 4;     /* How many mss's outstanding */
 338 static const int32_t bbr_min_req_free = 2;      /* The min we must have on the
 339                                                  * free list */
 340 static int32_t bbr_tlp_thresh = 1;
 341 static int32_t bbr_reorder_thresh = 2;
 342 static int32_t bbr_reorder_fade = 60000000;     /* 0 - never fade, def
 343                                                  * 60,000,000 - 60 seconds */
 344 static int32_t bbr_pkt_delay = 1000;
 345 static int32_t bbr_min_to = 1000;       /* Number of usec's minimum timeout */
 346 static int32_t bbr_incr_timers = 1;
 347
 348 static int32_t bbr_tlp_min = 10000;     /* 10ms in usecs */
 349 static int32_t bbr_delayed_ack_time = 200000;   /* 200ms in usecs */
 350 static int32_t bbr_exit_startup_at_loss = 1;
 351
 352 /*
 353  * bbr_lt_bw_ratio is 1/8th
 354  * bbr_lt_bw_diff is  < 4 Kbit/sec
 355  */
 356 static uint64_t bbr_lt_bw_diff = 4000 / 8;      /* In bytes per second */
 357 static uint64_t bbr_lt_bw_ratio = 8;    /* For 1/8th */
 358 static uint32_t bbr_lt_bw_max_rtts = 48;        /* How many rtt's do we use
 359                                                  * the lt_bw for */
 360 static uint32_t bbr_lt_intvl_min_rtts = 4;      /* Min num of RTT's to measure
 361                                                  * lt_bw */
 362 static int32_t bbr_lt_intvl_fp = 0;             /* False positive epoch diff */
 363 static int32_t bbr_lt_loss_thresh = 196;        /* Lost vs delivered % */
 364 static int32_t bbr_lt_fd_thresh = 100;          /* false detection % */
 365
 366 static int32_t bbr_verbose_logging = 0;
 367 /*
 368  * Currently regular tcp has a rto_min of 30ms
 369  * the backoff goes 12 times so that ends up
 370  * being a total of 122.850 seconds before a
 371  * connection is killed.
 372  */
 373 static int32_t bbr_rto_min_ms = 30;     /* 30ms same as main freebsd */
 374 static int32_t bbr_rto_max_sec = 4;     /* 4 seconds */
 375
 376 /****************************************************/
 377 /* DEFAULT TSO SIZING  (cpu performance impacting)  */
 378 /****************************************************/
 379 /* What amount is our formula using to get TSO size */
 380 static int32_t bbr_hptsi_per_second = 1000;
 381
 382 /*
 383  * For hptsi under bbr_cross_over connections what is delay
 384  * target 7ms (in usec) combined with a seg_max of 2
 385  * gets us close to identical google behavior in
 386  * TSO size selection (possibly more 1MSS sends).
 387  */
 388 static int32_t bbr_hptsi_segments_delay_tar = 7000;
 389
 390 /* Does pacing delay include overhead's in its time calculations? */
 391 static int32_t bbr_include_enet_oh = 0;
 392 static int32_t bbr_include_ip_oh = 1;
 393 static int32_t bbr_include_tcp_oh = 1;
 394 static int32_t bbr_google_discount = 10;
 395
 396 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */
 397 static int32_t bbr_state_is_pkt_epoch = 0;
 398 static int32_t bbr_state_drain_2_tar = 1;
 399 /* What is the max the 0 - bbr_cross_over MBPS TSO target
 400  * can reach using our delay target. Note that this
 401  * value becomes the floor for the cross over
 402  * algorithm.
 403  */
 404 static int32_t bbr_hptsi_segments_max = 2;
 405 static int32_t bbr_hptsi_segments_floor = 1;
 406 static int32_t bbr_hptsi_utter_max = 0;
 407
 408 /* What is the min the 0 - bbr_cross-over MBPS  TSO target can be */
 409 static int32_t bbr_hptsi_bytes_min = 1460;
 410 static int32_t bbr_all_get_min = 0;
 411
 412 /* Cross over point from algo-a to algo-b */
 413 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS;
 414
 415 /* Do we deal with our restart state? */
 416 static int32_t bbr_uses_idle_restart = 0;
 417 static int32_t bbr_idle_restart_threshold = 100000;     /* 100ms in useconds */
 418
 419 /* Do we allow hardware pacing? */
 420 static int32_t bbr_allow_hdwr_pacing = 0;
 421 static int32_t bbr_hdwr_pace_adjust = 2;        /* multipler when we calc the tso size */
 422 static int32_t bbr_hdwr_pace_floor = 1;
 423 static int32_t bbr_hdwr_pacing_delay_cnt = 10;
 424
 425 /****************************************************/
 426 static int32_t bbr_resends_use_tso = 0;
 427 static int32_t bbr_tlp_max_resend = 2;
 428 static int32_t bbr_sack_block_limit = 128;
 429
 430 #define  BBR_MAX_STAT 19
 431 counter_u64_t bbr_state_time[BBR_MAX_STAT];
 432 counter_u64_t bbr_state_lost[BBR_MAX_STAT];
 433 counter_u64_t bbr_state_resend[BBR_MAX_STAT];
 434 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE];
 435 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE];
 436 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE];
 437 counter_u64_t bbr_flows_whdwr_pacing;
 438 counter_u64_t bbr_flows_nohdwr_pacing;
 439
 440 counter_u64_t bbr_nohdwr_pacing_enobuf;
 441 counter_u64_t bbr_hdwr_pacing_enobuf;
 442
 443 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr);
 444
 445 /*
 446  * Static defintions we need for forward declarations.
 447  */
 448 static uint32_t
 449 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain,
 450     uint32_t useconds_time, uint64_t bw);
 451 static uint32_t
 452 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain);
 453 static void
 454      bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win);
 455 static void
 456 bbr_set_probebw_gains(struct tcp_bbr *bbr,  uint32_t cts, uint32_t losses);
 457 static void
 458 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line,
 459                     int dolog);
 460 static uint32_t
 461 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain);
 462 static void
 463 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch,
 464                  int32_t pkt_epoch, uint32_t losses);
 465 static uint32_t
 466 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm);
 467 static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp);
 468 static uint32_t
 469 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
 470     struct bbr_sendmap *rsm, uint32_t srtt,
 471     uint32_t cts);
 472 static void
 473 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts,
 474     int32_t line);
 475 static void
 476      bbr_set_state_target(struct tcp_bbr *bbr, int line);
 477 static void
 478      bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line);
 479
 480 static void
 481      bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line);
 482
 483 static void
 484      tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts);
 485
 486 static void
 487      bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts);
 488
 489 static void
 490      bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt,
 491                          uint32_t line, uint8_t is_start, uint16_t set);
 492
 493 static struct bbr_sendmap *
 494             bbr_find_lowest_rsm(struct tcp_bbr *bbr);
 495 static __inline uint32_t
 496 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type);
 497 static void
 498      bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which);
 499
 500 static void
 501 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
 502     uint32_t thresh, uint32_t to);
 503 static void
 504      bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag);
 505
 506 static void
 507 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot,
 508     uint32_t del_by, uint32_t cts, uint32_t sloton, uint32_t prev_delay);
 509
 510 static void
 511 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr,
 512     uint32_t cts, int32_t line);
 513 static void
 514      bbr_stop_all_timers(struct tcpcb *tp);
 515 static void
 516      bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts);
 517 static void
 518      bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts);
 519 static void
 520      bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts);
 521
 522
 523 static void
 524 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
 525     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod);
 526
 527 static inline uint8_t
 528 bbr_state_val(struct tcp_bbr *bbr)
 529 {
 530         return(bbr->rc_bbr_substate);
 531 }
 532
 533 static inline uint32_t
 534 get_min_cwnd(struct tcp_bbr *bbr)
 535 {
 536         int mss;
 537
 538         mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
 539         if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED)
 540                 return (bbr_cwnd_min_val_hs * mss);
 541         else
 542                 return (bbr_cwnd_min_val * mss);
 543 }
 544
 545 static uint32_t
 546 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr)
 547 {
 548         uint64_t srtt, var;
 549         uint64_t ret_val;
 550
 551         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
 552         if (tp->t_srtt == 0) {
 553                 srtt = (uint64_t)BBR_INITIAL_RTO;
 554                 var = 0;
 555         } else {
 556                 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
 557                 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT);
 558         }
 559         TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]),
 560             bbr_persist_min, bbr_persist_max);
 561         return ((uint32_t)ret_val);
 562 }
 563
 564 static uint32_t
 565 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
 566 {
 567         /*
 568          * Start the FR timer, we do this based on getting the first one in
 569          * the rc_tmap. Note that if its NULL we must stop the timer. in all
 570          * events we need to stop the running timer (if its running) before
 571          * starting the new one.
 572          */
 573         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
 574         int32_t idx;
 575         int32_t is_tlp_timer = 0;
 576         struct bbr_sendmap *rsm;
 577
 578         if (bbr->rc_all_timers_stopped) {
 579                 /* All timers have been stopped none are to run */
 580                 return (0);
 581         }
 582         if (bbr->rc_in_persist) {
 583                 /* We can't start any timer in persists */
 584                 return (bbr_get_persists_timer_val(tp, bbr));
 585         }
 586         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 587         if ((rsm == NULL) ||
 588             ((tp->t_flags & TF_SACK_PERMIT) == 0) ||
 589             (tp->t_state < TCPS_ESTABLISHED)) {
 590                 /* Nothing on the send map */
 591 activate_rxt:
 592                 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
 593                         uint64_t tov;
 594
 595                         time_since_sent = 0;
 596                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 597                         if (rsm) {
 598                                 idx = rsm->r_rtr_cnt - 1;
 599                                 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
 600                                         tstmp_touse = rsm->r_tim_lastsent[idx];
 601                                 else
 602                                         tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
 603                                 if (TSTMP_GT(tstmp_touse, cts))
 604                                     time_since_sent = cts - tstmp_touse;
 605                         }
 606                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
 607                         if (tp->t_srtt == 0)
 608                                 tov = BBR_INITIAL_RTO;
 609                         else
 610                                 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) +
 611                                     ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT);
 612                         if (tp->t_rxtshift)
 613                                 tov *= tcp_backoff[tp->t_rxtshift];
 614                         if (tov > time_since_sent)
 615                                 tov -= time_since_sent;
 616                         else
 617                                 tov = bbr->r_ctl.rc_min_to;
 618                         TCPT_RANGESET_NOSLOP(to, tov,
 619                             (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC),
 620                             (bbr->rc_max_rto_sec * USECS_IN_SECOND));
 621                         bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to);
 622                         return (to);
 623                 }
 624                 return (0);
 625         }
 626         if (rsm->r_flags & BBR_ACKED) {
 627                 rsm = bbr_find_lowest_rsm(bbr);
 628                 if (rsm == NULL) {
 629                         /* No lowest? */
 630                         goto activate_rxt;
 631                 }
 632         }
 633         /* Convert from ms to usecs */
 634         if (rsm->r_flags & BBR_SACK_PASSED) {
 635                 if ((tp->t_flags & TF_SENTFIN) &&
 636                     ((tp->snd_max - tp->snd_una) == 1) &&
 637                     (rsm->r_flags & BBR_HAS_FIN)) {
 638                         /*
 639                          * We don't start a bbr rack timer if all we have is
 640                          * a FIN outstanding.
 641                          */
 642                         goto activate_rxt;
 643                 }
 644                 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK);
 645                 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm);
 646                 idx = rsm->r_rtr_cnt - 1;
 647                 exp = rsm->r_tim_lastsent[idx] + thresh;
 648                 if (SEQ_GEQ(exp, cts)) {
 649                         to = exp - cts;
 650                         if (to < bbr->r_ctl.rc_min_to) {
 651                                 to = bbr->r_ctl.rc_min_to;
 652                         }
 653                 } else {
 654                         to = bbr->r_ctl.rc_min_to;
 655                 }
 656         } else {
 657                 /* Ok we need to do a TLP not RACK */
 658                 if (bbr->rc_tlp_in_progress != 0) {
 659                         /*
 660                          * The previous send was a TLP.
 661                          */
 662                         goto activate_rxt;
 663                 }
 664                 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
 665                 if (rsm == NULL) {
 666                         /* We found no rsm to TLP with. */
 667                         goto activate_rxt;
 668                 }
 669                 if (rsm->r_flags & BBR_HAS_FIN) {
 670                         /* If its a FIN we don't do TLP */
 671                         rsm = NULL;
 672                         goto activate_rxt;
 673                 }
 674                 time_since_sent = 0;
 675                 idx = rsm->r_rtr_cnt - 1;
 676                 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
 677                         tstmp_touse = rsm->r_tim_lastsent[idx];
 678                 else
 679                         tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
 680                 if (TSTMP_GT(tstmp_touse, cts))
 681                     time_since_sent = cts - tstmp_touse;
 682                 is_tlp_timer = 1;
 683                 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use);
 684                 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts);
 685                 if (thresh > time_since_sent)
 686                         to = thresh - time_since_sent;
 687                 else
 688                         to = bbr->r_ctl.rc_min_to;
 689                 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
 690                         /*
 691                          * If the TLP time works out to larger than the max
 692                          * RTO lets not do TLP.. just RTO.
 693                          */
 694                         goto activate_rxt;
 695                 }
 696                 if ((bbr->rc_tlp_rtx_out == 1) &&
 697                     (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) {
 698                         /*
 699                          * Second retransmit of the same TLP
 700                          * lets not.
 701                          */
 702                         bbr->rc_tlp_rtx_out = 0;
 703                         goto activate_rxt;
 704                 }
 705                 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) {
 706                         /*
 707                          * The tail is no longer the last one I did a probe
 708                          * on
 709                          */
 710                         bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
 711                         bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
 712                 }
 713         }
 714         if (is_tlp_timer == 0) {
 715                 BBR_STAT_INC(bbr_to_arm_rack);
 716                 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
 717         } else {
 718                 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to);
 719                 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
 720                         /*
 721                          * We have exceeded how many times we can retran the
 722                          * current TLP timer, switch to the RTO timer.
 723                          */
 724                         goto activate_rxt;
 725                 } else {
 726                         BBR_STAT_INC(bbr_to_arm_tlp);
 727                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
 728                 }
 729         }
 730         return (to);
 731 }
 732
 733 static inline int32_t
 734 bbr_minseg(struct tcp_bbr *bbr)
 735 {
 736         return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options);
 737 }
 738
 739 static void
 740 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len)
 741 {
 742         struct inpcb *inp;
 743         struct hpts_diag diag;
 744         uint32_t delayed_ack = 0;
 745         uint32_t left = 0;
 746         uint32_t hpts_timeout;
 747         uint8_t stopped;
 748         int32_t delay_calc = 0;
 749         uint32_t prev_delay = 0;
 750
 751         inp = tp->t_inpcb;
 752         if (inp->inp_in_hpts) {
 753                 /* A previous call is already set up */
 754                 return;
 755         }
 756         if ((tp->t_state == TCPS_CLOSED) ||
 757             (tp->t_state == TCPS_LISTEN)) {
 758                 return;
 759         }
 760         stopped = bbr->rc_tmr_stopped;
 761         if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) {
 762                 left = bbr->r_ctl.rc_timer_exp - cts;
 763         }
 764         bbr->r_ctl.rc_hpts_flags = 0;
 765         bbr->r_ctl.rc_timer_exp = 0;
 766         prev_delay = bbr->r_ctl.rc_last_delay_val;
 767         if (bbr->r_ctl.rc_last_delay_val &&
 768             (slot == 0)) {
 769                 /*
 770                  * If a previous pacer delay was in place we
 771                  * are not coming from the output side (where
 772                  * we calculate a delay, more likely a timer).
 773                  */
 774                 slot = bbr->r_ctl.rc_last_delay_val;
 775                 if (TSTMP_GT(cts, bbr->rc_pacer_started)) {
 776                         /* Compensate for time passed  */
 777                         delay_calc = cts - bbr->rc_pacer_started;
 778                         if (delay_calc <= slot)
 779                                 slot -= delay_calc;
 780                 }
 781         }
 782         /* Do we have early to make up for by pushing out the pacing time? */
 783         if (bbr->r_agg_early_set) {
 784                 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2);
 785                 slot += bbr->r_ctl.rc_agg_early;
 786                 bbr->r_ctl.rc_agg_early = 0;
 787                 bbr->r_agg_early_set = 0;
 788         }
 789         /* Are we running a total debt that needs to be compensated for? */
 790         if (bbr->r_ctl.rc_hptsi_agg_delay) {
 791                 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) {
 792                         /* We nuke the delay */
 793                         slot -= bbr->r_ctl.rc_hptsi_agg_delay;
 794                         bbr->r_ctl.rc_hptsi_agg_delay = 0;
 795                 } else {
 796                         /* We nuke some of the delay, put in a minimal 100usecs  */
 797                         bbr->r_ctl.rc_hptsi_agg_delay -= slot;
 798                         bbr->r_ctl.rc_last_delay_val = slot = 100;
 799                 }
 800         }
 801         bbr->r_ctl.rc_last_delay_val = slot;
 802         hpts_timeout = bbr_timer_start(tp, bbr, cts);
 803         if (tp->t_flags & TF_DELACK) {
 804                 if (bbr->rc_in_persist == 0) {
 805                         delayed_ack = bbr_delack_time;
 806                 } else {
 807                         /*
 808                          * We are in persists and have
 809                          * gotten a new data element.
 810                          */
 811                         if (hpts_timeout > bbr_delack_time) {
 812                                 /*
 813                                  * Lets make the persists timer (which acks)
 814                                  * be the smaller of hpts_timeout and bbr_delack_time.
 815                                  */
 816                                 hpts_timeout = bbr_delack_time;
 817                         }
 818                 }
 819         }
 820         if (delayed_ack &&
 821             ((hpts_timeout == 0) ||
 822              (delayed_ack < hpts_timeout))) {
 823                 /* We need a Delayed ack timer */
 824                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
 825                 hpts_timeout = delayed_ack;
 826         }
 827         if (slot) {
 828                 /* Mark that we have a pacing timer up */
 829                 BBR_STAT_INC(bbr_paced_segments);
 830                 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
 831         }
 832         /*
 833          * If no timers are going to run and we will fall off thfe hptsi
 834          * wheel, we resort to a keep-alive timer if its configured.
 835          */
 836         if ((hpts_timeout == 0) &&
 837             (slot == 0)) {
 838                 if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
 839                     (tp->t_state <= TCPS_CLOSING)) {
 840                         /*
 841                          * Ok we have no timer (persists, rack, tlp, rxt  or
 842                          * del-ack), we don't have segments being paced. So
 843                          * all that is left is the keepalive timer.
 844                          */
 845                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
 846                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
 847                         } else {
 848                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
 849                         }
 850                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
 851                 }
 852         }
 853         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
 854             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
 855                 /*
 856                  * RACK, TLP, persists and RXT timers all are restartable
 857                  * based on actions input .. i.e we received a packet (ack
 858                  * or sack) and that changes things (rw, or snd_una etc).
 859                  * Thus we can restart them with a new value. For
 860                  * keep-alive, delayed_ack we keep track of what was left
 861                  * and restart the timer with a smaller value.
 862                  */
 863                 if (left < hpts_timeout)
 864                         hpts_timeout = left;
 865         }
 866         if (bbr->r_ctl.rc_incr_tmrs && slot &&
 867             (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
 868                 /*
 869                  * If configured to do so, and the timer is either
 870                  * the TLP or RXT timer, we need to increase the timeout
 871                  * by the pacing time. Consider the bottleneck at my
 872                  * machine as an example, we are sending something
 873                  * to start a TLP on. The last packet won't be emitted
 874                  * fully until the pacing time (the bottleneck will hold
 875                  * the data in place). Once the packet is emitted that
 876                  * is when we want to start waiting for the TLP. This
 877                  * is most evident with hardware pacing (where the nic
 878                  * is holding the packet(s) before emitting). But it
 879                  * can also show up in the network so we do it for all
 880                  * cases. Technically we would take off one packet from
 881                  * this extra delay but this is easier and being more
 882                  * conservative is probably better.
 883                  */
 884                 hpts_timeout += slot;
 885         }
 886         if (hpts_timeout) {
 887                 /*
 888                  * Hack alert for now we can't time-out over 2147 seconds (a
 889                  * bit more than 35min)
 890                  */
 891                 if (hpts_timeout > 0x7ffffffe)
 892                         hpts_timeout = 0x7ffffffe;
 893                 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
 894         } else
 895                 bbr->r_ctl.rc_timer_exp = 0;
 896         if ((slot) &&
 897             (bbr->rc_use_google ||
 898              bbr->output_error_seen ||
 899              (slot <= hpts_timeout))  ) {
 900                 /*
 901                  * Tell LRO that it can queue packets while
 902                  * we pace.
 903                  */
 904                 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
 905                 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
 906                     (bbr->rc_cwnd_limited == 0)) {
 907                         /*
 908                          * If we are not cwnd limited and we
 909                          * are running a rack timer we put on
 910                          * the do not disturbe even for sack.
 911                          */
 912                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
 913                 } else
 914                         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
 915                 bbr->rc_pacer_started = cts;
 916
 917                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
 918                                            __LINE__, &diag);
 919                 bbr->rc_timer_first = 0;
 920                 bbr->bbr_timer_src = frm;
 921                 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1);
 922                 bbr_log_hpts_diag(bbr, cts, &diag);
 923         } else if (hpts_timeout) {
 924                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
 925                                            __LINE__, &diag);
 926                 /*
 927                  * We add the flag here as well if the slot is set,
 928                  * since hpts will call in to clear the queue first before
 929                  * calling the output routine (which does our timers).
 930                  * We don't want to set the flag if its just a timer
 931                  * else the arrival of data might (that causes us
 932                  * to send more) might get delayed. Imagine being
 933                  * on a keep-alive timer and a request comes in for
 934                  * more data.
 935                  */
 936                 if (slot)
 937                         bbr->rc_pacer_started = cts;
 938                 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
 939                     (bbr->rc_cwnd_limited == 0)) {
 940                         /*
 941                          * For a rack timer, don't wake us even
 942                          * if a sack arrives as long as we are
 943                          * not cwnd limited.
 944                          */
 945                         bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
 946                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
 947                 } else {
 948                         /* All other timers wake us up */
 949                         bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
 950                         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
 951                 }
 952                 bbr->bbr_timer_src = frm;
 953                 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0);
 954                 bbr_log_hpts_diag(bbr, cts, &diag);
 955                 bbr->rc_timer_first = 1;
 956         }
 957         bbr->rc_tmr_stopped = 0;
 958         bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay);
 959 }
 960
 961 static void
 962 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb)
 963 {
 964         /*
 965          * We received an ack, and then did not call send or were bounced
 966          * out due to the hpts was running. Now a timer is up as well, is it
 967          * the right timer?
 968          */
 969         struct inpcb *inp;
 970         struct bbr_sendmap *rsm;
 971         uint32_t hpts_timeout;
 972         int tmr_up;
 973
 974         tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
 975         if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
 976                 return;
 977         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 978         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
 979             (tmr_up == PACE_TMR_RXT)) {
 980                 /* Should be an RXT */
 981                 return;
 982         }
 983         inp = bbr->rc_inp;
 984         if (rsm == NULL) {
 985                 /* Nothing outstanding? */
 986                 if (tp->t_flags & TF_DELACK) {
 987                         if (tmr_up == PACE_TMR_DELACK)
 988                                 /*
 989                                  * We are supposed to have delayed ack up
 990                                  * and we do
 991                                  */
 992                                 return;
 993                 } else if (sbavail(&inp->inp_socket->so_snd) &&
 994                     (tmr_up == PACE_TMR_RXT)) {
 995                         /*
 996                          * if we hit enobufs then we would expect the
 997                          * possiblity of nothing outstanding and the RXT up
 998                          * (and the hptsi timer).
 999                          */
1000                         return;
1001                 } else if (((tcp_always_keepalive ||
1002                             inp->inp_socket->so_options & SO_KEEPALIVE) &&
1003                             (tp->t_state <= TCPS_CLOSING)) &&
1004                             (tmr_up == PACE_TMR_KEEP) &&
1005                     (tp->snd_max == tp->snd_una)) {
1006                         /* We should have keep alive up and we do */
1007                         return;
1008                 }
1009         }
1010         if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) {
1011                 if ((tp->t_flags & TF_SENTFIN) &&
1012                     ((tp->snd_max - tp->snd_una) == 1) &&
1013                     (rsm->r_flags & BBR_HAS_FIN)) {
1014                         /* needs to be a RXT */
1015                         if (tmr_up == PACE_TMR_RXT)
1016                                 return;
1017                         else
1018                                 goto wrong_timer;
1019                 } else if (tmr_up == PACE_TMR_RACK)
1020                         return;
1021                 else
1022                         goto wrong_timer;
1023         } else if (rsm && (tmr_up == PACE_TMR_RACK)) {
1024                 /* Rack timer has priority if we have data out */
1025                 return;
1026         } else if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1027                     ((tmr_up == PACE_TMR_TLP) ||
1028             (tmr_up == PACE_TMR_RXT))) {
1029                 /*
1030                  * Either a TLP or RXT is fine if no sack-passed is in place
1031                  * and data is outstanding.
1032                  */
1033                 return;
1034         } else if (tmr_up == PACE_TMR_DELACK) {
1035                 /*
1036                  * If the delayed ack was going to go off before the
1037                  * rtx/tlp/rack timer were going to expire, then that would
1038                  * be the timer in control. Note we don't check the time
1039                  * here trusting the code is correct.
1040                  */
1041                 return;
1042         }
1043         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1044             ((tmr_up == PACE_TMR_RXT) ||
1045              (tmr_up == PACE_TMR_TLP) ||
1046              (tmr_up == PACE_TMR_RACK))) {
1047                 /*
1048                  * We have outstanding data and
1049                  * we *do* have a RACK, TLP or RXT
1050                  * timer running. We won't restart
1051                  * anything here since thats probably ok we
1052                  * will get called with some timer here shortly.
1053                  */
1054                 return;
1055         }
1056         /*
1057          * Ok the timer originally started is not what we want now. We will
1058          * force the hpts to be stopped if any, and restart with the slot
1059          * set to what was in the saved slot.
1060          */
1061 wrong_timer:
1062         if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) {
1063                 if (inp->inp_in_hpts)
1064                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
1065                 bbr_timer_cancel(bbr, __LINE__, cts);
1066                 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val,
1067                     0);
1068         } else {
1069                 /*
1070                  * Output is hptsi so we just need to switch the type of
1071                  * timer. We don't bother with keep-alive, since when we
1072                  * jump through the output, it will start the keep-alive if
1073                  * nothing is sent.
1074                  *
1075                  * We only need a delayed-ack added and or the hpts_timeout.
1076                  */
1077                 hpts_timeout = bbr_timer_start(tp, bbr, cts);
1078                 if (tp->t_flags & TF_DELACK) {
1079                         if (hpts_timeout == 0) {
1080                                 hpts_timeout = bbr_delack_time;
1081                                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1082                         }
1083                         else if (hpts_timeout > bbr_delack_time) {
1084                                 hpts_timeout = bbr_delack_time;
1085                                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1086                         }
1087                 }
1088                 if (hpts_timeout) {
1089                         if (hpts_timeout > 0x7ffffffe)
1090                                 hpts_timeout = 0x7ffffffe;
1091                         bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
1092                 }
1093         }
1094 }
1095
1096 int32_t bbr_clear_lost = 0;
1097
1098 /*
1099  * Considers the two time values now (cts) and earlier.
1100  * If cts is smaller than earlier, we could have
1101  * had a sequence wrap (our counter wraps every
1102  * 70 min or so) or it could be just clock skew
1103  * getting us two differnt time values. Clock skew
1104  * will show up within 10ms or so. So in such
1105  * a case (where cts is behind earlier time by
1106  * less than 10ms) we return 0. Otherwise we
1107  * return the true difference between them.
1108  */
1109 static inline uint32_t
1110 bbr_calc_time(uint32_t cts, uint32_t earlier_time) {
1111         /*
1112          * Given two timestamps, the current time stamp cts, and some other
1113          * time-stamp taken in theory earlier return the difference. The
1114          * trick is here sometimes locking will get the other timestamp
1115          * after the cts. If this occurs we need to return 0.
1116          */
1117         if (TSTMP_GEQ(cts, earlier_time))
1118                 return (cts - earlier_time);
1119         /*
1120          * cts is behind earlier_time if its less than 10ms consider it 0.
1121          * If its more than 10ms difference then we had a time wrap. Else
1122          * its just the normal locking foo. I wonder if we should not go to
1123          * 64bit TS and get rid of this issue.
1124          */
1125         if (TSTMP_GEQ((cts + 10000), earlier_time))
1126                 return (0);
1127         /*
1128          * Ok the time must have wrapped. So we need to answer a large
1129          * amount of time, which the normal subtraction should do.
1130          */
1131         return (cts - earlier_time);
1132 }
1133
1134
1135
1136 static int
1137 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS)
1138 {
1139         uint32_t stat;
1140         int32_t error;
1141
1142         error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t));
1143         if (error || req->newptr == NULL)
1144                 return error;
1145
1146         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
1147         if (error)
1148                 return (error);
1149         if (stat == 1) {
1150 #ifdef BBR_INVARIANTS
1151                 printf("Clearing BBR lost counters\n");
1152 #endif
1153                 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT);
1154                 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT);
1155                 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT);
1156         } else if (stat == 2) {
1157 #ifdef BBR_INVARIANTS
1158                 printf("Clearing BBR option counters\n");
1159 #endif
1160                 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE);
1161         } else if (stat == 3) {
1162 #ifdef BBR_INVARIANTS
1163                 printf("Clearing BBR stats counters\n");
1164 #endif
1165                 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE);
1166         } else if (stat == 4) {
1167 #ifdef BBR_INVARIANTS
1168                 printf("Clearing BBR out-size counters\n");
1169 #endif
1170                 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE);
1171         }
1172         bbr_clear_lost = 0;
1173         return (0);
1174 }
1175
1176 static void
1177 bbr_init_sysctls(void)
1178 {
1179         struct sysctl_oid *bbr_probertt;
1180         struct sysctl_oid *bbr_hptsi;
1181         struct sysctl_oid *bbr_measure;
1182         struct sysctl_oid *bbr_cwnd;
1183         struct sysctl_oid *bbr_timeout;
1184         struct sysctl_oid *bbr_states;
1185         struct sysctl_oid *bbr_startup;
1186         struct sysctl_oid *bbr_policer;
1187
1188         /* Probe rtt controls */
1189         bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1190             SYSCTL_CHILDREN(bbr_sysctl_root),
1191             OID_AUTO,
1192             "probertt",
1193             CTLFLAG_RW, 0,
1194             "");
1195         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1196             SYSCTL_CHILDREN(bbr_probertt),
1197             OID_AUTO, "gain", CTLFLAG_RW,
1198             &bbr_rttprobe_gain, 192,
1199             "What is the filter gain drop in probe_rtt (0=disable)?");
1200         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1201             SYSCTL_CHILDREN(bbr_probertt),
1202             OID_AUTO, "cwnd", CTLFLAG_RW,
1203             &bbr_rtt_probe_cwndtarg, 4,
1204             "How many mss's are outstanding during probe-rtt");
1205         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1206             SYSCTL_CHILDREN(bbr_probertt),
1207             OID_AUTO, "int", CTLFLAG_RW,
1208             &bbr_rtt_probe_limit, 4000000,
1209             "If RTT has not shrank in this many micro-seconds enter probe-rtt");
1210         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1211             SYSCTL_CHILDREN(bbr_probertt),
1212             OID_AUTO, "mintime", CTLFLAG_RW,
1213             &bbr_rtt_probe_time, 200000,
1214             "How many microseconds in probe-rtt");
1215         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1216             SYSCTL_CHILDREN(bbr_probertt),
1217             OID_AUTO, "filter_len_sec", CTLFLAG_RW,
1218             &bbr_filter_len_sec, 6,
1219             "How long in seconds does the rttProp filter run?");
1220         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1221             SYSCTL_CHILDREN(bbr_probertt),
1222             OID_AUTO, "drain_rtt", CTLFLAG_RW,
1223             &bbr_drain_rtt, BBR_SRTT,
1224             "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?");
1225         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1226             SYSCTL_CHILDREN(bbr_probertt),
1227             OID_AUTO, "can_force", CTLFLAG_RW,
1228             &bbr_can_force_probertt, 0,
1229             "If we keep setting new low rtt's but delay going in probe-rtt can we force in??");
1230         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1231             SYSCTL_CHILDREN(bbr_probertt),
1232             OID_AUTO, "enter_sets_force", CTLFLAG_RW,
1233             &bbr_probertt_sets_rtt, 0,
1234             "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?");
1235         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1236             SYSCTL_CHILDREN(bbr_probertt),
1237             OID_AUTO, "can_adjust", CTLFLAG_RW,
1238             &bbr_can_adjust_probertt, 1,
1239             "Can we dynamically adjust the probe-rtt limits and times?");
1240         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1241             SYSCTL_CHILDREN(bbr_probertt),
1242             OID_AUTO, "is_ratio", CTLFLAG_RW,
1243             &bbr_is_ratio, 0,
1244             "is the limit to filter a ratio?");
1245         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1246             SYSCTL_CHILDREN(bbr_probertt),
1247             OID_AUTO, "use_cwnd", CTLFLAG_RW,
1248             &bbr_prtt_slam_cwnd, 0,
1249             "Should we set/recover cwnd?");
1250         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1251             SYSCTL_CHILDREN(bbr_probertt),
1252             OID_AUTO, "can_use_ts", CTLFLAG_RW,
1253             &bbr_can_use_ts_for_rtt, 1,
1254             "Can we use the ms timestamp if available for retransmistted rtt calculations?");
1255
1256         /* Pacing controls */
1257         bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1258             SYSCTL_CHILDREN(bbr_sysctl_root),
1259             OID_AUTO,
1260             "pacing",
1261             CTLFLAG_RW, 0,
1262             "");
1263         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1264             SYSCTL_CHILDREN(bbr_hptsi),
1265             OID_AUTO, "hw_pacing", CTLFLAG_RW,
1266             &bbr_allow_hdwr_pacing, 1,
1267             "Do we allow hardware pacing?");
1268         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1269             SYSCTL_CHILDREN(bbr_hptsi),
1270             OID_AUTO, "hw_pacing_limit", CTLFLAG_RW,
1271             &bbr_hardware_pacing_limit, 4000,
1272             "Do we have a limited number of connections for pacing chelsio (0=no limit)?");
1273         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1274             SYSCTL_CHILDREN(bbr_hptsi),
1275             OID_AUTO, "hw_pacing_adj", CTLFLAG_RW,
1276             &bbr_hdwr_pace_adjust, 2,
1277             "Multiplier to calculated tso size?");
1278         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1279             SYSCTL_CHILDREN(bbr_hptsi),
1280             OID_AUTO, "hw_pacing_floor", CTLFLAG_RW,
1281             &bbr_hdwr_pace_floor, 1,
1282             "Do we invoke the hardware pacing floor?");
1283         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1284             SYSCTL_CHILDREN(bbr_hptsi),
1285             OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW,
1286             &bbr_hdwr_pacing_delay_cnt, 10,
1287             "How many packets must be sent after hdwr pacing is enabled");
1288         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1289             SYSCTL_CHILDREN(bbr_hptsi),
1290             OID_AUTO, "bw_cross", CTLFLAG_RW,
1291             &bbr_cross_over, 3000000,
1292             "What is the point where we cross over to linux like TSO size set");
1293         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1294             SYSCTL_CHILDREN(bbr_hptsi),
1295             OID_AUTO, "seg_deltarg", CTLFLAG_RW,
1296             &bbr_hptsi_segments_delay_tar, 7000,
1297             "What is the worse case delay target for hptsi < 48Mbp connections");
1298         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1299             SYSCTL_CHILDREN(bbr_hptsi),
1300             OID_AUTO, "enet_oh", CTLFLAG_RW,
1301             &bbr_include_enet_oh, 0,
1302             "Do we include the ethernet overhead in calculating pacing delay?");
1303         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1304             SYSCTL_CHILDREN(bbr_hptsi),
1305             OID_AUTO, "ip_oh", CTLFLAG_RW,
1306             &bbr_include_ip_oh, 1,
1307             "Do we include the IP overhead in calculating pacing delay?");
1308         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1309             SYSCTL_CHILDREN(bbr_hptsi),
1310             OID_AUTO, "tcp_oh", CTLFLAG_RW,
1311             &bbr_include_tcp_oh, 0,
1312             "Do we include the TCP overhead in calculating pacing delay?");
1313         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1314             SYSCTL_CHILDREN(bbr_hptsi),
1315             OID_AUTO, "google_discount", CTLFLAG_RW,
1316             &bbr_google_discount, 10,
1317             "What is the default google discount percentage wise for pacing (11 = 1.1%%)?");
1318         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1319             SYSCTL_CHILDREN(bbr_hptsi),
1320             OID_AUTO, "all_get_min", CTLFLAG_RW,
1321             &bbr_all_get_min, 0,
1322             "If you are less than a MSS do you just get the min?");
1323         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1324             SYSCTL_CHILDREN(bbr_hptsi),
1325             OID_AUTO, "tso_min", CTLFLAG_RW,
1326             &bbr_hptsi_bytes_min, 1460,
1327             "For 0 -> 24Mbps what is floor number of segments for TSO");
1328         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1329             SYSCTL_CHILDREN(bbr_hptsi),
1330             OID_AUTO, "seg_tso_max", CTLFLAG_RW,
1331             &bbr_hptsi_segments_max, 6,
1332             "For 0 -> 24Mbps what is top number of segments for TSO");
1333         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1334             SYSCTL_CHILDREN(bbr_hptsi),
1335             OID_AUTO, "seg_floor", CTLFLAG_RW,
1336             &bbr_hptsi_segments_floor, 1,
1337             "Minimum TSO size we will fall too in segments");
1338         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1339             SYSCTL_CHILDREN(bbr_hptsi),
1340             OID_AUTO, "utter_max", CTLFLAG_RW,
1341             &bbr_hptsi_utter_max, 0,
1342             "The absolute maximum that any pacing (outside of hardware) can be");
1343         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1344             SYSCTL_CHILDREN(bbr_hptsi),
1345             OID_AUTO, "seg_divisor", CTLFLAG_RW,
1346             &bbr_hptsi_per_second, 100,
1347             "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps ");
1348         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1349             SYSCTL_CHILDREN(bbr_hptsi),
1350             OID_AUTO, "srtt_mul", CTLFLAG_RW,
1351             &bbr_hptsi_max_mul, 1,
1352             "The multiplier for pace len max");
1353         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1354             SYSCTL_CHILDREN(bbr_hptsi),
1355             OID_AUTO, "srtt_div", CTLFLAG_RW,
1356             &bbr_hptsi_max_div, 2,
1357             "The divisor for pace len max");
1358         /* Measurement controls */
1359         bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1360             SYSCTL_CHILDREN(bbr_sysctl_root),
1361             OID_AUTO,
1362             "measure",
1363             CTLFLAG_RW, 0,
1364             "Measurement controls");
1365         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1366             SYSCTL_CHILDREN(bbr_measure),
1367             OID_AUTO, "min_i_bw", CTLFLAG_RW,
1368             &bbr_initial_bw_bps, 62500,
1369             "Minimum initial b/w in bytes per second");
1370         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1371             SYSCTL_CHILDREN(bbr_measure),
1372             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1373             &bbr_sack_not_required, 0,
1374             "Do we allow bbr to run on connections not supporting SACK?");
1375         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1376             SYSCTL_CHILDREN(bbr_measure),
1377             OID_AUTO, "use_google", CTLFLAG_RW,
1378             &bbr_use_google_algo, 0,
1379             "Use has close to google V1.0 has possible?");
1380         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1381             SYSCTL_CHILDREN(bbr_measure),
1382             OID_AUTO, "ts_limiting", CTLFLAG_RW,
1383             &bbr_ts_limiting, 1,
1384             "Do we attempt to use the peers timestamp to limit b/w caculations?");
1385         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1386             SYSCTL_CHILDREN(bbr_measure),
1387             OID_AUTO, "ts_can_raise", CTLFLAG_RW,
1388             &bbr_ts_can_raise, 0,
1389             "Can we raise the b/w via timestamp b/w calculation?");
1390         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1391             SYSCTL_CHILDREN(bbr_measure),
1392             OID_AUTO, "ts_delta", CTLFLAG_RW,
1393             &bbr_min_usec_delta, 20000,
1394             "How long in usec between ts of our sends in ts validation code?");
1395         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1396             SYSCTL_CHILDREN(bbr_measure),
1397             OID_AUTO, "ts_peer_delta", CTLFLAG_RW,
1398             &bbr_min_peer_delta, 20,
1399             "What min numerical value should be between the peer deltas?");
1400         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1401             SYSCTL_CHILDREN(bbr_measure),
1402             OID_AUTO, "ts_delta_percent", CTLFLAG_RW,
1403             &bbr_delta_percent, 150,
1404             "What percentage (150 = 15.0) do we allow variance for?");
1405         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1406             SYSCTL_CHILDREN(bbr_measure),
1407             OID_AUTO, "min_measure_good_bw", CTLFLAG_RW,
1408             &bbr_min_measurements_req, 1,
1409             "What is the minimum measurment count we need before we switch to our b/w estimate");
1410         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1411             SYSCTL_CHILDREN(bbr_measure),
1412             OID_AUTO, "min_measure_before_pace", CTLFLAG_RW,
1413             &bbr_no_pacing_until, 4,
1414             "How many pkt-epoch's (0 is off) do we need before pacing is on?");
1415         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1416             SYSCTL_CHILDREN(bbr_measure),
1417             OID_AUTO, "quanta", CTLFLAG_RW,
1418             &bbr_quanta, 2,
1419             "Extra quanta to add when calculating the target (ID section 4.2.3.2).");
1420         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1421             SYSCTL_CHILDREN(bbr_measure),
1422             OID_AUTO, "noretran", CTLFLAG_RW,
1423             &bbr_no_retran, 0,
1424             "Should google mode not use retransmission measurements for the b/w estimation?");
1425         /* State controls */
1426         bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1427             SYSCTL_CHILDREN(bbr_sysctl_root),
1428             OID_AUTO,
1429             "states",
1430             CTLFLAG_RW, 0,
1431             "State controls");
1432         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1433             SYSCTL_CHILDREN(bbr_states),
1434             OID_AUTO, "idle_restart", CTLFLAG_RW,
1435             &bbr_uses_idle_restart, 0,
1436             "Do we use a new special idle_restart state to ramp back up quickly?");
1437         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1438             SYSCTL_CHILDREN(bbr_states),
1439             OID_AUTO, "idle_restart_threshold", CTLFLAG_RW,
1440             &bbr_idle_restart_threshold, 100000,
1441             "How long must we be idle before we restart??");
1442         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1443             SYSCTL_CHILDREN(bbr_states),
1444             OID_AUTO, "use_pkt_epoch", CTLFLAG_RW,
1445             &bbr_state_is_pkt_epoch, 0,
1446             "Do we use a pkt-epoch for substate if 0 rttProp?");
1447         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1448             SYSCTL_CHILDREN(bbr_states),
1449             OID_AUTO, "startup_rtt_gain", CTLFLAG_RW,
1450             &bbr_rtt_gain_thresh, 0,
1451             "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?");
1452         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1453             SYSCTL_CHILDREN(bbr_states),
1454             OID_AUTO, "drain_floor", CTLFLAG_RW,
1455             &bbr_drain_floor, 88,
1456             "What is the lowest we can drain (pg) too?");
1457         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1458             SYSCTL_CHILDREN(bbr_states),
1459             OID_AUTO, "drain_2_target", CTLFLAG_RW,
1460             &bbr_state_drain_2_tar, 1,
1461             "Do we drain to target in drain substate?");
1462         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1463             SYSCTL_CHILDREN(bbr_states),
1464             OID_AUTO, "gain_2_target", CTLFLAG_RW,
1465             &bbr_gain_to_target, 1,
1466             "Does probe bw gain to target??");
1467         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1468             SYSCTL_CHILDREN(bbr_states),
1469             OID_AUTO, "gain_extra_time", CTLFLAG_RW,
1470             &bbr_gain_gets_extra_too, 1,
1471             "Does probe bw gain get the extra time too?");
1472         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1473             SYSCTL_CHILDREN(bbr_states),
1474             OID_AUTO, "ld_div", CTLFLAG_RW,
1475             &bbr_drain_drop_div, 5,
1476             "Long drain drop divider?");
1477         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1478             SYSCTL_CHILDREN(bbr_states),
1479             OID_AUTO, "ld_mul", CTLFLAG_RW,
1480             &bbr_drain_drop_mul, 4,
1481             "Long drain drop multiplier?");
1482         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1483             SYSCTL_CHILDREN(bbr_states),
1484             OID_AUTO, "rand_ot_disc", CTLFLAG_RW,
1485             &bbr_rand_ot, 50,
1486             "Random discount of the ot?");
1487         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1488             SYSCTL_CHILDREN(bbr_states),
1489             OID_AUTO, "dr_filter_life", CTLFLAG_RW,
1490             &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT,
1491             "How many packet-epochs does the b/w delivery rate last?");
1492         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1493             SYSCTL_CHILDREN(bbr_states),
1494             OID_AUTO, "subdrain_applimited", CTLFLAG_RW,
1495             &bbr_sub_drain_app_limit, 0,
1496             "Does our sub-state drain invoke app limited if its long?");
1497         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1498             SYSCTL_CHILDREN(bbr_states),
1499             OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW,
1500             &bbr_sub_drain_slam_cwnd, 0,
1501             "Should we set/recover cwnd for sub-state drain?");
1502         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1503             SYSCTL_CHILDREN(bbr_states),
1504             OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW,
1505             &bbr_slam_cwnd_in_main_drain, 0,
1506             "Should we set/recover cwnd for main-state drain?");
1507         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1508             SYSCTL_CHILDREN(bbr_states),
1509             OID_AUTO, "google_gets_earlyout", CTLFLAG_RW,
1510             &google_allow_early_out, 1,
1511             "Should we allow google probe-bw/drain to exit early at flight target?");
1512         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1513             SYSCTL_CHILDREN(bbr_states),
1514             OID_AUTO, "google_exit_loss", CTLFLAG_RW,
1515             &google_consider_lost, 1,
1516             "Should we have losses exit gain of probebw in google mode??");
1517         /* Startup controls */
1518         bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1519             SYSCTL_CHILDREN(bbr_sysctl_root),
1520             OID_AUTO,
1521             "startup",
1522             CTLFLAG_RW, 0,
1523             "Startup controls");
1524         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1525             SYSCTL_CHILDREN(bbr_startup),
1526             OID_AUTO, "cheat_iwnd", CTLFLAG_RW,
1527             &bbr_sends_full_iwnd, 1,
1528             "Do we not pace but burst out initial windows has our TSO size?");
1529         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1530             SYSCTL_CHILDREN(bbr_startup),
1531             OID_AUTO, "loss_threshold", CTLFLAG_RW,
1532             &bbr_startup_loss_thresh, 2000,
1533             "In startup what is the loss threshold in a pe that will exit us from startup?");
1534         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1535             SYSCTL_CHILDREN(bbr_startup),
1536             OID_AUTO, "use_lowerpg", CTLFLAG_RW,
1537             &bbr_use_lower_gain_in_startup, 1,
1538             "Should we use a lower hptsi gain if we see loss in startup?");
1539         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1540             SYSCTL_CHILDREN(bbr_startup),
1541             OID_AUTO, "gain", CTLFLAG_RW,
1542             &bbr_start_exit, 25,
1543             "What gain percent do we need to see to stay in startup??");
1544         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1545             SYSCTL_CHILDREN(bbr_startup),
1546             OID_AUTO, "low_gain", CTLFLAG_RW,
1547             &bbr_low_start_exit, 15,
1548             "What gain percent do we need to see to stay in the lower gain startup??");
1549         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1550             SYSCTL_CHILDREN(bbr_startup),
1551             OID_AUTO, "loss_exit", CTLFLAG_RW,
1552             &bbr_exit_startup_at_loss, 1,
1553             "Should we exit startup at loss in an epoch if we are not gaining?");
1554         /* CWND controls */
1555         bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1556             SYSCTL_CHILDREN(bbr_sysctl_root),
1557             OID_AUTO,
1558             "cwnd",
1559             CTLFLAG_RW, 0,
1560             "Cwnd controls");
1561         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1562             SYSCTL_CHILDREN(bbr_cwnd),
1563             OID_AUTO, "tar_rtt", CTLFLAG_RW,
1564             &bbr_cwndtarget_rtt_touse, 0,
1565             "Target cwnd rtt measurment to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?");
1566         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1567             SYSCTL_CHILDREN(bbr_cwnd),
1568             OID_AUTO, "may_shrink", CTLFLAG_RW,
1569             &bbr_cwnd_may_shrink, 0,
1570             "Can the cwnd shrink if it would grow to more than the target?");
1571         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1572             SYSCTL_CHILDREN(bbr_cwnd),
1573             OID_AUTO, "max_target_limit", CTLFLAG_RW,
1574             &bbr_target_cwnd_mult_limit, 8,
1575             "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?");
1576         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1577             SYSCTL_CHILDREN(bbr_cwnd),
1578             OID_AUTO, "highspeed_min", CTLFLAG_RW,
1579             &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS,
1580             "What is the high-speed min cwnd (rttProp under 1ms)");
1581         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1582             SYSCTL_CHILDREN(bbr_cwnd),
1583             OID_AUTO, "lowspeed_min", CTLFLAG_RW,
1584             &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS,
1585             "What is the min cwnd (rttProp > 1ms)");
1586         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1587             SYSCTL_CHILDREN(bbr_cwnd),
1588             OID_AUTO, "initwin", CTLFLAG_RW,
1589             &bbr_def_init_win, 10,
1590             "What is the BBR initial window, if 0 use tcp version");
1591         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1592             SYSCTL_CHILDREN(bbr_cwnd),
1593             OID_AUTO, "do_loss_red", CTLFLAG_RW,
1594             &bbr_do_red, 600,
1595             "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?");
1596         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1597             SYSCTL_CHILDREN(bbr_cwnd),
1598             OID_AUTO, "red_scale", CTLFLAG_RW,
1599             &bbr_red_scale, 20000,
1600             "What RTT do we scale with?");
1601         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1602             SYSCTL_CHILDREN(bbr_cwnd),
1603             OID_AUTO, "red_growslow", CTLFLAG_RW,
1604             &bbr_red_growth_restrict, 1,
1605             "Do we restrict cwnd growth for whats in flight?");
1606         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1607             SYSCTL_CHILDREN(bbr_cwnd),
1608             OID_AUTO, "red_div", CTLFLAG_RW,
1609             &bbr_red_div, 2,
1610             "If we reduce whats the divisor?");
1611         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1612             SYSCTL_CHILDREN(bbr_cwnd),
1613             OID_AUTO, "red_mul", CTLFLAG_RW,
1614             &bbr_red_mul, 1,
1615             "If we reduce whats the mulitiplier?");
1616         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1617             SYSCTL_CHILDREN(bbr_cwnd),
1618             OID_AUTO, "target_is_unit", CTLFLAG_RW,
1619             &bbr_target_is_bbunit, 0,
1620             "Is the state target the pacing_gain or BBR_UNIT?");
1621         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1622             SYSCTL_CHILDREN(bbr_cwnd),
1623             OID_AUTO, "drop_limit", CTLFLAG_RW,
1624             &bbr_drop_limit, 0,
1625             "Number of segments limit for drop (0=use min_cwnd w/flight)?");
1626
1627         /* Timeout controls */
1628         bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1629             SYSCTL_CHILDREN(bbr_sysctl_root),
1630             OID_AUTO,
1631             "timeout",
1632             CTLFLAG_RW, 0,
1633             "Time out controls");
1634         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1635             SYSCTL_CHILDREN(bbr_timeout),
1636             OID_AUTO, "delack", CTLFLAG_RW,
1637             &bbr_delack_time, 100000,
1638             "BBR's delayed ack time");
1639         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1640             SYSCTL_CHILDREN(bbr_timeout),
1641             OID_AUTO, "tlp_uses", CTLFLAG_RW,
1642             &bbr_tlp_type_to_use, 3,
1643             "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt");
1644         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1645             SYSCTL_CHILDREN(bbr_timeout),
1646             OID_AUTO, "persmin", CTLFLAG_RW,
1647             &bbr_persist_min, 250000,
1648             "What is the minimum time in microseconds between persists");
1649         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1650             SYSCTL_CHILDREN(bbr_timeout),
1651             OID_AUTO, "persmax", CTLFLAG_RW,
1652             &bbr_persist_max, 1000000,
1653             "What is the largest delay in microseconds between persists");
1654         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1655             SYSCTL_CHILDREN(bbr_timeout),
1656             OID_AUTO, "tlp_minto", CTLFLAG_RW,
1657             &bbr_tlp_min, 10000,
1658             "TLP Min timeout in usecs");
1659         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1660             SYSCTL_CHILDREN(bbr_timeout),
1661             OID_AUTO, "tlp_dack_time", CTLFLAG_RW,
1662             &bbr_delayed_ack_time, 200000,
1663             "TLP delayed ack compensation value");
1664         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1665             SYSCTL_CHILDREN(bbr_sysctl_root),
1666             OID_AUTO, "minrto", CTLFLAG_RW,
1667             &bbr_rto_min_ms, 30,
1668             "Minimum RTO in ms");
1669         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1670             SYSCTL_CHILDREN(bbr_timeout),
1671             OID_AUTO, "maxrto", CTLFLAG_RW,
1672             &bbr_rto_max_sec, 4,
1673             "Maxiumum RTO in seconds -- should be at least as large as min_rto");
1674         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1675             SYSCTL_CHILDREN(bbr_timeout),
1676             OID_AUTO, "tlp_retry", CTLFLAG_RW,
1677             &bbr_tlp_max_resend, 2,
1678             "How many times does TLP retry a single segment or multiple with no ACK");
1679         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1680             SYSCTL_CHILDREN(bbr_timeout),
1681             OID_AUTO, "minto", CTLFLAG_RW,
1682             &bbr_min_to, 1000,
1683             "Minimum rack timeout in useconds");
1684         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1685             SYSCTL_CHILDREN(bbr_timeout),
1686             OID_AUTO, "pktdelay", CTLFLAG_RW,
1687             &bbr_pkt_delay, 1000,
1688             "Extra RACK time (in useconds) besides reordering thresh");
1689         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1690             SYSCTL_CHILDREN(bbr_timeout),
1691             OID_AUTO, "incr_tmrs", CTLFLAG_RW,
1692             &bbr_incr_timers, 1,
1693             "Increase the RXT/TLP timer by the pacing time used?");
1694         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1695             SYSCTL_CHILDREN(bbr_timeout),
1696             OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW,
1697             &bbr_marks_rxt_sack_passed, 0,
1698             "Mark sack passed on all those not ack'd when a RXT hits?");
1699         /* Policer controls */
1700         bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1701             SYSCTL_CHILDREN(bbr_sysctl_root),
1702             OID_AUTO,
1703             "policer",
1704             CTLFLAG_RW, 0,
1705             "Policer controls");
1706         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1707             SYSCTL_CHILDREN(bbr_policer),
1708             OID_AUTO, "detect_enable", CTLFLAG_RW,
1709             &bbr_policer_detection_enabled, 1,
1710             "Is policer detection enabled??");
1711         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1712             SYSCTL_CHILDREN(bbr_policer),
1713             OID_AUTO, "min_pes", CTLFLAG_RW,
1714             &bbr_lt_intvl_min_rtts, 4,
1715             "Minimum number of PE's?");
1716         SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1717             SYSCTL_CHILDREN(bbr_policer),
1718             OID_AUTO, "bwdiff", CTLFLAG_RW,
1719             &bbr_lt_bw_diff, (4000/8),
1720             "Minimal bw diff?");
1721         SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1722             SYSCTL_CHILDREN(bbr_policer),
1723             OID_AUTO, "bwratio", CTLFLAG_RW,
1724             &bbr_lt_bw_ratio, 8,
1725             "Minimal bw diff?");
1726         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1727             SYSCTL_CHILDREN(bbr_policer),
1728             OID_AUTO, "from_rack_rxt", CTLFLAG_RW,
1729             &bbr_policer_call_from_rack_to, 0,
1730             "Do we call the policer detection code from a rack-timeout?");
1731         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1732             SYSCTL_CHILDREN(bbr_policer),
1733             OID_AUTO, "false_postive", CTLFLAG_RW,
1734             &bbr_lt_intvl_fp, 0,
1735             "What packet epoch do we do false-postive detection at (0=no)?");
1736         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1737             SYSCTL_CHILDREN(bbr_policer),
1738             OID_AUTO, "loss_thresh", CTLFLAG_RW,
1739             &bbr_lt_loss_thresh, 196,
1740             "Loss threshold 196 = 19.6%?");
1741         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1742             SYSCTL_CHILDREN(bbr_policer),
1743             OID_AUTO, "false_postive_thresh", CTLFLAG_RW,
1744             &bbr_lt_fd_thresh, 100,
1745             "What percentage is the false detection threshold (150=15.0)?");
1746         /* All the rest */
1747         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1748             SYSCTL_CHILDREN(bbr_sysctl_root),
1749             OID_AUTO, "cheat_rxt", CTLFLAG_RW,
1750             &bbr_use_rack_resend_cheat, 0,
1751             "Do we burst 1ms between sends on retransmissions (like rack)?");
1752         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1753             SYSCTL_CHILDREN(bbr_sysctl_root),
1754             OID_AUTO, "error_paceout", CTLFLAG_RW,
1755             &bbr_error_base_paceout, 10000,
1756             "When we hit an error what is the min to pace out in usec's?");
1757         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1758             SYSCTL_CHILDREN(bbr_sysctl_root),
1759             OID_AUTO, "kill_paceout", CTLFLAG_RW,
1760             &bbr_max_net_error_cnt, 10,
1761             "When we hit this many errors in a row, kill the session?");
1762         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1763             SYSCTL_CHILDREN(bbr_sysctl_root),
1764             OID_AUTO, "data_after_close", CTLFLAG_RW,
1765             &bbr_ignore_data_after_close, 1,
1766             "Do we hold off sending a RST until all pending data is ack'd");
1767         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1768             SYSCTL_CHILDREN(bbr_sysctl_root),
1769             OID_AUTO, "resend_use_tso", CTLFLAG_RW,
1770             &bbr_resends_use_tso, 0,
1771             "Can resends use TSO?");
1772         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1773             SYSCTL_CHILDREN(bbr_sysctl_root),
1774             OID_AUTO, "sblklimit", CTLFLAG_RW,
1775             &bbr_sack_block_limit, 128,
1776             "When do we start ignoring small sack blocks");
1777         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1778             SYSCTL_CHILDREN(bbr_sysctl_root),
1779             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1780             &bbr_verbose_logging, 0,
1781             "Should BBR black box logging be verbose");
1782         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1783             SYSCTL_CHILDREN(bbr_sysctl_root),
1784             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1785             &bbr_reorder_thresh, 2,
1786             "What factor for rack will be added when seeing reordering (shift right)");
1787         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1788             SYSCTL_CHILDREN(bbr_sysctl_root),
1789             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1790             &bbr_reorder_fade, 0,
1791             "Does reorder detection fade, if so how many ms (0 means never)");
1792         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1793             SYSCTL_CHILDREN(bbr_sysctl_root),
1794             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1795             &bbr_tlp_thresh, 1,
1796             "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1797         /* Stats and counters */
1798         /* The pacing counters for hdwr/software can't be in the array */
1799         bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1800         bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1801         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1802             SYSCTL_CHILDREN(bbr_sysctl_root),
1803             OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD,
1804             &bbr_hdwr_pacing_enobuf,
1805             "Total number of enobufs for hardware paced flows");
1806         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1807             SYSCTL_CHILDREN(bbr_sysctl_root),
1808             OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD,
1809             &bbr_nohdwr_pacing_enobuf,
1810             "Total number of enobufs for non-hardware paced flows");
1811
1812
1813         bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK);
1814         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1815             SYSCTL_CHILDREN(bbr_sysctl_root),
1816             OID_AUTO, "hdwr_pacing", CTLFLAG_RD,
1817             &bbr_flows_whdwr_pacing,
1818             "Total number of hardware paced flows");
1819         bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK);
1820         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1821             SYSCTL_CHILDREN(bbr_sysctl_root),
1822             OID_AUTO, "software_pacing", CTLFLAG_RD,
1823             &bbr_flows_nohdwr_pacing,
1824             "Total number of software paced flows");
1825         COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK);
1826         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1827             OID_AUTO, "stats", CTLFLAG_RD,
1828             bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats");
1829         COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK);
1830         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1831             OID_AUTO, "opts", CTLFLAG_RD,
1832             bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats");
1833         COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK);
1834         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1835             OID_AUTO, "lost", CTLFLAG_RD,
1836             bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur");
1837         COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK);
1838         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1839             OID_AUTO, "stateresend", CTLFLAG_RD,
1840             bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend");
1841         COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK);
1842         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1843             OID_AUTO, "statetime", CTLFLAG_RD,
1844             bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states");
1845         COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1846         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1847             OID_AUTO, "outsize", CTLFLAG_RD,
1848             bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls");
1849         SYSCTL_ADD_PROC(&bbr_sysctl_ctx,
1850             SYSCTL_CHILDREN(bbr_sysctl_root),
1851             OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1852             &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters");
1853 }
1854
1855 static inline int32_t
1856 bbr_progress_timeout_check(struct tcp_bbr *bbr)
1857 {
1858         if (bbr->rc_tp->t_maxunacktime && bbr->rc_tp->t_acktime &&
1859             TSTMP_GT(ticks, bbr->rc_tp->t_acktime)) {
1860                 if ((((uint32_t)ticks - bbr->rc_tp->t_acktime)) >= bbr->rc_tp->t_maxunacktime) {
1861                         /*
1862                          * There is an assumption here that the caller will
1863                          * drop the connection, so we increment the
1864                          * statistics.
1865                          */
1866                         bbr_log_progress_event(bbr, bbr->rc_tp, ticks, PROGRESS_DROP, __LINE__);
1867                         BBR_STAT_INC(bbr_progress_drops);
1868 #ifdef NETFLIX_STATS
1869                         TCPSTAT_INC(tcps_progdrops);
1870 #endif
1871                         return (1);
1872                 }
1873         }
1874         return (0);
1875 }
1876
1877 static void
1878 bbr_counter_destroy(void)
1879 {
1880         COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE);
1881         COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE);
1882         COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE);
1883         COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT);
1884         COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT);
1885         COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT);
1886         counter_u64_free(bbr_flows_whdwr_pacing);
1887         counter_u64_free(bbr_flows_nohdwr_pacing);
1888
1889 }
1890
1891 static __inline void
1892 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts)
1893 {
1894         memset(l, 0, sizeof(union tcp_log_stackspecific));
1895         l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate;
1896         l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate);
1897         l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
1898         l->bw_inuse = bbr_get_bw(bbr);
1899         l->inflight = ctf_flight_size(bbr->rc_tp,
1900                           (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
1901         l->applimited = bbr->r_ctl.r_app_limited_until;
1902         l->delivered = bbr->r_ctl.rc_delivered;
1903         l->timeStamp = cts;
1904         l->lost = bbr->r_ctl.rc_lost;
1905         l->bbr_state = bbr->rc_bbr_state;
1906         l->bbr_substate = bbr_state_val(bbr);
1907         l->epoch = bbr->r_ctl.rc_rtt_epoch;
1908         l->lt_epoch = bbr->r_ctl.rc_lt_epoch;
1909         l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
1910         l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain;
1911         l->inhpts = bbr->rc_inp->inp_in_hpts;
1912         l->ininput = bbr->rc_inp->inp_in_input;
1913         l->use_lt_bw = bbr->rc_lt_use_bw;
1914         l->pkts_out = bbr->r_ctl.rc_flight_at_input;
1915         l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch;
1916 }
1917
1918 static void
1919 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason)
1920 {
1921         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1922                 union tcp_log_stackspecific log;
1923
1924                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1925                 log.u_bbr.flex1 = 0;
1926                 log.u_bbr.flex2 = 0;
1927                 log.u_bbr.flex5 = 0;
1928                 log.u_bbr.flex3 = 0;
1929                 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate;
1930                 log.u_bbr.flex7 = reason;
1931                 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt;
1932                 log.u_bbr.flex8 = 0;
1933                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1934                     &bbr->rc_inp->inp_socket->so_rcv,
1935                     &bbr->rc_inp->inp_socket->so_snd,
1936                     BBR_LOG_BW_RED_EV, 0,
1937                     0, &log, false, &bbr->rc_tv);
1938         }
1939 }
1940
1941 static void
1942 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count)
1943 {
1944         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1945                 union tcp_log_stackspecific log;
1946
1947                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1948                 log.u_bbr.flex1 = seq;
1949                 log.u_bbr.flex2 = count;
1950                 log.u_bbr.flex8 = mode;
1951                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1952                     &bbr->rc_inp->inp_socket->so_rcv,
1953                     &bbr->rc_inp->inp_socket->so_snd,
1954                     BBR_LOG_LOWGAIN, 0,
1955                     0, &log, false, &bbr->rc_tv);
1956         }
1957 }
1958
1959
1960
1961 static void
1962 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling,
1963     uint8_t reason, uint32_t p_maxseg, int len)
1964 {
1965         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1966                 union tcp_log_stackspecific log;
1967
1968                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1969                 log.u_bbr.flex1 = p_maxseg;
1970                 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags;
1971                 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
1972                 log.u_bbr.flex4 = reason;
1973                 log.u_bbr.flex5 = bbr->rc_in_persist;
1974                 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val;
1975                 log.u_bbr.flex7 = p_maxseg;
1976                 log.u_bbr.flex8 = bbr->rc_in_persist;
1977                 log.u_bbr.pkts_out = 0;
1978                 log.u_bbr.applimited = len;
1979                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1980                     &bbr->rc_inp->inp_socket->so_rcv,
1981                     &bbr->rc_inp->inp_socket->so_snd,
1982                     BBR_LOG_JUSTRET, 0,
1983                     tlen, &log, false, &bbr->rc_tv);
1984         }
1985 }
1986
1987
1988 static void
1989 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq)
1990 {
1991         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1992                 union tcp_log_stackspecific log;
1993
1994                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1995                 log.u_bbr.flex1 = seq;
1996                 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
1997                 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start;
1998                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1999                     &bbr->rc_inp->inp_socket->so_rcv,
2000                     &bbr->rc_inp->inp_socket->so_snd,
2001                     BBR_LOG_ENTREC, 0,
2002                     0, &log, false, &bbr->rc_tv);
2003         }
2004 }
2005
2006 static void
2007 bbr_log_msgsize_fail(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t len, uint32_t maxseg, uint32_t mtu, int32_t csum_flags, int32_t tso, uint32_t cts)
2008 {
2009         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2010                 union tcp_log_stackspecific log;
2011
2012                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2013                 log.u_bbr.flex1 = tso;
2014                 log.u_bbr.flex2 = maxseg;
2015                 log.u_bbr.flex3 = mtu;
2016                 log.u_bbr.flex4 = csum_flags;
2017                 TCP_LOG_EVENTP(tp, NULL,
2018                     &bbr->rc_inp->inp_socket->so_rcv,
2019                     &bbr->rc_inp->inp_socket->so_snd,
2020                     BBR_LOG_MSGSIZE, 0,
2021                     0, &log, false, &bbr->rc_tv);
2022         }
2023 }
2024
2025 static void
2026 bbr_log_flowend(struct tcp_bbr *bbr)
2027 {
2028         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2029                 union tcp_log_stackspecific log;
2030                 struct sockbuf *r, *s;
2031                 struct timeval tv;
2032
2033                 if (bbr->rc_inp->inp_socket) {
2034                         r = &bbr->rc_inp->inp_socket->so_rcv;
2035                         s = &bbr->rc_inp->inp_socket->so_snd;
2036                 } else {
2037                         r = s = NULL;
2038                 }
2039                 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv));
2040                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2041                     r, s,
2042                     TCP_LOG_FLOWEND, 0,
2043                     0, &log, false, &tv);
2044         }
2045 }
2046
2047 static void
2048 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line,
2049     uint32_t lost, uint32_t del)
2050 {
2051         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2052                 union tcp_log_stackspecific log;
2053
2054                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2055                 log.u_bbr.flex1 = lost;
2056                 log.u_bbr.flex2 = del;
2057                 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw;
2058                 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt;
2059                 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2060                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2061                 log.u_bbr.flex7 = line;
2062                 log.u_bbr.flex8 = 0;
2063                 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count;
2064                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2065                     &bbr->rc_inp->inp_socket->so_rcv,
2066                     &bbr->rc_inp->inp_socket->so_snd,
2067                     BBR_LOG_PKT_EPOCH, 0,
2068                     0, &log, false, &bbr->rc_tv);
2069         }
2070 }
2071
2072 static void
2073 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time)
2074 {
2075         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2076                 union tcp_log_stackspecific log;
2077
2078                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2079                 log.u_bbr.flex1 = bbr->r_ctl.rc_lost;
2080                 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat;
2081                 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat;
2082                 log.u_bbr.flex7 = line;
2083                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2084                     &bbr->rc_inp->inp_socket->so_rcv,
2085                     &bbr->rc_inp->inp_socket->so_snd,
2086                     BBR_LOG_TIME_EPOCH, 0,
2087                     0, &log, false, &bbr->rc_tv);
2088         }
2089 }
2090
2091 static void
2092 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth)
2093 {
2094         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2095                 union tcp_log_stackspecific log;
2096
2097                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2098                 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2099                 log.u_bbr.flex2 = new_tar;
2100                 log.u_bbr.flex3 = line;
2101                 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2102                 log.u_bbr.flex5 = bbr_quanta;
2103                 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs;
2104                 log.u_bbr.flex7 = bbr->rc_last_options;
2105                 log.u_bbr.flex8 = meth;
2106                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2107                     &bbr->rc_inp->inp_socket->so_rcv,
2108                     &bbr->rc_inp->inp_socket->so_snd,
2109                     BBR_LOG_STATE_TARGET, 0,
2110                     0, &log, false, &bbr->rc_tv);
2111         }
2112
2113 }
2114
2115 static void
2116 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2117 {
2118         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2119                 union tcp_log_stackspecific log;
2120
2121                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2122                 log.u_bbr.flex1 = line;
2123                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2124                 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2125                 if (bbr_state_is_pkt_epoch)
2126                         log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
2127                 else
2128                         log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP);
2129                 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2130                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2131                 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000);
2132                 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra;
2133                 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state;
2134                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2135                     &bbr->rc_inp->inp_socket->so_rcv,
2136                     &bbr->rc_inp->inp_socket->so_snd,
2137                     BBR_LOG_STATE, 0,
2138                     0, &log, false, &bbr->rc_tv);
2139         }
2140 }
2141
2142 static void
2143 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied,
2144                     uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond)
2145 {
2146         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2147                 union tcp_log_stackspecific log;
2148
2149                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2150                 log.u_bbr.flex1 = line;
2151                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2152                 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt;
2153                 log.u_bbr.flex4 = applied;
2154                 log.u_bbr.flex5 = rtt;
2155                 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2156                 log.u_bbr.flex7 = cond;
2157                 log.u_bbr.flex8 = reas;
2158                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2159                     &bbr->rc_inp->inp_socket->so_rcv,
2160                     &bbr->rc_inp->inp_socket->so_snd,
2161                     BBR_LOG_RTT_SHRINKS, 0,
2162                     0, &log, false, &bbr->rc_tv);
2163         }
2164 }
2165
2166 static void
2167 bbr_log_type_exit_rec(struct tcp_bbr *bbr)
2168 {
2169         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2170                 union tcp_log_stackspecific log;
2171
2172                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2173                 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start;
2174                 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
2175                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2176                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2177                     &bbr->rc_inp->inp_socket->so_rcv,
2178                     &bbr->rc_inp->inp_socket->so_snd,
2179                     BBR_LOG_EXITREC, 0,
2180                     0, &log, false, &bbr->rc_tv);
2181         }
2182 }
2183
2184 static void
2185 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg,
2186     uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line)
2187 {
2188         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2189                 union tcp_log_stackspecific log;
2190
2191                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2192                 log.u_bbr.flex1 = line;
2193                 log.u_bbr.flex2 = prev_acked;
2194                 log.u_bbr.flex3 = bytes_this_ack;
2195                 log.u_bbr.flex4 = chg;
2196                 log.u_bbr.flex5 = th_ack;
2197                 log.u_bbr.flex6 = target;
2198                 log.u_bbr.flex8 = meth;
2199                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2200                     &bbr->rc_inp->inp_socket->so_rcv,
2201                     &bbr->rc_inp->inp_socket->so_snd,
2202                     BBR_LOG_CWND, 0,
2203                     0, &log, false, &bbr->rc_tv);
2204         }
2205 }
2206
2207 static void
2208 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin)
2209 {
2210         /*
2211          * Log the rtt sample we are applying to the srtt algorithm in
2212          * useconds.
2213          */
2214         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2215                 union tcp_log_stackspecific log;
2216
2217                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2218                 log.u_bbr.flex1 = rtt;
2219                 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time;
2220                 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay;
2221                 log.u_bbr.flex4 = bbr->rc_tp->ts_offset;
2222                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2223                 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv);
2224                 log.u_bbr.flex6 = tsin;
2225                 log.u_bbr.flex7 = 0;
2226                 log.u_bbr.flex8 = bbr->rc_ack_was_delayed;
2227                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2228                     &bbr->rc_inp->inp_socket->so_rcv,
2229                     &bbr->rc_inp->inp_socket->so_snd,
2230                     TCP_LOG_RTT, 0,
2231                     0, &log, false, &bbr->rc_tv);
2232         }
2233 }
2234
2235 static void
2236 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit)
2237 {
2238         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2239                 union tcp_log_stackspecific log;
2240
2241                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2242                 log.u_bbr.flex1 = time_in;
2243                 log.u_bbr.flex2 = line;
2244                 log.u_bbr.flex8 = enter_exit;
2245                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2246                     &bbr->rc_inp->inp_socket->so_rcv,
2247                     &bbr->rc_inp->inp_socket->so_snd,
2248                     BBR_LOG_PERSIST, 0,
2249                     0, &log, false, &bbr->rc_tv);
2250         }
2251 }
2252 static void
2253 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts)
2254 {
2255         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2256                 union tcp_log_stackspecific log;
2257
2258                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2259                 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age;
2260                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2261                 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2262                 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time;
2263                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2264                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2265                     &bbr->rc_inp->inp_socket->so_rcv,
2266                     &bbr->rc_inp->inp_socket->so_snd,
2267                     BBR_LOG_ACKCLEAR, 0,
2268                     0, &log, false, &bbr->rc_tv);
2269         }
2270 }
2271
2272 static void
2273 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen,
2274                   uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m)
2275 {
2276         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2277                 union tcp_log_stackspecific log;
2278                 struct timeval tv;
2279
2280                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2281                 log.u_bbr.flex1 = nsegs;
2282                 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes;
2283                 if (m) {
2284                         struct timespec ts;
2285
2286                         log.u_bbr.flex3 = m->m_flags;
2287                         if (m->m_flags & M_TSTMP) {
2288                                 mbuf_tstmp2timespec(m, &ts);
2289                                 tv.tv_sec = ts.tv_sec;
2290                                 tv.tv_usec = ts.tv_nsec / 1000;
2291                                 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv);
2292                         } else {
2293                                 log.u_bbr.lt_epoch = 0;
2294                         }
2295                         if (m->m_flags & M_TSTMP_LRO) {
2296                                 tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000;
2297                                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000;
2298                                 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv);
2299                         } else {
2300                                 /* No arrival timestamp */
2301                                 log.u_bbr.flex5 = 0;
2302                         }
2303
2304                         log.u_bbr.pkts_out = tcp_get_usecs(&tv);
2305                 } else {
2306                         log.u_bbr.flex3 = 0;
2307                         log.u_bbr.flex5 = 0;
2308                         log.u_bbr.flex6 = 0;
2309                         log.u_bbr.pkts_out = 0;
2310                 }
2311                 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state;
2312                 log.u_bbr.flex7 = bbr->r_wanted_output;
2313                 log.u_bbr.flex8 = bbr->rc_in_persist;
2314                 TCP_LOG_EVENTP(bbr->rc_tp, th,
2315                     &bbr->rc_inp->inp_socket->so_rcv,
2316                     &bbr->rc_inp->inp_socket->so_snd,
2317                     TCP_LOG_IN, 0,
2318                     tlen, &log, true, &bbr->rc_tv);
2319         }
2320 }
2321
2322 static void
2323 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out)
2324 {
2325         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2326                 union tcp_log_stackspecific log;
2327
2328                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2329                 log.u_bbr.flex1 = did_out;
2330                 log.u_bbr.flex2 = nxt_pkt;
2331                 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val;
2332                 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2333                 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp;
2334                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes;
2335                 log.u_bbr.flex7 = bbr->r_wanted_output;
2336                 log.u_bbr.flex8 = bbr->rc_in_persist;
2337                 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay;
2338                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2339                     &bbr->rc_inp->inp_socket->so_rcv,
2340                     &bbr->rc_inp->inp_socket->so_snd,
2341                     BBR_LOG_DOSEG_DONE, 0,
2342                     0, &log, true, &bbr->rc_tv);
2343         }
2344 }
2345
2346 static void
2347 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts,
2348     int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz)
2349 {
2350         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2351                 union tcp_log_stackspecific log;
2352
2353                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2354                 log.u_bbr.flex1 = line;
2355                 log.u_bbr.flex2 = o_len;
2356                 log.u_bbr.flex3 = segcnt;
2357                 log.u_bbr.flex4 = segsiz;
2358                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2359                     &bbr->rc_inp->inp_socket->so_rcv,
2360                     &bbr->rc_inp->inp_socket->so_snd,
2361                     BBR_LOG_ENOBUF_JMP, ENOBUFS,
2362                     len, &log, true, &bbr->rc_tv);
2363         }
2364 }
2365
2366 static void
2367 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling)
2368 {
2369         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2370                 union tcp_log_stackspecific log;
2371
2372                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2373                 log.u_bbr.flex1 = timers;
2374                 log.u_bbr.flex2 = ret;
2375                 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
2376                 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2377                 log.u_bbr.flex5 = cts;
2378                 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2379                 log.u_bbr.flex8 = hpts_calling;
2380                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2381                     &bbr->rc_inp->inp_socket->so_rcv,
2382                     &bbr->rc_inp->inp_socket->so_snd,
2383                     BBR_LOG_TO_PROCESS, 0,
2384                     0, &log, false, &bbr->rc_tv);
2385         }
2386 }
2387
2388 static void
2389 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num)
2390 {
2391         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2392                 union tcp_log_stackspecific log;
2393                 uint64_t ar;
2394
2395                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2396                 log.u_bbr.flex1 = bbr->bbr_timer_src;
2397                 log.u_bbr.flex2 = 0;
2398                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2399                 ar = (uint64_t)(bbr->r_ctl.rc_resend);
2400                 ar >>= 32;
2401                 ar &= 0x00000000ffffffff;
2402                 log.u_bbr.flex4 = (uint32_t)ar;
2403                 ar = (uint64_t)bbr->r_ctl.rc_resend;
2404                 ar &= 0x00000000ffffffff;
2405                 log.u_bbr.flex5 = (uint32_t)ar;
2406                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2407                 log.u_bbr.flex8 = to_num;
2408                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2409                     &bbr->rc_inp->inp_socket->so_rcv,
2410                     &bbr->rc_inp->inp_socket->so_snd,
2411                     BBR_LOG_RTO, 0,
2412                     0, &log, false, &bbr->rc_tv);
2413         }
2414 }
2415
2416 static void
2417 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason)
2418 {
2419         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2420                 union tcp_log_stackspecific log;
2421
2422                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2423                 log.u_bbr.flex1 = flex1;
2424                 log.u_bbr.flex2 = flex2;
2425                 log.u_bbr.flex3 = flex3;
2426                 log.u_bbr.flex4 = 0;
2427                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2428                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2429                 log.u_bbr.flex8 = reason;
2430                 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw;
2431                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2432                     &bbr->rc_inp->inp_socket->so_rcv,
2433                     &bbr->rc_inp->inp_socket->so_snd,
2434                     BBR_LOG_REDUCE, 0,
2435                     0, &log, false, &bbr->rc_tv);
2436         }
2437 }
2438
2439 static void
2440 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag)
2441 {
2442         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2443                 union tcp_log_stackspecific log;
2444
2445                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2446                 log.u_bbr.flex1 = diag->p_nxt_slot;
2447                 log.u_bbr.flex2 = diag->p_cur_slot;
2448                 log.u_bbr.flex3 = diag->slot_req;
2449                 log.u_bbr.flex4 = diag->inp_hptsslot;
2450                 log.u_bbr.flex5 = diag->slot_remaining;
2451                 log.u_bbr.flex6 = diag->need_new_to;
2452                 log.u_bbr.flex7 = diag->p_hpts_active;
2453                 log.u_bbr.flex8 = diag->p_on_min_sleep;
2454                 /* Hijack other fields as needed  */
2455                 log.u_bbr.epoch = diag->have_slept;
2456                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
2457                 log.u_bbr.pkts_out = diag->co_ret;
2458                 log.u_bbr.applimited = diag->hpts_sleep_time;
2459                 log.u_bbr.delivered = diag->p_prev_slot;
2460                 log.u_bbr.inflight = diag->p_runningtick;
2461                 log.u_bbr.bw_inuse = diag->wheel_tick;
2462                 log.u_bbr.rttProp = diag->wheel_cts;
2463                 log.u_bbr.delRate = diag->maxticks;
2464                 log.u_bbr.cur_del_rate = diag->p_curtick;
2465                 log.u_bbr.cur_del_rate <<= 32;
2466                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
2467                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2468                     &bbr->rc_inp->inp_socket->so_rcv,
2469                     &bbr->rc_inp->inp_socket->so_snd,
2470                     BBR_LOG_HPTSDIAG, 0,
2471                     0, &log, false, &bbr->rc_tv);
2472         }
2473 }
2474
2475 static void
2476 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
2477     uint32_t thresh, uint32_t to)
2478 {
2479         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2480                 union tcp_log_stackspecific log;
2481
2482                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2483                 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar;
2484                 log.u_bbr.flex2 = time_since_sent;
2485                 log.u_bbr.flex3 = srtt;
2486                 log.u_bbr.flex4 = thresh;
2487                 log.u_bbr.flex5 = to;
2488                 log.u_bbr.flex6 = bbr->rc_tp->t_srtt;
2489                 log.u_bbr.flex8 = mode;
2490                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2491                     &bbr->rc_inp->inp_socket->so_rcv,
2492                     &bbr->rc_inp->inp_socket->so_snd,
2493                     BBR_LOG_TIMERPREP, 0,
2494                     0, &log, false, &bbr->rc_tv);
2495         }
2496 }
2497
2498 static void
2499 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
2500     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod)
2501 {
2502         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2503                 union tcp_log_stackspecific log;
2504
2505                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2506                 log.u_bbr.flex1 = usecs;
2507                 log.u_bbr.flex2 = len;
2508                 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff);
2509                 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff);
2510                 if (override)
2511                         log.u_bbr.flex5 = (1 << 2);
2512                 else
2513                         log.u_bbr.flex5 = 0;
2514                 log.u_bbr.flex6 = override;
2515                 log.u_bbr.flex7 = gain;
2516                 log.u_bbr.flex8 = mod;
2517                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2518                     &bbr->rc_inp->inp_socket->so_rcv,
2519                     &bbr->rc_inp->inp_socket->so_snd,
2520                     BBR_LOG_HPTSI_CALC, 0,
2521                     len, &log, false, &bbr->rc_tv);
2522         }
2523 }
2524
2525 static void
2526 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2527 {
2528         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2529                 union tcp_log_stackspecific log;
2530
2531                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2532
2533                 log.u_bbr.flex1 = bbr->bbr_timer_src;
2534                 log.u_bbr.flex2 = to;
2535                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2536                 log.u_bbr.flex4 = slot;
2537                 log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot;
2538                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2539                 log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2;
2540                 log.u_bbr.flex8 = which;
2541                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2542                     &bbr->rc_inp->inp_socket->so_rcv,
2543                     &bbr->rc_inp->inp_socket->so_snd,
2544                     BBR_LOG_TIMERSTAR, 0,
2545                     0, &log, false, &bbr->rc_tv);
2546         }
2547 }
2548
2549 static void
2550 bbr_log_thresh_choice(struct tcp_bbr *bbr, uint32_t cts, uint32_t thresh, uint32_t lro, uint32_t srtt, struct bbr_sendmap *rsm, uint8_t frm)
2551 {
2552         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2553                 union tcp_log_stackspecific log;
2554
2555                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2556                 log.u_bbr.flex1 = thresh;
2557                 log.u_bbr.flex2 = lro;
2558                 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts;
2559                 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
2560                 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2561                 log.u_bbr.flex6 = srtt;
2562                 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift;
2563                 log.u_bbr.flex8 = frm;
2564                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2565                     &bbr->rc_inp->inp_socket->so_rcv,
2566                     &bbr->rc_inp->inp_socket->so_snd,
2567                     BBR_LOG_THRESH_CALC, 0,
2568                     0, &log, false, &bbr->rc_tv);
2569         }
2570 }
2571
2572 static void
2573 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed)
2574 {
2575         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2576                 union tcp_log_stackspecific log;
2577
2578                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2579                 log.u_bbr.flex1 = line;
2580                 log.u_bbr.flex2 = bbr->bbr_timer_src;
2581                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2582                 log.u_bbr.flex4 = bbr->rc_in_persist;
2583                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2584                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2585                 log.u_bbr.flex8 = hpts_removed;
2586                 log.u_bbr.pkts_out = bbr->rc_pacer_started;
2587                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2588                     &bbr->rc_inp->inp_socket->so_rcv,
2589                     &bbr->rc_inp->inp_socket->so_snd,
2590                     BBR_LOG_TIMERCANC, 0,
2591                     0, &log, false, &bbr->rc_tv);
2592         }
2593 }
2594
2595
2596 static void
2597 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta)
2598 {
2599         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2600                 union tcp_log_stackspecific log;
2601
2602                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2603                 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio;
2604                 log.u_bbr.flex2 = (peer_delta >> 32);
2605                 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff);
2606                 log.u_bbr.flex4 = (delta >> 32);
2607                 log.u_bbr.flex5 = (delta & 0x00000000ffffffff);
2608                 log.u_bbr.flex7 = bbr->rc_ts_clock_set;
2609                 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used;
2610                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2611                     &bbr->rc_inp->inp_socket->so_rcv,
2612                     &bbr->rc_inp->inp_socket->so_snd,
2613                     BBR_LOG_TSTMP_VAL, 0,
2614                     0, &log, false, &bbr->rc_tv);
2615
2616         }
2617 }
2618
2619 static void
2620 bbr_log_type_tsosize(struct tcp_bbr *bbr, uint32_t cts, uint32_t tsosz, uint32_t tls, uint32_t old_val, uint32_t maxseg, int hdwr)
2621 {
2622         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2623                 union tcp_log_stackspecific log;
2624
2625                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2626                 log.u_bbr.flex1 = tsosz;
2627                 log.u_bbr.flex2 = tls;
2628                 log.u_bbr.flex3 = tcp_min_hptsi_time;
2629                 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min;
2630                 log.u_bbr.flex5 = old_val;
2631                 log.u_bbr.flex6 = maxseg;
2632                 log.u_bbr.flex7 = bbr->rc_no_pacing;
2633                 log.u_bbr.flex7 <<= 1;
2634                 log.u_bbr.flex7 |= bbr->rc_past_init_win;
2635                 if (hdwr)
2636                         log.u_bbr.flex8 = 0x80 | bbr->rc_use_google;
2637                 else
2638                         log.u_bbr.flex8 = bbr->rc_use_google;
2639                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2640                     &bbr->rc_inp->inp_socket->so_rcv,
2641                     &bbr->rc_inp->inp_socket->so_snd,
2642                     BBR_LOG_BBRTSO, 0,
2643                     0, &log, false, &bbr->rc_tv);
2644         }
2645 }
2646
2647 static void
2648 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm,
2649                       uint32_t flags, uint32_t line)
2650 {
2651         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2652                 union tcp_log_stackspecific log;
2653
2654                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2655                 log.u_bbr.flex1 = line;
2656                 log.u_bbr.flex2 = rsm->r_start;
2657                 log.u_bbr.flex3 = rsm->r_end;
2658                 log.u_bbr.flex4 = rsm->r_delivered;
2659                 log.u_bbr.flex5 = rsm->r_rtr_cnt;
2660                 log.u_bbr.flex6 = rsm->r_dupack;
2661                 log.u_bbr.flex7 = rsm->r_tim_lastsent[0];
2662                 log.u_bbr.flex8 = rsm->r_flags;
2663                 /* Hijack the pkts_out fids */
2664                 log.u_bbr.applimited = flags;
2665                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2666                     &bbr->rc_inp->inp_socket->so_rcv,
2667                     &bbr->rc_inp->inp_socket->so_snd,
2668                     BBR_RSM_CLEARED, 0,
2669                     0, &log, false, &bbr->rc_tv);
2670         }
2671 }
2672
2673 static void
2674 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts,
2675     uint32_t flex3, uint32_t flex2, uint32_t flex5,
2676     uint32_t flex6, uint32_t pkts_out, int flex7,
2677     uint32_t flex4, uint32_t flex1)
2678 {
2679
2680         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2681                 union tcp_log_stackspecific log;
2682
2683                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2684                 log.u_bbr.flex1 = flex1;
2685                 log.u_bbr.flex2 = flex2;
2686                 log.u_bbr.flex3 = flex3;
2687                 log.u_bbr.flex4 = flex4;
2688                 log.u_bbr.flex5 = flex5;
2689                 log.u_bbr.flex6 = flex6;
2690                 log.u_bbr.flex7 = flex7;
2691                 /* Hijack the pkts_out fids */
2692                 log.u_bbr.pkts_out = pkts_out;
2693                 log.u_bbr.flex8 = flex8;
2694                 if (bbr->rc_ack_was_delayed)
2695                         log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay;
2696                 else
2697                         log.u_bbr.epoch = 0;
2698                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2699                     &bbr->rc_inp->inp_socket->so_rcv,
2700                     &bbr->rc_inp->inp_socket->so_snd,
2701                     BBR_LOG_BBRUPD, 0,
2702                     flex2, &log, false, &bbr->rc_tv);
2703         }
2704 }
2705
2706
2707 static void
2708 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason,
2709         uint32_t newbw, uint32_t obw, uint32_t diff,
2710         uint32_t tim)
2711 {
2712         if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2713                 union tcp_log_stackspecific log;
2714
2715                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2716                 log.u_bbr.flex1 = reason;
2717                 log.u_bbr.flex2 = newbw;
2718                 log.u_bbr.flex3 = obw;
2719                 log.u_bbr.flex4 = diff;
2720                 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost;
2721                 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del;
2722                 log.u_bbr.flex7 = bbr->rc_lt_is_sampling;
2723                 log.u_bbr.pkts_out = tim;
2724                 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw;
2725                 if (bbr->rc_lt_use_bw == 0)
2726                         log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
2727                 else
2728                         log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
2729                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2730                     &bbr->rc_inp->inp_socket->so_rcv,
2731                     &bbr->rc_inp->inp_socket->so_snd,
2732                     BBR_LOG_BWSAMP, 0,
2733                     0, &log, false, &bbr->rc_tv);
2734         }
2735 }
2736
2737 static inline void
2738 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line)
2739 {
2740         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2741                 union tcp_log_stackspecific log;
2742
2743                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2744                 log.u_bbr.flex1 = line;
2745                 log.u_bbr.flex2 = tick;
2746                 log.u_bbr.flex3 = tp->t_maxunacktime;
2747                 log.u_bbr.flex4 = tp->t_acktime;
2748                 log.u_bbr.flex8 = event;
2749                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2750                     &bbr->rc_inp->inp_socket->so_rcv,
2751                     &bbr->rc_inp->inp_socket->so_snd,
2752                     BBR_LOG_PROGRESS, 0,
2753                     0, &log, false, &bbr->rc_tv);
2754         }
2755 }
2756
2757 static void
2758 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp,
2759                          uint64_t rate, uint64_t hw_rate, int line, uint32_t cts,
2760                          int error)
2761 {
2762         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2763                 union tcp_log_stackspecific log;
2764
2765                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2766                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2767                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2768                 log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2769                 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2770                 log.u_bbr.bw_inuse = rate;
2771                 log.u_bbr.flex5 = line;
2772                 log.u_bbr.flex6 = error;
2773                 log.u_bbr.flex8 = bbr->skip_gain;
2774                 log.u_bbr.flex8 <<= 1;
2775                 log.u_bbr.flex8 |= bbr->gain_is_limited;
2776                 log.u_bbr.flex8 <<= 1;
2777                 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing;
2778                 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
2779                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2780                     &bbr->rc_inp->inp_socket->so_rcv,
2781                     &bbr->rc_inp->inp_socket->so_snd,
2782                     BBR_LOG_HDWR_PACE, 0,
2783                     0, &log, false, &bbr->rc_tv);
2784         }
2785 }
2786
2787 static void
2788 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, uint32_t del_by, uint32_t cts, uint32_t line, uint32_t prev_delay)
2789 {
2790         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2791                 union tcp_log_stackspecific log;
2792
2793                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2794                 log.u_bbr.flex1 = slot;
2795                 log.u_bbr.flex2 = del_by;
2796                 log.u_bbr.flex3 = prev_delay;
2797                 log.u_bbr.flex4 = line;
2798                 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val;
2799                 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay;
2800                 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags);
2801                 log.u_bbr.flex8 = bbr->rc_in_persist;
2802                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2803                     &bbr->rc_inp->inp_socket->so_rcv,
2804                     &bbr->rc_inp->inp_socket->so_snd,
2805                     BBR_LOG_BBRSND, 0,
2806                     len, &log, false, &bbr->rc_tv);
2807         }
2808 }
2809
2810 static void
2811 bbr_log_type_bbrrttprop(struct tcp_bbr *bbr, uint32_t t, uint32_t end, uint32_t tsconv, uint32_t cts, int32_t match, uint32_t seq, uint8_t flags)
2812 {
2813         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2814                 union tcp_log_stackspecific log;
2815
2816                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2817                 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered;
2818                 log.u_bbr.flex2 = 0;
2819                 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt;
2820                 log.u_bbr.flex4 = end;
2821                 log.u_bbr.flex5 = seq;
2822                 log.u_bbr.flex6 = t;
2823                 log.u_bbr.flex7 = match;
2824                 log.u_bbr.flex8 = flags;
2825                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2826                     &bbr->rc_inp->inp_socket->so_rcv,
2827                     &bbr->rc_inp->inp_socket->so_snd,
2828                     BBR_LOG_BBRRTT, 0,
2829                     0, &log, false, &bbr->rc_tv);
2830         }
2831 }
2832
2833 static void
2834 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method)
2835 {
2836         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2837                 union tcp_log_stackspecific log;
2838
2839                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2840                 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2841                 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
2842                 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch;
2843                 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2844                 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs;
2845                 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight;
2846                 log.u_bbr.flex7 = 0;
2847                 log.u_bbr.flex8 = entry_method;
2848                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2849                     &bbr->rc_inp->inp_socket->so_rcv,
2850                     &bbr->rc_inp->inp_socket->so_snd,
2851                     BBR_LOG_EXIT_GAIN, 0,
2852                     0, &log, false, &bbr->rc_tv);
2853         }
2854 }
2855
2856 static void
2857 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired)
2858 {
2859         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2860                 union tcp_log_stackspecific log;
2861
2862                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2863                 /* R-HU */
2864                 log.u_bbr.flex1 = 0;
2865                 log.u_bbr.flex2 = 0;
2866                 log.u_bbr.flex3 = 0;
2867                 log.u_bbr.flex4 = 0;
2868                 log.u_bbr.flex7 = 0;
2869                 log.u_bbr.flex8 = settings_desired;
2870
2871                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2872                     &bbr->rc_inp->inp_socket->so_rcv,
2873                     &bbr->rc_inp->inp_socket->so_snd,
2874                     BBR_LOG_SETTINGS_CHG, 0,
2875                     0, &log, false, &bbr->rc_tv);
2876         }
2877 }
2878
2879 /*
2880  * Returns the bw from the our filter.
2881  */
2882 static inline uint64_t
2883 bbr_get_full_bw(struct tcp_bbr *bbr)
2884 {
2885         uint64_t bw;
2886
2887         bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2888
2889         return (bw);
2890 }
2891
2892 static inline void
2893 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2894 {
2895         uint64_t calclr;
2896         uint32_t lost, del;
2897
2898         if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch)
2899                 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch;
2900         else
2901                 lost = 0;
2902         del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del;
2903         if (lost == 0)  {
2904                 calclr = 0;
2905         } else if (del) {
2906                 calclr = lost;
2907                 calclr *= (uint64_t)1000;
2908                 calclr /= (uint64_t)del;
2909         } else {
2910                 /* Nothing delivered? 100.0% loss */
2911                 calclr = 1000;
2912         }
2913         bbr->r_ctl.rc_pkt_epoch_loss_rate =  (uint32_t)calclr;
2914         if (IN_RECOVERY(bbr->rc_tp->t_flags))
2915                 bbr->r_ctl.recovery_lr += (uint32_t)calclr;
2916         bbr->r_ctl.rc_pkt_epoch++;
2917         if (bbr->rc_no_pacing &&
2918             (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) {
2919                 bbr->rc_no_pacing = 0;
2920                 tcp_bbr_tso_size_check(bbr, cts);
2921         }
2922         bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time);
2923         bbr->r_ctl.rc_pkt_epoch_time = cts;
2924         /* What was our loss rate */
2925         bbr_log_pkt_epoch(bbr, cts, line, lost, del);
2926         bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered;
2927         bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost;
2928 }
2929
2930 static inline void
2931 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2932 {
2933         uint32_t epoch_time;
2934
2935         /* Tick the RTT clock */
2936         bbr->r_ctl.rc_rtt_epoch++;
2937         epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start;
2938         bbr_log_time_epoch(bbr, cts, line, epoch_time);
2939         bbr->r_ctl.rc_rcv_epoch_start = cts;
2940 }
2941
2942
2943 static inline void
2944 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked)
2945 {
2946         if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) {
2947                 bbr->rc_is_pkt_epoch_now = 1;
2948         }
2949 }
2950
2951 /*
2952  * Returns the bw from either the b/w filter
2953  * or from the lt_bw (if the connection is being
2954  * policed).
2955  */
2956 static inline uint64_t
2957 __bbr_get_bw(struct tcp_bbr *bbr)
2958 {
2959         uint64_t bw, min_bw;
2960         uint64_t rtt;
2961         int gm_measure_cnt = 1;
2962
2963         /*
2964          * For startup we make, like google, a
2965          * minimum b/w. This is generated from the
2966          * IW and the rttProp. We do fall back to srtt
2967          * if for some reason (initial handshake) we don't
2968          * have a rttProp. We, in the worst case, fall back
2969          * to the configured min_bw (rc_initial_hptsi_bw).
2970          */
2971         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
2972                 /* Attempt first to use rttProp */
2973                 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2974                 if (rtt && (rtt < 0xffffffff)) {
2975 measure:
2976                         min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2977                                 ((uint64_t)1000000);
2978                         min_bw /= rtt;
2979                         if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2980                                 min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2981                         }
2982
2983                 } else if (bbr->rc_tp->t_srtt != 0) {
2984                         /* No rttProp, use srtt? */
2985                         rtt = bbr_get_rtt(bbr, BBR_SRTT);
2986                         goto measure;
2987                 } else {
2988                         min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2989                 }
2990         } else
2991                 min_bw = 0;
2992
2993         if ((bbr->rc_past_init_win == 0) &&
2994             (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp)))
2995                 bbr->rc_past_init_win = 1;
2996         if ((bbr->rc_use_google)  && (bbr->r_ctl.r_measurement_count >= 1))
2997                 gm_measure_cnt = 0;
2998         if (gm_measure_cnt &&
2999             ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) ||
3000              (bbr->rc_past_init_win == 0))) {
3001                 /* For google we use our guess rate until we get 1 measurement */
3002
3003 use_initial_window:
3004                 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
3005                 if (rtt && (rtt < 0xffffffff)) {
3006                         /*
3007                          * We have an RTT measurment. Use that in
3008                          * combination with our initial window to calculate
3009                          * a b/w.
3010                          */
3011                         bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
3012                                 ((uint64_t)1000000);
3013                         bw /= rtt;
3014                         if (bw < bbr->r_ctl.rc_initial_hptsi_bw) {
3015                                 bw = bbr->r_ctl.rc_initial_hptsi_bw;
3016                         }
3017                 } else {
3018                         /* Drop back to the 40 and punt to a default */
3019                         bw = bbr->r_ctl.rc_initial_hptsi_bw;
3020                 }
3021                 if (bw < 1)
3022                         /* Probably should panic */
3023                         bw = 1;
3024                 if (bw > min_bw)
3025                         return (bw);
3026                 else
3027                         return (min_bw);
3028         }
3029         if (bbr->rc_lt_use_bw)
3030                 bw = bbr->r_ctl.rc_lt_bw;
3031         else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0))
3032                 bw = bbr->r_ctl.red_bw;
3033         else
3034                 bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3035         if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) {
3036                 /*
3037                  * Enforce user set rate limit, keep in mind that
3038                  * t_peakrate_thr is in B/s already
3039                  */
3040                 bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw);
3041         }
3042         if (bw == 0) {
3043                 /* We should not be at 0, go to the initial window then  */
3044                 goto use_initial_window;
3045         }
3046         if (bw < 1)
3047                 /* Probably should panic */
3048                 bw = 1;
3049         if (bw < min_bw)
3050                 bw = min_bw;
3051         return (bw);
3052 }
3053
3054 static inline uint64_t
3055 bbr_get_bw(struct tcp_bbr *bbr)
3056 {
3057         uint64_t bw;
3058
3059         bw = __bbr_get_bw(bbr);
3060         return (bw);
3061 }
3062
3063 static inline void
3064 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts)
3065 {
3066         bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch;
3067         bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time;
3068         bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3069         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3070 }
3071
3072 static inline void
3073 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts)
3074 {
3075         bbr->rc_lt_is_sampling = 0;
3076         bbr->rc_lt_use_bw = 0;
3077         bbr->r_ctl.rc_lt_bw = 0;
3078         bbr_reset_lt_bw_interval(bbr, cts);
3079 }
3080
3081 static inline void
3082 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin)
3083 {
3084         uint64_t diff;
3085
3086         /* Do we have a previous sample? */
3087         if (bbr->r_ctl.rc_lt_bw) {
3088                 /* Get the diff in bytes per second */
3089                 if (bbr->r_ctl.rc_lt_bw > bw)
3090                         diff = bbr->r_ctl.rc_lt_bw - bw;
3091                 else
3092                         diff = bw - bbr->r_ctl.rc_lt_bw;
3093                 if ((diff <= bbr_lt_bw_diff) ||
3094                     (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) {
3095                         /* Consider us policed */
3096                         uint32_t saved_bw;
3097
3098                         saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw;
3099                         bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2;   /* average of two */
3100                         bbr->rc_lt_use_bw = 1;
3101                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
3102                         /*
3103                          * Use pkt based epoch for measuring length of
3104                          * policer up
3105                          */
3106                         bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch;
3107                         /*
3108                          * reason 4 is we need to start consider being
3109                          * policed
3110                          */
3111                         bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin);
3112                         return;
3113                 }
3114         }
3115         bbr->r_ctl.rc_lt_bw = bw;
3116         bbr_reset_lt_bw_interval(bbr, cts);
3117         bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin);
3118 }
3119
3120 /*
3121  * RRS: Copied from user space!
3122  * Calculate a uniformly distributed random number less than upper_bound
3123  * avoiding "modulo bias".
3124  *
3125  * Uniformity is achieved by generating new random numbers until the one
3126  * returned is outside the range [0, 2**32 % upper_bound).  This
3127  * guarantees the selected random number will be inside
3128  * [2**32 % upper_bound, 2**32) which maps back to [0, upper_bound)
3129  * after reduction modulo upper_bound.
3130  */
3131 static uint32_t
3132 arc4random_uniform(uint32_t upper_bound)
3133 {
3134         uint32_t r, min;
3135
3136         if (upper_bound < 2)
3137                 return 0;
3138
3139         /* 2**32 % x == (2**32 - x) % x */
3140         min = -upper_bound % upper_bound;
3141
3142         /*
3143          * This could theoretically loop forever but each retry has
3144          * p > 0.5 (worst case, usually far better) of selecting a
3145          * number inside the range we need, so it should rarely need
3146          * to re-roll.
3147          */
3148         for (;;) {
3149                 r = arc4random();
3150                 if (r >= min)
3151                         break;
3152         }
3153
3154         return r % upper_bound;
3155 }
3156
3157 static void
3158 bbr_randomize_extra_state_time(struct tcp_bbr *bbr)
3159 {
3160         uint32_t ran, deduct;
3161
3162         ran = arc4random_uniform(bbr_rand_ot);
3163         if (ran) {
3164                 deduct = bbr->r_ctl.rc_level_state_extra / ran;
3165                 bbr->r_ctl.rc_level_state_extra -= deduct;
3166         }
3167 }
3168 /*
3169  * Return randomly the starting state
3170  * to use in probebw.
3171  */
3172 static uint8_t
3173 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts)
3174 {
3175         uint32_t ran;
3176         uint8_t ret_val;
3177
3178         /* Initialize the offset to 0 */
3179         bbr->r_ctl.rc_exta_time_gd = 0;
3180         bbr->rc_hit_state_1 = 0;
3181         bbr->r_ctl.rc_level_state_extra = 0;
3182         ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1));
3183         /*
3184          * The math works funny here :) the return value is used to set the
3185          * substate and then the state change is called which increments by
3186          * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when
3187          * we fully enter the state. Note that the (8 - 1 - ran) assures that
3188          * we return 1 - 7, so we dont return 0 and end up starting in
3189          * state 1 (DRAIN).
3190          */
3191         ret_val = BBR_SUBSTATE_COUNT - 1 - ran;
3192         /* Set an epoch */
3193         if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP))
3194                 bbr_set_epoch(bbr, cts, __LINE__);
3195
3196         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
3197         return (ret_val);
3198 }
3199
3200 static void
3201 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected)
3202 {
3203         uint32_t diff, d_time;
3204         uint64_t del_time, bw, lost, delivered;
3205
3206         if (bbr->r_use_policer == 0)
3207                 return;
3208         if (bbr->rc_lt_use_bw) {
3209                 /* We are using lt bw do we stop yet? */
3210                 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
3211                 if (diff > bbr_lt_bw_max_rtts) {
3212                         /* Reset it all */
3213 reset_all:
3214                         bbr_reset_lt_bw_sampling(bbr, cts);
3215                         if (bbr->rc_filled_pipe) {
3216                                 bbr_set_epoch(bbr, cts, __LINE__);
3217                                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
3218                                 bbr_substate_change(bbr, cts, __LINE__, 0);
3219                                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
3220                                 bbr_log_type_statechange(bbr, cts, __LINE__);
3221                         } else {
3222                                 /*
3223                                  * This should not happen really
3224                                  * unless we remove the startup/drain
3225                                  * restrictions above.
3226                                  */
3227                                 bbr->rc_bbr_state = BBR_STATE_STARTUP;
3228                                 bbr_set_epoch(bbr, cts, __LINE__);
3229                                 bbr->r_ctl.rc_bbr_state_time = cts;
3230                                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
3231                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
3232                                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
3233                                 bbr_set_state_target(bbr, __LINE__);
3234                                 bbr_log_type_statechange(bbr, cts, __LINE__);
3235                         }
3236                         /* reason 0 is to stop using lt-bw */
3237                         bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0);
3238                         return;
3239                 }
3240                 if (bbr_lt_intvl_fp == 0) {
3241                         /* Not doing false-postive detection */
3242                         return;
3243                 }
3244                 /* False positive detection */
3245                 if (diff == bbr_lt_intvl_fp) {
3246                         /* At bbr_lt_intvl_fp we record the lost */
3247                         bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3248                         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3249                 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) {
3250                         /* Now is our loss rate still high? */
3251                         lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3252                         delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3253                         if ((delivered == 0) ||
3254                             (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) {
3255                                 /* No still below our threshold */
3256                                 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0);
3257                         } else {
3258                                 /* Yikes its still high, it must be a false positive */
3259                                 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0);
3260                                 goto reset_all;
3261                         }
3262                 }
3263                 return;
3264         }
3265         /*
3266          * Wait for the first loss before sampling, to let the policer
3267          * exhaust its tokens and estimate the steady-state rate allowed by
3268          * the policer. Starting samples earlier includes bursts that
3269          * over-estimate the bw.
3270          */
3271         if (bbr->rc_lt_is_sampling == 0) {
3272                 /* reason 1 is to begin doing the sampling  */
3273                 if (loss_detected == 0)
3274                         return;
3275                 bbr_reset_lt_bw_interval(bbr, cts);
3276                 bbr->rc_lt_is_sampling = 1;
3277                 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0);
3278                 return;
3279         }
3280         /* Now how long were we delivering long term last> */
3281         if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time))
3282                 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time;
3283         else
3284                 d_time = 0;
3285
3286         /* To avoid underestimates, reset sampling if we run out of data. */
3287         if (bbr->r_ctl.r_app_limited_until) {
3288                 /* Can not measure in app-limited state */
3289                 bbr_reset_lt_bw_sampling(bbr, cts);
3290                 /* reason 2 is to reset sampling due to app limits  */
3291                 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time);
3292                 return;
3293         }
3294         diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
3295         if (diff < bbr_lt_intvl_min_rtts) {
3296                 /*
3297                  * need more samples (we don't
3298                  * start on a round like linux so
3299                  * we need 1 more).
3300                  */
3301                 /* 6 is not_enough time or no-loss */
3302                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3303                 return;
3304         }
3305         if (diff > (4 * bbr_lt_intvl_min_rtts)) {
3306                 /*
3307                  * For now if we wait too long, reset all sampling. We need
3308                  * to do some research here, its possible that we should
3309                  * base this on how much loss as occurred.. something like
3310                  * if its under 10% (or some thresh) reset all otherwise
3311                  * don't.  Thats for phase II I guess.
3312                  */
3313                 bbr_reset_lt_bw_sampling(bbr, cts);
3314                 /* reason 3 is to reset sampling due too long of sampling */
3315                 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3316                 return;
3317         }
3318         /*
3319          * End sampling interval when a packet is lost, so we estimate the
3320          * policer tokens were exhausted. Stopping the sampling before the
3321          * tokens are exhausted under-estimates the policed rate.
3322          */
3323         if (loss_detected == 0) {
3324                 /* 6 is not_enough time or no-loss */
3325                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3326                 return;
3327         }
3328         /* Calculate packets lost and delivered in sampling interval. */
3329         lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3330         delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3331         if ((delivered == 0) ||
3332             (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) {
3333                 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time);
3334                 return;
3335         }
3336         if (d_time < 1000) {
3337                 /* Not enough time. wait */
3338                 /* 6 is not_enough time or no-loss */
3339                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3340                 return;
3341         }
3342         if (d_time >= (0xffffffff / USECS_IN_MSEC)) {
3343                 /* Too long */
3344                 bbr_reset_lt_bw_sampling(bbr, cts);
3345                 /* reason 3 is to reset sampling due too long of sampling */
3346                 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3347                 return;
3348         }
3349         del_time = d_time;
3350         bw = delivered;
3351         bw *= (uint64_t)USECS_IN_SECOND;
3352         bw /= del_time;
3353         bbr_lt_bw_samp_done(bbr, bw, cts, d_time);
3354 }
3355
3356 /*
3357  * Allocate a sendmap from our zone.
3358  */
3359 static struct bbr_sendmap *
3360 bbr_alloc(struct tcp_bbr *bbr)
3361 {
3362         struct bbr_sendmap *rsm;
3363
3364         BBR_STAT_INC(bbr_to_alloc);
3365         rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO));
3366         if (rsm) {
3367                 bbr->r_ctl.rc_num_maps_alloced++;
3368                 return (rsm);
3369         }
3370         if (bbr->r_ctl.rc_free_cnt) {
3371                 BBR_STAT_INC(bbr_to_alloc_emerg);
3372                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
3373                 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
3374                 bbr->r_ctl.rc_free_cnt--;
3375                 return (rsm);
3376         }
3377         BBR_STAT_INC(bbr_to_alloc_failed);
3378         return (NULL);
3379 }
3380
3381 static struct bbr_sendmap *
3382 bbr_alloc_full_limit(struct tcp_bbr *bbr)
3383 {
3384         if ((bbr_tcp_map_entries_limit > 0) &&
3385             (bbr->r_ctl.rc_num_maps_alloced >= bbr_tcp_map_entries_limit)) {
3386                 BBR_STAT_INC(bbr_alloc_limited);
3387                 if (!bbr->alloc_limit_reported) {
3388                         bbr->alloc_limit_reported = 1;
3389                         BBR_STAT_INC(bbr_alloc_limited_conns);
3390                 }
3391                 return (NULL);
3392         }
3393         return (bbr_alloc(bbr));
3394 }
3395
3396
3397 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3398 static struct bbr_sendmap *
3399 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type)
3400 {
3401         struct bbr_sendmap *rsm;
3402
3403         if (limit_type) {
3404                 /* currently there is only one limit type */
3405                 if (bbr_tcp_map_split_limit > 0 &&
3406                     bbr->r_ctl.rc_num_split_allocs >= bbr_tcp_map_split_limit) {
3407                         BBR_STAT_INC(bbr_split_limited);
3408                         if (!bbr->alloc_limit_reported) {
3409                                 bbr->alloc_limit_reported = 1;
3410                                 BBR_STAT_INC(bbr_alloc_limited_conns);
3411                         }
3412                         return (NULL);
3413                 }
3414         }
3415
3416         /* allocate and mark in the limit type, if set */
3417         rsm = bbr_alloc(bbr);
3418         if (rsm != NULL && limit_type) {
3419                 rsm->r_limit_type = limit_type;
3420                 bbr->r_ctl.rc_num_split_allocs++;
3421         }
3422         return (rsm);
3423 }
3424
3425 static void
3426 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
3427 {
3428         if (rsm->r_limit_type) {
3429                 /* currently there is only one limit type */
3430                 bbr->r_ctl.rc_num_split_allocs--;
3431         }
3432         if (rsm->r_is_smallmap)
3433                 bbr->r_ctl.rc_num_small_maps_alloced--;
3434         if (bbr->r_ctl.rc_tlp_send == rsm)
3435                 bbr->r_ctl.rc_tlp_send = NULL;
3436         if (bbr->r_ctl.rc_resend == rsm) {
3437                 bbr->r_ctl.rc_resend = NULL;
3438         }
3439         if (bbr->r_ctl.rc_next == rsm)
3440                 bbr->r_ctl.rc_next = NULL;
3441         if (bbr->r_ctl.rc_sacklast == rsm)
3442                 bbr->r_ctl.rc_sacklast = NULL;
3443         if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
3444                 memset(rsm, 0, sizeof(struct bbr_sendmap));
3445                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
3446                 rsm->r_limit_type = 0;
3447                 bbr->r_ctl.rc_free_cnt++;
3448                 return;
3449         }
3450         bbr->r_ctl.rc_num_maps_alloced--;
3451         uma_zfree(bbr_zone, rsm);
3452 }
3453
3454 /*
3455  * Returns the BDP.
3456  */
3457 static uint64_t
3458 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) {
3459         /*
3460          * Calculate the bytes in flight needed given the bw (in bytes per
3461          * second) and the specifyed rtt in useconds. We need to put out the
3462          * returned value per RTT to match that rate. Gain will normaly
3463          * raise it up from there.
3464          *
3465          * This should not overflow as long as the bandwidth is below 1
3466          * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller
3467          * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30).
3468          */
3469         uint64_t usec_per_sec;
3470
3471         usec_per_sec = USECS_IN_SECOND;
3472         return ((rtt * bw) / usec_per_sec);
3473 }
3474
3475 /*
3476  * Return the initial cwnd.
3477  */
3478 static uint32_t
3479 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp)
3480 {
3481         uint32_t i_cwnd;
3482
3483         if (bbr->rc_init_win) {
3484                 i_cwnd = bbr->rc_init_win * tp->t_maxseg;
3485         } else if (V_tcp_initcwnd_segments)
3486                 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg),
3487                     max(2 * tp->t_maxseg, 14600));
3488         else if (V_tcp_do_rfc3390)
3489                 i_cwnd = min(4 * tp->t_maxseg,
3490                     max(2 * tp->t_maxseg, 4380));
3491         else {
3492                 /* Per RFC5681 Section 3.1 */
3493                 if (tp->t_maxseg > 2190)
3494                         i_cwnd = 2 * tp->t_maxseg;
3495                 else if (tp->t_maxseg > 1095)
3496                         i_cwnd = 3 * tp->t_maxseg;
3497                 else
3498                         i_cwnd = 4 * tp->t_maxseg;
3499         }
3500         return (i_cwnd);
3501 }
3502
3503 /*
3504  * Given a specified gain, return the target
3505  * cwnd based on that gain.
3506  */
3507 static uint32_t
3508 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw)
3509 {
3510         uint64_t bdp, rtt;
3511         uint32_t cwnd;
3512
3513         if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) ||
3514             (bbr_get_full_bw(bbr) == 0)) {
3515                 /* No measurements yet */
3516                 return (bbr_initial_cwnd(bbr, bbr->rc_tp));
3517         }
3518         /*
3519          * Get bytes per RTT needed (rttProp is normally in
3520          * bbr_cwndtarget_rtt_touse)
3521          */
3522         rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse);
3523         /* Get the bdp from the two values */
3524         bdp = bbr_get_bw_delay_prod(rtt, bw);
3525         /* Now apply the gain */
3526         cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT));
3527
3528         return (cwnd);
3529 }
3530
3531 static uint32_t
3532 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain)
3533 {
3534         uint32_t cwnd, mss;
3535
3536         mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
3537         /* Get the base cwnd with gain rounded to a mss */
3538         cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss);
3539         /*
3540          * Add in N (2 default since we do not have a
3541          * fq layer to trap packets in) quanta's per the I-D
3542          * section 4.2.3.2 quanta adjust.
3543          */
3544         cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs);
3545         if (bbr->rc_use_google) {
3546                 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3547                    (bbr_state_val(bbr) == BBR_SUB_GAIN)) {
3548                         /*
3549                          * The linux implementation adds
3550                          * an extra 2 x mss in gain cycle which
3551                          * is documented no-where except in the code.
3552                          * so we add more for Neal undocumented feature
3553                          */
3554                         cwnd += 2 * mss;
3555                 }
3556                 if ((cwnd / mss) & 0x1) {
3557                         /* Round up for odd num mss */
3558                         cwnd += mss;
3559                 }
3560         }
3561         /* Are we below the min cwnd? */
3562         if (cwnd < get_min_cwnd(bbr))
3563                 return (get_min_cwnd(bbr));
3564         return (cwnd);
3565 }
3566
3567 static uint16_t
3568 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain)
3569 {
3570         if (gain < 1)
3571                 gain = 1;
3572         return (gain);
3573 }
3574
3575 static uint32_t
3576 bbr_get_header_oh(struct tcp_bbr *bbr)
3577 {
3578         int seg_oh;
3579
3580         seg_oh = 0;
3581         if (bbr->r_ctl.rc_inc_tcp_oh) {
3582                 /* Do we include TCP overhead? */
3583                 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr));
3584         }
3585         if (bbr->r_ctl.rc_inc_ip_oh) {
3586                 /* Do we include IP overhead? */
3587 #ifdef INET6
3588                 if (bbr->r_is_v6)
3589                         seg_oh += sizeof(struct ip6_hdr);
3590                 else
3591 #endif
3592 #ifdef INET
3593                         seg_oh += sizeof(struct ip);
3594 #endif
3595         }
3596         if (bbr->r_ctl.rc_inc_enet_oh) {
3597                 /* Do we include the ethernet overhead?  */
3598                 seg_oh += sizeof(struct ether_header);
3599         }
3600         return(seg_oh);
3601 }
3602
3603
3604 static uint32_t
3605 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw)
3606 {
3607         uint64_t divor, res, tim;
3608
3609         if (useconds_time == 0)
3610                 return (0);
3611         gain = bbr_gain_adjust(bbr, gain);
3612         divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT;
3613         tim = useconds_time;
3614         res = (tim * bw * gain) / divor;
3615         if (res == 0)
3616                 res = 1;
3617         return ((uint32_t)res);
3618 }
3619
3620 /*
3621  * Given a gain and a length return the delay in useconds that
3622  * should be used to evenly space out packets
3623  * on the connection (based on the gain factor).
3624  */
3625 static uint32_t
3626 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog)
3627 {
3628         uint64_t bw, lentim, res;
3629         uint32_t usecs, srtt, over = 0;
3630         uint32_t seg_oh, num_segs, maxseg;
3631
3632         if (len == 0)
3633                 return (0);
3634
3635         maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
3636         num_segs = (len + maxseg - 1) / maxseg;
3637         if (bbr->rc_use_google == 0) {
3638                 seg_oh = bbr_get_header_oh(bbr);
3639                 len += (num_segs * seg_oh);
3640         }
3641         gain = bbr_gain_adjust(bbr, gain);
3642         bw = bbr_get_bw(bbr);
3643         if (bbr->rc_use_google) {
3644                 uint64_t cbw;
3645
3646                 /*
3647                  * Reduce the b/w by the google discount
3648                  * factor 10 = 1%.
3649                  */
3650                 cbw = bw *  (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount);
3651                 cbw /= (uint64_t)1000;
3652                 /* We don't apply a discount if it results in 0 */
3653                 if (cbw > 0)
3654                         bw = cbw;
3655         }
3656         lentim = ((uint64_t)len *
3657                   (uint64_t)USECS_IN_SECOND *
3658                   (uint64_t)BBR_UNIT);
3659         res = lentim / ((uint64_t)gain * bw);
3660         if (res == 0)
3661                 res = 1;
3662         usecs = (uint32_t)res;
3663         srtt = bbr_get_rtt(bbr, BBR_SRTT);
3664         if (bbr_hptsi_max_mul && bbr_hptsi_max_div &&
3665             (bbr->rc_use_google == 0) &&
3666             (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) {
3667                 /*
3668                  * We cannot let the delay be more than 1/2 the srtt time.
3669                  * Otherwise we cannot pace out or send properly.
3670                  */
3671                 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div;
3672                 BBR_STAT_INC(bbr_hpts_min_time);
3673         }
3674         if (!nolog)
3675                 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1);
3676         return (usecs);
3677 }
3678
3679 static void
3680 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack,
3681                  uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses)
3682 {
3683         INP_WLOCK_ASSERT(tp->t_inpcb);
3684         uint64_t bw;
3685         uint32_t cwnd, target_cwnd, saved_bytes, maxseg;
3686         int32_t meth;
3687
3688 #ifdef NETFLIX_STATS
3689         if ((tp->t_flags & TF_GPUTINPROG) &&
3690             SEQ_GEQ(th->th_ack, tp->gput_ack)) {
3691                 /*
3692                  * Strech acks and compressed acks will cause this to
3693                  * oscillate but we are doing it the same way as the main
3694                  * stack so it will be compariable (though possibly not
3695                  * ideal).
3696                  */
3697                 int32_t cgput;
3698                 int64_t gput, time_stamp;
3699
3700                 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8;
3701                 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000));
3702                 cgput = gput / time_stamp;
3703                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
3704                                          cgput);
3705                 if (tp->t_stats_gput_prev > 0)
3706                         stats_voi_update_abs_s32(tp->t_stats,
3707                                                  VOI_TCP_GPUT_ND,
3708                                                  ((gput - tp->t_stats_gput_prev) * 100) /
3709                                                  tp->t_stats_gput_prev);
3710                 tp->t_flags &= ~TF_GPUTINPROG;
3711                 tp->t_stats_gput_prev = cgput;
3712         }
3713 #endif
3714         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3715             ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
3716                 /* We don't change anything in probe-rtt */
3717                 return;
3718         }
3719         maxseg = tp->t_maxseg - bbr->rc_last_options;
3720         saved_bytes = bytes_this_ack;
3721         bytes_this_ack += sack_changed;
3722         if (bytes_this_ack > prev_acked) {
3723                 bytes_this_ack -= prev_acked;
3724                 /*
3725                  * A byte ack'd gives us a full mss
3726                  * to be like linux i.e. they count packets.
3727                  */
3728                 if ((bytes_this_ack < maxseg) && bbr->rc_use_google)
3729                         bytes_this_ack = maxseg;
3730         } else {
3731                 /* Unlikely */
3732                 bytes_this_ack = 0;
3733         }
3734         cwnd = tp->snd_cwnd;
3735         bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3736         if (bw)
3737                 target_cwnd = bbr_get_target_cwnd(bbr,
3738                                                   bw,
3739                                                   (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain);
3740         else
3741                 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp);
3742         if (IN_RECOVERY(tp->t_flags) &&
3743             (bbr->bbr_prev_in_rec == 0)) {
3744                 /*
3745                  * We are entering recovery and
3746                  * thus packet conservation.
3747                  */
3748                 bbr->pkt_conservation = 1;
3749                 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime;
3750                 cwnd = ctf_flight_size(tp,
3751                                        (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3752                         bytes_this_ack;
3753         }
3754         if (IN_RECOVERY(tp->t_flags)) {
3755                 uint32_t flight;
3756
3757                 bbr->bbr_prev_in_rec = 1;
3758                 if (cwnd > losses) {
3759                         cwnd -= losses;
3760                         if (cwnd < maxseg)
3761                                 cwnd = maxseg;
3762                 } else
3763                         cwnd = maxseg;
3764                 flight = ctf_flight_size(tp,
3765                                          (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3766                 bbr_log_type_cwndupd(bbr, flight, 0,
3767                                      losses, 10, 0, 0, line);
3768                 if (bbr->pkt_conservation) {
3769                         uint32_t time_in;
3770
3771                         if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start))
3772                                 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start;
3773                         else
3774                                 time_in = 0;
3775
3776                         if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
3777                                 /* Clear packet conservation after an rttProp */
3778                                 bbr->pkt_conservation = 0;
3779                         } else {
3780                                 if ((flight + bytes_this_ack) > cwnd)
3781                                         cwnd = flight + bytes_this_ack;
3782                                 if (cwnd < get_min_cwnd(bbr))
3783                                         cwnd = get_min_cwnd(bbr);
3784                                 tp->snd_cwnd = cwnd;
3785                                 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed,
3786                                                      prev_acked, 1, target_cwnd, th->th_ack, line);
3787                                 return;
3788                         }
3789                 }
3790         } else
3791                 bbr->bbr_prev_in_rec = 0;
3792         if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) {
3793                 bbr->r_ctl.restrict_growth--;
3794                 if (bytes_this_ack > maxseg)
3795                         bytes_this_ack = maxseg;
3796         }
3797         if (bbr->rc_filled_pipe) {
3798                 /*
3799                  * Here we have exited startup and filled the pipe. We will
3800                  * thus allow the cwnd to shrink to the target. We hit here
3801                  * mostly.
3802                  */
3803                 uint32_t s_cwnd;
3804
3805                 meth = 2;
3806                 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd);
3807                 if (s_cwnd > cwnd)
3808                         cwnd = s_cwnd;
3809                 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing)
3810                         cwnd = s_cwnd;
3811         } else {
3812                 /*
3813                  * Here we are still in startup, we increase cwnd by what
3814                  * has been acked.
3815                  */
3816                 if ((cwnd < target_cwnd) ||
3817                     (bbr->rc_past_init_win == 0)) {
3818                         meth = 3;
3819                         cwnd += bytes_this_ack;
3820                 } else {
3821                         /*
3822                          * Method 4 means we are at target so no gain in
3823                          * startup and past the initial window.
3824                          */
3825                         meth = 4;
3826                 }
3827         }
3828         tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr));
3829         bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line);
3830 }
3831
3832 static void
3833 tcp_bbr_partialack(struct tcpcb *tp)
3834 {
3835         struct tcp_bbr *bbr;
3836
3837         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3838         INP_WLOCK_ASSERT(tp->t_inpcb);
3839         if (ctf_flight_size(tp,
3840                 (bbr->r_ctl.rc_sacked  + bbr->r_ctl.rc_lost_bytes)) <=
3841             tp->snd_cwnd) {
3842                 bbr->r_wanted_output = 1;
3843         }
3844 }
3845
3846 static void
3847 bbr_post_recovery(struct tcpcb *tp)
3848 {
3849         struct tcp_bbr *bbr;
3850         uint32_t  flight;
3851
3852         INP_WLOCK_ASSERT(tp->t_inpcb);
3853         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3854         /*
3855          * Here we just exit recovery.
3856          */
3857         EXIT_RECOVERY(tp->t_flags);
3858         /* Lock in our b/w reduction for the specified number of pkt-epochs */
3859         bbr->r_recovery_bw = 0;
3860         tp->snd_recover = tp->snd_una;
3861         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3862         bbr->pkt_conservation = 0;
3863         if (bbr->rc_use_google == 0) {
3864                 /*
3865                  * For non-google mode lets
3866                  * go ahead and make sure we clear
3867                  * the recovery state so if we
3868                  * bounce back in to recovery we
3869                  * will do PC.
3870                  */
3871                 bbr->bbr_prev_in_rec = 0;
3872         }
3873         bbr_log_type_exit_rec(bbr);
3874         if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
3875                 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
3876                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__);
3877         } else {
3878                 /* For probe-rtt case lets fix up its saved_cwnd */
3879                 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) {
3880                         bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent;
3881                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__);
3882                 }
3883         }
3884         flight = ctf_flight_size(tp,
3885                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3886         if ((bbr->rc_use_google == 0) &&
3887             bbr_do_red) {
3888                 uint64_t val, lr2use;
3889                 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd;
3890                 uint32_t *cwnd_p;
3891
3892                 if (bbr_get_rtt(bbr, BBR_SRTT)) {
3893                         val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000);
3894                         val /= bbr_get_rtt(bbr, BBR_SRTT);
3895                         ratio = (uint32_t)val;
3896                 } else
3897                         ratio = 1000;
3898
3899                 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div,
3900                                      bbr->r_ctl.recovery_lr, 21,
3901                                      ratio,
3902                                      bbr->r_ctl.rc_red_cwnd_pe,
3903                                      __LINE__);
3904                 if ((ratio < bbr_do_red) || (bbr_do_red == 0))
3905                         goto done;
3906                 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3907                      bbr_prtt_slam_cwnd) ||
3908                     (bbr_sub_drain_slam_cwnd &&
3909                      (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3910                      bbr->rc_hit_state_1 &&
3911                      (bbr_state_val(bbr) == BBR_SUB_DRAIN)) ||
3912                     ((bbr->rc_bbr_state == BBR_STATE_DRAIN) &&
3913                      bbr_slam_cwnd_in_main_drain)) {
3914                         /*
3915                          * Here we must poke at the saved cwnd
3916                          * as well as the cwnd.
3917                          */
3918                         cwnd = bbr->r_ctl.rc_saved_cwnd;
3919                         cwnd_p = &bbr->r_ctl.rc_saved_cwnd;
3920                 } else {
3921                         cwnd = tp->snd_cwnd;
3922                         cwnd_p = &tp->snd_cwnd;
3923                 }
3924                 maxseg = tp->t_maxseg - bbr->rc_last_options;
3925                 /* Add the overall lr with the recovery lr */
3926                 if (bbr->r_ctl.rc_lost == 0)
3927                         lr2use = 0;
3928                 else if (bbr->r_ctl.rc_delivered == 0)
3929                         lr2use = 1000;
3930                 else {
3931                         lr2use = bbr->r_ctl.rc_lost * 1000;
3932                         lr2use /= bbr->r_ctl.rc_delivered;
3933                 }
3934                 lr2use += bbr->r_ctl.recovery_lr;
3935                 acks_inflight = (flight / (maxseg * 2));
3936                 if (bbr_red_scale) {
3937                         lr2use *= bbr_get_rtt(bbr, BBR_SRTT);
3938                         lr2use /= bbr_red_scale;
3939                         if ((bbr_red_growth_restrict) &&
3940                             ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1))
3941                             bbr->r_ctl.restrict_growth += acks_inflight;
3942                 }
3943                 if (lr2use) {
3944                         val = (uint64_t)cwnd * lr2use;
3945                         val /= 1000;
3946                         if (cwnd > val)
3947                                 newcwnd = roundup((cwnd - val), maxseg);
3948                         else
3949                                 newcwnd = maxseg;
3950                 } else {
3951                         val = (uint64_t)cwnd * (uint64_t)bbr_red_mul;
3952                         val /= (uint64_t)bbr_red_div;
3953                         newcwnd = roundup((uint32_t)val, maxseg);
3954                 }
3955                 /* with standard delayed acks how many acks can I expect? */
3956                 if (bbr_drop_limit == 0) {
3957                         /*
3958                          * Anticpate how much we will
3959                          * raise the cwnd based on the acks.
3960                          */
3961                         if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) {
3962                                 /* We do enforce the min (with the acks) */
3963                                 newcwnd = (get_min_cwnd(bbr) - acks_inflight);
3964                         }
3965                 } else {
3966                         /*
3967                          * A strict drop limit of N is is inplace
3968                          */
3969                         if (newcwnd < (bbr_drop_limit * maxseg)) {
3970                                 newcwnd = bbr_drop_limit * maxseg;
3971                         }
3972                 }
3973                 /* For the next N acks do we restrict the growth */
3974                 *cwnd_p = newcwnd;
3975                 if (tp->snd_cwnd > newcwnd)
3976                         tp->snd_cwnd = newcwnd;
3977                 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22,
3978                                      (uint32_t)lr2use,
3979                                      bbr_get_rtt(bbr, BBR_SRTT), __LINE__);
3980                 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch;
3981         }
3982 done:
3983         bbr->r_ctl.recovery_lr = 0;
3984         if (flight <= tp->snd_cwnd) {
3985                 bbr->r_wanted_output = 1;
3986         }
3987         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3988 }
3989
3990 static void
3991 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts)
3992 {
3993         bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3994         /* Limit the drop in b/w to 1/2 our current filter. */
3995         if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate)
3996                 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate;
3997         if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2))
3998                 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2;
3999         tcp_bbr_tso_size_check(bbr, cts);
4000 }
4001
4002 static void
4003 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm)
4004 {
4005         struct tcp_bbr *bbr;
4006
4007         INP_WLOCK_ASSERT(tp->t_inpcb);
4008         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
4009         switch (type) {
4010         case CC_NDUPACK:
4011                 if (!IN_RECOVERY(tp->t_flags)) {
4012                         tp->snd_recover = tp->snd_max;
4013                         /* Start a new epoch */
4014                         bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
4015                         if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) {
4016                                 /*
4017                                  * Move forward the lt epoch
4018                                  * so it won't count the truncated
4019                                  * epoch.
4020                                  */
4021                                 bbr->r_ctl.rc_lt_epoch++;
4022                         }
4023                         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
4024                                 /*
4025                                  * Just like the policer detection code
4026                                  * if we are in startup we must push
4027                                  * forward the last startup epoch
4028                                  * to hide the truncated PE.
4029                                  */
4030                                 bbr->r_ctl.rc_bbr_last_startup_epoch++;
4031                         }
4032                         bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd;
4033                         ENTER_RECOVERY(tp->t_flags);
4034                         bbr->rc_tlp_rtx_out = 0;
4035                         bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate;
4036                         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
4037                         if (bbr->rc_inp->inp_in_hpts &&
4038                             ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) {
4039                                 /*
4040                                  * When we enter recovery, we need to restart
4041                                  * any timers. This may mean we gain an agg
4042                                  * early, which will be made up for at the last
4043                                  * rxt out.
4044                                  */
4045                                 bbr->rc_timer_first = 1;
4046                                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
4047                         }
4048                         /*
4049                          * Calculate a new cwnd based on to the current
4050                          * delivery rate with no gain. We get the bdp
4051                          * without gaining it up like we normally would and
4052                          * we use the last cur_del_rate.
4053                          */
4054                         if ((bbr->rc_use_google == 0) &&
4055                             (bbr->r_ctl.bbr_rttprobe_gain_val ||
4056                              (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) {
4057                                 tp->snd_cwnd = ctf_flight_size(tp,
4058                                                    (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
4059                                         (tp->t_maxseg - bbr->rc_last_options);
4060                                 if (tp->snd_cwnd < get_min_cwnd(bbr)) {
4061                                         /* We always gate to min cwnd */
4062                                         tp->snd_cwnd = get_min_cwnd(bbr);
4063                                 }
4064                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__);
4065                         }
4066                         bbr_log_type_enter_rec(bbr, rsm->r_start);
4067                 }
4068                 break;
4069         case CC_RTO_ERR:
4070                 TCPSTAT_INC(tcps_sndrexmitbad);
4071                 /* RTO was unnecessary, so reset everything. */
4072                 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime);
4073                 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
4074                         tp->snd_cwnd = tp->snd_cwnd_prev;
4075                         tp->snd_ssthresh = tp->snd_ssthresh_prev;
4076                         tp->snd_recover = tp->snd_recover_prev;
4077                         tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
4078                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__);
4079                 }
4080                 tp->t_badrxtwin = 0;
4081                 break;
4082         }
4083 }
4084
4085 /*
4086  * Indicate whether this ack should be delayed.  We can delay the ack if
4087  * following conditions are met:
4088  *      - There is no delayed ack timer in progress.
4089  *      - Our last ack wasn't a 0-sized window. We never want to delay
4090  *        the ack that opens up a 0-sized window.
4091  *      - LRO wasn't used for this segment. We make sure by checking that the
4092  *        segment size is not larger than the MSS.
4093  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4094  *        connection.
4095  *      - The data being acked is less than a full segment (a stretch ack
4096  *        of more than a segment we should ack.
4097  *      - nsegs is 1 (if its more than that we received more than 1 ack).
4098  */
4099 #define DELAY_ACK(tp, bbr, nsegs)                               \
4100         (((tp->t_flags & TF_RXWIN0SENT) == 0) &&                \
4101          ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) &&  \
4102          (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4103
4104 /*
4105  * Return the lowest RSM in the map of
4106  * packets still in flight that is not acked.
4107  * This should normally find on the first one
4108  * since we remove packets from the send
4109  * map after they are marked ACKED.
4110  */
4111 static struct bbr_sendmap *
4112 bbr_find_lowest_rsm(struct tcp_bbr *bbr)
4113 {
4114         struct bbr_sendmap *rsm;
4115
4116         /*
4117          * Walk the time-order transmitted list looking for an rsm that is
4118          * not acked. This will be the one that was sent the longest time
4119          * ago that is still outstanding.
4120          */
4121         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) {
4122                 if (rsm->r_flags & BBR_ACKED) {
4123                         continue;
4124                 }
4125                 goto finish;
4126         }
4127 finish:
4128         return (rsm);
4129 }
4130
4131 static struct bbr_sendmap *
4132 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
4133 {
4134         struct bbr_sendmap *prsm;
4135
4136         /*
4137          * Walk the sequence order list backward until we hit and arrive at
4138          * the highest seq not acked. In theory when this is called it
4139          * should be the last segment (which it was not).
4140          */
4141         prsm = rsm;
4142         TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4143                 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) {
4144                         continue;
4145                 }
4146                 return (prsm);
4147         }
4148         return (NULL);
4149 }
4150
4151 /*
4152  * Returns to the caller the number of microseconds that
4153  * the packet can be outstanding before we think we
4154  * should have had an ack returned.
4155  */
4156 static uint32_t
4157 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm)
4158 {
4159         /*
4160          * lro is the flag we use to determine if we have seen reordering.
4161          * If it gets set we have seen reordering. The reorder logic either
4162          * works in one of two ways:
4163          *
4164          * If reorder-fade is configured, then we track the last time we saw
4165          * re-ordering occur. If we reach the point where enough time as
4166          * passed we no longer consider reordering has occuring.
4167          *
4168          * Or if reorder-face is 0, then once we see reordering we consider
4169          * the connection to alway be subject to reordering and just set lro
4170          * to 1.
4171          *
4172          * In the end if lro is non-zero we add the extra time for
4173          * reordering in.
4174          */
4175         int32_t lro;
4176         uint32_t thresh, t_rxtcur;
4177
4178         if (srtt == 0)
4179                 srtt = 1;
4180         if (bbr->r_ctl.rc_reorder_ts) {
4181                 if (bbr->r_ctl.rc_reorder_fade) {
4182                         if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) {
4183                                 lro = cts - bbr->r_ctl.rc_reorder_ts;
4184                                 if (lro == 0) {
4185                                         /*
4186                                          * No time as passed since the last
4187                                          * reorder, mark it as reordering.
4188                                          */
4189                                         lro = 1;
4190                                 }
4191                         } else {
4192                                 /* Negative time? */
4193                                 lro = 0;
4194                         }
4195                         if (lro > bbr->r_ctl.rc_reorder_fade) {
4196                                 /* Turn off reordering seen too */
4197                                 bbr->r_ctl.rc_reorder_ts = 0;
4198                                 lro = 0;
4199                         }
4200                 } else {
4201                         /* Reodering does not fade */
4202                         lro = 1;
4203                 }
4204         } else {
4205                 lro = 0;
4206         }
4207         thresh = srtt + bbr->r_ctl.rc_pkt_delay;
4208         if (lro) {
4209                 /* It must be set, if not you get 1/4 rtt */
4210                 if (bbr->r_ctl.rc_reorder_shift)
4211                         thresh += (srtt >> bbr->r_ctl.rc_reorder_shift);
4212                 else
4213                         thresh += (srtt >> 2);
4214         } else {
4215                 thresh += 1000;
4216         }
4217         /* We don't let the rack timeout be above a RTO */
4218         if ((bbr->rc_tp)->t_srtt == 0)
4219                 t_rxtcur = BBR_INITIAL_RTO;
4220         else
4221                 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
4222         if (thresh > t_rxtcur) {
4223                 thresh = t_rxtcur;
4224         }
4225         /* And we don't want it above the RTO max either */
4226         if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4227                 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4228         }
4229         bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK);
4230         return (thresh);
4231 }
4232
4233 /*
4234  * Return to the caller the amount of time in mico-seconds
4235  * that should be used for the TLP timer from the last
4236  * send time of this packet.
4237  */
4238 static uint32_t
4239 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
4240     struct bbr_sendmap *rsm, uint32_t srtt,
4241     uint32_t cts)
4242 {
4243         uint32_t thresh, len, maxseg, t_rxtcur;
4244         struct bbr_sendmap *prsm;
4245
4246         if (srtt == 0)
4247                 srtt = 1;
4248         if (bbr->rc_tlp_threshold)
4249                 thresh = srtt + (srtt / bbr->rc_tlp_threshold);
4250         else
4251                 thresh = (srtt * 2);
4252         maxseg = tp->t_maxseg - bbr->rc_last_options;
4253         /* Get the previous sent packet, if any  */
4254         len = rsm->r_end - rsm->r_start;
4255
4256         /* 2.1 behavior */
4257         prsm = TAILQ_PREV(rsm, bbr_head, r_tnext);
4258         if (prsm && (len <= maxseg)) {
4259                 /*
4260                  * Two packets outstanding, thresh should be (2*srtt) +
4261                  * possible inter-packet delay (if any).
4262                  */
4263                 uint32_t inter_gap = 0;
4264                 int idx, nidx;
4265
4266                 idx = rsm->r_rtr_cnt - 1;
4267                 nidx = prsm->r_rtr_cnt - 1;
4268                 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
4269                         /* Yes it was sent later (or at the same time) */
4270                         inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
4271                 }
4272                 thresh += inter_gap;
4273         } else if (len <= maxseg) {
4274                 /*
4275                  * Possibly compensate for delayed-ack.
4276                  */
4277                 uint32_t alt_thresh;
4278
4279                 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time;
4280                 if (alt_thresh > thresh)
4281                         thresh = alt_thresh;
4282         }
4283         /* Not above the current  RTO */
4284         if (tp->t_srtt == 0)
4285                 t_rxtcur = BBR_INITIAL_RTO;
4286         else
4287                 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur);
4288
4289         bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP);
4290         /* Not above an RTO */
4291         if (thresh > t_rxtcur) {
4292                 thresh = t_rxtcur;
4293         }
4294         /* Not above a RTO max */
4295         if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4296                 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4297         }
4298         /* And now apply the user TLP min */
4299         if (thresh < bbr_tlp_min) {
4300                 thresh = bbr_tlp_min;
4301         }
4302         return (thresh);
4303 }
4304
4305 /*
4306  * Return one of three RTTs to use (in microseconds).
4307  */
4308 static __inline uint32_t
4309 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type)
4310 {
4311         uint32_t f_rtt;
4312         uint32_t srtt;
4313
4314         f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
4315         if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) {
4316                 /* We have no rtt at all */
4317                 if (bbr->rc_tp->t_srtt == 0)
4318                         f_rtt = BBR_INITIAL_RTO;
4319                 else
4320                         f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4321                 /*
4322                  * Since we don't know how good the rtt is apply a
4323                  * delayed-ack min
4324                  */
4325                 if (f_rtt < bbr_delayed_ack_time) {
4326                         f_rtt = bbr_delayed_ack_time;
4327                 }
4328         }
4329         /* Take the filter version or last measured pkt-rtt */
4330         if (rtt_type == BBR_RTT_PROP) {
4331                 srtt = f_rtt;
4332         } else if (rtt_type == BBR_RTT_PKTRTT) {
4333                 if (bbr->r_ctl.rc_pkt_epoch_rtt) {
4334                         srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
4335                 } else {
4336                         /* No pkt rtt yet */
4337                         srtt = f_rtt;
4338                 }
4339         } else if (rtt_type == BBR_RTT_RACK) {
4340                 srtt = bbr->r_ctl.rc_last_rtt;
4341                 /* We need to add in any internal delay for our timer */
4342                 if (bbr->rc_ack_was_delayed)
4343                         srtt += bbr->r_ctl.rc_ack_hdwr_delay;
4344         } else if (rtt_type == BBR_SRTT) {
4345                 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4346         } else {
4347                 /* TSNH */
4348                 srtt = f_rtt;
4349 #ifdef BBR_INVARIANTS
4350                 panic("Unknown rtt request type %d", rtt_type);
4351 #endif
4352         }
4353         return (srtt);
4354 }
4355
4356 static int
4357 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts)
4358 {
4359         uint32_t thresh;
4360
4361
4362         thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK),
4363                                       cts, rsm);
4364         if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) {
4365                 /* It is lost (past time) */
4366                 return (1);
4367         }
4368         return (0);
4369 }
4370
4371 /*
4372  * Return a sendmap if we need to retransmit something.
4373  */
4374 static struct bbr_sendmap *
4375 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4376 {
4377         /*
4378          * Check to see that we don't need to fall into recovery. We will
4379          * need to do so if our oldest transmit is past the time we should
4380          * have had an ack.
4381          */
4382
4383         struct bbr_sendmap *rsm;
4384         int32_t idx;
4385
4386         if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) {
4387                 /* Nothing outstanding that we know of */
4388                 return (NULL);
4389         }
4390         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
4391         if (rsm == NULL) {
4392                 /* Nothing in the transmit map */
4393                 return (NULL);
4394         }
4395         if (tp->t_flags & TF_SENTFIN) {
4396                 /* Fin restricted, don't find anything once a fin is sent */
4397                 return (NULL);
4398         }
4399         if (rsm->r_flags & BBR_ACKED) {
4400                 /*
4401                  * Ok the first one is acked (this really should not happen
4402                  * since we remove the from the tmap once they are acked)
4403                  */
4404                 rsm = bbr_find_lowest_rsm(bbr);
4405                 if (rsm == NULL)
4406                         return (NULL);
4407         }
4408         idx = rsm->r_rtr_cnt - 1;
4409         if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) {
4410                 /* Send timestamp is the same or less? can't be ready */
4411                 return (NULL);
4412         }
4413         /* Get our RTT time */
4414         if (bbr_is_lost(bbr, rsm, cts) &&
4415             ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
4416              (rsm->r_flags & BBR_SACK_PASSED))) {
4417                 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4418                         rsm->r_flags |= BBR_MARKED_LOST;
4419                         bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4420                         bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4421                 }
4422                 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm);
4423 #ifdef BBR_INVARIANTS
4424                 if ((rsm->r_end - rsm->r_start) == 0)
4425                         panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm);
4426 #endif
4427                 return (rsm);
4428         }
4429         return (NULL);
4430 }
4431
4432 /*
4433  * RACK Timer, here we simply do logging and house keeping.
4434  * the normal bbr_output_wtime() function will call the
4435  * appropriate thing to check if we need to do a RACK retransmit.
4436  * We return 1, saying don't proceed with bbr_output_wtime only
4437  * when all timers have been stopped (destroyed PCB?).
4438  */
4439 static int
4440 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4441 {
4442         /*
4443          * This timer simply provides an internal trigger to send out data.
4444          * The check_recovery_mode call will see if there are needed
4445          * retransmissions, if so we will enter fast-recovery. The output
4446          * call may or may not do the same thing depending on sysctl
4447          * settings.
4448          */
4449         uint32_t lost;
4450
4451         if (bbr->rc_all_timers_stopped) {
4452                 return (1);
4453         }
4454         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4455                 /* Its not time yet */
4456                 return (0);
4457         }
4458         BBR_STAT_INC(bbr_to_tot);
4459         lost = bbr->r_ctl.rc_lost;
4460         if (bbr->r_state && (bbr->r_state != tp->t_state))
4461                 bbr_set_state(tp, bbr, 0);
4462         bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK);
4463         if (bbr->r_ctl.rc_resend == NULL) {
4464                 /* Lets do the check here */
4465                 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
4466         }
4467         if (bbr_policer_call_from_rack_to)
4468                 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4469         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
4470         return (0);
4471 }
4472
4473 static __inline void
4474 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start)
4475 {
4476         int idx;
4477
4478         nrsm->r_start = start;
4479         nrsm->r_end = rsm->r_end;
4480         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
4481         nrsm->r_flags = rsm->r_flags;
4482         /* We don't transfer forward the SYN flag */
4483         nrsm->r_flags &= ~BBR_HAS_SYN;
4484         /* We move forward the FIN flag, not that this should happen */
4485         rsm->r_flags &= ~BBR_HAS_FIN;
4486         nrsm->r_dupack = rsm->r_dupack;
4487         nrsm->r_rtr_bytes = 0;
4488         nrsm->r_is_gain = rsm->r_is_gain;
4489         nrsm->r_is_drain = rsm->r_is_drain;
4490         nrsm->r_delivered = rsm->r_delivered;
4491         nrsm->r_ts_valid = rsm->r_ts_valid;
4492         nrsm->r_del_ack_ts = rsm->r_del_ack_ts;
4493         nrsm->r_del_time = rsm->r_del_time;
4494         nrsm->r_app_limited = rsm->r_app_limited;
4495         nrsm->r_first_sent_time = rsm->r_first_sent_time;
4496         nrsm->r_flight_at_send = rsm->r_flight_at_send;
4497         /* We split a piece the lower section looses any just_ret flag. */
4498         nrsm->r_bbr_state = rsm->r_bbr_state;
4499         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
4500                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
4501         }
4502         rsm->r_end = nrsm->r_start;
4503         idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
4504         idx /= 8;
4505         /* Check if we got too small */
4506         if ((rsm->r_is_smallmap == 0) &&
4507             ((rsm->r_end - rsm->r_start) <= idx)) {
4508                 bbr->r_ctl.rc_num_small_maps_alloced++;
4509                 rsm->r_is_smallmap = 1;
4510         }
4511         /* Check the new one as well */
4512         if ((nrsm->r_end - nrsm->r_start) <= idx) {
4513                 bbr->r_ctl.rc_num_small_maps_alloced++;
4514                 nrsm->r_is_smallmap = 1;
4515         }
4516 }
4517
4518 static int
4519 bbr_sack_mergable(struct bbr_sendmap *at,
4520                   uint32_t start, uint32_t end)
4521 {
4522         /*
4523          * Given a sack block defined by
4524          * start and end, and a current postion
4525          * at. Return 1 if either side of at
4526          * would show that the block is mergable
4527          * to that side. A block to be mergable
4528          * must have overlap with the start/end
4529          * and be in the SACK'd state.
4530          */
4531         struct bbr_sendmap *l_rsm;
4532         struct bbr_sendmap *r_rsm;
4533
4534         /* first get the either side blocks */
4535         l_rsm = TAILQ_PREV(at, bbr_head, r_next);
4536         r_rsm = TAILQ_NEXT(at, r_next);
4537         if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) {
4538                 /* Potentially mergeable */
4539                 if ((l_rsm->r_end == start) ||
4540                     (SEQ_LT(start, l_rsm->r_end) &&
4541                      SEQ_GT(end, l_rsm->r_end))) {
4542                             /*
4543                              * map blk   |------|
4544                              * sack blk         |------|
4545                              * <or>
4546                              * map blk   |------|
4547                              * sack blk      |------|
4548                              */
4549                             return (1);
4550                     }
4551         }
4552         if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) {
4553                 /* Potentially mergeable */
4554                 if ((r_rsm->r_start == end) ||
4555                     (SEQ_LT(start, r_rsm->r_start) &&
4556                      SEQ_GT(end, r_rsm->r_start))) {
4557                         /*
4558                          * map blk          |---------|
4559                          * sack blk    |----|
4560                          * <or>
4561                          * map blk          |---------|
4562                          * sack blk    |-------|
4563                          */
4564                         return (1);
4565                 }
4566         }
4567         return (0);
4568 }
4569
4570 static struct bbr_sendmap *
4571 bbr_merge_rsm(struct tcp_bbr *bbr,
4572               struct bbr_sendmap *l_rsm,
4573               struct bbr_sendmap *r_rsm)
4574 {
4575         /*
4576          * We are merging two ack'd RSM's,
4577          * the l_rsm is on the left (lower seq
4578          * values) and the r_rsm is on the right
4579          * (higher seq value). The simplest way
4580          * to merge these is to move the right
4581          * one into the left. I don't think there
4582          * is any reason we need to try to find
4583          * the oldest (or last oldest retransmitted).
4584          */
4585         l_rsm->r_end = r_rsm->r_end;
4586         if (l_rsm->r_dupack < r_rsm->r_dupack)
4587                 l_rsm->r_dupack = r_rsm->r_dupack;
4588         if (r_rsm->r_rtr_bytes)
4589                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
4590         if (r_rsm->r_in_tmap) {
4591                 /* This really should not happen */
4592                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext);
4593         }
4594         if (r_rsm->r_app_limited)
4595                 l_rsm->r_app_limited = r_rsm->r_app_limited;
4596         /* Now the flags */
4597         if (r_rsm->r_flags & BBR_HAS_FIN)
4598                 l_rsm->r_flags |= BBR_HAS_FIN;
4599         if (r_rsm->r_flags & BBR_TLP)
4600                 l_rsm->r_flags |= BBR_TLP;
4601         if (r_rsm->r_flags & BBR_RWND_COLLAPSED)
4602                 l_rsm->r_flags |= BBR_RWND_COLLAPSED;
4603         if (r_rsm->r_flags & BBR_MARKED_LOST) {
4604                 /* This really should not happen */
4605                 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start;
4606         }
4607         TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next);
4608         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
4609                 /* Transfer the split limit to the map we free */
4610                 r_rsm->r_limit_type = l_rsm->r_limit_type;
4611                 l_rsm->r_limit_type = 0;
4612         }
4613         bbr_free(bbr, r_rsm);
4614         return(l_rsm);
4615 }
4616
4617 /*
4618  * TLP Timer, here we simply setup what segment we want to
4619  * have the TLP expire on, the normal bbr_output_wtime() will then
4620  * send it out.
4621  *
4622  * We return 1, saying don't proceed with bbr_output_wtime only
4623  * when all timers have been stopped (destroyed PCB?).
4624  */
4625 static int
4626 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4627 {
4628         /*
4629          * Tail Loss Probe.
4630          */
4631         struct bbr_sendmap *rsm = NULL;
4632         struct socket *so;
4633         uint32_t amm;
4634         uint32_t out, avail;
4635         uint32_t maxseg;
4636         int collapsed_win = 0;
4637
4638         if (bbr->rc_all_timers_stopped) {
4639                 return (1);
4640         }
4641         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4642                 /* Its not time yet */
4643                 return (0);
4644         }
4645         if (bbr_progress_timeout_check(bbr)) {
4646                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4647                 return (1);
4648         }
4649         /* Did we somehow get into persists? */
4650         if (bbr->rc_in_persist) {
4651                 return (0);
4652         }
4653         if (bbr->r_state && (bbr->r_state != tp->t_state))
4654                 bbr_set_state(tp, bbr, 0);
4655         BBR_STAT_INC(bbr_tlp_tot);
4656         maxseg = tp->t_maxseg - bbr->rc_last_options;
4657 #ifdef KERN_TLS
4658         if (bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) {
4659                 /*
4660                  * For hardware TLS we do *not* want to send
4661                  * new data.
4662                  */
4663                 goto need_retran;
4664         }
4665 #endif
4666         /*
4667          * A TLP timer has expired. We have been idle for 2 rtts. So we now
4668          * need to figure out how to force a full MSS segment out.
4669          */
4670         so = tp->t_inpcb->inp_socket;
4671         avail = sbavail(&so->so_snd);
4672         out = ctf_outstanding(tp);
4673         if (out > tp->snd_wnd) {
4674                 /* special case, we need a retransmission */
4675                 collapsed_win = 1;
4676                 goto need_retran;
4677         }
4678         if (avail > out) {
4679                 /* New data is available */
4680                 amm = avail - out;
4681                 if (amm > maxseg) {
4682                         amm = maxseg;
4683                 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) {
4684                         /* not enough to fill a MTU and no-delay is off */
4685                         goto need_retran;
4686                 }
4687                 /* Set the send-new override */
4688                 if ((out + amm) <= tp->snd_wnd) {
4689                         bbr->rc_tlp_new_data = 1;
4690                 } else {
4691                         goto need_retran;
4692                 }
4693                 bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4694                 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max;
4695                 bbr->r_ctl.rc_tlp_send = NULL;
4696                 /* cap any slots */
4697                 BBR_STAT_INC(bbr_tlp_newdata);
4698                 goto send;
4699         }
4700 need_retran:
4701         /*
4702          * Ok we need to arrange the last un-acked segment to be re-sent, or
4703          * optionally the first un-acked segment.
4704          */
4705         if (collapsed_win == 0) {
4706                 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
4707                 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) {
4708                         rsm = bbr_find_high_nonack(bbr, rsm);
4709                 }
4710                 if (rsm == NULL) {
4711                         goto restore;
4712                 }
4713         } else {
4714                 /*
4715                  * We must find the last segment
4716                  * that was acceptable by the client.
4717                  */
4718                 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4719                         if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) {
4720                                 /* Found one */
4721                                 break;
4722                         }
4723                 }
4724                 if (rsm == NULL) {
4725                         /* None? if so send the first */
4726                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4727                         if (rsm == NULL)
4728                                 goto restore;
4729                 }
4730         }
4731         if ((rsm->r_end - rsm->r_start) > maxseg) {
4732                 /*
4733                  * We need to split this the last segment in two.
4734                  */
4735                 struct bbr_sendmap *nrsm;
4736
4737                 nrsm = bbr_alloc_full_limit(bbr);
4738                 if (nrsm == NULL) {
4739                         /*
4740                          * We can't get memory to split, we can either just
4741                          * not split it. Or retransmit the whole piece, lets
4742                          * do the large send (BTLP :-) ).
4743                          */
4744                         goto go_for_it;
4745                 }
4746                 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg));
4747                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
4748                 if (rsm->r_in_tmap) {
4749                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
4750                         nrsm->r_in_tmap = 1;
4751                 }
4752                 rsm->r_flags &= (~BBR_HAS_FIN);
4753                 rsm = nrsm;
4754         }
4755 go_for_it:
4756         bbr->r_ctl.rc_tlp_send = rsm;
4757         bbr->rc_tlp_rtx_out = 1;
4758         if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) {
4759                 bbr->r_ctl.rc_tlp_seg_send_cnt++;
4760                 tp->t_rxtshift++;
4761         } else {
4762                 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
4763                 bbr->r_ctl.rc_tlp_seg_send_cnt = 1;
4764         }
4765 send:
4766         if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
4767                 /*
4768                  * Can't [re]/transmit a segment we have retranmitted the
4769                  * max times. We need the retransmit timer to take over.
4770                  */
4771 restore:
4772                 bbr->rc_tlp_new_data = 0;
4773                 bbr->r_ctl.rc_tlp_send = NULL;
4774                 if (rsm)
4775                         rsm->r_flags &= ~BBR_TLP;
4776                 BBR_STAT_INC(bbr_tlp_retran_fail);
4777                 return (0);
4778         } else if (rsm) {
4779                 rsm->r_flags |= BBR_TLP;
4780         }
4781         if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) &&
4782             (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) {
4783                 /*
4784                  * We have retransmitted to many times for TLP. Switch to
4785                  * the regular RTO timer
4786                  */
4787                 goto restore;
4788         }
4789         bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP);
4790         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
4791         return (0);
4792 }
4793
4794 /*
4795  * Delayed ack Timer, here we simply need to setup the
4796  * ACK_NOW flag and remove the DELACK flag. From there
4797  * the output routine will send the ack out.
4798  *
4799  * We only return 1, saying don't proceed, if all timers
4800  * are stopped (destroyed PCB?).
4801  */
4802 static int
4803 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4804 {
4805         if (bbr->rc_all_timers_stopped) {
4806                 return (1);
4807         }
4808         bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK);
4809         tp->t_flags &= ~TF_DELACK;
4810         tp->t_flags |= TF_ACKNOW;
4811         TCPSTAT_INC(tcps_delack);
4812         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
4813         return (0);
4814 }
4815
4816 /*
4817  * Persists timer, here we simply need to setup the
4818  * FORCE-DATA flag the output routine will send
4819  * the one byte send.
4820  *
4821  * We only return 1, saying don't proceed, if all timers
4822  * are stopped (destroyed PCB?).
4823  */
4824 static int
4825 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4826 {
4827         struct tcptemp *t_template;
4828         int32_t retval = 1;
4829
4830         if (bbr->rc_all_timers_stopped) {
4831                 return (1);
4832         }
4833         if (bbr->rc_in_persist == 0)
4834                 return (0);
4835         KASSERT(tp->t_inpcb != NULL,
4836             ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
4837         /*
4838          * Persistence timer into zero window. Force a byte to be output, if
4839          * possible.
4840          */
4841         bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST);
4842         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
4843         TCPSTAT_INC(tcps_persisttimeo);
4844         /*
4845          * Have we exceeded the user specified progress time?
4846          */
4847         if (bbr_progress_timeout_check(bbr)) {
4848                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4849                 goto out;
4850         }
4851         /*
4852          * Hack: if the peer is dead/unreachable, we do not time out if the
4853          * window is closed.  After a full backoff, drop the connection if
4854          * the idle time (no responses to probes) reaches the maximum
4855          * backoff that we would use if retransmitting.
4856          */
4857         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
4858             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
4859             ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
4860                 TCPSTAT_INC(tcps_persistdrop);
4861                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4862                 goto out;
4863         }
4864         if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) &&
4865             tp->snd_una == tp->snd_max) {
4866                 bbr_exit_persist(tp, bbr, cts, __LINE__);
4867                 retval = 0;
4868                 goto out;
4869         }
4870         /*
4871          * If the user has closed the socket then drop a persisting
4872          * connection after a much reduced timeout.
4873          */
4874         if (tp->t_state > TCPS_CLOSE_WAIT &&
4875             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
4876                 TCPSTAT_INC(tcps_persistdrop);
4877                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4878                 goto out;
4879         }
4880         t_template = tcpip_maketemplate(bbr->rc_inp);
4881         if (t_template) {
4882                 tcp_respond(tp, t_template->tt_ipgen,
4883                             &t_template->tt_t, (struct mbuf *)NULL,
4884                             tp->rcv_nxt, tp->snd_una - 1, 0);
4885                 /* This sends an ack */
4886                 if (tp->t_flags & TF_DELACK)
4887                         tp->t_flags &= ~TF_DELACK;
4888                 free(t_template, M_TEMP);
4889         }
4890         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
4891                 tp->t_rxtshift++;
4892         bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0);
4893 out:
4894         return (retval);
4895 }
4896
4897 /*
4898  * If a keepalive goes off, we had no other timers
4899  * happening. We always return 1 here since this
4900  * routine either drops the connection or sends
4901  * out a segment with respond.
4902  */
4903 static int
4904 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4905 {
4906         struct tcptemp *t_template;
4907         struct inpcb *inp;
4908
4909         if (bbr->rc_all_timers_stopped) {
4910                 return (1);
4911         }
4912         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
4913         inp = tp->t_inpcb;
4914         bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP);
4915         /*
4916          * Keep-alive timer went off; send something or drop connection if
4917          * idle for too long.
4918          */
4919         TCPSTAT_INC(tcps_keeptimeo);
4920         if (tp->t_state < TCPS_ESTABLISHED)
4921                 goto dropit;
4922         if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
4923             tp->t_state <= TCPS_CLOSING) {
4924                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
4925                         goto dropit;
4926                 /*
4927                  * Send a packet designed to force a response if the peer is
4928                  * up and reachable: either an ACK if the connection is
4929                  * still alive, or an RST if the peer has closed the
4930                  * connection due to timeout or reboot. Using sequence
4931                  * number tp->snd_una-1 causes the transmitted zero-length
4932                  * segment to lie outside the receive window; by the
4933                  * protocol spec, this requires the correspondent TCP to
4934                  * respond.
4935                  */
4936                 TCPSTAT_INC(tcps_keepprobe);
4937                 t_template = tcpip_maketemplate(inp);
4938                 if (t_template) {
4939                         tcp_respond(tp, t_template->tt_ipgen,
4940                             &t_template->tt_t, (struct mbuf *)NULL,
4941                             tp->rcv_nxt, tp->snd_una - 1, 0);
4942                         free(t_template, M_TEMP);
4943                 }
4944         }
4945         bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0);
4946         return (1);
4947 dropit:
4948         TCPSTAT_INC(tcps_keepdrops);
4949         tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4950         return (1);
4951 }
4952
4953 /*
4954  * Retransmit helper function, clear up all the ack
4955  * flags and take care of important book keeping.
4956  */
4957 static void
4958 bbr_remxt_tmr(struct tcpcb *tp)
4959 {
4960         /*
4961          * The retransmit timer went off, all sack'd blocks must be
4962          * un-acked.
4963          */
4964         struct bbr_sendmap *rsm, *trsm = NULL;
4965         struct tcp_bbr *bbr;
4966         uint32_t cts, lost;
4967
4968         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
4969         cts = tcp_get_usecs(&bbr->rc_tv);
4970         lost = bbr->r_ctl.rc_lost;
4971         if (bbr->r_state && (bbr->r_state != tp->t_state))
4972                 bbr_set_state(tp, bbr, 0);
4973
4974         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
4975                 if (rsm->r_flags & BBR_ACKED) {
4976                         uint32_t old_flags;
4977
4978                         rsm->r_dupack = 0;
4979                         if (rsm->r_in_tmap == 0) {
4980                                 /* We must re-add it back to the tlist */
4981                                 if (trsm == NULL) {
4982                                         TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
4983                                 } else {
4984                                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext);
4985                                 }
4986                                 rsm->r_in_tmap = 1;
4987                         }
4988                         old_flags = rsm->r_flags;
4989                         rsm->r_flags |= BBR_RXT_CLEARED;
4990                         rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS);
4991                         bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
4992                 } else {
4993                         if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4994                                 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4995                                 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4996                         }
4997                         if (bbr_marks_rxt_sack_passed) {
4998                                 /*
4999                                  * With this option, we will rack out
5000                                  * in 1ms increments the rest of the packets.
5001                                  */
5002                                 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST;
5003                                 rsm->r_flags &= ~BBR_WAS_SACKPASS;
5004                         } else {
5005                                 /*
5006                                  * With this option we only mark them lost
5007                                  * and remove all sack'd markings. We will run
5008                                  * another RXT or a TLP. This will cause
5009                                  * us to eventually send more based on what
5010                                  * ack's come in.
5011                                  */
5012                                 rsm->r_flags |= BBR_MARKED_LOST;
5013                                 rsm->r_flags &= ~BBR_WAS_SACKPASS;
5014                                 rsm->r_flags &= ~BBR_SACK_PASSED;
5015                         }
5016                 }
5017                 trsm = rsm;
5018         }
5019         bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map);
5020         /* Clear the count (we just un-acked them) */
5021         bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR);
5022         bbr->rc_tlp_new_data = 0;
5023         bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
5024         /* zap the behindness on a rxt */
5025         bbr->r_ctl.rc_hptsi_agg_delay = 0;
5026         bbr->r_agg_early_set = 0;
5027         bbr->r_ctl.rc_agg_early = 0;
5028         bbr->rc_tlp_rtx_out = 0;
5029         bbr->r_ctl.rc_sacked = 0;
5030         bbr->r_ctl.rc_sacklast = NULL;
5031         bbr->r_timer_override = 1;
5032         bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
5033 }
5034
5035 /*
5036  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
5037  * we will setup to retransmit the lowest seq number outstanding.
5038  */
5039 static int
5040 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
5041 {
5042         int32_t rexmt;
5043         int32_t retval = 0;
5044
5045         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
5046         if (bbr->rc_all_timers_stopped) {
5047                 return (1);
5048         }
5049         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
5050             (tp->snd_una == tp->snd_max)) {
5051                 /* Nothing outstanding .. nothing to do */
5052                 return (0);
5053         }
5054         /*
5055          * Retransmission timer went off.  Message has not been acked within
5056          * retransmit interval.  Back off to a longer retransmit interval
5057          * and retransmit one segment.
5058          */
5059         if (bbr_progress_timeout_check(bbr)) {
5060                 retval = 1;
5061                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
5062                 goto out;
5063         }
5064         bbr_remxt_tmr(tp);
5065         if ((bbr->r_ctl.rc_resend == NULL) ||
5066             ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) {
5067                 /*
5068                  * If the rwnd collapsed on
5069                  * the one we are retransmitting
5070                  * it does not count against the
5071                  * rxt count.
5072                  */
5073                 tp->t_rxtshift++;
5074         }
5075         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
5076                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
5077                 TCPSTAT_INC(tcps_timeoutdrop);
5078                 retval = 1;
5079                 tcp_set_inp_to_drop(bbr->rc_inp,
5080                     (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
5081                 goto out;
5082         }
5083         if (tp->t_state == TCPS_SYN_SENT) {
5084                 /*
5085                  * If the SYN was retransmitted, indicate CWND to be limited
5086                  * to 1 segment in cc_conn_init().
5087                  */
5088                 tp->snd_cwnd = 1;
5089         } else if (tp->t_rxtshift == 1) {
5090                 /*
5091                  * first retransmit; record ssthresh and cwnd so they can be
5092                  * recovered if this turns out to be a "bad" retransmit. A
5093                  * retransmit is considered "bad" if an ACK for this segment
5094                  * is received within RTT/2 interval; the assumption here is
5095                  * that the ACK was already in flight.  See "On Estimating
5096                  * End-to-End Network Path Properties" by Allman and Paxson
5097                  * for more details.
5098                  */
5099                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5100                 if (!IN_RECOVERY(tp->t_flags)) {
5101                         tp->snd_cwnd_prev = tp->snd_cwnd;
5102                         tp->snd_ssthresh_prev = tp->snd_ssthresh;
5103                         tp->snd_recover_prev = tp->snd_recover;
5104                         tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
5105                         tp->t_flags |= TF_PREVVALID;
5106                 } else {
5107                         tp->t_flags &= ~TF_PREVVALID;
5108                 }
5109                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5110         } else {
5111                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5112                 tp->t_flags &= ~TF_PREVVALID;
5113         }
5114         TCPSTAT_INC(tcps_rexmttimeo);
5115         if ((tp->t_state == TCPS_SYN_SENT) ||
5116             (tp->t_state == TCPS_SYN_RECEIVED))
5117                 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift];
5118         else
5119                 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
5120         TCPT_RANGESET(tp->t_rxtcur, rexmt,
5121             MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms),
5122             MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
5123         /*
5124          * We enter the path for PLMTUD if connection is established or, if
5125          * connection is FIN_WAIT_1 status, reason for the last is that if
5126          * amount of data we send is very small, we could send it in couple
5127          * of packets and process straight to FIN. In that case we won't
5128          * catch ESTABLISHED state.
5129          */
5130         if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED))
5131             || (tp->t_state == TCPS_FIN_WAIT_1))) {
5132 #ifdef INET6
5133                 int32_t isipv6;
5134 #endif
5135
5136                 /*
5137                  * Idea here is that at each stage of mtu probe (usually,
5138                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
5139                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
5140                  * should take care of that.
5141                  */
5142                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
5143                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
5144                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
5145                     tp->t_rxtshift % 2 == 0)) {
5146                         /*
5147                          * Enter Path MTU Black-hole Detection mechanism: -
5148                          * Disable Path MTU Discovery (IP "DF" bit). -
5149                          * Reduce MTU to lower value than what we negotiated
5150                          * with peer.
5151                          */
5152                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
5153                                 /*
5154                                  * Record that we may have found a black
5155                                  * hole.
5156                                  */
5157                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
5158                                 /* Keep track of previous MSS. */
5159                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
5160                         }
5161                         /*
5162                          * Reduce the MSS to blackhole value or to the
5163                          * default in an attempt to retransmit.
5164                          */
5165 #ifdef INET6
5166                         isipv6 = bbr->r_is_v6;
5167                         if (isipv6 &&
5168                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
5169                                 /* Use the sysctl tuneable blackhole MSS. */
5170                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
5171                                 TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5172                         } else if (isipv6) {
5173                                 /* Use the default MSS. */
5174                                 tp->t_maxseg = V_tcp_v6mssdflt;
5175                                 /*
5176                                  * Disable Path MTU Discovery when we switch
5177                                  * to minmss.
5178                                  */
5179                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5180                                 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5181                         }
5182 #endif
5183 #if defined(INET6) && defined(INET)
5184                         else
5185 #endif
5186 #ifdef INET
5187                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
5188                                 /* Use the sysctl tuneable blackhole MSS. */
5189                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
5190                                 TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5191                         } else {
5192                                 /* Use the default MSS. */
5193                                 tp->t_maxseg = V_tcp_mssdflt;
5194                                 /*
5195                                  * Disable Path MTU Discovery when we switch
5196                                  * to minmss.
5197                                  */
5198                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5199                                 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5200                         }
5201 #endif
5202                 } else {
5203                         /*
5204                          * If further retransmissions are still unsuccessful
5205                          * with a lowered MTU, maybe this isn't a blackhole
5206                          * and we restore the previous MSS and blackhole
5207                          * detection flags. The limit '6' is determined by
5208                          * giving each probe stage (1448, 1188, 524) 2
5209                          * chances to recover.
5210                          */
5211                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
5212                             (tp->t_rxtshift >= 6)) {
5213                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
5214                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
5215                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
5216                                 TCPSTAT_INC(tcps_pmtud_blackhole_failed);
5217                         }
5218                 }
5219         }
5220         /*
5221          * Disable RFC1323 and SACK if we haven't got any response to our
5222          * third SYN to work-around some broken terminal servers (most of
5223          * which have hopefully been retired) that have bad VJ header
5224          * compression code which trashes TCP segments containing
5225          * unknown-to-them TCP options.
5226          */
5227         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
5228             (tp->t_rxtshift == 3))
5229                 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT);
5230         /*
5231          * If we backed off this far, our srtt estimate is probably bogus.
5232          * Clobber it so we'll take the next rtt measurement as our srtt;
5233          * move the current srtt into rttvar to keep the current retransmit
5234          * times until then.
5235          */
5236         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
5237 #ifdef INET6
5238                 if (bbr->r_is_v6)
5239                         in6_losing(tp->t_inpcb);
5240                 else
5241 #endif
5242                         in_losing(tp->t_inpcb);
5243                 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
5244                 tp->t_srtt = 0;
5245         }
5246         sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
5247         tp->snd_recover = tp->snd_max;
5248         tp->t_flags |= TF_ACKNOW;
5249         tp->t_rtttime = 0;
5250 out:
5251         return (retval);
5252 }
5253
5254 static int
5255 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling)
5256 {
5257         int32_t ret = 0;
5258         int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
5259
5260         if (timers == 0) {
5261                 return (0);
5262         }
5263         if (tp->t_state == TCPS_LISTEN) {
5264                 /* no timers on listen sockets */
5265                 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
5266                         return (0);
5267                 return (1);
5268         }
5269         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
5270                 uint32_t left;
5271
5272                 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
5273                         ret = -1;
5274                         bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5275                         return (0);
5276                 }
5277                 if (hpts_calling == 0) {
5278                         ret = -2;
5279                         bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5280                         return (0);
5281                 }
5282                 /*
5283                  * Ok our timer went off early and we are not paced false
5284                  * alarm, go back to sleep.
5285                  */
5286                 left = bbr->r_ctl.rc_timer_exp - cts;
5287                 ret = -3;
5288                 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling);
5289                 tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left));
5290                 return (1);
5291         }
5292         bbr->rc_tmr_stopped = 0;
5293         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
5294         if (timers & PACE_TMR_DELACK) {
5295                 ret = bbr_timeout_delack(tp, bbr, cts);
5296         } else if (timers & PACE_TMR_PERSIT) {
5297                 ret = bbr_timeout_persist(tp, bbr, cts);
5298         } else if (timers & PACE_TMR_RACK) {
5299                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5300                 ret = bbr_timeout_rack(tp, bbr, cts);
5301         } else if (timers & PACE_TMR_TLP) {
5302                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5303                 ret = bbr_timeout_tlp(tp, bbr, cts);
5304         } else if (timers & PACE_TMR_RXT) {
5305                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5306                 ret = bbr_timeout_rxt(tp, bbr, cts);
5307         } else if (timers & PACE_TMR_KEEP) {
5308                 ret = bbr_timeout_keepalive(tp, bbr, cts);
5309         }
5310         bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling);
5311         return (ret);
5312 }
5313
5314 static void
5315 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts)
5316 {
5317         if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
5318                 uint8_t hpts_removed = 0;
5319
5320                 if (bbr->rc_inp->inp_in_hpts &&
5321                     (bbr->rc_timer_first == 1)) {
5322                         /*
5323                          * If we are canceling timer's when we have the
5324                          * timer ahead of the output being paced. We also
5325                          * must remove ourselves from the hpts.
5326                          */
5327                         hpts_removed = 1;
5328                         tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
5329                         if (bbr->r_ctl.rc_last_delay_val) {
5330                                 /* Update the last hptsi delay too */
5331                                 uint32_t time_since_send;
5332
5333                                 if (TSTMP_GT(cts, bbr->rc_pacer_started))
5334                                         time_since_send = cts - bbr->rc_pacer_started;
5335                                 else
5336                                         time_since_send = 0;
5337                                 if (bbr->r_ctl.rc_last_delay_val > time_since_send) {
5338                                         /* Cut down our slot time */
5339                                         bbr->r_ctl.rc_last_delay_val -= time_since_send;
5340                                 } else {
5341                                         bbr->r_ctl.rc_last_delay_val = 0;
5342                                 }
5343                                 bbr->rc_pacer_started = cts;
5344                         }
5345                 }
5346                 bbr->rc_timer_first = 0;
5347                 bbr_log_to_cancel(bbr, line, cts, hpts_removed);
5348                 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
5349                 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
5350         }
5351 }
5352
5353 static void
5354 bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type)
5355 {
5356         struct tcp_bbr *bbr;
5357
5358         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
5359         bbr->rc_all_timers_stopped = 1;
5360         return;
5361 }
5362
5363 /*
5364  * stop all timers always returning 0.
5365  */
5366 static int
5367 bbr_stopall(struct tcpcb *tp)
5368 {
5369         return (0);
5370 }
5371
5372 static void
5373 bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
5374 {
5375         return;
5376 }
5377
5378 /*
5379  * return true if a bbr timer (rack or tlp) is active.
5380  */
5381 static int
5382 bbr_timer_active(struct tcpcb *tp, uint32_t timer_type)
5383 {
5384         return (0);
5385 }
5386
5387 static uint32_t
5388 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts)
5389 {
5390         struct bbr_sendmap *rsm;
5391
5392         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
5393         if ((rsm == NULL) || (u_rsm == rsm))
5394                 return (cts);
5395         return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
5396 }
5397
5398 static void
5399 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr,
5400      struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time)
5401 {
5402         int32_t idx;
5403
5404         rsm->r_rtr_cnt++;
5405         rsm->r_dupack = 0;
5406         if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) {
5407                 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS;
5408                 rsm->r_flags |= BBR_OVERMAX;
5409         }
5410         if (rsm->r_flags & BBR_RWND_COLLAPSED) {
5411                 /* Take off the collapsed flag at rxt */
5412                 rsm->r_flags &= ~BBR_RWND_COLLAPSED;
5413         }
5414         if (rsm->r_flags & BBR_MARKED_LOST) {
5415                 /* We have retransmitted, its no longer lost */
5416                 rsm->r_flags &= ~BBR_MARKED_LOST;
5417                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
5418         }
5419         if (rsm->r_flags & BBR_RXT_CLEARED) {
5420                 /*
5421                  * We hit a RXT timer on it and
5422                  * we cleared the "acked" flag.
5423                  * We now have it going back into
5424                  * flight, we can remove the cleared
5425                  * flag and possibly do accounting on
5426                  * this piece.
5427                  */
5428                 rsm->r_flags &= ~BBR_RXT_CLEARED;
5429         }
5430         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) {
5431                 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
5432                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
5433         }
5434         idx = rsm->r_rtr_cnt - 1;
5435         rsm->r_tim_lastsent[idx] = cts;
5436         rsm->r_pacing_delay = pacing_time;
5437         rsm->r_delivered = bbr->r_ctl.rc_delivered;
5438         rsm->r_ts_valid = bbr->rc_ts_valid;
5439         if (bbr->rc_ts_valid)
5440                 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
5441         if (bbr->r_ctl.r_app_limited_until)
5442                 rsm->r_app_limited = 1;
5443         else
5444                 rsm->r_app_limited = 0;
5445         if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
5446                 rsm->r_bbr_state = bbr_state_val(bbr);
5447         else
5448                 rsm->r_bbr_state = 8;
5449         if (rsm->r_flags & BBR_ACKED) {
5450                 /* Problably MTU discovery messing with us */
5451                 uint32_t old_flags;
5452
5453                 old_flags = rsm->r_flags;
5454                 rsm->r_flags &= ~BBR_ACKED;
5455                 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
5456                 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
5457                 if (bbr->r_ctl.rc_sacked == 0)
5458                         bbr->r_ctl.rc_sacklast = NULL;
5459         }
5460         if (rsm->r_in_tmap) {
5461                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5462         }
5463         TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5464         rsm->r_in_tmap = 1;
5465         if (rsm->r_flags & BBR_SACK_PASSED) {
5466                 /* We have retransmitted due to the SACK pass */
5467                 rsm->r_flags &= ~BBR_SACK_PASSED;
5468                 rsm->r_flags |= BBR_WAS_SACKPASS;
5469         }
5470         rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
5471         rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
5472                                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
5473         bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
5474         if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
5475                 rsm->r_is_gain = 1;
5476                 rsm->r_is_drain = 0;
5477         } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
5478                 rsm->r_is_drain = 1;
5479                 rsm->r_is_gain = 0;
5480         } else {
5481                 rsm->r_is_drain = 0;
5482                 rsm->r_is_gain = 0;
5483         }
5484         rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */
5485 }
5486
5487 /*
5488  * Returns 0, or the sequence where we stopped
5489  * updating. We also update the lenp to be the amount
5490  * of data left.
5491  */
5492
5493 static uint32_t
5494 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr,
5495     struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time)
5496 {
5497         /*
5498          * We (re-)transmitted starting at rsm->r_start for some length
5499          * (possibly less than r_end.
5500          */
5501         struct bbr_sendmap *nrsm;
5502         uint32_t c_end;
5503         int32_t len;
5504
5505         len = *lenp;
5506         c_end = rsm->r_start + len;
5507         if (SEQ_GEQ(c_end, rsm->r_end)) {
5508                 /*
5509                  * We retransmitted the whole piece or more than the whole
5510                  * slopping into the next rsm.
5511                  */
5512                 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5513                 if (c_end == rsm->r_end) {
5514                         *lenp = 0;
5515                         return (0);
5516                 } else {
5517                         int32_t act_len;
5518
5519                         /* Hangs over the end return whats left */
5520                         act_len = rsm->r_end - rsm->r_start;
5521                         *lenp = (len - act_len);
5522                         return (rsm->r_end);
5523                 }
5524                 /* We don't get out of this block. */
5525         }
5526         /*
5527          * Here we retransmitted less than the whole thing which means we
5528          * have to split this into what was transmitted and what was not.
5529          */
5530         nrsm = bbr_alloc_full_limit(bbr);
5531         if (nrsm == NULL) {
5532                 *lenp = 0;
5533                 return (0);
5534         }
5535         /*
5536          * So here we are going to take the original rsm and make it what we
5537          * retransmitted. nrsm will be the tail portion we did not
5538          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
5539          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
5540          * 1, 6 and the new piece will be 6, 11.
5541          */
5542         bbr_clone_rsm(bbr, nrsm, rsm, c_end);
5543         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
5544         nrsm->r_dupack = 0;
5545         if (rsm->r_in_tmap) {
5546                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
5547                 nrsm->r_in_tmap = 1;
5548         }
5549         rsm->r_flags &= (~BBR_HAS_FIN);
5550         bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5551         *lenp = 0;
5552         return (0);
5553 }
5554
5555 static uint64_t
5556 bbr_get_hardware_rate(struct tcp_bbr *bbr)
5557 {
5558         uint64_t bw;
5559
5560         bw = bbr_get_bw(bbr);
5561         bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN];
5562         bw /= (uint64_t)BBR_UNIT;
5563         return(bw);
5564 }
5565
5566 static void
5567 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts,
5568                        uint64_t act_rate, uint64_t rate_wanted)
5569 {
5570         /*
5571          * We could not get a full gains worth
5572          * of rate.
5573          */
5574         if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) {
5575                 /* we can't even get the real rate */
5576                 uint64_t red;
5577
5578                 bbr->skip_gain = 1;
5579                 bbr->gain_is_limited = 0;
5580                 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate;
5581                 if (red)
5582                         filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts);
5583         } else {
5584                 /* We can use a lower gain */
5585                 bbr->skip_gain = 0;
5586                 bbr->gain_is_limited = 1;
5587         }
5588 }
5589
5590 static void
5591 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts)
5592 {
5593         const struct tcp_hwrate_limit_table *nrte;
5594         int error, rate = -1;
5595
5596         if (bbr->r_ctl.crte == NULL)
5597                 return;
5598         if ((bbr->rc_inp->inp_route.ro_rt == NULL) ||
5599             (bbr->rc_inp->inp_route.ro_rt->rt_ifp == NULL)) {
5600                 /* Lost our routes? */
5601                 /* Clear the way for a re-attempt */
5602                 bbr->bbr_attempt_hdwr_pace = 0;
5603 lost_rate:
5604                 bbr->gain_is_limited = 0;
5605                 bbr->skip_gain = 0;
5606                 bbr->bbr_hdrw_pacing = 0;
5607                 counter_u64_add(bbr_flows_whdwr_pacing, -1);
5608                 counter_u64_add(bbr_flows_nohdwr_pacing, 1);
5609                 tcp_bbr_tso_size_check(bbr, cts);
5610                 return;
5611         }
5612         rate = bbr_get_hardware_rate(bbr);
5613         nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte,
5614                                    bbr->rc_tp,
5615                                    bbr->rc_inp->inp_route.ro_rt->rt_ifp,
5616                                    rate,
5617                                    (RS_PACING_GEQ|RS_PACING_SUB_OK),
5618                                    &error);
5619         if (nrte == NULL) {
5620                 goto lost_rate;
5621         }
5622         if (nrte != bbr->r_ctl.crte) {
5623                 bbr->r_ctl.crte = nrte;
5624                 if (error == 0)  {
5625                         BBR_STAT_INC(bbr_hdwr_rl_mod_ok);
5626                         if (bbr->r_ctl.crte->rate < rate) {
5627                                 /* We have a problem */
5628                                 bbr_setup_less_of_rate(bbr, cts,
5629                                                        bbr->r_ctl.crte->rate, rate);
5630                         } else {
5631                                 /* We are good */
5632                                 bbr->gain_is_limited = 0;
5633                                 bbr->skip_gain = 0;
5634                         }
5635                 } else {
5636                         /* A failure should release the tag */
5637                         BBR_STAT_INC(bbr_hdwr_rl_mod_fail);
5638                         bbr->gain_is_limited = 0;
5639                         bbr->skip_gain = 0;
5640                         bbr->bbr_hdrw_pacing = 0;
5641                 }
5642                 bbr_type_log_hdwr_pacing(bbr,
5643                                          bbr->r_ctl.crte->ptbl->rs_ifp,
5644                                          rate,
5645                                          ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate),
5646                                          __LINE__,
5647                                          cts,
5648                                          error);
5649         }
5650 }
5651
5652 static void
5653 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts)
5654 {
5655         /*
5656          * If we have hardware pacing support
5657          * we need to factor that in for our
5658          * TSO size.
5659          */
5660         const struct tcp_hwrate_limit_table *rlp;
5661         uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay;
5662
5663         if ((bbr->bbr_hdrw_pacing == 0) ||
5664             (IN_RECOVERY(bbr->rc_tp->t_flags)) ||
5665             (bbr->r_ctl.crte == NULL))
5666                 return;
5667         if (bbr->hw_pacing_set == 0) {
5668                 /* Not yet by the hdwr pacing count delay */
5669                 return;
5670         }
5671         if (bbr_hdwr_pace_adjust == 0) {
5672                 /* No adjustment */
5673                 return;
5674         }
5675         rlp = bbr->r_ctl.crte;
5676         if (bbr->rc_tp->t_maxseg > bbr->rc_last_options)
5677                 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5678         else
5679                 maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5680         /*
5681          * So lets first get the
5682          * time we will take between
5683          * TSO sized sends currently without
5684          * hardware help.
5685          */
5686         cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT,
5687                         bbr->r_ctl.rc_pace_max_segs, cts, 1);
5688         hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg;
5689         hdwr_delay *= rlp->time_between;
5690         if (cur_delay > hdwr_delay)
5691                 delta = cur_delay - hdwr_delay;
5692         else
5693                 delta = 0;
5694         bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay,
5695                              (bbr->r_ctl.rc_pace_max_segs / maxseg),
5696                              1);
5697         if (delta &&
5698             (delta < (max(rlp->time_between,
5699                           bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) {
5700                 /*
5701                  * Now lets divide by the pacing
5702                  * time between each segment the
5703                  * hardware sends rounding up and
5704                  * derive a bytes from that. We multiply
5705                  * that by bbr_hdwr_pace_adjust to get
5706                  * more bang for our buck.
5707                  *
5708                  * The goal is to have the software pacer
5709                  * waiting no more than an additional
5710                  * pacing delay if we can (without the
5711                  * compensation i.e. x bbr_hdwr_pace_adjust).
5712                  */
5713                 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between),
5714                              (bbr->r_ctl.rc_pace_max_segs/maxseg));
5715                 seg_sz *= bbr_hdwr_pace_adjust;
5716                 if (bbr_hdwr_pace_floor &&
5717                     (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5718                         /* Currently hardware paces
5719                          * out rs_min_seg segments at a time.
5720                          * We need to make sure we always send at least
5721                          * a full burst of bbr_hdwr_pace_floor down.
5722                          */
5723                         seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5724                 }
5725                 seg_sz *= maxseg;
5726         } else if (delta == 0) {
5727                 /*
5728                  * The highest pacing rate is
5729                  * above our b/w gained. This means
5730                  * we probably are going quite fast at
5731                  * the hardware highest rate. Lets just multiply
5732                  * the calculated TSO size by the
5733                  * multiplier factor (its probably
5734                  * 4 segments in the default config for
5735                  * mlx).
5736                  */
5737                 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust;
5738                 if (bbr_hdwr_pace_floor &&
5739                     (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5740                         /* Currently hardware paces
5741                          * out rs_min_seg segments at a time.
5742                          * We need to make sure we always send at least
5743                          * a full burst of bbr_hdwr_pace_floor down.
5744                          */
5745                         seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5746                 }
5747         } else {
5748                 /*
5749                  * The pacing time difference is so
5750                  * big that the hardware will
5751                  * pace out more rapidly then we
5752                  * really want and then we
5753                  * will have a long delay. Lets just keep
5754                  * the same TSO size so its as if
5755                  * we were not using hdwr pacing (we
5756                  * just gain a bit of spacing from the
5757                  * hardware if seg_sz > 1).
5758                  */
5759                 seg_sz = bbr->r_ctl.rc_pace_max_segs;
5760         }
5761         if (seg_sz > bbr->r_ctl.rc_pace_max_segs)
5762                 new_tso = seg_sz;
5763         else
5764                 new_tso = bbr->r_ctl.rc_pace_max_segs;
5765         if (new_tso >= (PACE_MAX_IP_BYTES-maxseg))
5766                 new_tso = PACE_MAX_IP_BYTES - maxseg;
5767
5768         if (new_tso != bbr->r_ctl.rc_pace_max_segs) {
5769                 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0);
5770                 bbr->r_ctl.rc_pace_max_segs = new_tso;
5771         }
5772 }
5773
5774 static void
5775 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts)
5776 {
5777         uint64_t bw;
5778         uint32_t old_tso = 0, new_tso;
5779         uint32_t maxseg, bytes;
5780         uint32_t tls_seg=0;
5781         /*
5782          * Google/linux uses the following algorithm to determine
5783          * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18):
5784          *
5785          *  bytes = bw_in_bytes_per_second / 1000
5786          *  bytes = min(bytes, 64k)
5787          *  tso_segs = bytes / MSS
5788          *  if (bw < 1.2Mbs)
5789          *      min_tso_segs = 1
5790          *  else
5791          *      min_tso_segs = 2
5792          * tso_segs = max(tso_segs, min_tso_segs)
5793          *
5794          * * Note apply a device specific limit (we apply this in the
5795          *   tcp_m_copym).
5796          * Note that before the initial measurement is made google bursts out
5797          * a full iwnd just like new-reno/cubic.
5798          *
5799          * We do not use this algorithm. Instead we
5800          * use a two phased approach:
5801          *
5802          *  if ( bw <= per-tcb-cross-over)
5803          *     goal_tso =  calculate how much with this bw we
5804          *                 can send in goal-time seconds.
5805          *     if (goal_tso > mss)
5806          *         seg = goal_tso / mss
5807          *         tso = seg * mss
5808          *     else
5809          *         tso = mss
5810          *     if (tso > per-tcb-max)
5811          *         tso = per-tcb-max
5812          *  else if ( bw > 512Mbps)
5813          *     tso = max-tso (64k/mss)
5814          *  else
5815          *     goal_tso = bw / per-tcb-divsor
5816          *     seg = (goal_tso + mss-1)/mss
5817          *     tso = seg * mss
5818          *
5819          * if (tso < per-tcb-floor)
5820          *    tso = per-tcb-floor
5821          * if (tso > per-tcb-utter_max)
5822          *    tso = per-tcb-utter_max
5823          *
5824          * Note the default per-tcb-divisor is 1000 (same as google).
5825          * the goal cross over is 30Mbps however. To recreate googles
5826          * algorithm you need to set:
5827          *
5828          * cross-over = 23,168,000 bps
5829          * goal-time = 18000
5830          * per-tcb-max = 2
5831          * per-tcb-divisor = 1000
5832          * per-tcb-floor = 1
5833          *
5834          * This will get you "google bbr" behavior with respect to tso size.
5835          *
5836          * Note we do set anything TSO size until we are past the initial
5837          * window. Before that we gnerally use either a single MSS
5838          * or we use the full IW size (so we burst a IW at a time)
5839          * Also note that Hardware-TLS is special and does alternate
5840          * things to minimize PCI Bus Bandwidth use.
5841          */
5842
5843         if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) {
5844                 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5845         } else {
5846                 maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5847         }
5848 #ifdef KERN_TLS
5849         if (bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) {
5850                 tls_seg =  ctf_get_opt_tls_size(bbr->rc_inp->inp_socket, bbr->rc_tp->snd_wnd);
5851                 bbr->r_ctl.rc_pace_min_segs = (tls_seg + bbr->rc_last_options);
5852         }
5853 #endif
5854         old_tso = bbr->r_ctl.rc_pace_max_segs;
5855         if (bbr->rc_past_init_win == 0) {
5856                 /*
5857                  * Not enough data has been acknowledged to make a
5858                  * judgement unless we are hardware TLS. Set up
5859                  * the inital TSO based on if we are sending a
5860                  * full IW at once or not.
5861                  */
5862                 if (bbr->rc_use_google)
5863                         bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2);
5864                 else if (bbr->bbr_init_win_cheat)
5865                         bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp);
5866                 else
5867                         bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5868                 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg)
5869                         bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg;
5870 #ifdef KERN_TLS
5871                 if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) && tls_seg) {
5872                         /*
5873                          * For hardware TLS we set our min to the tls_seg size.
5874                          */
5875                         bbr->r_ctl.rc_pace_max_segs = tls_seg;
5876                         bbr->r_ctl.rc_pace_min_segs = tls_seg + bbr->rc_last_options;
5877                 }
5878 #endif
5879                 if (bbr->r_ctl.rc_pace_max_segs == 0) {
5880                         bbr->r_ctl.rc_pace_max_segs = maxseg;
5881                 }
5882                 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0);
5883 #ifdef KERN_TLS
5884                 if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) == 0)
5885 #endif
5886                         bbr_adjust_for_hw_pacing(bbr, cts);
5887                 return;
5888         }
5889         /**
5890          * Now lets set the TSO goal based on our delivery rate in
5891          * bytes per second. Note we only do this if
5892          * we have acked at least the initial cwnd worth of data.
5893          */
5894         bw = bbr_get_bw(bbr);
5895         if (IN_RECOVERY(bbr->rc_tp->t_flags) &&
5896              (bbr->rc_use_google == 0)) {
5897                 /* We clamp to one MSS in recovery */
5898                 new_tso = maxseg;
5899         } else if (bbr->rc_use_google) {
5900                 int min_tso_segs;
5901
5902                 /* Google considers the gain too */
5903                 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) {
5904                         bw *= bbr->r_ctl.rc_bbr_hptsi_gain;
5905                         bw /= BBR_UNIT;
5906                 }
5907                 bytes = bw / 1024;
5908                 if (bytes > (64 * 1024))
5909                         bytes = 64 * 1024;
5910                 new_tso = bytes / maxseg;
5911                 if (bw < ONE_POINT_TWO_MEG)
5912                         min_tso_segs = 1;
5913                 else
5914                         min_tso_segs = 2;
5915                 if (new_tso < min_tso_segs)
5916                         new_tso = min_tso_segs;
5917                 new_tso *= maxseg;
5918         } else if (bbr->rc_no_pacing) {
5919                 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg;
5920         } else if (bw <= bbr->r_ctl.bbr_cross_over) {
5921                 /*
5922                  * Calculate the worse case b/w TSO if we are inserting no
5923                  * more than a delay_target number of TSO's.
5924                  */
5925                 uint32_t tso_len, min_tso;
5926
5927                 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw);
5928                 if (tso_len > maxseg) {
5929                         new_tso = tso_len / maxseg;
5930                         if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max)
5931                                 new_tso = bbr->r_ctl.bbr_hptsi_segments_max;
5932                         new_tso *= maxseg;
5933                 } else {
5934                         /*
5935                          * less than a full sized frame yikes.. long rtt or
5936                          * low bw?
5937                          */
5938                         min_tso = bbr_minseg(bbr);
5939                         if ((tso_len > min_tso) && (bbr_all_get_min == 0))
5940                                 new_tso = rounddown(tso_len, min_tso);
5941                         else
5942                                 new_tso = min_tso;
5943                 }
5944         } else if (bw > FIVETWELVE_MBPS) {
5945                 /*
5946                  * This guy is so fast b/w wise that we can TSO as large as
5947                  * possible of segments that the NIC will allow.
5948                  */
5949                 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5950         } else {
5951                 /*
5952                  * This formula is based on attempting to send a segment or
5953                  * more every bbr_hptsi_per_second. The default is 1000
5954                  * which means you are targeting what you can send every 1ms
5955                  * based on the peers bw.
5956                  *
5957                  * If the number drops to say 500, then you are looking more
5958                  * at 2ms and you will raise how much we send in a single
5959                  * TSO thus saving CPU (less bbr_output_wtime() calls). The
5960                  * trade off of course is you will send more at once and
5961                  * thus tend to clump up the sends into larger "bursts"
5962                  * building a queue.
5963                  */
5964                 bw /= bbr->r_ctl.bbr_hptsi_per_second;
5965                 new_tso = roundup(bw, (uint64_t)maxseg);
5966                 /*
5967                  * Gate the floor to match what our lower than 48Mbps
5968                  * algorithm does. The ceiling (bbr_hptsi_segments_max) thus
5969                  * becomes the floor for this calculation.
5970                  */
5971                 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg))
5972                         new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg);
5973         }
5974         if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor)))
5975                 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor;
5976         if (new_tso > PACE_MAX_IP_BYTES)
5977                 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5978         /* Enforce an utter maximum if we are not HW-TLS */
5979 #ifdef KERN_TLS
5980         if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) == 0)
5981 #endif
5982                 if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) {
5983                         new_tso = bbr->r_ctl.bbr_utter_max * maxseg;
5984                 }
5985 #ifdef KERN_TLS
5986         if (tls_seg) {
5987                 /*
5988                  * Lets move the output size
5989                  * up to 1 or more TLS record sizes.
5990                  */
5991                 uint32_t temp;
5992
5993                 temp = roundup(new_tso, tls_seg);
5994                 new_tso = temp;
5995                 /* Back down if needed to under a full frame */
5996                 while (new_tso > PACE_MAX_IP_BYTES)
5997                         new_tso -= tls_seg;
5998         }
5999 #endif
6000         if (old_tso != new_tso) {
6001                 /* Only log changes */
6002                 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0);
6003                 bbr->r_ctl.rc_pace_max_segs = new_tso;
6004         }
6005 #ifdef KERN_TLS
6006         if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) &&
6007              tls_seg) {
6008                 bbr->r_ctl.rc_pace_min_segs = tls_seg + bbr->rc_last_options;
6009         } else
6010 #endif
6011                 /* We have hardware pacing and not hardware TLS! */
6012                 bbr_adjust_for_hw_pacing(bbr, cts);
6013 }
6014
6015 static void
6016 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len,
6017     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t cts,
6018     struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc,
6019     struct sockbuf *sb)
6020 {
6021
6022         struct bbr_sendmap *rsm, *nrsm;
6023         register uint32_t snd_max, snd_una;
6024         uint32_t pacing_time;
6025         /*
6026          * Add to the RACK log of packets in flight or retransmitted. If
6027          * there is a TS option we will use the TS echoed, if not we will
6028          * grab a TS.
6029          *
6030          * Retransmissions will increment the count and move the ts to its
6031          * proper place. Note that if options do not include TS's then we
6032          * won't be able to effectively use the ACK for an RTT on a retran.
6033          *
6034          * Notes about r_start and r_end. Lets consider a send starting at
6035          * sequence 1 for 10 bytes. In such an example the r_start would be
6036          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
6037          * This means that r_end is actually the first sequence for the next
6038          * slot (11).
6039          *
6040          */
6041         INP_WLOCK_ASSERT(tp->t_inpcb);
6042         if (err) {
6043                 /*
6044                  * We don't log errors -- we could but snd_max does not
6045                  * advance in this case either.
6046                  */
6047                 return;
6048         }
6049         if (th_flags & TH_RST) {
6050                 /*
6051                  * We don't log resets and we return immediately from
6052                  * sending
6053                  */
6054                 *abandon = 1;
6055                 return;
6056         }
6057         snd_una = tp->snd_una;
6058         if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) {
6059                 /*
6060                  * The call to bbr_log_output is made before bumping
6061                  * snd_max. This means we can record one extra byte on a SYN
6062                  * or FIN if seq_out is adding more on and a FIN is present
6063                  * (and we are not resending).
6064                  */
6065                 if (th_flags & TH_SYN)
6066                         len++;
6067                 if (th_flags & TH_FIN)
6068                         len++;
6069         }
6070         if (SEQ_LEQ((seq_out + len), snd_una)) {
6071                 /* Are sending an old segment to induce an ack (keep-alive)? */
6072                 return;
6073         }
6074         if (SEQ_LT(seq_out, snd_una)) {
6075                 /* huh? should we panic? */
6076                 uint32_t end;
6077
6078                 end = seq_out + len;
6079                 seq_out = snd_una;
6080                 len = end - seq_out;
6081         }
6082         snd_max = tp->snd_max;
6083         if (len == 0) {
6084                 /* We don't log zero window probes */
6085                 return;
6086         }
6087         pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1);
6088         /* First question is it a retransmission? */
6089         if (seq_out == snd_max) {
6090 again:
6091                 rsm = bbr_alloc(bbr);
6092                 if (rsm == NULL) {
6093                         return;
6094                 }
6095                 rsm->r_flags = 0;
6096                 if (th_flags & TH_SYN)
6097                         rsm->r_flags |= BBR_HAS_SYN;
6098                 if (th_flags & TH_FIN)
6099                         rsm->r_flags |= BBR_HAS_FIN;
6100                 rsm->r_tim_lastsent[0] = cts;
6101                 rsm->r_rtr_cnt = 1;
6102                 rsm->r_rtr_bytes = 0;
6103                 rsm->r_start = seq_out;
6104                 rsm->r_end = rsm->r_start + len;
6105                 rsm->r_dupack = 0;
6106                 rsm->r_delivered = bbr->r_ctl.rc_delivered;
6107                 rsm->r_pacing_delay = pacing_time;
6108                 rsm->r_ts_valid = bbr->rc_ts_valid;
6109                 if (bbr->rc_ts_valid)
6110                         rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
6111                 rsm->r_del_time = bbr->r_ctl.rc_del_time;
6112                 if (bbr->r_ctl.r_app_limited_until)
6113                         rsm->r_app_limited = 1;
6114                 else
6115                         rsm->r_app_limited = 0;
6116                 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
6117                 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
6118                                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
6119                 /*
6120                  * Here we must also add in this rsm since snd_max
6121                  * is updated after we return from a new send.
6122                  */
6123                 rsm->r_flight_at_send += len;
6124                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
6125                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
6126                 rsm->r_in_tmap = 1;
6127                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
6128                         rsm->r_bbr_state = bbr_state_val(bbr);
6129                 else
6130                         rsm->r_bbr_state = 8;
6131                 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
6132                         rsm->r_is_gain = 1;
6133                         rsm->r_is_drain = 0;
6134                 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
6135                         rsm->r_is_drain = 1;
6136                         rsm->r_is_gain = 0;
6137                 } else {
6138                         rsm->r_is_drain = 0;
6139                         rsm->r_is_gain = 0;
6140                 }
6141                 return;
6142         }
6143         /*
6144          * If we reach here its a retransmission and we need to find it.
6145          */
6146 more:
6147         if (hintrsm && (hintrsm->r_start == seq_out)) {
6148                 rsm = hintrsm;
6149                 hintrsm = NULL;
6150         } else if (bbr->r_ctl.rc_next) {
6151                 /* We have a hint from a previous run */
6152                 rsm = bbr->r_ctl.rc_next;
6153         } else {
6154                 /* No hints sorry */
6155                 rsm = NULL;
6156         }
6157         if ((rsm) && (rsm->r_start == seq_out)) {
6158                 /*
6159                  * We used rc_next or hintrsm  to retransmit, hopefully the
6160                  * likely case.
6161                  */
6162                 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6163                 if (len == 0) {
6164                         return;
6165                 } else {
6166                         goto more;
6167                 }
6168         }
6169         /* Ok it was not the last pointer go through it the hard way. */
6170         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6171                 if (rsm->r_start == seq_out) {
6172                         seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6173                         bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
6174                         if (len == 0) {
6175                                 return;
6176                         } else {
6177                                 continue;
6178                         }
6179                 }
6180                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
6181                         /* Transmitted within this piece */
6182                         /*
6183                          * Ok we must split off the front and then let the
6184                          * update do the rest
6185                          */
6186                         nrsm = bbr_alloc_full_limit(bbr);
6187                         if (nrsm == NULL) {
6188                                 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
6189                                 return;
6190                         }
6191                         /*
6192                          * copy rsm to nrsm and then trim the front of rsm
6193                          * to not include this part.
6194                          */
6195                         bbr_clone_rsm(bbr, nrsm, rsm, seq_out);
6196                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
6197                         if (rsm->r_in_tmap) {
6198                                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6199                                 nrsm->r_in_tmap = 1;
6200                         }
6201                         rsm->r_flags &= (~BBR_HAS_FIN);
6202                         seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time);
6203                         if (len == 0) {
6204                                 return;
6205                         }
6206                 }
6207         }
6208         /*
6209          * Hmm not found in map did they retransmit both old and on into the
6210          * new?
6211          */
6212         if (seq_out == tp->snd_max) {
6213                 goto again;
6214         } else if (SEQ_LT(seq_out, tp->snd_max)) {
6215 #ifdef BBR_INVARIANTS
6216                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
6217                     seq_out, len, tp->snd_una, tp->snd_max);
6218                 printf("Starting Dump of all rack entries\n");
6219                 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6220                         printf("rsm:%p start:%u end:%u\n",
6221                             rsm, rsm->r_start, rsm->r_end);
6222                 }
6223                 printf("Dump complete\n");
6224                 panic("seq_out not found rack:%p tp:%p",
6225                     bbr, tp);
6226 #endif
6227         } else {
6228 #ifdef BBR_INVARIANTS
6229                 /*
6230                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
6231                  * flag)
6232                  */
6233                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
6234                     seq_out, len, tp->snd_max, tp);
6235 #endif
6236         }
6237 }
6238
6239 static void
6240 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt)
6241 {
6242         /*
6243          * Collapse timeout back the cum-ack moved.
6244          */
6245         tp->t_rxtshift = 0;
6246         tp->t_softerror = 0;
6247 }
6248
6249
6250 static void
6251 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin)
6252 {
6253         bbr->rtt_valid = 1;
6254         bbr->r_ctl.cur_rtt = rtt_usecs;
6255         bbr->r_ctl.ts_in = tsin;
6256         if (rsm_send_time)
6257                 bbr->r_ctl.cur_rtt_send_time = rsm_send_time;
6258 }
6259
6260 static void
6261 bbr_make_timestamp_determination(struct tcp_bbr *bbr)
6262 {
6263         /**
6264          * We have in our bbr control:
6265          * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp).
6266          * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts).
6267          * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts)
6268          * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time)
6269          *
6270          * Now we can calculate the time between the sends by doing:
6271          *
6272          * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts
6273          *
6274          * And the peer's time between receiving them by doing:
6275          *
6276          * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp
6277          *
6278          * We want to figure out if the timestamp values are in msec, 10msec or usec.
6279          * We also may find that we can't use the timestamps if say we see
6280          * that the peer_delta indicates that though we may have taken 10ms to
6281          * pace out the data, it only saw 1ms between the two packets. This would
6282          * indicate that somewhere on the path is a batching entity that is giving
6283          * out time-slices of the actual b/w. This would mean we could not use
6284          * reliably the peers timestamps.
6285          *
6286          * We expect delta > peer_delta initially. Until we figure out the
6287          * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio.
6288          * If we place 1000 there then its a ms vs our usec. If we place 10000 there
6289          * then its 10ms vs our usec. If the peer is running a usec clock we would
6290          * put a 1 there. If the value is faster then ours, we will disable the
6291          * use of timestamps (though we could revist this later if we find it to be not
6292          * just an isolated one or two flows)).
6293          *
6294          * To detect the batching middle boxes we will come up with our compensation and
6295          * if with it in place, we find the peer is drastically off (by some margin) in
6296          * the smaller direction, then we will assume the worst case and disable use of timestamps.
6297          *
6298          */
6299         uint64_t delta, peer_delta, delta_up;
6300
6301         delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts;
6302         if (delta < bbr_min_usec_delta) {
6303                 /*
6304                  * Have not seen a min amount of time
6305                  * between our send times so we can
6306                  * make a determination of the timestamp
6307                  * yet.
6308                  */
6309                 return;
6310         }
6311         peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp;
6312         if (peer_delta < bbr_min_peer_delta) {
6313                 /*
6314                  * We may have enough in the form of
6315                  * our delta but the peers number
6316                  * has not changed that much. It could
6317                  * be its clock ratio is such that
6318                  * we need more data (10ms tick) or
6319                  * there may be other compression scenarios
6320                  * going on. In any event we need the
6321                  * spread to be larger.
6322                  */
6323                 return;
6324         }
6325         /* Ok lets first see which way our delta is going */
6326         if (peer_delta > delta) {
6327                 /* Very unlikely, the peer without
6328                  * compensation shows that it saw
6329                  * the two sends arrive further apart
6330                  * then we saw then in micro-seconds.
6331                  */
6332                 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) {
6333                         /* well it looks like the peer is a micro-second clock. */
6334                         bbr->rc_ts_clock_set = 1;
6335                         bbr->r_ctl.bbr_peer_tsratio = 1;
6336                 } else {
6337                         bbr->rc_ts_cant_be_used = 1;
6338                         bbr->rc_ts_clock_set = 1;
6339                 }
6340                 return;
6341         }
6342         /* Ok we know that the peer_delta is smaller than our send distance */
6343         bbr->rc_ts_clock_set = 1;
6344         /* First question is it within the percentage that they are using usec time? */
6345         delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent;
6346         if ((peer_delta + delta_up) >= delta) {
6347                 /* Its a usec clock */
6348                 bbr->r_ctl.bbr_peer_tsratio = 1;
6349                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6350                 return;
6351         }
6352         /* Ok if not usec, what about 10usec (though unlikely)? */
6353         delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent;
6354         if (((peer_delta * 10) + delta_up) >= delta) {
6355                 bbr->r_ctl.bbr_peer_tsratio = 10;
6356                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6357                 return;
6358         }
6359         /* And what about 100usec (though again unlikely)? */
6360         delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent;
6361         if (((peer_delta * 100) + delta_up) >= delta) {
6362                 bbr->r_ctl.bbr_peer_tsratio = 100;
6363                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6364                 return;
6365         }
6366         /* And how about 1 msec (the most likely one)? */
6367         delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent;
6368         if (((peer_delta * 1000) + delta_up) >= delta) {
6369                 bbr->r_ctl.bbr_peer_tsratio = 1000;
6370                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6371                 return;
6372         }
6373         /* Ok if not msec could it be 10 msec? */
6374         delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent;
6375         if (((peer_delta * 10000) + delta_up) >= delta) {
6376                 bbr->r_ctl.bbr_peer_tsratio = 10000;
6377                 return;
6378         }
6379         /* If we fall down here the clock tick so slowly we can't use it */
6380         bbr->rc_ts_cant_be_used = 1;
6381         bbr->r_ctl.bbr_peer_tsratio = 0;
6382         bbr_log_tstmp_validation(bbr, peer_delta, delta);
6383 }
6384
6385 /*
6386  * Collect new round-trip time estimate
6387  * and update averages and current timeout.
6388  */
6389 static void
6390 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts)
6391 {
6392         int32_t delta;
6393         uint32_t rtt, tsin;
6394         int32_t rtt_ticks;
6395
6396
6397         if (bbr->rtt_valid == 0)
6398                 /* No valid sample */
6399                 return;
6400
6401         rtt = bbr->r_ctl.cur_rtt;
6402         tsin = bbr->r_ctl.ts_in;
6403         if (bbr->rc_prtt_set_ts) {
6404                 /*
6405                  * We are to force feed the rttProp filter due
6406                  * to an entry into PROBE_RTT. This assures
6407                  * that the times are sync'd between when we
6408                  * go into PROBE_RTT and the filter expiration.
6409                  *
6410                  * Google does not use a true filter, so they do
6411                  * this implicitly since they only keep one value
6412                  * and when they enter probe-rtt they update the
6413                  * value to the newest rtt.
6414                  */
6415                 uint32_t rtt_prop;
6416
6417                 bbr->rc_prtt_set_ts = 0;
6418                 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
6419                 if (rtt > rtt_prop)
6420                         filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts);
6421                 else
6422                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6423         }
6424         if (bbr->rc_ack_was_delayed)
6425                 rtt += bbr->r_ctl.rc_ack_hdwr_delay;
6426
6427         if (rtt < bbr->r_ctl.rc_lowest_rtt)
6428                 bbr->r_ctl.rc_lowest_rtt = rtt;
6429         bbr_log_rtt_sample(bbr, rtt, tsin);
6430         if (bbr->r_init_rtt) {
6431                 /*
6432                  * The initial rtt is not-trusted, nuke it and lets get
6433                  * our first valid measurement in.
6434                  */
6435                 bbr->r_init_rtt = 0;
6436                 tp->t_srtt = 0;
6437         }
6438         if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) {
6439                 /*
6440                  * So we have not yet figured out
6441                  * what the peers TSTMP value is
6442                  * in (most likely ms). We need a
6443                  * series of cum-ack's to determine
6444                  * this reliably.
6445                  */
6446                 if (bbr->rc_ack_is_cumack) {
6447                         if (bbr->rc_ts_data_set) {
6448                                 /* Lets attempt to determine the timestamp granularity. */
6449                                 bbr_make_timestamp_determination(bbr);
6450                         } else {
6451                                 bbr->rc_ts_data_set = 1;
6452                                 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts;
6453                                 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time;
6454                         }
6455                 } else {
6456                         /*
6457                          * We have to have consecutive acks
6458                          * reset any "filled" state to none.
6459                          */
6460                         bbr->rc_ts_data_set = 0;
6461                 }
6462         }
6463         /* Round it up */
6464         rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1)));
6465         if (rtt_ticks == 0)
6466                 rtt_ticks = 1;
6467         if (tp->t_srtt != 0) {
6468                 /*
6469                  * srtt is stored as fixed point with 5 bits after the
6470                  * binary point (i.e., scaled by 8).  The following magic is
6471                  * equivalent to the smoothing algorithm in rfc793 with an
6472                  * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
6473                  * Adjust rtt to origin 0.
6474                  */
6475
6476                 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT)
6477                     - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
6478
6479                 tp->t_srtt += delta;
6480                 if (tp->t_srtt <= 0)
6481                         tp->t_srtt = 1;
6482
6483                 /*
6484                  * We accumulate a smoothed rtt variance (actually, a
6485                  * smoothed mean difference), then set the retransmit timer
6486                  * to smoothed rtt + 4 times the smoothed variance. rttvar
6487                  * is stored as fixed point with 4 bits after the binary
6488                  * point (scaled by 16).  The following is equivalent to
6489                  * rfc793 smoothing with an alpha of .75 (rttvar =
6490                  * rttvar*3/4 + |delta| / 4).  This replaces rfc793's
6491                  * wired-in beta.
6492                  */
6493                 if (delta < 0)
6494                         delta = -delta;
6495                 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
6496                 tp->t_rttvar += delta;
6497                 if (tp->t_rttvar <= 0)
6498                         tp->t_rttvar = 1;
6499                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
6500                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6501         } else {
6502                 /*
6503                  * No rtt measurement yet - use the unsmoothed rtt. Set the
6504                  * variance to half the rtt (so our first retransmit happens
6505                  * at 3*rtt).
6506                  */
6507                 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT;
6508                 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1);
6509                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6510         }
6511         TCPSTAT_INC(tcps_rttupdated);
6512         tp->t_rttupdated++;
6513 #ifdef NETFLIX_STATS
6514         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks));
6515 #endif
6516         /*
6517          * the retransmit should happen at rtt + 4 * rttvar. Because of the
6518          * way we do the smoothing, srtt and rttvar will each average +1/2
6519          * tick of bias.  When we compute the retransmit timer, we want 1/2
6520          * tick of rounding and 1 extra tick because of +-1/2 tick
6521          * uncertainty in the firing of the timer.  The bias will give us
6522          * exactly the 1.5 tick we need.  But, because the bias is
6523          * statistical, we have to test that we don't drop below the minimum
6524          * feasible timer (which is 2 ticks).
6525          */
6526         TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
6527             max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2),
6528             MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
6529
6530         /*
6531          * We received an ack for a packet that wasn't retransmitted; it is
6532          * probably safe to discard any error indications we've received
6533          * recently.  This isn't quite right, but close enough for now (a
6534          * route might have failed after we sent a segment, and the return
6535          * path might not be symmetrical).
6536          */
6537         tp->t_softerror = 0;
6538         rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
6539         if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt)
6540                 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt;
6541 }
6542
6543 static void
6544 bbr_earlier_retran(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm,
6545                    uint32_t t, uint32_t cts, int ack_type)
6546 {
6547         /*
6548          * For this RSM, we acknowledged the data from a previous
6549          * transmission, not the last one we made. This means we did a false
6550          * retransmit.
6551          */
6552         if (rsm->r_flags & BBR_HAS_FIN) {
6553                 /*
6554                  * The sending of the FIN often is multiple sent when we
6555                  * have everything outstanding ack'd. We ignore this case
6556                  * since its over now.
6557                  */
6558                 return;
6559         }
6560         if (rsm->r_flags & BBR_TLP) {
6561                 /*
6562                  * We expect TLP's to have this occur often
6563                  */
6564                 bbr->rc_tlp_rtx_out = 0;
6565                 return;
6566         }
6567         if (ack_type != BBR_CUM_ACKED) {
6568                 /*
6569                  * If it was not a cum-ack we
6570                  * don't really know for sure since
6571                  * the timestamp could be from some
6572                  * other transmission.
6573                  */
6574                 return;
6575         }
6576
6577         if (rsm->r_flags & BBR_WAS_SACKPASS) {
6578                 /*
6579                  * We retransmitted based on a sack and the earlier
6580                  * retransmission ack'd it - re-ordering is occuring.
6581                  */
6582                 BBR_STAT_INC(bbr_reorder_seen);
6583                 bbr->r_ctl.rc_reorder_ts = cts;
6584         }
6585         /* Back down the loss count */
6586         if (rsm->r_flags & BBR_MARKED_LOST) {
6587                 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
6588                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
6589                 rsm->r_flags &= ~BBR_MARKED_LOST;
6590                 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
6591                         /* LT sampling also needs adjustment */
6592                         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
6593         }
6594         /***** RRS HERE ************************/
6595         /* Do we need to do this???            */
6596         /* bbr_reset_lt_bw_sampling(bbr, cts); */
6597         /***** RRS HERE ************************/
6598         BBR_STAT_INC(bbr_badfr);
6599         BBR_STAT_ADD(bbr_badfr_bytes, (rsm->r_end - rsm->r_start));
6600 }
6601
6602
6603 static void
6604 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line)
6605 {
6606         bbr->r_ctl.rc_rtt_shrinks = cts;
6607         if (bbr_can_force_probertt &&
6608             (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
6609             ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
6610                 /*
6611                  * We should enter probe-rtt its been too long
6612                  * since we have been there.
6613                  */
6614                 bbr_enter_probe_rtt(bbr, cts, __LINE__);
6615         } else
6616                 bbr_check_probe_rtt_limits(bbr, cts);
6617 }
6618
6619 static void
6620 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts)
6621 {
6622         uint64_t orig_bw;
6623
6624         if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) {
6625                 /* We never apply a zero measurment */
6626                 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0,
6627                                     0, 0, 0, 0, 0, 0);
6628                 return;
6629         }
6630         if (bbr->r_ctl.r_measurement_count < 0xffffffff)
6631                 bbr->r_ctl.r_measurement_count++;
6632         orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
6633         apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch);
6634         bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw,
6635                             (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate),
6636                             0, 0, 0, 0, 0, 0);
6637         if (orig_bw &&
6638             (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) {
6639                 if (bbr->bbr_hdrw_pacing) {
6640                         /*
6641                          * Apply a new rate to the hardware
6642                          * possibly.
6643                          */
6644                         bbr_update_hardware_pacing_rate(bbr, cts);
6645                 }
6646                 bbr_set_state_target(bbr, __LINE__);
6647                 tcp_bbr_tso_size_check(bbr, cts);
6648                 if (bbr->r_recovery_bw)  {
6649                         bbr_setup_red_bw(bbr, cts);
6650                         bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW);
6651                 }
6652         } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate))
6653                 tcp_bbr_tso_size_check(bbr, cts);
6654 }
6655
6656 static void
6657 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6658 {
6659         if (bbr->rc_in_persist == 0) {
6660                 /* We log only when not in persist */
6661                 /* Translate to a Bytes Per Second */
6662                 uint64_t tim, bw, ts_diff, ts_bw;
6663                 uint32_t upper, lower, delivered;
6664
6665                 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6666                         tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6667                 else
6668                         tim = 1;
6669                 /*
6670                  * Now that we have processed the tim (skipping the sample
6671                  * or possibly updating the time, go ahead and
6672                  * calculate the cdr.
6673                  */
6674                 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6675                 bw = (uint64_t)delivered;
6676                 bw *= (uint64_t)USECS_IN_SECOND;
6677                 bw /= tim;
6678                 if (bw == 0) {
6679                         /* We must have a calculatable amount */
6680                         return;
6681                 }
6682                 upper = (bw >> 32) & 0x00000000ffffffff;
6683                 lower = bw & 0x00000000ffffffff;
6684                 /*
6685                  * If we are using this b/w shove it in now so we
6686                  * can see in the trace viewer if it gets over-ridden.
6687                  */
6688                 if (rsm->r_ts_valid &&
6689                     bbr->rc_ts_valid &&
6690                     bbr->rc_ts_clock_set &&
6691                     (bbr->rc_ts_cant_be_used == 0) &&
6692                     bbr->rc_use_ts_limit) {
6693                         ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1);
6694                         ts_diff *= bbr->r_ctl.bbr_peer_tsratio;
6695                         if ((delivered == 0) ||
6696                             (rtt < 1000)) {
6697                                 /* Can't use the ts */
6698                                 bbr_log_type_bbrupd(bbr, 61, cts,
6699                                                     ts_diff,
6700                                                     bbr->r_ctl.last_inbound_ts,
6701                                                     rsm->r_del_ack_ts, 0,
6702                                                     0, 0, 0, delivered);
6703                         } else {
6704                                 ts_bw = (uint64_t)delivered;
6705                                 ts_bw *= (uint64_t)USECS_IN_SECOND;
6706                                 ts_bw /= ts_diff;
6707                                 bbr_log_type_bbrupd(bbr, 62, cts,
6708                                                     (ts_bw >> 32),
6709                                                     (ts_bw & 0xffffffff), 0, 0,
6710                                                     0, 0, ts_diff, delivered);
6711                                 if ((bbr->ts_can_raise) &&
6712                                     (ts_bw > bw)) {
6713                                         bbr_log_type_bbrupd(bbr, 8, cts,
6714                                                             delivered,
6715                                                             ts_diff,
6716                                                             (bw >> 32),
6717                                                             (bw & 0x00000000ffffffff),
6718                                                             0, 0, 0, 0);
6719                                         bw = ts_bw;
6720                                 } else if (ts_bw && (ts_bw < bw)) {
6721                                         bbr_log_type_bbrupd(bbr, 7, cts,
6722                                                             delivered,
6723                                                             ts_diff,
6724                                                             (bw >> 32),
6725                                                             (bw & 0x00000000ffffffff),
6726                                                             0, 0, 0, 0);
6727                                         bw = ts_bw;
6728                                 }
6729                         }
6730                 }
6731                 if (rsm->r_first_sent_time &&
6732                     TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6733                         uint64_t sbw, sti;
6734                         /*
6735                          * We use what was in flight at the time of our
6736                          * send  and the size of this send to figure
6737                          * out what we have been sending at (amount).
6738                          * For the time we take from the time of
6739                          * the send of the first send outstanding
6740                          * until this send plus this sends pacing
6741                          * time. This gives us a good calculation
6742                          * as to the rate we have been sending at.
6743                          */
6744
6745                         sbw = (uint64_t)(rsm->r_flight_at_send);
6746                         sbw *= (uint64_t)USECS_IN_SECOND;
6747                         sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6748                         sti += rsm->r_pacing_delay;
6749                         sbw /= sti;
6750                         if (sbw < bw) {
6751                                 bbr_log_type_bbrupd(bbr, 6, cts,
6752                                                     delivered,
6753                                                     (uint32_t)sti,
6754                                                     (bw >> 32),
6755                                                     (uint32_t)bw,
6756                                                     rsm->r_first_sent_time, 0, (sbw >> 32),
6757                                                     (uint32_t)sbw);
6758                                 bw = sbw;
6759                         }
6760                 }
6761                 /* Use the google algorithm for b/w measurements */
6762                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6763                 if ((rsm->r_app_limited == 0) ||
6764                     (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) {
6765                         tcp_bbr_commit_bw(bbr, cts);
6766                         bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6767                                             0, 0, 0, 0,  bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6768                 }
6769         }
6770 }
6771
6772 static void
6773 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6774 {
6775         if (bbr->rc_in_persist == 0) {
6776                 /* We log only when not in persist */
6777                 /* Translate to a Bytes Per Second */
6778                 uint64_t tim, bw;
6779                 uint32_t upper, lower, delivered;
6780                 int no_apply = 0;
6781
6782                 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6783                         tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6784                 else
6785                         tim = 1;
6786                 /*
6787                  * Now that we have processed the tim (skipping the sample
6788                  * or possibly updating the time, go ahead and
6789                  * calculate the cdr.
6790                  */
6791                 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6792                 bw = (uint64_t)delivered;
6793                 bw *= (uint64_t)USECS_IN_SECOND;
6794                 bw /= tim;
6795                 if (tim < bbr->r_ctl.rc_lowest_rtt) {
6796                         bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6797                                             tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6798
6799                         no_apply = 1;
6800                 }
6801                 upper = (bw >> 32) & 0x00000000ffffffff;
6802                 lower = bw & 0x00000000ffffffff;
6803                 /*
6804                  * If we are using this b/w shove it in now so we
6805                  * can see in the trace viewer if it gets over-ridden.
6806                  */
6807                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6808                 /* Gate by the sending rate */
6809                 if (rsm->r_first_sent_time &&
6810                     TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6811                         uint64_t sbw, sti;
6812                         /*
6813                          * We use what was in flight at the time of our
6814                          * send  and the size of this send to figure
6815                          * out what we have been sending at (amount).
6816                          * For the time we take from the time of
6817                          * the send of the first send outstanding
6818                          * until this send plus this sends pacing
6819                          * time. This gives us a good calculation
6820                          * as to the rate we have been sending at.
6821                          */
6822
6823                         sbw = (uint64_t)(rsm->r_flight_at_send);
6824                         sbw *= (uint64_t)USECS_IN_SECOND;
6825                         sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6826                         sti += rsm->r_pacing_delay;
6827                         sbw /= sti;
6828                         if (sbw < bw) {
6829                                 bbr_log_type_bbrupd(bbr, 6, cts,
6830                                                     delivered,
6831                                                     (uint32_t)sti,
6832                                                     (bw >> 32),
6833                                                     (uint32_t)bw,
6834                                                     rsm->r_first_sent_time, 0, (sbw >> 32),
6835                                                     (uint32_t)sbw);
6836                                 bw = sbw;
6837                         }
6838                         if ((sti > tim) &&
6839                             (sti < bbr->r_ctl.rc_lowest_rtt)) {
6840                                 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6841                                                     (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6842                                 no_apply = 1;
6843                         } else
6844                                 no_apply = 0;
6845                 }
6846                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6847                 if ((no_apply == 0) &&
6848                     ((rsm->r_app_limited == 0) ||
6849                      (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) {
6850                         tcp_bbr_commit_bw(bbr, cts);
6851                         bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6852                                             0, 0, 0, 0, bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6853                 }
6854         }
6855 }
6856
6857
6858 static void
6859 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin,
6860     uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to)
6861 {
6862         uint64_t old_rttprop;
6863
6864         /* Update our delivery time and amount */
6865         bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start);
6866         bbr->r_ctl.rc_del_time = cts;
6867         if (rtt == 0) {
6868                 /*
6869                  * 0 means its a retransmit, for now we don't use these for
6870                  * the rest of BBR.
6871                  */
6872                 return;
6873         }
6874         if ((bbr->rc_use_google == 0) &&
6875             (match != BBR_RTT_BY_EXACTMATCH) &&
6876             (match != BBR_RTT_BY_TIMESTAMP)){
6877                 /*
6878                  * We get a lot of rtt updates, lets not pay attention to
6879                  * any that are not an exact match. That way we don't have
6880                  * to worry about timestamps and the whole nonsense of
6881                  * unsure if its a retransmission etc (if we ever had the
6882                  * timestamp fixed to always have the last thing sent this
6883                  * would not be a issue).
6884                  */
6885                 return;
6886         }
6887         if ((bbr_no_retran && bbr->rc_use_google) &&
6888             (match != BBR_RTT_BY_EXACTMATCH) &&
6889             (match != BBR_RTT_BY_TIMESTAMP)){
6890                 /*
6891                  * We only do measurements in google mode
6892                  * with bbr_no_retran on for sure things.
6893                  */
6894                 return;
6895         }
6896         /* Only update srtt if we know by exact match */
6897         tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin);
6898         if (ack_type == BBR_CUM_ACKED)
6899                 bbr->rc_ack_is_cumack = 1;
6900         else
6901                 bbr->rc_ack_is_cumack = 0;
6902         old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP);
6903         /*
6904          * Note the following code differs to the original
6905          * BBR spec. It calls for <= not <. However after a
6906          * long discussion in email with Neal, he acknowledged
6907          * that it should be < than so that we will have flows
6908          * going into probe-rtt (we were seeing cases where that
6909          * did not happen and caused ugly things to occur). We
6910          * have added this agreed upon fix to our code base.
6911          */
6912         if (rtt < old_rttprop) {
6913                 /* Update when we last saw a rtt drop */
6914                 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0);
6915                 bbr_set_reduced_rtt(bbr, cts, __LINE__);
6916         }
6917         bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts,
6918             match, rsm->r_start, rsm->r_flags);
6919         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6920         if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) {
6921                 /*
6922                  * The RTT-prop moved, reset the target (may be a
6923                  * nop for some states).
6924                  */
6925                 bbr_set_state_target(bbr, __LINE__);
6926                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)
6927                         bbr_log_rtt_shrinks(bbr, cts, 0, 0,
6928                                             __LINE__, BBR_RTTS_NEW_TARGET, 0);
6929                 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP))
6930                         /* It went up */
6931                         bbr_check_probe_rtt_limits(bbr, cts);
6932         }
6933         if ((bbr->rc_use_google == 0) &&
6934             (match == BBR_RTT_BY_TIMESTAMP)) {
6935                 /*
6936                  * We don't do b/w update with
6937                  * these since they are not really
6938                  * reliable.
6939                  */
6940                 return;
6941         }
6942         if (bbr->r_ctl.r_app_limited_until &&
6943             (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) {
6944                 /* We are no longer app-limited */
6945                 bbr->r_ctl.r_app_limited_until = 0;
6946         }
6947         if (bbr->rc_use_google) {
6948                 bbr_google_measurement(bbr, rsm, rtt, cts);
6949         } else {
6950                 bbr_nf_measurement(bbr, rsm, rtt, cts);
6951         }
6952 }
6953
6954 /*
6955  * Convert a timestamp that the main stack
6956  * uses (milliseconds) into one that bbr uses
6957  * (microseconds). Return that converted timestamp.
6958  */
6959 static uint32_t
6960 bbr_ts_convert(uint32_t cts) {
6961         uint32_t sec, msec;
6962
6963         sec = cts / MS_IN_USEC;
6964         msec = cts - (MS_IN_USEC * sec);
6965         return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC));
6966 }
6967
6968 /*
6969  * Return 0 if we did not update the RTT time, return
6970  * 1 if we did.
6971  */
6972 static int
6973 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr,
6974     struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack)
6975 {
6976         int32_t i;
6977         uint32_t t, uts = 0;
6978
6979         if ((rsm->r_flags & BBR_ACKED) ||
6980             (rsm->r_flags & BBR_WAS_RENEGED) ||
6981             (rsm->r_flags & BBR_RXT_CLEARED)) {
6982                 /* Already done */
6983                 return (0);
6984         }
6985         if (rsm->r_rtr_cnt == 1) {
6986                 /*
6987                  * Only one transmit. Hopefully the normal case.
6988                  */
6989                 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0]))
6990                         t = cts - rsm->r_tim_lastsent[0];
6991                 else
6992                         t = 1;
6993                 if ((int)t <= 0)
6994                         t = 1;
6995                 bbr->r_ctl.rc_last_rtt = t;
6996                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6997                                     BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to);
6998                 return (1);
6999         }
7000         /* Convert to usecs */
7001         if ((bbr_can_use_ts_for_rtt == 1) &&
7002             (bbr->rc_use_google == 1) &&
7003             (ack_type == BBR_CUM_ACKED) &&
7004             (to->to_flags & TOF_TS) &&
7005             (to->to_tsecr != 0)) {
7006
7007                 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr;
7008                 if (t < 1)
7009                         t = 1;
7010                 t *= MS_IN_USEC;
7011                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
7012                                     BBR_RTT_BY_TIMESTAMP,
7013                                     rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)],
7014                                     ack_type, to);
7015                 return (1);
7016         }
7017         uts = bbr_ts_convert(to->to_tsecr);
7018         if ((to->to_flags & TOF_TS) &&
7019             (to->to_tsecr != 0) &&
7020             (ack_type == BBR_CUM_ACKED) &&
7021             ((rsm->r_flags & BBR_OVERMAX) == 0)) {
7022                 /*
7023                  * Now which timestamp does it match? In this block the ACK
7024                  * may be coming from a previous transmission.
7025                  */
7026                 uint32_t fudge;
7027
7028                 fudge = BBR_TIMER_FUDGE;
7029                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7030                         if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) &&
7031                             (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) {
7032                                 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
7033                                         t = cts - rsm->r_tim_lastsent[i];
7034                                 else
7035                                         t = 1;
7036                                 if ((int)t <= 0)
7037                                         t = 1;
7038                                 bbr->r_ctl.rc_last_rtt = t;
7039                                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING,
7040                                                     rsm->r_tim_lastsent[i], ack_type, to);
7041                                 if ((i + 1) < rsm->r_rtr_cnt) {
7042                                         /* Likely */
7043                                         bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type);
7044                                 } else if (rsm->r_flags & BBR_TLP) {
7045                                         bbr->rc_tlp_rtx_out = 0;
7046                                 }
7047                                 return (1);
7048                         }
7049                 }
7050                 /* Fall through if we can't find a matching timestamp */
7051         }
7052         /*
7053          * Ok its a SACK block that we retransmitted. or a windows
7054          * machine without timestamps. We can tell nothing from the
7055          * time-stamp since its not there or the time the peer last
7056          * recieved a segment that moved forward its cum-ack point.
7057          *
7058          * Lets look at the last retransmit and see what we can tell
7059          * (with BBR for space we only keep 2 note we have to keep
7060          * at least 2 so the map can not be condensed more).
7061          */
7062         i = rsm->r_rtr_cnt - 1;
7063         if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
7064                 t = cts - rsm->r_tim_lastsent[i];
7065         else
7066                 goto not_sure;
7067         if (t < bbr->r_ctl.rc_lowest_rtt) {
7068                 /*
7069                  * We retransmitted and the ack came back in less
7070                  * than the smallest rtt we have observed in the
7071                  * windowed rtt. We most likey did an improper
7072                  * retransmit as outlined in 4.2 Step 3 point 2 in
7073                  * the rack-draft.
7074                  *
7075                  * Use the prior transmission to update all the
7076                  * information as long as there is only one prior
7077                  * transmission.
7078                  */
7079                 if ((rsm->r_flags & BBR_OVERMAX) == 0) {
7080 #ifdef BBR_INVARIANTS
7081                         if (rsm->r_rtr_cnt == 1)
7082                                 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags);
7083 #endif
7084                         i = rsm->r_rtr_cnt - 2;
7085                         if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
7086                                 t = cts - rsm->r_tim_lastsent[i];
7087                         else
7088                                 t = 1;
7089                         bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET,
7090                                             rsm->r_tim_lastsent[i], ack_type, to);
7091                         bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type);
7092                 } else {
7093                         /*
7094                          * Too many prior transmissions, just
7095                          * updated BBR delivered
7096                          */
7097 not_sure:
7098                         bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
7099                                             BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
7100                 }
7101         } else {
7102                 /*
7103                  * We retransmitted it and the retransmit did the
7104                  * job.
7105                  */
7106                 if (rsm->r_flags & BBR_TLP)
7107                         bbr->rc_tlp_rtx_out = 0;
7108                 if ((rsm->r_flags & BBR_OVERMAX) == 0)
7109                         bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts,
7110                                             BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to);
7111                 else
7112                         bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
7113                                             BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
7114                 return (1);
7115         }
7116         return (0);
7117 }
7118
7119 /*
7120  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7121  */
7122 static void
7123 bbr_log_sack_passed(struct tcpcb *tp,
7124     struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
7125 {
7126         struct bbr_sendmap *nrsm;
7127
7128         nrsm = rsm;
7129         TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap,
7130             bbr_head, r_tnext) {
7131                 if (nrsm == rsm) {
7132                         /* Skip orginal segment he is acked */
7133                         continue;
7134                 }
7135                 if (nrsm->r_flags & BBR_ACKED) {
7136                         /* Skip ack'd segments */
7137                         continue;
7138                 }
7139                 if (nrsm->r_flags & BBR_SACK_PASSED) {
7140                         /*
7141                          * We found one that is already marked
7142                          * passed, we have been here before and
7143                          * so all others below this are marked.
7144                          */
7145                         break;
7146                 }
7147                 BBR_STAT_INC(bbr_sack_passed);
7148                 nrsm->r_flags |= BBR_SACK_PASSED;
7149                 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) &&
7150                     bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) {
7151                         bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start;
7152                         bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start;
7153                         nrsm->r_flags |= BBR_MARKED_LOST;
7154                 }
7155                 nrsm->r_flags &= ~BBR_WAS_SACKPASS;
7156         }
7157 }
7158
7159 /*
7160  * Returns the number of bytes that were
7161  * newly ack'd by sack blocks.
7162  */
7163 static uint32_t
7164 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack,
7165     struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts)
7166 {
7167         int32_t times = 0;
7168         uint32_t start, end, maxseg, changed = 0;
7169         struct bbr_sendmap *rsm, *nrsm;
7170         int32_t used_ref = 1;
7171         uint8_t went_back = 0, went_fwd = 0;
7172
7173         maxseg = tp->t_maxseg - bbr->rc_last_options;
7174         start = sack->start;
7175         end = sack->end;
7176         rsm = *prsm;
7177         if (rsm == NULL)
7178                 used_ref = 0;
7179
7180         /* Do we locate the block behind where we last were? */
7181         if (rsm && SEQ_LT(start, rsm->r_start)) {
7182                 went_back = 1;
7183                 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
7184                         if (SEQ_GEQ(start, rsm->r_start) &&
7185                             SEQ_LT(start, rsm->r_end)) {
7186                                 goto do_rest_ofb;
7187                         }
7188                 }
7189         }
7190 start_at_beginning:
7191         went_fwd = 1;
7192         /*
7193          * Ok lets locate the block where this guy is fwd from rsm (if its
7194          * set)
7195          */
7196         TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) {
7197                 if (SEQ_GEQ(start, rsm->r_start) &&
7198                     SEQ_LT(start, rsm->r_end)) {
7199                         break;
7200                 }
7201         }
7202 do_rest_ofb:
7203         if (rsm == NULL) {
7204                 /*
7205                  * This happens when we get duplicate sack blocks with the
7206                  * same end. For example SACK 4: 100 SACK 3: 100 The sort
7207                  * will not change there location so we would just start at
7208                  * the end of the first one and get lost.
7209                  */
7210                 if (tp->t_flags & TF_SENTFIN) {
7211                         /*
7212                          * Check to see if we have not logged the FIN that
7213                          * went out.
7214                          */
7215                         nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7216                         if (nrsm && (nrsm->r_end + 1) == tp->snd_max) {
7217                                 /*
7218                                  * Ok we did not get the FIN logged.
7219                                  */
7220                                 nrsm->r_end++;
7221                                 rsm = nrsm;
7222                                 goto do_rest_ofb;
7223                         }
7224                 }
7225                 if (times == 1) {
7226 #ifdef BBR_INVARIANTS
7227                         panic("tp:%p bbr:%p sack:%p to:%p prsm:%p",
7228                             tp, bbr, sack, to, prsm);
7229 #else
7230                         goto out;
7231 #endif
7232                 }
7233                 times++;
7234                 BBR_STAT_INC(bbr_sack_proc_restart);
7235                 rsm = NULL;
7236                 goto start_at_beginning;
7237         }
7238         /* Ok we have an ACK for some piece of rsm */
7239         if (rsm->r_start != start) {
7240                 /*
7241                  * Need to split this in two pieces the before and after.
7242                  */
7243                 if (bbr_sack_mergable(rsm, start, end))
7244                         nrsm = bbr_alloc_full_limit(bbr);
7245                 else
7246                         nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7247                 if (nrsm == NULL) {
7248                         /* We could not allocate ignore the sack */
7249                         struct sackblk blk;
7250
7251                         blk.start = start;
7252                         blk.end = end;
7253                         sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7254                         goto out;
7255                 }
7256                 bbr_clone_rsm(bbr, nrsm, rsm, start);
7257                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7258                 if (rsm->r_in_tmap) {
7259                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7260                         nrsm->r_in_tmap = 1;
7261                 }
7262                 rsm->r_flags &= (~BBR_HAS_FIN);
7263                 rsm = nrsm;
7264         }
7265         if (SEQ_GEQ(end, rsm->r_end)) {
7266                 /*
7267                  * The end of this block is either beyond this guy or right
7268                  * at this guy.
7269                  */
7270                 if ((rsm->r_flags & BBR_ACKED) == 0) {
7271                         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7272                         changed += (rsm->r_end - rsm->r_start);
7273                         bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7274                         bbr_log_sack_passed(tp, bbr, rsm);
7275                         if (rsm->r_flags & BBR_MARKED_LOST) {
7276                                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7277                         }
7278                         /* Is Reordering occuring? */
7279                         if (rsm->r_flags & BBR_SACK_PASSED) {
7280                                 BBR_STAT_INC(bbr_reorder_seen);
7281                                 bbr->r_ctl.rc_reorder_ts = cts;
7282                                 if (rsm->r_flags & BBR_MARKED_LOST) {
7283                                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7284                                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7285                                                 /* LT sampling also needs adjustment */
7286                                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7287                                 }
7288                         }
7289                         rsm->r_flags |= BBR_ACKED;
7290                         rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7291                         if (rsm->r_in_tmap) {
7292                                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7293                                 rsm->r_in_tmap = 0;
7294                         }
7295                 }
7296                 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7297                 if (end == rsm->r_end) {
7298                         /* This block only - done */
7299                         goto out;
7300                 }
7301                 /* There is more not coverend by this rsm move on */
7302                 start = rsm->r_end;
7303                 nrsm = TAILQ_NEXT(rsm, r_next);
7304                 rsm = nrsm;
7305                 times = 0;
7306                 goto do_rest_ofb;
7307         }
7308         if (rsm->r_flags & BBR_ACKED) {
7309                 /* Been here done that */
7310                 goto out;
7311         }
7312         /* Ok we need to split off this one at the tail */
7313         if (bbr_sack_mergable(rsm, start, end))
7314                 nrsm = bbr_alloc_full_limit(bbr);
7315         else
7316                 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7317         if (nrsm == NULL) {
7318                 /* failed XXXrrs what can we do but loose the sack info? */
7319                 struct sackblk blk;
7320
7321                 blk.start = start;
7322                 blk.end = end;
7323                 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7324                 goto out;
7325         }
7326         /* Clone it */
7327         bbr_clone_rsm(bbr, nrsm, rsm, end);
7328         /* The sack block does not cover this guy fully */
7329         rsm->r_flags &= (~BBR_HAS_FIN);
7330         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7331         if (rsm->r_in_tmap) {
7332                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7333                 nrsm->r_in_tmap = 1;
7334         }
7335         nrsm->r_dupack = 0;
7336         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7337         bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7338         changed += (rsm->r_end - rsm->r_start);
7339         bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7340         bbr_log_sack_passed(tp, bbr, rsm);
7341         /* Is Reordering occuring? */
7342         if (rsm->r_flags & BBR_MARKED_LOST) {
7343                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7344         }
7345         if (rsm->r_flags & BBR_SACK_PASSED) {
7346                 BBR_STAT_INC(bbr_reorder_seen);
7347                 bbr->r_ctl.rc_reorder_ts = cts;
7348                 if (rsm->r_flags & BBR_MARKED_LOST) {
7349                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7350                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7351                                 /* LT sampling also needs adjustment */
7352                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7353                 }
7354         }
7355         rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7356         rsm->r_flags |= BBR_ACKED;
7357         if (rsm->r_in_tmap) {
7358                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7359                 rsm->r_in_tmap = 0;
7360         }
7361 out:
7362         if (rsm && (rsm->r_flags & BBR_ACKED)) {
7363                 /*
7364                  * Now can we merge this newly acked
7365                  * block with either the previous or
7366                  * next block?
7367                  */
7368                 nrsm = TAILQ_NEXT(rsm, r_next);
7369                 if (nrsm &&
7370                     (nrsm->r_flags & BBR_ACKED)) {
7371                         /* yep this and next can be merged */
7372                         rsm = bbr_merge_rsm(bbr, rsm, nrsm);
7373                 }
7374                 /* Now what about the previous? */
7375                 nrsm = TAILQ_PREV(rsm, bbr_head, r_next);
7376                 if (nrsm &&
7377                     (nrsm->r_flags & BBR_ACKED)) {
7378                         /* yep the previous and this can be merged */
7379                         rsm = bbr_merge_rsm(bbr, nrsm, rsm);
7380                 }
7381         }
7382         if (used_ref == 0) {
7383                 BBR_STAT_INC(bbr_sack_proc_all);
7384         } else {
7385                 BBR_STAT_INC(bbr_sack_proc_short);
7386         }
7387         if (went_fwd && went_back) {
7388                 BBR_STAT_INC(bbr_sack_search_both);
7389         } else if (went_fwd) {
7390                 BBR_STAT_INC(bbr_sack_search_fwd);
7391         } else if (went_back) {
7392                 BBR_STAT_INC(bbr_sack_search_back);
7393         }
7394         /* Save off where the next seq is */
7395         if (rsm)
7396                 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next);
7397         else
7398                 bbr->r_ctl.rc_sacklast = NULL;
7399         *prsm = rsm;
7400         return (changed);
7401 }
7402
7403
7404 static void inline
7405 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack)
7406 {
7407         struct bbr_sendmap *tmap;
7408
7409         BBR_STAT_INC(bbr_reneges_seen);
7410         tmap = NULL;
7411         while (rsm && (rsm->r_flags & BBR_ACKED)) {
7412                 /* Its no longer sacked, mark it so */
7413                 uint32_t oflags;
7414                 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7415 #ifdef BBR_INVARIANTS
7416                 if (rsm->r_in_tmap) {
7417                         panic("bbr:%p rsm:%p flags:0x%x in tmap?",
7418                             bbr, rsm, rsm->r_flags);
7419                 }
7420 #endif
7421                 oflags = rsm->r_flags;
7422                 if (rsm->r_flags & BBR_MARKED_LOST) {
7423                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7424                         bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7425                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7426                                 /* LT sampling also needs adjustment */
7427                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7428                 }
7429                 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST);
7430                 rsm->r_flags |= BBR_WAS_RENEGED;
7431                 rsm->r_flags |= BBR_RXT_CLEARED;
7432                 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__);
7433                 /* Rebuild it into our tmap */
7434                 if (tmap == NULL) {
7435                         TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7436                         tmap = rsm;
7437                 } else {
7438                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext);
7439                         tmap = rsm;
7440                 }
7441                 tmap->r_in_tmap = 1;
7442                 /*
7443                  * XXXrrs Delivered? Should we do anything here?
7444                  *
7445                  * Of course we don't on a rxt timeout so maybe its ok that
7446                  * we don't?
7447                  *
7448                  * For now lets not.
7449                  */
7450                 rsm = TAILQ_NEXT(rsm, r_next);
7451         }
7452         /*
7453          * Now lets possibly clear the sack filter so we start recognizing
7454          * sacks that cover this area.
7455          */
7456         sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack);
7457 }
7458
7459 static void
7460 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to)
7461 {
7462         struct tcp_bbr *bbr;
7463         struct bbr_sendmap *rsm;
7464         uint32_t cts;
7465
7466         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7467         cts = bbr->r_ctl.rc_rcvtime;
7468         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7469         if (rsm && (rsm->r_flags & BBR_HAS_SYN)) {
7470                 if ((rsm->r_end - rsm->r_start) <= 1) {
7471                         /* Log out the SYN completely */
7472                         bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7473                         rsm->r_rtr_bytes = 0;
7474                         TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7475                         if (rsm->r_in_tmap) {
7476                                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7477                                 rsm->r_in_tmap = 0;
7478                         }
7479                         if (bbr->r_ctl.rc_next == rsm) {
7480                                 /* scoot along the marker */
7481                                 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7482                         }
7483                         if (to != NULL)
7484                                 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0);
7485                         bbr_free(bbr, rsm);
7486                 } else {
7487                         /* There is more (Fast open)? strip out SYN. */
7488                         rsm->r_flags &= ~BBR_HAS_SYN;
7489                         rsm->r_start++;
7490                 }
7491         }
7492 }
7493
7494 /*
7495  * Returns the number of bytes that were
7496  * acknowledged by SACK blocks.
7497  */
7498
7499 static uint32_t
7500 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th,
7501     uint32_t *prev_acked)
7502 {
7503         uint32_t changed, last_seq, entered_recovery = 0;
7504         struct tcp_bbr *bbr;
7505         struct bbr_sendmap *rsm;
7506         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
7507         register uint32_t th_ack;
7508         int32_t i, j, k, new_sb, num_sack_blks = 0;
7509         uint32_t cts, acked, ack_point, sack_changed = 0;
7510         uint32_t p_maxseg, maxseg, p_acked = 0;
7511
7512         INP_WLOCK_ASSERT(tp->t_inpcb);
7513         if (th->th_flags & TH_RST) {
7514                 /* We don't log resets */
7515                 return (0);
7516         }
7517         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7518         cts = bbr->r_ctl.rc_rcvtime;
7519
7520         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7521         changed = 0;
7522         maxseg = tp->t_maxseg - bbr->rc_last_options;
7523         p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg);
7524         th_ack = th->th_ack;
7525         if (SEQ_GT(th_ack, tp->snd_una)) {
7526                 acked = th_ack - tp->snd_una;
7527                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__);
7528                 bbr->rc_tp->t_acktime = ticks;
7529         } else
7530                 acked = 0;
7531         if (SEQ_LEQ(th_ack, tp->snd_una)) {
7532                 /* Only sent here for sack processing */
7533                 goto proc_sack;
7534         }
7535         if (rsm && SEQ_GT(th_ack, rsm->r_start)) {
7536                 changed = th_ack - rsm->r_start;
7537         } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) {
7538                 /*
7539                  * For the SYN incoming case we will not have called
7540                  * tcp_output for the sending of the SYN, so there will be
7541                  * no map. All other cases should probably be a panic.
7542                  */
7543                 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) {
7544                         /*
7545                          * We have a timestamp that can be used to generate
7546                          * an initial RTT.
7547                          */
7548                         uint32_t ts, now, rtt;
7549
7550                         ts = bbr_ts_convert(to->to_tsecr);
7551                         now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv));
7552                         rtt = now - ts;
7553                         if (rtt < 1)
7554                                 rtt = 1;
7555                         bbr_log_type_bbrrttprop(bbr, rtt,
7556                                                 tp->iss, 0, cts,
7557                                                 BBR_RTT_BY_TIMESTAMP, tp->iss, 0);
7558                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
7559                         changed = 1;
7560                         bbr->r_wanted_output = 1;
7561                         goto out;
7562                 }
7563                 goto proc_sack;
7564         } else if (rsm == NULL) {
7565                 goto out;
7566         }
7567         if (changed) {
7568                 /*
7569                  * The ACK point is advancing to th_ack, we must drop off
7570                  * the packets in the rack log and calculate any eligble
7571                  * RTT's.
7572                  */
7573                 bbr->r_wanted_output = 1;
7574 more:
7575                 if (rsm == NULL) {
7576
7577                         if (tp->t_flags & TF_SENTFIN) {
7578                                 /* if we send a FIN we will not hav a map */
7579                                 goto proc_sack;
7580                         }
7581 #ifdef BBR_INVARIANTS
7582                         panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n",
7583                             tp,
7584                             th, tp->t_state, bbr,
7585                             tp->snd_una, tp->snd_max, changed);
7586 #endif
7587                         goto proc_sack;
7588                 }
7589         }
7590         if (SEQ_LT(th_ack, rsm->r_start)) {
7591                 /* Huh map is missing this */
7592 #ifdef BBR_INVARIANTS
7593                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n",
7594                     rsm->r_start,
7595                     th_ack, tp->t_state,
7596                     bbr->r_state, bbr);
7597                 panic("th-ack is bad bbr:%p tp:%p", bbr, tp);
7598 #endif
7599                 goto proc_sack;
7600         } else if (th_ack == rsm->r_start) {
7601                 /* None here to ack */
7602                 goto proc_sack;
7603         }
7604         /*
7605          * Clear the dup ack counter, it will
7606          * either be freed or if there is some
7607          * remaining we need to start it at zero.
7608          */
7609         rsm->r_dupack = 0;
7610         /* Now do we consume the whole thing? */
7611         if (SEQ_GEQ(th_ack, rsm->r_end)) {
7612                 /* Its all consumed. */
7613                 uint32_t left;
7614
7615                 if (rsm->r_flags & BBR_ACKED) {
7616                         /*
7617                          * It was acked on the scoreboard -- remove it from
7618                          * total
7619                          */
7620                         p_acked += (rsm->r_end - rsm->r_start);
7621                         bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7622                         if (bbr->r_ctl.rc_sacked == 0)
7623                                 bbr->r_ctl.rc_sacklast = NULL;
7624                 } else {
7625                         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack);
7626                         if (rsm->r_flags & BBR_MARKED_LOST) {
7627                                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7628                         }
7629                         if (rsm->r_flags & BBR_SACK_PASSED) {
7630                                 /*
7631                                  * There are acked segments ACKED on the
7632                                  * scoreboard further up. We are seeing
7633                                  * reordering.
7634                                  */
7635                                 BBR_STAT_INC(bbr_reorder_seen);
7636                                 bbr->r_ctl.rc_reorder_ts = cts;
7637                                 if (rsm->r_flags & BBR_MARKED_LOST) {
7638                                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7639                                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7640                                                 /* LT sampling also needs adjustment */
7641                                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7642                                 }
7643                         }
7644                         rsm->r_flags &= ~BBR_MARKED_LOST;
7645                 }
7646                 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7647                 rsm->r_rtr_bytes = 0;
7648                 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7649                 if (rsm->r_in_tmap) {
7650                         TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7651                         rsm->r_in_tmap = 0;
7652                 }
7653                 if (bbr->r_ctl.rc_next == rsm) {
7654                         /* scoot along the marker */
7655                         bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7656                 }
7657                 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7658                 /* Adjust the packet counts */
7659                 left = th_ack - rsm->r_end;
7660                 /* Free back to zone */
7661                 bbr_free(bbr, rsm);
7662                 if (left) {
7663                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7664                         goto more;
7665                 }
7666                 goto proc_sack;
7667         }
7668         if (rsm->r_flags & BBR_ACKED) {
7669                 /*
7670                  * It was acked on the scoreboard -- remove it from total
7671                  * for the part being cum-acked.
7672                  */
7673                 p_acked += (rsm->r_end - rsm->r_start);
7674                 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
7675                 if (bbr->r_ctl.rc_sacked == 0)
7676                         bbr->r_ctl.rc_sacklast = NULL;
7677         } else {
7678                 /*
7679                  * It was acked up to th_ack point for the first time
7680                  */
7681                 struct bbr_sendmap lrsm;
7682
7683                 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap));
7684                 lrsm.r_end = th_ack;
7685                 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack);
7686         }
7687         if ((rsm->r_flags & BBR_MARKED_LOST) &&
7688             ((rsm->r_flags & BBR_ACKED) == 0)) {
7689                 /*
7690                  * It was marked lost and partly ack'd now
7691                  * for the first time. We lower the rc_lost_bytes
7692                  * and still leave it MARKED.
7693                  */
7694                 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start;
7695         }
7696         bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7697         bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7698         rsm->r_rtr_bytes = 0;
7699         /* adjust packet count */
7700         rsm->r_start = th_ack;
7701 proc_sack:
7702         /* Check for reneging */
7703         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7704         if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) {
7705                 /*
7706                  * The peer has moved snd_una up to the edge of this send,
7707                  * i.e. one that it had previously acked. The only way that
7708                  * can be true if the peer threw away data (space issues)
7709                  * that it had previously sacked (else it would have given
7710                  * us snd_una up to (rsm->r_end). We need to undo the acked
7711                  * markings here.
7712                  *
7713                  * Note we have to look to make sure th_ack is our
7714                  * rsm->r_start in case we get an old ack where th_ack is
7715                  * behind snd_una.
7716                  */
7717                 bbr_peer_reneges(bbr, rsm, th->th_ack);
7718         }
7719         if ((to->to_flags & TOF_SACK) == 0) {
7720                 /* We are done nothing left to log */
7721                 goto out;
7722         }
7723         rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7724         if (rsm) {
7725                 last_seq = rsm->r_end;
7726         } else {
7727                 last_seq = tp->snd_max;
7728         }
7729         /* Sack block processing */
7730         if (SEQ_GT(th_ack, tp->snd_una))
7731                 ack_point = th_ack;
7732         else
7733                 ack_point = tp->snd_una;
7734         for (i = 0; i < to->to_nsacks; i++) {
7735                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
7736                     &sack, sizeof(sack));
7737                 sack.start = ntohl(sack.start);
7738                 sack.end = ntohl(sack.end);
7739                 if (SEQ_GT(sack.end, sack.start) &&
7740                     SEQ_GT(sack.start, ack_point) &&
7741                     SEQ_LT(sack.start, tp->snd_max) &&
7742                     SEQ_GT(sack.end, ack_point) &&
7743                     SEQ_LEQ(sack.end, tp->snd_max)) {
7744                         if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) &&
7745                             (SEQ_LT(sack.end, last_seq)) &&
7746                             ((sack.end - sack.start) < (p_maxseg / 8))) {
7747                                 /*
7748                                  * Not the last piece and its smaller than
7749                                  * 1/8th of a p_maxseg. We ignore this.
7750                                  */
7751                                 BBR_STAT_INC(bbr_runt_sacks);
7752                                 continue;
7753                         }
7754                         sack_blocks[num_sack_blks] = sack;
7755                         num_sack_blks++;
7756 #ifdef NETFLIX_STATS
7757                 } else if (SEQ_LEQ(sack.start, th_ack) &&
7758                     SEQ_LEQ(sack.end, th_ack)) {
7759                         /*
7760                          * Its a D-SACK block.
7761                          */
7762                         tcp_record_dsack(sack.start, sack.end);
7763 #endif
7764                 }
7765         }
7766         if (num_sack_blks == 0)
7767                 goto out;
7768         /*
7769          * Sort the SACK blocks so we can update the rack scoreboard with
7770          * just one pass.
7771          */
7772         new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks,
7773                                   num_sack_blks, th->th_ack);
7774         ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks);
7775         BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks);
7776         BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb));
7777         num_sack_blks = new_sb;
7778         if (num_sack_blks < 2) {
7779                 goto do_sack_work;
7780         }
7781         /* Sort the sacks */
7782         for (i = 0; i < num_sack_blks; i++) {
7783                 for (j = i + 1; j < num_sack_blks; j++) {
7784                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
7785                                 sack = sack_blocks[i];
7786                                 sack_blocks[i] = sack_blocks[j];
7787                                 sack_blocks[j] = sack;
7788                         }
7789                 }
7790         }
7791         /*
7792          * Now are any of the sack block ends the same (yes some
7793          * implememtations send these)?
7794          */
7795 again:
7796         if (num_sack_blks > 1) {
7797                 for (i = 0; i < num_sack_blks; i++) {
7798                         for (j = i + 1; j < num_sack_blks; j++) {
7799                                 if (sack_blocks[i].end == sack_blocks[j].end) {
7800                                         /*
7801                                          * Ok these two have the same end we
7802                                          * want the smallest end and then
7803                                          * throw away the larger and start
7804                                          * again.
7805                                          */
7806                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
7807                                                 /*
7808                                                  * The second block covers
7809                                                  * more area use that
7810                                                  */
7811                                                 sack_blocks[i].start = sack_blocks[j].start;
7812                                         }
7813                                         /*
7814                                          * Now collapse out the dup-sack and
7815                                          * lower the count
7816                                          */
7817                                         for (k = (j + 1); k < num_sack_blks; k++) {
7818                                                 sack_blocks[j].start = sack_blocks[k].start;
7819                                                 sack_blocks[j].end = sack_blocks[k].end;
7820                                                 j++;
7821                                         }
7822                                         num_sack_blks--;
7823                                         goto again;
7824                                 }
7825                         }
7826                 }
7827         }
7828 do_sack_work:
7829         rsm = bbr->r_ctl.rc_sacklast;
7830         for (i = 0; i < num_sack_blks; i++) {
7831                 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts);
7832                 if (acked) {
7833                         bbr->r_wanted_output = 1;
7834                         changed += acked;
7835                         sack_changed += acked;
7836                 }
7837         }
7838 out:
7839         *prev_acked = p_acked;
7840         if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) {
7841                 /*
7842                  * Ok we have a high probability that we need to go in to
7843                  * recovery since we have data sack'd
7844                  */
7845                 struct bbr_sendmap *rsm;
7846
7847                 rsm = bbr_check_recovery_mode(tp, bbr, cts);
7848                 if (rsm) {
7849                         /* Enter recovery */
7850                         entered_recovery = 1;
7851                         bbr->r_wanted_output = 1;
7852                         /*
7853                          * When we enter recovery we need to assure we send
7854                          * one packet.
7855                          */
7856                         if (bbr->r_ctl.rc_resend == NULL) {
7857                                 bbr->r_ctl.rc_resend = rsm;
7858                         }
7859                 }
7860         }
7861         if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) {
7862                 /*
7863                  * See if we need to rack-retransmit anything if so set it
7864                  * up as the thing to resend assuming something else is not
7865                  * already in that position.
7866                  */
7867                 if (bbr->r_ctl.rc_resend == NULL) {
7868                         bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
7869                 }
7870         }
7871         /*
7872          * We return the amount that changed via sack, this is used by the
7873          * ack-received code to augment what was changed between th_ack <->
7874          * snd_una.
7875          */
7876         return (sack_changed);
7877 }
7878
7879 static void
7880 bbr_strike_dupack(struct tcp_bbr *bbr)
7881 {
7882         struct bbr_sendmap *rsm;
7883
7884         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
7885         if (rsm && (rsm->r_dupack < 0xff)) {
7886                 rsm->r_dupack++;
7887                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD)
7888                         bbr->r_wanted_output = 1;
7889         }
7890 }
7891
7892 /*
7893  * Return value of 1, we do not need to call bbr_process_data().
7894  * return value of 0, bbr_process_data can be called.
7895  * For ret_val if its 0 the TCB is locked and valid, if its non-zero
7896  * its unlocked and probably unsafe to touch the TCB.
7897  */
7898 static int
7899 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
7900     struct tcpcb *tp, struct tcpopt *to,
7901     uint32_t tiwin, int32_t tlen,
7902     int32_t * ofia, int32_t thflags, int32_t * ret_val)
7903 {
7904         int32_t ourfinisacked = 0;
7905         int32_t acked_amount;
7906         uint16_t nsegs;
7907         int32_t acked;
7908         uint32_t lost, sack_changed = 0;
7909         struct mbuf *mfree;
7910         struct tcp_bbr *bbr;
7911         uint32_t prev_acked = 0;
7912
7913         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7914         lost = bbr->r_ctl.rc_lost;
7915         nsegs = max(1, m->m_pkthdr.lro_nsegs);
7916         if (SEQ_GT(th->th_ack, tp->snd_max)) {
7917                 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
7918                 bbr->r_wanted_output = 1;
7919                 return (1);
7920         }
7921         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
7922                 /* Process the ack */
7923                 if (bbr->rc_in_persist)
7924                         tp->t_rxtshift = 0;
7925                 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd))
7926                         bbr_strike_dupack(bbr);
7927                 sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
7928         }
7929         bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost));
7930         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
7931                 /*
7932                  * Old ack, behind the last one rcv'd or a duplicate ack
7933                  * with SACK info.
7934                  */
7935                 if (th->th_ack == tp->snd_una) {
7936                         bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0);
7937                         if (bbr->r_state == TCPS_SYN_SENT) {
7938                                 /*
7939                                  * Special case on where we sent SYN. When
7940                                  * the SYN-ACK is processed in syn_sent
7941                                  * state it bumps the snd_una. This causes
7942                                  * us to hit here even though we did ack 1
7943                                  * byte.
7944                                  *
7945                                  * Go through the nothing left case so we
7946                                  * send data.
7947                                  */
7948                                 goto nothing_left;
7949                         }
7950                 }
7951                 return (0);
7952         }
7953         /*
7954          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
7955          * something we sent.
7956          */
7957         if (tp->t_flags & TF_NEEDSYN) {
7958                 /*
7959                  * T/TCP: Connection was half-synchronized, and our SYN has
7960                  * been ACK'd (so connection is now fully synchronized).  Go
7961                  * to non-starred state, increment snd_una for ACK of SYN,
7962                  * and check if we can do window scaling.
7963                  */
7964                 tp->t_flags &= ~TF_NEEDSYN;
7965                 tp->snd_una++;
7966                 /* Do window scaling? */
7967                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
7968                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
7969                         tp->rcv_scale = tp->request_r_scale;
7970                         /* Send window already scaled. */
7971                 }
7972         }
7973         INP_WLOCK_ASSERT(tp->t_inpcb);
7974
7975         acked = BYTES_THIS_ACK(tp, th);
7976         TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
7977         TCPSTAT_ADD(tcps_rcvackbyte, acked);
7978
7979         /*
7980          * If we just performed our first retransmit, and the ACK arrives
7981          * within our recovery window, then it was a mistake to do the
7982          * retransmit in the first place.  Recover our original cwnd and
7983          * ssthresh, and proceed to transmit where we left off.
7984          */
7985         if (tp->t_flags & TF_PREVVALID) {
7986                 tp->t_flags &= ~TF_PREVVALID;
7987                 if (tp->t_rxtshift == 1 &&
7988                     (int)(ticks - tp->t_badrxtwin) < 0)
7989                         bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
7990         }
7991         SOCKBUF_LOCK(&so->so_snd);
7992         acked_amount = min(acked, (int)sbavail(&so->so_snd));
7993         tp->snd_wnd -= acked_amount;
7994         mfree = sbcut_locked(&so->so_snd, acked_amount);
7995         /* NB: sowwakeup_locked() does an implicit unlock. */
7996         sowwakeup_locked(so);
7997         m_freem(mfree);
7998         if (SEQ_GT(th->th_ack, tp->snd_una)) {
7999                 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
8000         }
8001         tp->snd_una = th->th_ack;
8002         bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost));
8003         if (IN_RECOVERY(tp->t_flags)) {
8004                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
8005                     (SEQ_LT(th->th_ack, tp->snd_max))) {
8006                         tcp_bbr_partialack(tp);
8007                 } else {
8008                         bbr_post_recovery(tp);
8009                 }
8010         }
8011         if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
8012                 tp->snd_recover = tp->snd_una;
8013         }
8014         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
8015                 tp->snd_nxt = tp->snd_max;
8016         }
8017         if (tp->snd_una == tp->snd_max) {
8018                 /* Nothing left outstanding */
8019 nothing_left:
8020                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
8021                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
8022                         bbr->rc_tp->t_acktime = 0;
8023                 if ((sbused(&so->so_snd) == 0) &&
8024                     (tp->t_flags & TF_SENTFIN)) {
8025                         ourfinisacked = 1;
8026                 }
8027                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8028                 if (bbr->rc_in_persist == 0) {
8029                         bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
8030                 }
8031                 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
8032                 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
8033                 /*
8034                  * We invalidate the last ack here since we
8035                  * don't want to transfer forward the time
8036                  * for our sum's calculations.
8037                  */
8038                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
8039                     (sbavail(&so->so_snd) == 0) &&
8040                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
8041                         /*
8042                          * The socket was gone and the peer sent data, time
8043                          * to reset him.
8044                          */
8045                         *ret_val = 1;
8046                         tp = tcp_close(tp);
8047                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
8048                         BBR_STAT_INC(bbr_dropped_af_data);
8049                         return (1);
8050                 }
8051                 /* Set need output so persist might get set */
8052                 bbr->r_wanted_output = 1;
8053         }
8054         if (ofia)
8055                 *ofia = ourfinisacked;
8056         return (0);
8057 }
8058
8059 static void
8060 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
8061 {
8062         if (bbr->rc_in_persist == 0) {
8063                 bbr_timer_cancel(bbr, __LINE__, cts);
8064                 bbr->r_ctl.rc_last_delay_val = 0;
8065                 tp->t_rxtshift = 0;
8066                 bbr->rc_in_persist = 1;
8067                 bbr->r_ctl.rc_went_idle_time = cts;
8068                 /* We should be capped when rw went to 0 but just in case */
8069                 bbr_log_type_pesist(bbr, cts, 0, line, 1);
8070                 /* Time freezes for the state, so do the accounting now */
8071                 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
8072                         uint32_t time_in;
8073
8074                         time_in = cts - bbr->r_ctl.rc_bbr_state_time;
8075                         if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
8076                                 int32_t idx;
8077
8078                                 idx = bbr_state_val(bbr);
8079                                 counter_u64_add(bbr_state_time[(idx + 5)], time_in);
8080                         } else {
8081                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
8082                         }
8083                 }
8084                 bbr->r_ctl.rc_bbr_state_time = cts;
8085         }
8086 }
8087
8088 static void
8089 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time)
8090 {
8091         /*
8092          * Note that if idle time does not exceed our
8093          * threshold, we do nothing continuing the state
8094          * transitions we were last walking through.
8095          */
8096         if (idle_time >= bbr_idle_restart_threshold) {
8097                 if (bbr->rc_use_idle_restart) {
8098                         bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT;
8099                         /*
8100                          * Set our target using BBR_UNIT, so
8101                          * we increase at a dramatic rate but
8102                          * we stop when we get the pipe
8103                          * full again for our current b/w estimate.
8104                          */
8105                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
8106                         bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
8107                         bbr_set_state_target(bbr, __LINE__);
8108                         /* Now setup our gains to ramp up */
8109                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
8110                         bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
8111                         bbr_log_type_statechange(bbr, cts, __LINE__);
8112                 } else {
8113                         bbr_substate_change(bbr, cts, __LINE__, 1);
8114                 }
8115         }
8116 }
8117
8118 static void
8119 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
8120 {
8121         uint32_t idle_time;
8122
8123         if (bbr->rc_in_persist == 0)
8124                 return;
8125         idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time);
8126         bbr->rc_in_persist = 0;
8127         bbr->rc_hit_state_1 = 0;
8128         tp->t_flags &= ~TF_FORCEDATA;
8129         bbr->r_ctl.rc_del_time = cts;
8130         /*
8131          * We invalidate the last ack here since we
8132          * don't want to transfer forward the time
8133          * for our sum's calculations.
8134          */
8135         if (bbr->rc_inp->inp_in_hpts) {
8136                 tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
8137                 bbr->rc_timer_first = 0;
8138                 bbr->r_ctl.rc_hpts_flags = 0;
8139                 bbr->r_ctl.rc_last_delay_val = 0;
8140                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
8141                 bbr->r_agg_early_set = 0;
8142                 bbr->r_ctl.rc_agg_early = 0;
8143         }
8144         bbr_log_type_pesist(bbr, cts, idle_time, line, 0);
8145         if (idle_time >= bbr_rtt_probe_time) {
8146                 /*
8147                  * This qualifies as a RTT_PROBE session since we drop the
8148                  * data outstanding to nothing and waited more than
8149                  * bbr_rtt_probe_time.
8150                  */
8151                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0);
8152                 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts;
8153         }
8154         tp->t_rxtshift = 0;
8155         /*
8156          * If in probeBW and we have persisted more than an RTT lets do
8157          * special handling.
8158          */
8159         /* Force a time based epoch */
8160         bbr_set_epoch(bbr, cts, __LINE__);
8161         /*
8162          * Setup the lost so we don't count anything against the guy
8163          * we have been stuck with during persists.
8164          */
8165         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
8166         /* Time un-freezes for the state */
8167         bbr->r_ctl.rc_bbr_state_time = cts;
8168         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) ||
8169             (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) {
8170                 /*
8171                  * If we are going back to probe-bw
8172                  * or probe_rtt, we may need to possibly
8173                  * do a fast restart.
8174                  */
8175                 bbr_restart_after_idle(bbr, cts, idle_time);
8176         }
8177 }
8178
8179 static void
8180 bbr_collapsed_window(struct tcp_bbr *bbr)
8181 {
8182         /*
8183          * Now we must walk the
8184          * send map and divide the
8185          * ones left stranded. These
8186          * guys can't cause us to abort
8187          * the connection and are really
8188          * "unsent". However if a buggy
8189          * client actually did keep some
8190          * of the data i.e. collapsed the win
8191          * and refused to ack and then opened
8192          * the win and acked that data. We would
8193          * get into an ack war, the simplier
8194          * method then of just pretending we
8195          * did not send those segments something
8196          * won't work.
8197          */
8198         struct bbr_sendmap *rsm, *nrsm;
8199         tcp_seq max_seq;
8200         uint32_t maxseg;
8201         int can_split = 0;
8202         int fnd = 0;
8203
8204         maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
8205         max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd;
8206         bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0);
8207         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
8208                 /* Find the first seq past or at maxseq */
8209                 if (rsm->r_flags & BBR_RWND_COLLAPSED)
8210                         rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8211                 if (SEQ_GEQ(max_seq, rsm->r_start) &&
8212                     SEQ_GEQ(rsm->r_end, max_seq)) {
8213                         fnd = 1;
8214                         break;
8215                 }
8216         }
8217         bbr->rc_has_collapsed = 0;
8218         if (!fnd) {
8219                 /* Nothing to do strange */
8220                 return;
8221         }
8222         /*
8223          * Now can we split?
8224          *
8225          * We don't want to split if splitting
8226          * would generate too many small segments
8227          * less we let an attacker fragment our
8228          * send_map and leave us out of memory.
8229          */
8230         if ((max_seq != rsm->r_start) &&
8231             (max_seq != rsm->r_end)){
8232                 /* can we split? */
8233                 int res1, res2;
8234
8235                 res1 = max_seq - rsm->r_start;
8236                 res2 = rsm->r_end - max_seq;
8237                 if ((res1 >= (maxseg/8)) &&
8238                     (res2 >= (maxseg/8))) {
8239                         /* No small pieces here */
8240                         can_split = 1;
8241                 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) {
8242                         /* We are under the limit */
8243                         can_split = 1;
8244                 }
8245         }
8246         /* Ok do we need to split this rsm? */
8247         if (max_seq == rsm->r_start) {
8248                 /* It's this guy no split required */
8249                 nrsm = rsm;
8250         } else if (max_seq == rsm->r_end) {
8251                 /* It's the next one no split required. */
8252                 nrsm = TAILQ_NEXT(rsm, r_next);
8253                 if (nrsm == NULL) {
8254                         /* Huh? */
8255                         return;
8256                 }
8257         } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) {
8258                 /* yep we need to split it */
8259                 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
8260                 if (nrsm == NULL) {
8261                         /* failed XXXrrs what can we do mark the whole? */
8262                         nrsm = rsm;
8263                         goto no_split;
8264                 }
8265                 /* Clone it */
8266                 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0);
8267                 bbr_clone_rsm(bbr, nrsm, rsm, max_seq);
8268                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
8269                 if (rsm->r_in_tmap) {
8270                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8271                         nrsm->r_in_tmap = 1;
8272                 }
8273         } else {
8274                 /*
8275                  * Split not allowed just start here just
8276                  * use this guy.
8277                  */
8278                 nrsm = rsm;
8279         }
8280 no_split:
8281         BBR_STAT_INC(bbr_collapsed_win);
8282         /* reuse fnd as a count */
8283         fnd = 0;
8284         TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) {
8285                 nrsm->r_flags |= BBR_RWND_COLLAPSED;
8286                 fnd++;
8287                 bbr->rc_has_collapsed = 1;
8288         }
8289         bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd);
8290 }
8291
8292 static void
8293 bbr_un_collapse_window(struct tcp_bbr *bbr)
8294 {
8295         struct bbr_sendmap *rsm;
8296         int cleared = 0;
8297
8298         TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
8299                 if (rsm->r_flags & BBR_RWND_COLLAPSED) {
8300                         /* Clear the flag */
8301                         rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8302                         cleared++;
8303                 } else
8304                         break;
8305         }
8306         bbr_log_type_rwnd_collapse(bbr,
8307                                    (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared);
8308         bbr->rc_has_collapsed = 0;
8309 }
8310
8311 /*
8312  * Return value of 1, the TCB is unlocked and most
8313  * likely gone, return value of 0, the TCB is still
8314  * locked.
8315  */
8316 static int
8317 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
8318     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
8319     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8320 {
8321         /*
8322          * Update window information. Don't look at window if no ACK: TAC's
8323          * send garbage on first SYN.
8324          */
8325         uint16_t nsegs;
8326         int32_t tfo_syn;
8327         struct tcp_bbr *bbr;
8328
8329         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8330         INP_WLOCK_ASSERT(tp->t_inpcb);
8331         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8332         if ((thflags & TH_ACK) &&
8333             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
8334             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
8335             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
8336                 /* keep track of pure window updates */
8337                 if (tlen == 0 &&
8338                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
8339                         TCPSTAT_INC(tcps_rcvwinupd);
8340                 tp->snd_wnd = tiwin;
8341                 tp->snd_wl1 = th->th_seq;
8342                 tp->snd_wl2 = th->th_ack;
8343                 if (tp->snd_wnd > tp->max_sndwnd)
8344                         tp->max_sndwnd = tp->snd_wnd;
8345                 bbr->r_wanted_output = 1;
8346         } else if (thflags & TH_ACK) {
8347                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
8348                         tp->snd_wnd = tiwin;
8349                         tp->snd_wl1 = th->th_seq;
8350                         tp->snd_wl2 = th->th_ack;
8351                 }
8352         }
8353         if (tp->snd_wnd < ctf_outstanding(tp))
8354                 /* The peer collapsed its window on us */
8355                 bbr_collapsed_window(bbr);
8356         else if (bbr->rc_has_collapsed)
8357                 bbr_un_collapse_window(bbr);
8358         /* Was persist timer active and now we have window space? */
8359         if ((bbr->rc_in_persist != 0) &&
8360             (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8361                                 bbr_minseg(bbr)))) {
8362                 /*
8363                  * Make the rate persist at end of persist mode if idle long
8364                  * enough
8365                  */
8366                 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8367
8368                 /* Make sure we output to start the timer */
8369                 bbr->r_wanted_output = 1;
8370         }
8371         /* Do we need to enter persist? */
8372         if ((bbr->rc_in_persist == 0) &&
8373             (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8374             TCPS_HAVEESTABLISHED(tp->t_state) &&
8375             (tp->snd_max == tp->snd_una) &&
8376             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8377             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8378                 /* No send window.. we must enter persist */
8379                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8380         }
8381         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
8382                 m_freem(m);
8383                 return (0);
8384         }
8385         /*
8386          * Process segments with URG.
8387          */
8388         if ((thflags & TH_URG) && th->th_urp &&
8389             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8390                 /*
8391                  * This is a kludge, but if we receive and accept random
8392                  * urgent pointers, we'll crash in soreceive.  It's hard to
8393                  * imagine someone actually wanting to send this much urgent
8394                  * data.
8395                  */
8396                 SOCKBUF_LOCK(&so->so_rcv);
8397                 if (th->th_urp + sbavail(&so->so_rcv) > sb_max) {
8398                         th->th_urp = 0; /* XXX */
8399                         thflags &= ~TH_URG;     /* XXX */
8400                         SOCKBUF_UNLOCK(&so->so_rcv);    /* XXX */
8401                         goto dodata;    /* XXX */
8402                 }
8403                 /*
8404                  * If this segment advances the known urgent pointer, then
8405                  * mark the data stream.  This should not happen in
8406                  * CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since a
8407                  * FIN has been received from the remote side. In these
8408                  * states we ignore the URG.
8409                  *
8410                  * According to RFC961 (Assigned Protocols), the urgent
8411                  * pointer points to the last octet of urgent data.  We
8412                  * continue, however, to consider it to indicate the first
8413                  * octet of data past the urgent section as the original
8414                  * spec states (in one of two places).
8415                  */
8416                 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) {
8417                         tp->rcv_up = th->th_seq + th->th_urp;
8418                         so->so_oobmark = sbavail(&so->so_rcv) +
8419                             (tp->rcv_up - tp->rcv_nxt) - 1;
8420                         if (so->so_oobmark == 0)
8421                                 so->so_rcv.sb_state |= SBS_RCVATMARK;
8422                         sohasoutofband(so);
8423                         tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA);
8424                 }
8425                 SOCKBUF_UNLOCK(&so->so_rcv);
8426                 /*
8427                  * Remove out of band data so doesn't get presented to user.
8428                  * This can happen independent of advancing the URG pointer,
8429                  * but if two URG's are pending at once, some out-of-band
8430                  * data may creep in... ick.
8431                  */
8432                 if (th->th_urp <= (uint32_t)tlen &&
8433                     !(so->so_options & SO_OOBINLINE)) {
8434                         /* hdr drop is delayed */
8435                         tcp_pulloutofband(so, th, m, drop_hdrlen);
8436                 }
8437         } else {
8438                 /*
8439                  * If no out of band data is expected, pull receive urgent
8440                  * pointer along with the receive window.
8441                  */
8442                 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up))
8443                         tp->rcv_up = tp->rcv_nxt;
8444         }
8445 dodata:                         /* XXX */
8446         INP_WLOCK_ASSERT(tp->t_inpcb);
8447
8448         /*
8449          * Process the segment text, merging it into the TCP sequencing
8450          * queue, and arranging for acknowledgment of receipt if necessary.
8451          * This process logically involves adjusting tp->rcv_wnd as data is
8452          * presented to the user (this happens in tcp_usrreq.c, case
8453          * PRU_RCVD).  If a FIN has already been received on this connection
8454          * then we just ignore the text.
8455          */
8456         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
8457                    IS_FASTOPEN(tp->t_flags));
8458         if ((tlen || (thflags & TH_FIN) || tfo_syn) &&
8459             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8460                 tcp_seq save_start = th->th_seq;
8461                 tcp_seq save_rnxt  = tp->rcv_nxt;
8462                 int     save_tlen  = tlen;
8463
8464                 m_adj(m, drop_hdrlen);  /* delayed header drop */
8465                 /*
8466                  * Insert segment which includes th into TCP reassembly
8467                  * queue with control block tp.  Set thflags to whether
8468                  * reassembly now includes a segment with FIN.  This handles
8469                  * the common case inline (segment is the next to be
8470                  * received on an established connection, and the queue is
8471                  * empty), avoiding linkage into and removal from the queue
8472                  * and repetition of various conversions. Set DELACK for
8473                  * segments received in order, but ack immediately when
8474                  * segments are out of order (so fast retransmit can work).
8475                  */
8476                 if (th->th_seq == tp->rcv_nxt &&
8477                     SEGQ_EMPTY(tp) &&
8478                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
8479                     tfo_syn)) {
8480 #ifdef NETFLIX_SB_LIMITS
8481                         u_int mcnt, appended;
8482
8483                         if (so->so_rcv.sb_shlim) {
8484                                 mcnt = m_memcnt(m);
8485                                 appended = 0;
8486                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8487                                     CFO_NOSLEEP, NULL) == false) {
8488                                         counter_u64_add(tcp_sb_shlim_fails, 1);
8489                                         m_freem(m);
8490                                         return (0);
8491                                 }
8492                         }
8493 #endif
8494                         if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) {
8495                                 bbr->bbr_segs_rcvd += max(1, nsegs);
8496                                 tp->t_flags |= TF_DELACK;
8497                                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8498                         } else {
8499                                 bbr->r_wanted_output = 1;
8500                                 tp->t_flags |= TF_ACKNOW;
8501                         }
8502                         tp->rcv_nxt += tlen;
8503                         thflags = th->th_flags & TH_FIN;
8504                         TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8505                         TCPSTAT_ADD(tcps_rcvbyte, tlen);
8506                         SOCKBUF_LOCK(&so->so_rcv);
8507                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
8508                                 m_freem(m);
8509                         else
8510 #ifdef NETFLIX_SB_LIMITS
8511                                 appended =
8512 #endif
8513                                         sbappendstream_locked(&so->so_rcv, m, 0);
8514                         /* NB: sorwakeup_locked() does an implicit unlock. */
8515                         sorwakeup_locked(so);
8516 #ifdef NETFLIX_SB_LIMITS
8517                         if (so->so_rcv.sb_shlim && appended != mcnt)
8518                                 counter_fo_release(so->so_rcv.sb_shlim,
8519                                     mcnt - appended);
8520 #endif
8521                 } else {
8522                         /*
8523                          * XXX: Due to the header drop above "th" is
8524                          * theoretically invalid by now.  Fortunately
8525                          * m_adj() doesn't actually frees any mbufs when
8526                          * trimming from the head.
8527                          */
8528                         tcp_seq temp = save_start;
8529                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
8530                         tp->t_flags |= TF_ACKNOW;
8531                 }
8532                 if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0)) {
8533                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
8534                                 /*
8535                                  * DSACK actually handled in the fastpath
8536                                  * above.
8537                                  */
8538                                 tcp_update_sack_list(tp, save_start,
8539                                     save_start + save_tlen);
8540                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
8541                                 if ((tp->rcv_numsacks >= 1) &&
8542                                     (tp->sackblks[0].end == save_start)) {
8543                                         /*
8544                                          * Partial overlap, recorded at todrop
8545                                          * above.
8546                                          */
8547                                         tcp_update_sack_list(tp,
8548                                             tp->sackblks[0].start,
8549                                             tp->sackblks[0].end);
8550                                 } else {
8551                                         tcp_update_dsack_list(tp, save_start,
8552                                             save_start + save_tlen);
8553                                 }
8554                         } else if (tlen >= save_tlen) {
8555                                 /* Update of sackblks. */
8556                                 tcp_update_dsack_list(tp, save_start,
8557                                     save_start + save_tlen);
8558                         } else if (tlen > 0) {
8559                                 tcp_update_dsack_list(tp, save_start,
8560                                     save_start + tlen);
8561                         }
8562                 }
8563         } else {
8564                 m_freem(m);
8565                 thflags &= ~TH_FIN;
8566         }
8567
8568         /*
8569          * If FIN is received ACK the FIN and let the user know that the
8570          * connection is closing.
8571          */
8572         if (thflags & TH_FIN) {
8573                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8574                         socantrcvmore(so);
8575                         /*
8576                          * If connection is half-synchronized (ie NEEDSYN
8577                          * flag on) then delay ACK, so it may be piggybacked
8578                          * when SYN is sent. Otherwise, since we received a
8579                          * FIN then no more input can be expected, send ACK
8580                          * now.
8581                          */
8582                         if (tp->t_flags & TF_NEEDSYN) {
8583                                 tp->t_flags |= TF_DELACK;
8584                                 bbr_timer_cancel(bbr,
8585                                     __LINE__, bbr->r_ctl.rc_rcvtime);
8586                         } else {
8587                                 tp->t_flags |= TF_ACKNOW;
8588                         }
8589                         tp->rcv_nxt++;
8590                 }
8591                 switch (tp->t_state) {
8592
8593                         /*
8594                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
8595                          * CLOSE_WAIT state.
8596                          */
8597                 case TCPS_SYN_RECEIVED:
8598                         tp->t_starttime = ticks;
8599                         /* FALLTHROUGH */
8600                 case TCPS_ESTABLISHED:
8601                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
8602                         break;
8603
8604                         /*
8605                          * If still in FIN_WAIT_1 STATE FIN has not been
8606                          * acked so enter the CLOSING state.
8607                          */
8608                 case TCPS_FIN_WAIT_1:
8609                         tcp_state_change(tp, TCPS_CLOSING);
8610                         break;
8611
8612                         /*
8613                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
8614                          * starting the time-wait timer, turning off the
8615                          * other standard timers.
8616                          */
8617                 case TCPS_FIN_WAIT_2:
8618                         bbr->rc_timer_first = 1;
8619                         bbr_timer_cancel(bbr,
8620                             __LINE__, bbr->r_ctl.rc_rcvtime);
8621                         INP_WLOCK_ASSERT(tp->t_inpcb);
8622                         tcp_twstart(tp);
8623                         return (1);
8624                 }
8625         }
8626         /*
8627          * Return any desired output.
8628          */
8629         if ((tp->t_flags & TF_ACKNOW) ||
8630             (sbavail(&so->so_snd) > ctf_outstanding(tp))) {
8631                 bbr->r_wanted_output = 1;
8632         }
8633         INP_WLOCK_ASSERT(tp->t_inpcb);
8634         return (0);
8635 }
8636
8637 /*
8638  * Here nothing is really faster, its just that we
8639  * have broken out the fast-data path also just like
8640  * the fast-ack. Return 1 if we processed the packet
8641  * return 0 if you need to take the "slow-path".
8642  */
8643 static int
8644 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
8645     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8646     uint32_t tiwin, int32_t nxt_pkt)
8647 {
8648         uint16_t nsegs;
8649         int32_t newsize = 0;    /* automatic sockbuf scaling */
8650         struct tcp_bbr *bbr;
8651 #ifdef NETFLIX_SB_LIMITS
8652         u_int mcnt, appended;
8653 #endif
8654 #ifdef TCPDEBUG
8655         /*
8656          * The size of tcp_saveipgen must be the size of the max ip header,
8657          * now IPv6.
8658          */
8659         u_char tcp_saveipgen[IP6_HDR_LEN];
8660         struct tcphdr tcp_savetcp;
8661         short ostate = 0;
8662
8663 #endif
8664         /* On the hpts and we would have called output */
8665         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8666
8667         /*
8668          * If last ACK falls within this segment's sequence numbers, record
8669          * the timestamp. NOTE that the test is modified according to the
8670          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8671          */
8672         if (bbr->r_ctl.rc_resend != NULL) {
8673                 return (0);
8674         }
8675         if (tiwin && tiwin != tp->snd_wnd) {
8676                 return (0);
8677         }
8678         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
8679                 return (0);
8680         }
8681         if (__predict_false((to->to_flags & TOF_TS) &&
8682             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
8683                 return (0);
8684         }
8685         if (__predict_false((th->th_ack != tp->snd_una))) {
8686                 return (0);
8687         }
8688         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
8689                 return (0);
8690         }
8691         if ((to->to_flags & TOF_TS) != 0 &&
8692             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8693                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
8694                 tp->ts_recent = to->to_tsval;
8695         }
8696         /*
8697          * This is a pure, in-sequence data packet with nothing on the
8698          * reassembly queue and we have enough buffer space to take it.
8699          */
8700         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8701 #ifdef NETFLIX_SB_LIMITS
8702         if (so->so_rcv.sb_shlim) {
8703                 mcnt = m_memcnt(m);
8704                 appended = 0;
8705                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8706                     CFO_NOSLEEP, NULL) == false) {
8707                         counter_u64_add(tcp_sb_shlim_fails, 1);
8708                         m_freem(m);
8709                         return (1);
8710                 }
8711         }
8712 #endif
8713         /* Clean receiver SACK report if present */
8714         if (tp->rcv_numsacks)
8715                 tcp_clean_sackreport(tp);
8716         TCPSTAT_INC(tcps_preddat);
8717         tp->rcv_nxt += tlen;
8718         /*
8719          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
8720          */
8721         tp->snd_wl1 = th->th_seq;
8722         /*
8723          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
8724          */
8725         tp->rcv_up = tp->rcv_nxt;
8726         TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8727         TCPSTAT_ADD(tcps_rcvbyte, tlen);
8728 #ifdef TCPDEBUG
8729         if (so->so_options & SO_DEBUG)
8730                 tcp_trace(TA_INPUT, ostate, tp,
8731                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
8732 #endif
8733         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
8734
8735         /* Add data to socket buffer. */
8736         SOCKBUF_LOCK(&so->so_rcv);
8737         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8738                 m_freem(m);
8739         } else {
8740                 /*
8741                  * Set new socket buffer size. Give up when limit is
8742                  * reached.
8743                  */
8744                 if (newsize)
8745                         if (!sbreserve_locked(&so->so_rcv,
8746                             newsize, so, NULL))
8747                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
8748                 m_adj(m, drop_hdrlen);  /* delayed header drop */
8749 #ifdef NETFLIX_SB_LIMITS
8750                 appended =
8751 #endif
8752                         sbappendstream_locked(&so->so_rcv, m, 0);
8753                 ctf_calc_rwin(so, tp);
8754         }
8755         /* NB: sorwakeup_locked() does an implicit unlock. */
8756         sorwakeup_locked(so);
8757 #ifdef NETFLIX_SB_LIMITS
8758         if (so->so_rcv.sb_shlim && mcnt != appended)
8759                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
8760 #endif
8761         if (DELAY_ACK(tp, bbr, nsegs)) {
8762                 bbr->bbr_segs_rcvd += max(1, nsegs);
8763                 tp->t_flags |= TF_DELACK;
8764                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8765         } else {
8766                 bbr->r_wanted_output = 1;
8767                 tp->t_flags |= TF_ACKNOW;
8768         }
8769         return (1);
8770 }
8771
8772 /*
8773  * This subfunction is used to try to highly optimize the
8774  * fast path. We again allow window updates that are
8775  * in sequence to remain in the fast-path. We also add
8776  * in the __predict's to attempt to help the compiler.
8777  * Note that if we return a 0, then we can *not* process
8778  * it and the caller should push the packet into the
8779  * slow-path. If we return 1, then all is well and
8780  * the packet is fully processed.
8781  */
8782 static int
8783 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8784     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8785     uint32_t tiwin, int32_t nxt_pkt)
8786 {
8787         int32_t acked;
8788         uint16_t nsegs;
8789         uint32_t sack_changed;
8790 #ifdef TCPDEBUG
8791         /*
8792          * The size of tcp_saveipgen must be the size of the max ip header,
8793          * now IPv6.
8794          */
8795         u_char tcp_saveipgen[IP6_HDR_LEN];
8796         struct tcphdr tcp_savetcp;
8797         short ostate = 0;
8798
8799 #endif
8800         uint32_t prev_acked = 0;
8801         struct tcp_bbr *bbr;
8802
8803         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8804                 /* Old ack, behind (or duplicate to) the last one rcv'd */
8805                 return (0);
8806         }
8807         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
8808                 /* Above what we have sent? */
8809                 return (0);
8810         }
8811         if (__predict_false(tiwin == 0)) {
8812                 /* zero window */
8813                 return (0);
8814         }
8815         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
8816                 /* We need a SYN or a FIN, unlikely.. */
8817                 return (0);
8818         }
8819         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
8820                 /* Timestamp is behind .. old ack with seq wrap? */
8821                 return (0);
8822         }
8823         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
8824                 /* Still recovering */
8825                 return (0);
8826         }
8827         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8828         if (__predict_false(bbr->r_ctl.rc_resend != NULL)) {
8829                 /* We are retransmitting */
8830                 return (0);
8831         }
8832         if (__predict_false(bbr->rc_in_persist != 0)) {
8833                 /* In persist mode */
8834                 return (0);
8835         }
8836         if (bbr->r_ctl.rc_sacked) {
8837                 /* We have sack holes on our scoreboard */
8838                 return (0);
8839         }
8840         /* Ok if we reach here, we can process a fast-ack */
8841         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8842         sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
8843         /*
8844          * We never detect loss in fast ack [we can't
8845          * have a sack and can't be in recovery so
8846          * we always pass 0 (nothing detected)].
8847          */
8848         bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0);
8849         /* Did the window get updated? */
8850         if (tiwin != tp->snd_wnd) {
8851                 tp->snd_wnd = tiwin;
8852                 tp->snd_wl1 = th->th_seq;
8853                 if (tp->snd_wnd > tp->max_sndwnd)
8854                         tp->max_sndwnd = tp->snd_wnd;
8855         }
8856         /* Do we need to exit persists? */
8857         if ((bbr->rc_in_persist != 0) &&
8858             (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8859                                bbr_minseg(bbr)))) {
8860                 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8861                 bbr->r_wanted_output = 1;
8862         }
8863         /* Do we need to enter persists? */
8864         if ((bbr->rc_in_persist == 0) &&
8865             (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8866             TCPS_HAVEESTABLISHED(tp->t_state) &&
8867             (tp->snd_max == tp->snd_una) &&
8868             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8869             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8870                 /* No send window.. we must enter persist */
8871                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8872         }
8873         /*
8874          * If last ACK falls within this segment's sequence numbers, record
8875          * the timestamp. NOTE that the test is modified according to the
8876          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8877          */
8878         if ((to->to_flags & TOF_TS) != 0 &&
8879             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8880                 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime;
8881                 tp->ts_recent = to->to_tsval;
8882         }
8883         /*
8884          * This is a pure ack for outstanding data.
8885          */
8886         TCPSTAT_INC(tcps_predack);
8887
8888         /*
8889          * "bad retransmit" recovery.
8890          */
8891         if (tp->t_flags & TF_PREVVALID) {
8892                 tp->t_flags &= ~TF_PREVVALID;
8893                 if (tp->t_rxtshift == 1 &&
8894                     (int)(ticks - tp->t_badrxtwin) < 0)
8895                         bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
8896         }
8897         /*
8898          * Recalculate the transmit timer / rtt.
8899          *
8900          * Some boxes send broken timestamp replies during the SYN+ACK
8901          * phase, ignore timestamps of 0 or we could calculate a huge RTT
8902          * and blow up the retransmit timer.
8903          */
8904         acked = BYTES_THIS_ACK(tp, th);
8905
8906 #ifdef TCP_HHOOK
8907         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
8908         hhook_run_tcp_est_in(tp, th, to);
8909 #endif
8910
8911         TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
8912         TCPSTAT_ADD(tcps_rcvackbyte, acked);
8913         sbdrop(&so->so_snd, acked);
8914
8915         if (SEQ_GT(th->th_ack, tp->snd_una))
8916                 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
8917         tp->snd_una = th->th_ack;
8918         if (tp->snd_wnd < ctf_outstanding(tp))
8919                 /* The peer collapsed its window on us */
8920                 bbr_collapsed_window(bbr);
8921         else if (bbr->rc_has_collapsed)
8922                 bbr_un_collapse_window(bbr);
8923
8924         if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
8925                 tp->snd_recover = tp->snd_una;
8926         }
8927         bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0);
8928         /*
8929          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
8930          */
8931         tp->snd_wl2 = th->th_ack;
8932         m_freem(m);
8933         /*
8934          * If all outstanding data are acked, stop retransmit timer,
8935          * otherwise restart timer using current (possibly backed-off)
8936          * value. If process is waiting for space, wakeup/selwakeup/signal.
8937          * If data are ready to send, let tcp_output decide between more
8938          * output or persist.
8939          */
8940 #ifdef TCPDEBUG
8941         if (so->so_options & SO_DEBUG)
8942                 tcp_trace(TA_INPUT, ostate, tp,
8943                     (void *)tcp_saveipgen,
8944                     &tcp_savetcp, 0);
8945 #endif
8946         /* Wake up the socket if we have room to write more */
8947         sowwakeup(so);
8948         if (tp->snd_una == tp->snd_max) {
8949                 /* Nothing left outstanding */
8950                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
8951                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
8952                         bbr->rc_tp->t_acktime = 0;
8953                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8954                 if (bbr->rc_in_persist == 0) {
8955                         bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
8956                 }
8957                 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
8958                 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
8959                 /*
8960                  * We invalidate the last ack here since we
8961                  * don't want to transfer forward the time
8962                  * for our sum's calculations.
8963                  */
8964                 bbr->r_wanted_output = 1;
8965         }
8966         if (sbavail(&so->so_snd)) {
8967                 bbr->r_wanted_output = 1;
8968         }
8969         return (1);
8970 }
8971
8972 /*
8973  * Return value of 1, the TCB is unlocked and most
8974  * likely gone, return value of 0, the TCB is still
8975  * locked.
8976  */
8977 static int
8978 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
8979     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8980     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8981 {
8982         int32_t todrop;
8983         int32_t ourfinisacked = 0;
8984         struct tcp_bbr *bbr;
8985         int32_t ret_val = 0;
8986
8987         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8988         ctf_calc_rwin(so, tp);
8989         /*
8990          * If the state is SYN_SENT: if seg contains an ACK, but not for our
8991          * SYN, drop the input. if seg contains a RST, then drop the
8992          * connection. if seg does not contain SYN, then drop it. Otherwise
8993          * this is an acceptable SYN segment initialize tp->rcv_nxt and
8994          * tp->irs if seg contains ack then advance tp->snd_una. BRR does
8995          * not support ECN so we will not say we are capable. if SYN has
8996          * been acked change to ESTABLISHED else SYN_RCVD state arrange for
8997          * segment to be acked (eventually) continue processing rest of
8998          * data/controls, beginning with URG
8999          */
9000         if ((thflags & TH_ACK) &&
9001             (SEQ_LEQ(th->th_ack, tp->iss) ||
9002             SEQ_GT(th->th_ack, tp->snd_max))) {
9003                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9004                 return (1);
9005         }
9006         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
9007                 TCP_PROBE5(connect__refused, NULL, tp,
9008                     mtod(m, const char *), tp, th);
9009                 tp = tcp_drop(tp, ECONNREFUSED);
9010                 ctf_do_drop(m, tp);
9011                 return (1);
9012         }
9013         if (thflags & TH_RST) {
9014                 ctf_do_drop(m, tp);
9015                 return (1);
9016         }
9017         if (!(thflags & TH_SYN)) {
9018                 ctf_do_drop(m, tp);
9019                 return (1);
9020         }
9021         tp->irs = th->th_seq;
9022         tcp_rcvseqinit(tp);
9023         if (thflags & TH_ACK) {
9024                 int tfo_partial = 0;
9025
9026                 TCPSTAT_INC(tcps_connects);
9027                 soisconnected(so);
9028 #ifdef MAC
9029                 mac_socketpeer_set_from_mbuf(m, so);
9030 #endif
9031                 /* Do window scaling on this connection? */
9032                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9033                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9034                         tp->rcv_scale = tp->request_r_scale;
9035                 }
9036                 tp->rcv_adv += min(tp->rcv_wnd,
9037                     TCP_MAXWIN << tp->rcv_scale);
9038                 /*
9039                  * If not all the data that was sent in the TFO SYN
9040                  * has been acked, resend the remainder right away.
9041                  */
9042                 if (IS_FASTOPEN(tp->t_flags) &&
9043                     (tp->snd_una != tp->snd_max)) {
9044                         tp->snd_nxt = th->th_ack;
9045                         tfo_partial = 1;
9046                 }
9047                 /*
9048                  * If there's data, delay ACK; if there's also a FIN ACKNOW
9049                  * will be turned on later.
9050                  */
9051                 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && (tfo_partial == 0)) {
9052                         bbr->bbr_segs_rcvd += 1;
9053                         tp->t_flags |= TF_DELACK;
9054                         bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
9055                 } else {
9056                         bbr->r_wanted_output = 1;
9057                         tp->t_flags |= TF_ACKNOW;
9058                 }
9059                 if (SEQ_GT(th->th_ack, tp->iss)) {
9060                         /*
9061                          * The SYN is acked
9062                          * handle it specially.
9063                          */
9064                         bbr_log_syn(tp, to);
9065                 }
9066                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
9067                         /*
9068                          * We advance snd_una for the
9069                          * fast open case. If th_ack is
9070                          * acknowledging data beyond
9071                          * snd_una we can't just call
9072                          * ack-processing since the
9073                          * data stream in our send-map
9074                          * will start at snd_una + 1 (one
9075                          * beyond the SYN). If its just
9076                          * equal we don't need to do that
9077                          * and there is no send_map.
9078                          */
9079                         tp->snd_una++;
9080                 }
9081                 /*
9082                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
9083                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
9084                  */
9085                 tp->t_starttime = ticks;
9086                 if (tp->t_flags & TF_NEEDFIN) {
9087                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
9088                         tp->t_flags &= ~TF_NEEDFIN;
9089                         thflags &= ~TH_SYN;
9090                 } else {
9091                         tcp_state_change(tp, TCPS_ESTABLISHED);
9092                         TCP_PROBE5(connect__established, NULL, tp,
9093                             mtod(m, const char *), tp, th);
9094                         cc_conn_init(tp);
9095                 }
9096         } else {
9097                 /*
9098                  * Received initial SYN in SYN-SENT[*] state => simultaneous
9099                  * open.  If segment contains CC option and there is a
9100                  * cached CC, apply TAO test. If it succeeds, connection is *
9101                  * half-synchronized. Otherwise, do 3-way handshake:
9102                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
9103                  * there was no CC option, clear cached CC value.
9104                  */
9105                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
9106                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
9107         }
9108         INP_WLOCK_ASSERT(tp->t_inpcb);
9109         /*
9110          * Advance th->th_seq to correspond to first data byte. If data,
9111          * trim to stay within window, dropping FIN if necessary.
9112          */
9113         th->th_seq++;
9114         if (tlen > tp->rcv_wnd) {
9115                 todrop = tlen - tp->rcv_wnd;
9116                 m_adj(m, -todrop);
9117                 tlen = tp->rcv_wnd;
9118                 thflags &= ~TH_FIN;
9119                 TCPSTAT_INC(tcps_rcvpackafterwin);
9120                 TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
9121         }
9122         tp->snd_wl1 = th->th_seq - 1;
9123         tp->rcv_up = th->th_seq;
9124         /*
9125          * Client side of transaction: already sent SYN and data. If the
9126          * remote host used T/TCP to validate the SYN, our data will be
9127          * ACK'd; if so, enter normal data segment processing in the middle
9128          * of step 5, ack processing. Otherwise, goto step 6.
9129          */
9130         if (thflags & TH_ACK) {
9131                 if ((to->to_flags & TOF_TS) != 0) {
9132                         uint32_t t, rtt;
9133
9134                         t = tcp_tv_to_mssectick(&bbr->rc_tv);
9135                         if (TSTMP_GEQ(t, to->to_tsecr)) {
9136                                 rtt = t - to->to_tsecr;
9137                                 if (rtt == 0) {
9138                                         rtt = 1;
9139                                 }
9140                                 rtt *= MS_IN_USEC;
9141                                 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9142                                 apply_filter_min_small(&bbr->r_ctl.rc_rttprop,
9143                                                        rtt, bbr->r_ctl.rc_rcvtime);
9144                         }
9145                 }
9146                 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
9147                         return (ret_val);
9148                 /* We may have changed to FIN_WAIT_1 above */
9149                 if (tp->t_state == TCPS_FIN_WAIT_1) {
9150                         /*
9151                          * In FIN_WAIT_1 STATE in addition to the processing
9152                          * for the ESTABLISHED state if our FIN is now
9153                          * acknowledged then enter FIN_WAIT_2.
9154                          */
9155                         if (ourfinisacked) {
9156                                 /*
9157                                  * If we can't receive any more data, then
9158                                  * closing user can proceed. Starting the
9159                                  * timer is contrary to the specification,
9160                                  * but if we don't get a FIN we'll hang
9161                                  * forever.
9162                                  *
9163                                  * XXXjl: we should release the tp also, and
9164                                  * use a compressed state.
9165                                  */
9166                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9167                                         soisdisconnected(so);
9168                                         tcp_timer_activate(tp, TT_2MSL,
9169                                             (tcp_fast_finwait2_recycle ?
9170                                             tcp_finwait2_timeout :
9171                                             TP_MAXIDLE(tp)));
9172                                 }
9173                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
9174                         }
9175                 }
9176         }
9177         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9178             tiwin, thflags, nxt_pkt));
9179 }
9180
9181 /*
9182  * Return value of 1, the TCB is unlocked and most
9183  * likely gone, return value of 0, the TCB is still
9184  * locked.
9185  */
9186 static int
9187 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
9188                 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9189                 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
9190 {
9191         int32_t ourfinisacked = 0;
9192         int32_t ret_val;
9193         struct tcp_bbr *bbr;
9194
9195         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9196         ctf_calc_rwin(so, tp);
9197         if ((thflags & TH_ACK) &&
9198             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
9199              SEQ_GT(th->th_ack, tp->snd_max))) {
9200                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9201                 return (1);
9202         }
9203         if (IS_FASTOPEN(tp->t_flags)) {
9204                 /*
9205                  * When a TFO connection is in SYN_RECEIVED, the only valid
9206                  * packets are the initial SYN, a retransmit/copy of the
9207                  * initial SYN (possibly with a subset of the original
9208                  * data), a valid ACK, a FIN, or a RST.
9209                  */
9210                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
9211                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9212                         return (1);
9213                 } else if (thflags & TH_SYN) {
9214                         /* non-initial SYN is ignored */
9215                         if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
9216                             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
9217                             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
9218                                 ctf_do_drop(m, NULL);
9219                                 return (0);
9220                         }
9221                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
9222                         ctf_do_drop(m, NULL);
9223                         return (0);
9224                 }
9225         }
9226         if ((thflags & TH_RST) ||
9227             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9228                 return (ctf_process_rst(m, th, so, tp));
9229         /*
9230          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9231          * it's less than ts_recent, drop it.
9232          */
9233         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9234             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9235                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9236                         return (ret_val);
9237         }
9238         /*
9239          * In the SYN-RECEIVED state, validate that the packet belongs to
9240          * this connection before trimming the data to fit the receive
9241          * window.  Check the sequence number versus IRS since we know the
9242          * sequence numbers haven't wrapped.  This is a partial fix for the
9243          * "LAND" DoS attack.
9244          */
9245         if (SEQ_LT(th->th_seq, tp->irs)) {
9246                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9247                 return (1);
9248         }
9249         INP_WLOCK_ASSERT(tp->t_inpcb);
9250         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9251                 return (ret_val);
9252         }
9253         /*
9254          * If last ACK falls within this segment's sequence numbers, record
9255          * its timestamp. NOTE: 1) That the test incorporates suggestions
9256          * from the latest proposal of the tcplw@cray.com list (Braden
9257          * 1993/04/26). 2) That updating only on newer timestamps interferes
9258          * with our earlier PAWS tests, so this check should be solely
9259          * predicated on the sequence space of this segment. 3) That we
9260          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9261          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9262          * SEG.Len, This modified check allows us to overcome RFC1323's
9263          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9264          * p.869. In such cases, we can still calculate the RTT correctly
9265          * when RCV.NXT == Last.ACK.Sent.
9266          */
9267         if ((to->to_flags & TOF_TS) != 0 &&
9268             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9269             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9270                     ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9271                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9272                 tp->ts_recent = to->to_tsval;
9273         }
9274         tp->snd_wnd = tiwin;
9275         /*
9276          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9277          * is on (half-synchronized state), then queue data for later
9278          * processing; else drop segment and return.
9279          */
9280         if ((thflags & TH_ACK) == 0) {
9281                 if (IS_FASTOPEN(tp->t_flags)) {
9282                         cc_conn_init(tp);
9283                 }
9284                 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9285                                          tiwin, thflags, nxt_pkt));
9286         }
9287         TCPSTAT_INC(tcps_connects);
9288         soisconnected(so);
9289         /* Do window scaling? */
9290         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9291             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9292                 tp->rcv_scale = tp->request_r_scale;
9293         }
9294         /*
9295          * ok for the first time in lets see if we can use the ts to figure
9296          * out what the initial RTT was.
9297          */
9298         if ((to->to_flags & TOF_TS) != 0) {
9299                 uint32_t t, rtt;
9300
9301                 t = tcp_tv_to_mssectick(&bbr->rc_tv);
9302                 if (TSTMP_GEQ(t, to->to_tsecr)) {
9303                         rtt = t - to->to_tsecr;
9304                         if (rtt == 0) {
9305                                 rtt = 1;
9306                         }
9307                         rtt *= MS_IN_USEC;
9308                         tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9309                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime);
9310                 }
9311         }
9312         /* Drop off any SYN in the send map (probably not there)  */
9313         if (thflags & TH_ACK)
9314                 bbr_log_syn(tp, to);
9315         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
9316
9317                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
9318                 tp->t_tfo_pending = NULL;
9319                 /*
9320                  * Account for the ACK of our SYN prior to regular
9321                  * ACK processing below.
9322                  */
9323                 tp->snd_una++;
9324         }
9325         /*
9326          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
9327          * FIN-WAIT-1
9328          */
9329         tp->t_starttime = ticks;
9330         if (tp->t_flags & TF_NEEDFIN) {
9331                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
9332                 tp->t_flags &= ~TF_NEEDFIN;
9333         } else {
9334                 tcp_state_change(tp, TCPS_ESTABLISHED);
9335                 TCP_PROBE5(accept__established, NULL, tp,
9336                            mtod(m, const char *), tp, th);
9337                 /*
9338                  * TFO connections call cc_conn_init() during SYN
9339                  * processing.  Calling it again here for such connections
9340                  * is not harmless as it would undo the snd_cwnd reduction
9341                  * that occurs when a TFO SYN|ACK is retransmitted.
9342                  */
9343                 if (!IS_FASTOPEN(tp->t_flags))
9344                         cc_conn_init(tp);
9345         }
9346         /*
9347          * If segment contains data or ACK, will call tcp_reass() later; if
9348          * not, do so now to pass queued data to user.
9349          */
9350         if (tlen == 0 && (thflags & TH_FIN) == 0)
9351                 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
9352                         (struct mbuf *)0);
9353         tp->snd_wl1 = th->th_seq - 1;
9354         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9355                 return (ret_val);
9356         }
9357         if (tp->t_state == TCPS_FIN_WAIT_1) {
9358                 /* We could have went to FIN_WAIT_1 (or EST) above */
9359                 /*
9360                  * In FIN_WAIT_1 STATE in addition to the processing for the
9361                  * ESTABLISHED state if our FIN is now acknowledged then
9362                  * enter FIN_WAIT_2.
9363                  */
9364                 if (ourfinisacked) {
9365                         /*
9366                          * If we can't receive any more data, then closing
9367                          * user can proceed. Starting the timer is contrary
9368                          * to the specification, but if we don't get a FIN
9369                          * we'll hang forever.
9370                          *
9371                          * XXXjl: we should release the tp also, and use a
9372                          * compressed state.
9373                          */
9374                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9375                                 soisdisconnected(so);
9376                                 tcp_timer_activate(tp, TT_2MSL,
9377                                                    (tcp_fast_finwait2_recycle ?
9378                                                     tcp_finwait2_timeout :
9379                                                     TP_MAXIDLE(tp)));
9380                         }
9381                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
9382                 }
9383         }
9384         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9385                                  tiwin, thflags, nxt_pkt));
9386 }
9387
9388 /*
9389  * Return value of 1, the TCB is unlocked and most
9390  * likely gone, return value of 0, the TCB is still
9391  * locked.
9392  */
9393 static int
9394 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
9395     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9396     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
9397 {
9398         struct tcp_bbr *bbr;
9399         int32_t ret_val;
9400
9401         /*
9402          * Header prediction: check for the two common cases of a
9403          * uni-directional data xfer.  If the packet has no control flags,
9404          * is in-sequence, the window didn't change and we're not
9405          * retransmitting, it's a candidate.  If the length is zero and the
9406          * ack moved forward, we're the sender side of the xfer.  Just free
9407          * the data acked & wake any higher level process that was blocked
9408          * waiting for space.  If the length is non-zero and the ack didn't
9409          * move, we're the receiver side.  If we're getting packets in-order
9410          * (the reassembly queue is empty), add the data toc The socket
9411          * buffer and note that we need a delayed ack. Make sure that the
9412          * hidden state-flags are also off. Since we check for
9413          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
9414          */
9415         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9416         if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) {
9417                 /*
9418                  * If we have delived under 4 segments increase the initial
9419                  * window if raised by the peer. We use this to determine
9420                  * dynamic and static rwnd's at the end of a connection.
9421                  */
9422                 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd);
9423         }
9424         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
9425             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) &&
9426             __predict_true(SEGQ_EMPTY(tp)) &&
9427             __predict_true(th->th_seq == tp->rcv_nxt)) {
9428                 if (tlen == 0) {
9429                         if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
9430                             tiwin, nxt_pkt)) {
9431                                 return (0);
9432                         }
9433                 } else {
9434                         if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
9435                             tiwin, nxt_pkt)) {
9436                                 return (0);
9437                         }
9438                 }
9439         }
9440         ctf_calc_rwin(so, tp);
9441
9442         if ((thflags & TH_RST) ||
9443             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9444                 return (ctf_process_rst(m, th, so, tp));
9445         /*
9446          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9447          * synchronized state.
9448          */
9449         if (thflags & TH_SYN) {
9450                 ctf_challenge_ack(m, th, tp, &ret_val);
9451                 return (ret_val);
9452         }
9453         /*
9454          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9455          * it's less than ts_recent, drop it.
9456          */
9457         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9458             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9459                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9460                         return (ret_val);
9461         }
9462         INP_WLOCK_ASSERT(tp->t_inpcb);
9463         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9464                 return (ret_val);
9465         }
9466         /*
9467          * If last ACK falls within this segment's sequence numbers, record
9468          * its timestamp. NOTE: 1) That the test incorporates suggestions
9469          * from the latest proposal of the tcplw@cray.com list (Braden
9470          * 1993/04/26). 2) That updating only on newer timestamps interferes
9471          * with our earlier PAWS tests, so this check should be solely
9472          * predicated on the sequence space of this segment. 3) That we
9473          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9474          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9475          * SEG.Len, This modified check allows us to overcome RFC1323's
9476          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9477          * p.869. In such cases, we can still calculate the RTT correctly
9478          * when RCV.NXT == Last.ACK.Sent.
9479          */
9480         if ((to->to_flags & TOF_TS) != 0 &&
9481             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9482             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9483             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9484                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9485                 tp->ts_recent = to->to_tsval;
9486         }
9487         /*
9488          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9489          * is on (half-synchronized state), then queue data for later
9490          * processing; else drop segment and return.
9491          */
9492         if ((thflags & TH_ACK) == 0) {
9493                 if (tp->t_flags & TF_NEEDSYN) {
9494                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9495                             tiwin, thflags, nxt_pkt));
9496                 } else if (tp->t_flags & TF_ACKNOW) {
9497                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9498                         bbr->r_wanted_output = 1;
9499                         return (ret_val);
9500                 } else {
9501                         ctf_do_drop(m, NULL);
9502                         return (0);
9503                 }
9504         }
9505         /*
9506          * Ack processing.
9507          */
9508         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9509                 return (ret_val);
9510         }
9511         if (sbavail(&so->so_snd)) {
9512                 if (bbr_progress_timeout_check(bbr)) {
9513                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9514                         return (1);
9515                 }
9516         }
9517         /* State changes only happen in bbr_process_data() */
9518         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9519             tiwin, thflags, nxt_pkt));
9520 }
9521
9522 /*
9523  * Return value of 1, the TCB is unlocked and most
9524  * likely gone, return value of 0, the TCB is still
9525  * locked.
9526  */
9527 static int
9528 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
9529     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9530     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
9531 {
9532         struct tcp_bbr *bbr;
9533         int32_t ret_val;
9534
9535         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9536         ctf_calc_rwin(so, tp);
9537         if ((thflags & TH_RST) ||
9538             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9539                 return (ctf_process_rst(m, th, so, tp));
9540         /*
9541          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9542          * synchronized state.
9543          */
9544         if (thflags & TH_SYN) {
9545                 ctf_challenge_ack(m, th, tp, &ret_val);
9546                 return (ret_val);
9547         }
9548         /*
9549          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9550          * it's less than ts_recent, drop it.
9551          */
9552         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9553             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9554                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9555                         return (ret_val);
9556         }
9557         INP_WLOCK_ASSERT(tp->t_inpcb);
9558         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9559                 return (ret_val);
9560         }
9561         /*
9562          * If last ACK falls within this segment's sequence numbers, record
9563          * its timestamp. NOTE: 1) That the test incorporates suggestions
9564          * from the latest proposal of the tcplw@cray.com list (Braden
9565          * 1993/04/26). 2) That updating only on newer timestamps interferes
9566          * with our earlier PAWS tests, so this check should be solely
9567          * predicated on the sequence space of this segment. 3) That we
9568          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9569          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9570          * SEG.Len, This modified check allows us to overcome RFC1323's
9571          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9572          * p.869. In such cases, we can still calculate the RTT correctly
9573          * when RCV.NXT == Last.ACK.Sent.
9574          */
9575         if ((to->to_flags & TOF_TS) != 0 &&
9576             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9577             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9578             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9579                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9580                 tp->ts_recent = to->to_tsval;
9581         }
9582         /*
9583          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9584          * is on (half-synchronized state), then queue data for later
9585          * processing; else drop segment and return.
9586          */
9587         if ((thflags & TH_ACK) == 0) {
9588                 if (tp->t_flags & TF_NEEDSYN) {
9589                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9590                             tiwin, thflags, nxt_pkt));
9591                 } else if (tp->t_flags & TF_ACKNOW) {
9592                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9593                         bbr->r_wanted_output = 1;
9594                         return (ret_val);
9595                 } else {
9596                         ctf_do_drop(m, NULL);
9597                         return (0);
9598                 }
9599         }
9600         /*
9601          * Ack processing.
9602          */
9603         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9604                 return (ret_val);
9605         }
9606         if (sbavail(&so->so_snd)) {
9607                 if (bbr_progress_timeout_check(bbr)) {
9608                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9609                         return (1);
9610                 }
9611         }
9612         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9613             tiwin, thflags, nxt_pkt));
9614 }
9615
9616 static int
9617 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr,
9618     struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so)
9619 {
9620
9621         if (bbr->rc_allow_data_af_clo == 0) {
9622 close_now:
9623                 tp = tcp_close(tp);
9624                 TCPSTAT_INC(tcps_rcvafterclose);
9625                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
9626                 return (1);
9627         }
9628         if (sbavail(&so->so_snd) == 0)
9629                 goto close_now;
9630         /* Ok we allow data that is ignored and a followup reset */
9631         tp->rcv_nxt = th->th_seq + *tlen;
9632         tp->t_flags2 |= TF2_DROP_AF_DATA;
9633         bbr->r_wanted_output = 1;
9634         *tlen = 0;
9635         return (0);
9636 }
9637
9638 /*
9639  * Return value of 1, the TCB is unlocked and most
9640  * likely gone, return value of 0, the TCB is still
9641  * locked.
9642  */
9643 static int
9644 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
9645     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9646     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
9647 {
9648         int32_t ourfinisacked = 0;
9649         int32_t ret_val;
9650         struct tcp_bbr *bbr;
9651
9652         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9653         ctf_calc_rwin(so, tp);
9654         if ((thflags & TH_RST) ||
9655             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9656                 return (ctf_process_rst(m, th, so, tp));
9657         /*
9658          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9659          * synchronized state.
9660          */
9661         if (thflags & TH_SYN) {
9662                 ctf_challenge_ack(m, th, tp, &ret_val);
9663                 return (ret_val);
9664         }
9665         /*
9666          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9667          * it's less than ts_recent, drop it.
9668          */
9669         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9670             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9671                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9672                         return (ret_val);
9673         }
9674         INP_WLOCK_ASSERT(tp->t_inpcb);
9675         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9676                 return (ret_val);
9677         }
9678         /*
9679          * If new data are received on a connection after the user processes
9680          * are gone, then RST the other end.
9681          */
9682         if ((so->so_state & SS_NOFDREF) && tlen) {
9683                 /*
9684                  * We call a new function now so we might continue and setup
9685                  * to reset at all data being ack'd.
9686                  */
9687                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9688                         return (1);
9689         }
9690         /*
9691          * If last ACK falls within this segment's sequence numbers, record
9692          * its timestamp. NOTE: 1) That the test incorporates suggestions
9693          * from the latest proposal of the tcplw@cray.com list (Braden
9694          * 1993/04/26). 2) That updating only on newer timestamps interferes
9695          * with our earlier PAWS tests, so this check should be solely
9696          * predicated on the sequence space of this segment. 3) That we
9697          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9698          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9699          * SEG.Len, This modified check allows us to overcome RFC1323's
9700          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9701          * p.869. In such cases, we can still calculate the RTT correctly
9702          * when RCV.NXT == Last.ACK.Sent.
9703          */
9704         if ((to->to_flags & TOF_TS) != 0 &&
9705             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9706             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9707             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9708                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9709                 tp->ts_recent = to->to_tsval;
9710         }
9711         /*
9712          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9713          * is on (half-synchronized state), then queue data for later
9714          * processing; else drop segment and return.
9715          */
9716         if ((thflags & TH_ACK) == 0) {
9717                 if (tp->t_flags & TF_NEEDSYN) {
9718                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9719                             tiwin, thflags, nxt_pkt));
9720                 } else if (tp->t_flags & TF_ACKNOW) {
9721                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9722                         bbr->r_wanted_output = 1;
9723                         return (ret_val);
9724                 } else {
9725                         ctf_do_drop(m, NULL);
9726                         return (0);
9727                 }
9728         }
9729         /*
9730          * Ack processing.
9731          */
9732         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9733                 return (ret_val);
9734         }
9735         if (ourfinisacked) {
9736                 /*
9737                  * If we can't receive any more data, then closing user can
9738                  * proceed. Starting the timer is contrary to the
9739                  * specification, but if we don't get a FIN we'll hang
9740                  * forever.
9741                  *
9742                  * XXXjl: we should release the tp also, and use a
9743                  * compressed state.
9744                  */
9745                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9746                         soisdisconnected(so);
9747                         tcp_timer_activate(tp, TT_2MSL,
9748                             (tcp_fast_finwait2_recycle ?
9749                             tcp_finwait2_timeout :
9750                             TP_MAXIDLE(tp)));
9751                 }
9752                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
9753         }
9754         if (sbavail(&so->so_snd)) {
9755                 if (bbr_progress_timeout_check(bbr)) {
9756                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9757                         return (1);
9758                 }
9759         }
9760         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9761             tiwin, thflags, nxt_pkt));
9762 }
9763
9764 /*
9765  * Return value of 1, the TCB is unlocked and most
9766  * likely gone, return value of 0, the TCB is still
9767  * locked.
9768  */
9769 static int
9770 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
9771     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9772     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
9773 {
9774         int32_t ourfinisacked = 0;
9775         int32_t ret_val;
9776         struct tcp_bbr *bbr;
9777
9778         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9779         ctf_calc_rwin(so, tp);
9780         if ((thflags & TH_RST) ||
9781             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9782                 return (ctf_process_rst(m, th, so, tp));
9783         /*
9784          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9785          * synchronized state.
9786          */
9787         if (thflags & TH_SYN) {
9788                 ctf_challenge_ack(m, th, tp, &ret_val);
9789                 return (ret_val);
9790         }
9791         /*
9792          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9793          * it's less than ts_recent, drop it.
9794          */
9795         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9796             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9797                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9798                         return (ret_val);
9799         }
9800         INP_WLOCK_ASSERT(tp->t_inpcb);
9801         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9802                 return (ret_val);
9803         }
9804         /*
9805          * If new data are received on a connection after the user processes
9806          * are gone, then RST the other end.
9807          */
9808         if ((so->so_state & SS_NOFDREF) && tlen) {
9809                 /*
9810                  * We call a new function now so we might continue and setup
9811                  * to reset at all data being ack'd.
9812                  */
9813                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9814                         return (1);
9815         }
9816         /*
9817          * If last ACK falls within this segment's sequence numbers, record
9818          * its timestamp. NOTE: 1) That the test incorporates suggestions
9819          * from the latest proposal of the tcplw@cray.com list (Braden
9820          * 1993/04/26). 2) That updating only on newer timestamps interferes
9821          * with our earlier PAWS tests, so this check should be solely
9822          * predicated on the sequence space of this segment. 3) That we
9823          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9824          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9825          * SEG.Len, This modified check allows us to overcome RFC1323's
9826          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9827          * p.869. In such cases, we can still calculate the RTT correctly
9828          * when RCV.NXT == Last.ACK.Sent.
9829          */
9830         if ((to->to_flags & TOF_TS) != 0 &&
9831             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9832             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9833             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9834                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9835                 tp->ts_recent = to->to_tsval;
9836         }
9837         /*
9838          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9839          * is on (half-synchronized state), then queue data for later
9840          * processing; else drop segment and return.
9841          */
9842         if ((thflags & TH_ACK) == 0) {
9843                 if (tp->t_flags & TF_NEEDSYN) {
9844                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9845                             tiwin, thflags, nxt_pkt));
9846                 } else if (tp->t_flags & TF_ACKNOW) {
9847                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9848                         bbr->r_wanted_output = 1;
9849                         return (ret_val);
9850                 } else {
9851                         ctf_do_drop(m, NULL);
9852                         return (0);
9853                 }
9854         }
9855         /*
9856          * Ack processing.
9857          */
9858         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9859                 return (ret_val);
9860         }
9861         if (ourfinisacked) {
9862                 tcp_twstart(tp);
9863                 m_freem(m);
9864                 return (1);
9865         }
9866         if (sbavail(&so->so_snd)) {
9867                 if (bbr_progress_timeout_check(bbr)) {
9868                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9869                         return (1);
9870                 }
9871         }
9872         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9873             tiwin, thflags, nxt_pkt));
9874 }
9875
9876 /*
9877  * Return value of 1, the TCB is unlocked and most
9878  * likely gone, return value of 0, the TCB is still
9879  * locked.
9880  */
9881 static int
9882 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9883     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9884     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
9885 {
9886         int32_t ourfinisacked = 0;
9887         int32_t ret_val;
9888         struct tcp_bbr *bbr;
9889
9890         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9891         ctf_calc_rwin(so, tp);
9892         if ((thflags & TH_RST) ||
9893             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9894                 return (ctf_process_rst(m, th, so, tp));
9895         /*
9896          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9897          * synchronized state.
9898          */
9899         if (thflags & TH_SYN) {
9900                 ctf_challenge_ack(m, th, tp, &ret_val);
9901                 return (ret_val);
9902         }
9903         /*
9904          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9905          * it's less than ts_recent, drop it.
9906          */
9907         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9908             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9909                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9910                         return (ret_val);
9911         }
9912         INP_WLOCK_ASSERT(tp->t_inpcb);
9913         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9914                 return (ret_val);
9915         }
9916         /*
9917          * If new data are received on a connection after the user processes
9918          * are gone, then RST the other end.
9919          */
9920         if ((so->so_state & SS_NOFDREF) && tlen) {
9921                 /*
9922                  * We call a new function now so we might continue and setup
9923                  * to reset at all data being ack'd.
9924                  */
9925                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9926                         return (1);
9927         }
9928         /*
9929          * If last ACK falls within this segment's sequence numbers, record
9930          * its timestamp. NOTE: 1) That the test incorporates suggestions
9931          * from the latest proposal of the tcplw@cray.com list (Braden
9932          * 1993/04/26). 2) That updating only on newer timestamps interferes
9933          * with our earlier PAWS tests, so this check should be solely
9934          * predicated on the sequence space of this segment. 3) That we
9935          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9936          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9937          * SEG.Len, This modified check allows us to overcome RFC1323's
9938          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9939          * p.869. In such cases, we can still calculate the RTT correctly
9940          * when RCV.NXT == Last.ACK.Sent.
9941          */
9942         if ((to->to_flags & TOF_TS) != 0 &&
9943             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9944             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9945             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9946                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9947                 tp->ts_recent = to->to_tsval;
9948         }
9949         /*
9950          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9951          * is on (half-synchronized state), then queue data for later
9952          * processing; else drop segment and return.
9953          */
9954         if ((thflags & TH_ACK) == 0) {
9955                 if (tp->t_flags & TF_NEEDSYN) {
9956                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9957                             tiwin, thflags, nxt_pkt));
9958                 } else if (tp->t_flags & TF_ACKNOW) {
9959                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9960                         bbr->r_wanted_output = 1;
9961                         return (ret_val);
9962                 } else {
9963                         ctf_do_drop(m, NULL);
9964                         return (0);
9965                 }
9966         }
9967         /*
9968          * case TCPS_LAST_ACK: Ack processing.
9969          */
9970         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9971                 return (ret_val);
9972         }
9973         if (ourfinisacked) {
9974                 tp = tcp_close(tp);
9975                 ctf_do_drop(m, tp);
9976                 return (1);
9977         }
9978         if (sbavail(&so->so_snd)) {
9979                 if (bbr_progress_timeout_check(bbr)) {
9980                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9981                         return (1);
9982                 }
9983         }
9984         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9985             tiwin, thflags, nxt_pkt));
9986 }
9987
9988
9989 /*
9990  * Return value of 1, the TCB is unlocked and most
9991  * likely gone, return value of 0, the TCB is still
9992  * locked.
9993  */
9994 static int
9995 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
9996     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9997     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
9998 {
9999         int32_t ourfinisacked = 0;
10000         int32_t ret_val;
10001         struct tcp_bbr *bbr;
10002
10003         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10004         ctf_calc_rwin(so, tp);
10005         /* Reset receive buffer auto scaling when not in bulk receive mode. */
10006         if ((thflags & TH_RST) ||
10007             (tp->t_fin_is_rst && (thflags & TH_FIN)))
10008                 return (ctf_process_rst(m, th, so, tp));
10009
10010         /*
10011          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
10012          * synchronized state.
10013          */
10014         if (thflags & TH_SYN) {
10015                 ctf_challenge_ack(m, th, tp, &ret_val);
10016                 return (ret_val);
10017         }
10018         INP_WLOCK_ASSERT(tp->t_inpcb);
10019         /*
10020          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10021          * it's less than ts_recent, drop it.
10022          */
10023         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10024             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10025                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10026                         return (ret_val);
10027         }
10028         INP_WLOCK_ASSERT(tp->t_inpcb);
10029         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
10030                 return (ret_val);
10031         }
10032         /*
10033          * If new data are received on a connection after the user processes
10034          * are gone, then we may RST the other end depending on the outcome
10035          * of bbr_check_data_after_close.
10036          */
10037         if ((so->so_state & SS_NOFDREF) &&
10038             tlen) {
10039                 /*
10040                  * We call a new function now so we might continue and setup
10041                  * to reset at all data being ack'd.
10042                  */
10043                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
10044                         return (1);
10045         }
10046         INP_WLOCK_ASSERT(tp->t_inpcb);
10047         /*
10048          * If last ACK falls within this segment's sequence numbers, record
10049          * its timestamp. NOTE: 1) That the test incorporates suggestions
10050          * from the latest proposal of the tcplw@cray.com list (Braden
10051          * 1993/04/26). 2) That updating only on newer timestamps interferes
10052          * with our earlier PAWS tests, so this check should be solely
10053          * predicated on the sequence space of this segment. 3) That we
10054          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
10055          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
10056          * SEG.Len, This modified check allows us to overcome RFC1323's
10057          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
10058          * p.869. In such cases, we can still calculate the RTT correctly
10059          * when RCV.NXT == Last.ACK.Sent.
10060          */
10061         INP_WLOCK_ASSERT(tp->t_inpcb);
10062         if ((to->to_flags & TOF_TS) != 0 &&
10063             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
10064             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
10065             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
10066                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
10067                 tp->ts_recent = to->to_tsval;
10068         }
10069         /*
10070          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
10071          * is on (half-synchronized state), then queue data for later
10072          * processing; else drop segment and return.
10073          */
10074         if ((thflags & TH_ACK) == 0) {
10075                 if (tp->t_flags & TF_NEEDSYN) {
10076                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
10077                             tiwin, thflags, nxt_pkt));
10078                 } else if (tp->t_flags & TF_ACKNOW) {
10079                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
10080                         bbr->r_wanted_output = 1;
10081                         return (ret_val);
10082                 } else {
10083                         ctf_do_drop(m, NULL);
10084                         return (0);
10085                 }
10086         }
10087         /*
10088          * Ack processing.
10089          */
10090         INP_WLOCK_ASSERT(tp->t_inpcb);
10091         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
10092                 return (ret_val);
10093         }
10094         if (sbavail(&so->so_snd)) {
10095                 if (bbr_progress_timeout_check(bbr)) {
10096                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10097                         return (1);
10098                 }
10099         }
10100         INP_WLOCK_ASSERT(tp->t_inpcb);
10101         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
10102             tiwin, thflags, nxt_pkt));
10103 }
10104
10105 static void
10106 bbr_stop_all_timers(struct tcpcb *tp)
10107 {
10108         struct tcp_bbr *bbr;
10109
10110         /*
10111          * Assure no timers are running.
10112          */
10113         if (tcp_timer_active(tp, TT_PERSIST)) {
10114                 /* We enter in persists, set the flag appropriately */
10115                 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10116                 bbr->rc_in_persist = 1;
10117         }
10118         tcp_timer_suspend(tp, TT_PERSIST);
10119         tcp_timer_suspend(tp, TT_REXMT);
10120         tcp_timer_suspend(tp, TT_KEEP);
10121         tcp_timer_suspend(tp, TT_DELACK);
10122 }
10123
10124 static void
10125 bbr_google_mode_on(struct tcp_bbr *bbr)
10126 {
10127         bbr->rc_use_google = 1;
10128         bbr->rc_no_pacing = 0;
10129         bbr->r_ctl.bbr_google_discount = bbr_google_discount;
10130         bbr->r_use_policer = bbr_policer_detection_enabled;
10131         bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10132         bbr->bbr_use_rack_cheat = 0;
10133         bbr->r_ctl.rc_incr_tmrs = 0;
10134         bbr->r_ctl.rc_inc_tcp_oh = 0;
10135         bbr->r_ctl.rc_inc_ip_oh = 0;
10136         bbr->r_ctl.rc_inc_enet_oh = 0;
10137         reset_time(&bbr->r_ctl.rc_delrate,
10138                    BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10139         reset_time_small(&bbr->r_ctl.rc_rttprop,
10140                          (11 * USECS_IN_SECOND));
10141         tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
10142 }
10143
10144 static void
10145 bbr_google_mode_off(struct tcp_bbr *bbr)
10146 {
10147         bbr->rc_use_google = 0;
10148         bbr->r_ctl.bbr_google_discount = 0;
10149         bbr->no_pacing_until = bbr_no_pacing_until;
10150         bbr->r_use_policer = 0;
10151         if (bbr->no_pacing_until)
10152                 bbr->rc_no_pacing = 1;
10153         else
10154                 bbr->rc_no_pacing = 0;
10155         if (bbr_use_rack_resend_cheat)
10156                 bbr->bbr_use_rack_cheat = 1;
10157         else
10158                 bbr->bbr_use_rack_cheat = 0;
10159         if (bbr_incr_timers)
10160                 bbr->r_ctl.rc_incr_tmrs = 1;
10161         else
10162                 bbr->r_ctl.rc_incr_tmrs = 0;
10163         if (bbr_include_tcp_oh)
10164                 bbr->r_ctl.rc_inc_tcp_oh = 1;
10165         else
10166                 bbr->r_ctl.rc_inc_tcp_oh = 0;
10167         if (bbr_include_ip_oh)
10168                 bbr->r_ctl.rc_inc_ip_oh = 1;
10169         else
10170                 bbr->r_ctl.rc_inc_ip_oh = 0;
10171         if (bbr_include_enet_oh)
10172                 bbr->r_ctl.rc_inc_enet_oh = 1;
10173         else
10174                 bbr->r_ctl.rc_inc_enet_oh = 0;
10175         bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10176         reset_time(&bbr->r_ctl.rc_delrate,
10177                    bbr_num_pktepo_for_del_limit);
10178         reset_time_small(&bbr->r_ctl.rc_rttprop,
10179                          (bbr_filter_len_sec * USECS_IN_SECOND));
10180         tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
10181 }
10182 /*
10183  * Return 0 on success, non-zero on failure
10184  * which indicates the error (usually no memory).
10185  */
10186 static int
10187 bbr_init(struct tcpcb *tp)
10188 {
10189         struct tcp_bbr *bbr = NULL;
10190         struct inpcb *inp;
10191         uint32_t cts;
10192
10193         tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO));
10194         if (tp->t_fb_ptr == NULL) {
10195                 /*
10196                  * We need to allocate memory but cant. The INP and INP_INFO
10197                  * locks and they are recusive (happens during setup. So a
10198                  * scheme to drop the locks fails :(
10199                  *
10200                  */
10201                 return (ENOMEM);
10202         }
10203         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10204         bbr->rtt_valid = 0;
10205         inp = tp->t_inpcb;
10206         inp->inp_flags2 |= INP_CANNOT_DO_ECN;
10207         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
10208         TAILQ_INIT(&bbr->r_ctl.rc_map);
10209         TAILQ_INIT(&bbr->r_ctl.rc_free);
10210         TAILQ_INIT(&bbr->r_ctl.rc_tmap);
10211         bbr->rc_tp = tp;
10212         if (tp->t_inpcb) {
10213                 bbr->rc_inp = tp->t_inpcb;
10214         }
10215         cts = tcp_get_usecs(&bbr->rc_tv);
10216         tp->t_acktime = 0;
10217         bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close;
10218         bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade;
10219         bbr->rc_tlp_threshold = bbr_tlp_thresh;
10220         bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh;
10221         bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay;
10222         bbr->r_ctl.rc_min_to = bbr_min_to;
10223         bbr->rc_bbr_state = BBR_STATE_STARTUP;
10224         bbr->r_ctl.bbr_lost_at_state = 0;
10225         bbr->r_ctl.rc_lost_at_startup = 0;
10226         bbr->rc_all_timers_stopped = 0;
10227         bbr->r_ctl.rc_bbr_lastbtlbw = 0;
10228         bbr->r_ctl.rc_pkt_epoch_del = 0;
10229         bbr->r_ctl.rc_pkt_epoch = 0;
10230         bbr->r_ctl.rc_lowest_rtt = 0xffffffff;
10231         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain;
10232         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
10233         bbr->r_ctl.rc_went_idle_time = cts;
10234         bbr->rc_pacer_started = cts;
10235         bbr->r_ctl.rc_pkt_epoch_time = cts;
10236         bbr->r_ctl.rc_rcvtime = cts;
10237         bbr->r_ctl.rc_bbr_state_time = cts;
10238         bbr->r_ctl.rc_del_time = cts;
10239         bbr->r_ctl.rc_tlp_rxt_last_time = cts;
10240         bbr->r_ctl.last_in_probertt = cts;
10241         bbr->skip_gain = 0;
10242         bbr->gain_is_limited = 0;
10243         bbr->no_pacing_until = bbr_no_pacing_until;
10244         if (bbr->no_pacing_until)
10245                 bbr->rc_no_pacing = 1;
10246         if (bbr_use_google_algo) {
10247                 bbr->rc_no_pacing = 0;
10248                 bbr->rc_use_google = 1;
10249                 bbr->r_ctl.bbr_google_discount = bbr_google_discount;
10250                 bbr->r_use_policer = bbr_policer_detection_enabled;
10251         } else {
10252                 bbr->rc_use_google = 0;
10253                 bbr->r_ctl.bbr_google_discount = 0;
10254                 bbr->r_use_policer = 0;
10255         }
10256         if (bbr_ts_limiting)
10257                 bbr->rc_use_ts_limit = 1;
10258         else
10259                 bbr->rc_use_ts_limit = 0;
10260         if (bbr_ts_can_raise)
10261                 bbr->ts_can_raise = 1;
10262         else
10263                 bbr->ts_can_raise = 0;
10264         if (V_tcp_delack_enabled == 1)
10265                 tp->t_delayed_ack = 2;
10266         else if (V_tcp_delack_enabled == 0)
10267                 tp->t_delayed_ack = 0;
10268         else if (V_tcp_delack_enabled < 100)
10269                 tp->t_delayed_ack = V_tcp_delack_enabled;
10270         else
10271                 tp->t_delayed_ack = 2;
10272         if (bbr->rc_use_google == 0)
10273                 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10274         else
10275                 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10276         bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms;
10277         bbr->rc_max_rto_sec = bbr_rto_max_sec;
10278         bbr->rc_init_win = bbr_def_init_win;
10279         if (tp->t_flags & TF_REQ_TSTMP)
10280                 bbr->rc_last_options = TCP_TS_OVERHEAD;
10281         bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options;
10282         bbr->r_ctl.rc_high_rwnd = tp->snd_wnd;
10283         bbr->r_init_rtt = 1;
10284
10285         counter_u64_add(bbr_flows_nohdwr_pacing, 1);
10286         if (bbr_allow_hdwr_pacing)
10287                 bbr->bbr_hdw_pace_ena = 1;
10288         else
10289                 bbr->bbr_hdw_pace_ena = 0;
10290         if (bbr_sends_full_iwnd)
10291                 bbr->bbr_init_win_cheat = 1;
10292         else
10293                 bbr->bbr_init_win_cheat = 0;
10294         bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max;
10295         bbr->r_ctl.rc_drain_pg = bbr_drain_gain;
10296         bbr->r_ctl.rc_startup_pg = bbr_high_gain;
10297         bbr->rc_loss_exit = bbr_exit_startup_at_loss;
10298         bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain;
10299         bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second;
10300         bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar;
10301         bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max;
10302         bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor;
10303         bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min;
10304         bbr->r_ctl.bbr_cross_over = bbr_cross_over;
10305         bbr->r_ctl.rc_rtt_shrinks = cts;
10306         if (bbr->rc_use_google) {
10307                 setup_time_filter(&bbr->r_ctl.rc_delrate,
10308                                   FILTER_TYPE_MAX,
10309                                   BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10310                 setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10311                                         FILTER_TYPE_MIN, (11 * USECS_IN_SECOND));
10312         } else {
10313                 setup_time_filter(&bbr->r_ctl.rc_delrate,
10314                                   FILTER_TYPE_MAX,
10315                                   bbr_num_pktepo_for_del_limit);
10316                 setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10317                                         FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND));
10318         }
10319         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0);
10320         if (bbr_uses_idle_restart)
10321                 bbr->rc_use_idle_restart = 1;
10322         else
10323                 bbr->rc_use_idle_restart = 0;
10324         bbr->r_ctl.rc_bbr_cur_del_rate = 0;
10325         bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps;
10326         if (bbr_resends_use_tso)
10327                 bbr->rc_resends_use_tso = 1;
10328 #ifdef NETFLIX_PEAKRATE
10329         tp->t_peakrate_thr = tp->t_maxpeakrate;
10330 #endif
10331         if (tp->snd_una != tp->snd_max) {
10332                 /* Create a send map for the current outstanding data */
10333                 struct bbr_sendmap *rsm;
10334
10335                 rsm = bbr_alloc(bbr);
10336                 if (rsm == NULL) {
10337                         uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10338                         tp->t_fb_ptr = NULL;
10339                         return (ENOMEM);
10340                 }
10341                 rsm->r_flags = BBR_OVERMAX;
10342                 rsm->r_tim_lastsent[0] = cts;
10343                 rsm->r_rtr_cnt = 1;
10344                 rsm->r_rtr_bytes = 0;
10345                 rsm->r_start = tp->snd_una;
10346                 rsm->r_end = tp->snd_max;
10347                 rsm->r_dupack = 0;
10348                 rsm->r_delivered = bbr->r_ctl.rc_delivered;
10349                 rsm->r_ts_valid = 0;
10350                 rsm->r_del_ack_ts = tp->ts_recent;
10351                 rsm->r_del_time = cts;
10352                 if (bbr->r_ctl.r_app_limited_until)
10353                         rsm->r_app_limited = 1;
10354                 else
10355                         rsm->r_app_limited = 0;
10356                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
10357                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
10358                 rsm->r_in_tmap = 1;
10359                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
10360                         rsm->r_bbr_state = bbr_state_val(bbr);
10361                 else
10362                         rsm->r_bbr_state = 8;
10363         }
10364         if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0))
10365                 bbr->bbr_use_rack_cheat = 1;
10366         if (bbr_incr_timers && (bbr->rc_use_google == 0))
10367                 bbr->r_ctl.rc_incr_tmrs = 1;
10368         if (bbr_include_tcp_oh && (bbr->rc_use_google == 0))
10369                 bbr->r_ctl.rc_inc_tcp_oh = 1;
10370         if (bbr_include_ip_oh && (bbr->rc_use_google == 0))
10371                 bbr->r_ctl.rc_inc_ip_oh = 1;
10372         if (bbr_include_enet_oh && (bbr->rc_use_google == 0))
10373                 bbr->r_ctl.rc_inc_enet_oh = 1;
10374
10375         bbr_log_type_statechange(bbr, cts, __LINE__);
10376         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
10377             (tp->t_srtt)) {
10378                 uint32_t rtt;
10379
10380                 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
10381                 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
10382         }
10383         /* announce the settings and state */
10384         bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT);
10385         tcp_bbr_tso_size_check(bbr, cts);
10386         /*
10387          * Now call the generic function to start a timer. This will place
10388          * the TCB on the hptsi wheel if a timer is needed with appropriate
10389          * flags.
10390          */
10391         bbr_stop_all_timers(tp);
10392         bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0);
10393         return (0);
10394 }
10395
10396 /*
10397  * Return 0 if we can accept the connection. Return
10398  * non-zero if we can't handle the connection. A EAGAIN
10399  * means you need to wait until the connection is up.
10400  * a EADDRNOTAVAIL means we can never handle the connection
10401  * (no SACK).
10402  */
10403 static int
10404 bbr_handoff_ok(struct tcpcb *tp)
10405 {
10406         if ((tp->t_state == TCPS_CLOSED) ||
10407             (tp->t_state == TCPS_LISTEN)) {
10408                 /* Sure no problem though it may not stick */
10409                 return (0);
10410         }
10411         if ((tp->t_state == TCPS_SYN_SENT) ||
10412             (tp->t_state == TCPS_SYN_RECEIVED)) {
10413                 /*
10414                  * We really don't know you have to get to ESTAB or beyond
10415                  * to tell.
10416                  */
10417                 return (EAGAIN);
10418         }
10419         if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) {
10420                 return (0);
10421         }
10422         /*
10423          * If we reach here we don't do SACK on this connection so we can
10424          * never do rack.
10425          */
10426         return (EINVAL);
10427 }
10428
10429 static void
10430 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged)
10431 {
10432         if (tp->t_fb_ptr) {
10433                 uint32_t calc;
10434                 struct tcp_bbr *bbr;
10435                 struct bbr_sendmap *rsm;
10436
10437                 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10438                 if (bbr->r_ctl.crte)
10439                         tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
10440                 bbr_log_flowend(bbr);
10441                 bbr->rc_tp = NULL;
10442                 if (tp->t_inpcb) {
10443                         /* Backout any flags2 we applied */
10444                         tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN;
10445                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
10446                         tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
10447                 }
10448                 if (bbr->bbr_hdrw_pacing)
10449                         counter_u64_add(bbr_flows_whdwr_pacing, -1);
10450                 else
10451                         counter_u64_add(bbr_flows_nohdwr_pacing, -1);
10452                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10453                 while (rsm) {
10454                         TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
10455                         uma_zfree(bbr_zone, rsm);
10456                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10457                 }
10458                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10459                 while (rsm) {
10460                         TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
10461                         uma_zfree(bbr_zone, rsm);
10462                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10463                 }
10464                 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd;
10465                 if (calc > (bbr->r_ctl.rc_init_rwnd / 10))
10466                         BBR_STAT_INC(bbr_dynamic_rwnd);
10467                 else
10468                         BBR_STAT_INC(bbr_static_rwnd);
10469                 bbr->r_ctl.rc_free_cnt = 0;
10470                 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10471                 tp->t_fb_ptr = NULL;
10472         }
10473         /* Make sure snd_nxt is correctly set */
10474         tp->snd_nxt = tp->snd_max;
10475 }
10476
10477 static void
10478 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win)
10479 {
10480         switch (tp->t_state) {
10481         case TCPS_SYN_SENT:
10482                 bbr->r_state = TCPS_SYN_SENT;
10483                 bbr->r_substate = bbr_do_syn_sent;
10484                 break;
10485         case TCPS_SYN_RECEIVED:
10486                 bbr->r_state = TCPS_SYN_RECEIVED;
10487                 bbr->r_substate = bbr_do_syn_recv;
10488                 break;
10489         case TCPS_ESTABLISHED:
10490                 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd);
10491                 bbr->r_state = TCPS_ESTABLISHED;
10492                 bbr->r_substate = bbr_do_established;
10493                 break;
10494         case TCPS_CLOSE_WAIT:
10495                 bbr->r_state = TCPS_CLOSE_WAIT;
10496                 bbr->r_substate = bbr_do_close_wait;
10497                 break;
10498         case TCPS_FIN_WAIT_1:
10499                 bbr->r_state = TCPS_FIN_WAIT_1;
10500                 bbr->r_substate = bbr_do_fin_wait_1;
10501                 break;
10502         case TCPS_CLOSING:
10503                 bbr->r_state = TCPS_CLOSING;
10504                 bbr->r_substate = bbr_do_closing;
10505                 break;
10506         case TCPS_LAST_ACK:
10507                 bbr->r_state = TCPS_LAST_ACK;
10508                 bbr->r_substate = bbr_do_lastack;
10509                 break;
10510         case TCPS_FIN_WAIT_2:
10511                 bbr->r_state = TCPS_FIN_WAIT_2;
10512                 bbr->r_substate = bbr_do_fin_wait_2;
10513                 break;
10514         case TCPS_LISTEN:
10515         case TCPS_CLOSED:
10516         case TCPS_TIME_WAIT:
10517         default:
10518                 break;
10519         };
10520 }
10521
10522 static void
10523 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog)
10524 {
10525         /*
10526          * Now what state are we going into now? Is there adjustments
10527          * needed?
10528          */
10529         int32_t old_state, old_gain;
10530
10531
10532         old_state = bbr_state_val(bbr);
10533         old_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
10534         if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) {
10535                 /* Save the lowest srtt we saw in our end of the sub-state */
10536                 bbr->rc_hit_state_1 = 0;
10537                 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff)
10538                         bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state;
10539         }
10540         bbr->rc_bbr_substate++;
10541         if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) {
10542                 /* Cycle back to first state-> gain */
10543                 bbr->rc_bbr_substate = 0;
10544         }
10545         if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10546                 /*
10547                  * We enter the gain(5/4) cycle (possibly less if
10548                  * shallow buffer detection is enabled)
10549                  */
10550                 if (bbr->skip_gain) {
10551                         /*
10552                          * Hardware pacing has set our rate to
10553                          * the max and limited our b/w just
10554                          * do level i.e. no gain.
10555                          */
10556                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1];
10557                 } else if (bbr->gain_is_limited &&
10558                            bbr->bbr_hdrw_pacing &&
10559                            bbr->r_ctl.crte) {
10560                         /*
10561                          * We can't gain above the hardware pacing
10562                          * rate which is less than our rate + the gain
10563                          * calculate the gain needed to reach the hardware
10564                          * pacing rate..
10565                          */
10566                         uint64_t bw, rate, gain_calc;
10567
10568                         bw = bbr_get_bw(bbr);
10569                         rate = bbr->r_ctl.crte->rate;
10570                         if ((rate > bw) &&
10571                             (((bw *  (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) {
10572                                 gain_calc = (rate * BBR_UNIT) / bw;
10573                                 if (gain_calc < BBR_UNIT)
10574                                         gain_calc = BBR_UNIT;
10575                                 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc;
10576                         } else {
10577                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10578                         }
10579                 } else
10580                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10581                 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) {
10582                         bbr->r_ctl.rc_bbr_state_atflight = cts;
10583                 } else
10584                         bbr->r_ctl.rc_bbr_state_atflight = 0;
10585         } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10586                 bbr->rc_hit_state_1 = 1;
10587                 bbr->r_ctl.rc_exta_time_gd = 0;
10588                 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10589                                                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10590                 if (bbr_state_drain_2_tar) {
10591                         bbr->r_ctl.rc_bbr_state_atflight = 0;
10592                 } else
10593                         bbr->r_ctl.rc_bbr_state_atflight = cts;
10594                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN];
10595         } else {
10596                 /* All other cycles hit here 2-7 */
10597                 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) {
10598                         if (bbr_sub_drain_slam_cwnd &&
10599                             (bbr->rc_use_google == 0) &&
10600                             (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10601                                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10602                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10603                         }
10604                         if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP))
10605                                 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) -
10606                                                                bbr_get_rtt(bbr, BBR_RTT_PROP));
10607                         else
10608                                 bbr->r_ctl.rc_exta_time_gd = 0;
10609                         if (bbr->r_ctl.rc_exta_time_gd) {
10610                                 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd;
10611                                 /* Now chop up the time for each state (div by 7) */
10612                                 bbr->r_ctl.rc_level_state_extra /= 7;
10613                                 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) {
10614                                         /* Add a randomization */
10615                                         bbr_randomize_extra_state_time(bbr);
10616                                 }
10617                         }
10618                 }
10619                 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10620                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)];
10621         }
10622         if (bbr->rc_use_google) {
10623                 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10624         }
10625         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10626         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10627         if (dolog)
10628                 bbr_log_type_statechange(bbr, cts, line);
10629
10630         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10631                 uint32_t time_in;
10632
10633                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10634                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
10635                         counter_u64_add(bbr_state_time[(old_state + 5)], time_in);
10636                 } else {
10637                         counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10638                 }
10639         }
10640         bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
10641         bbr_set_state_target(bbr, __LINE__);
10642         if (bbr_sub_drain_slam_cwnd &&
10643             (bbr->rc_use_google == 0) &&
10644             (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10645                 /* Slam down the cwnd */
10646                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10647                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10648                 if (bbr_sub_drain_app_limit) {
10649                         /* Go app limited if we are on a long drain */
10650                         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered +
10651                                                           ctf_flight_size(bbr->rc_tp,
10652                                                               (bbr->r_ctl.rc_sacked +
10653                                                                bbr->r_ctl.rc_lost_bytes)));
10654                 }
10655                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10656         }
10657         if (bbr->rc_lt_use_bw) {
10658                 /* In policed mode we clamp pacing_gain to BBR_UNIT */
10659                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10660         }
10661         /* Google changes TSO size every cycle */
10662         if (bbr->rc_use_google)
10663                 tcp_bbr_tso_size_check(bbr, cts);
10664         bbr->r_ctl.gain_epoch = cts;
10665         bbr->r_ctl.rc_bbr_state_time = cts;
10666         bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch;
10667 }
10668
10669 static void
10670 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10671 {
10672         if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) &&
10673             (google_allow_early_out == 1) &&
10674             (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) {
10675                 /* We have reached out target flight size possibly early */
10676                 goto change_state;
10677         }
10678         if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10679                 return;
10680         }
10681         if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) {
10682                 /*
10683                  * Must be a rttProp movement forward before
10684                  * we can change states.
10685                  */
10686                 return;
10687         }
10688         if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10689                 /*
10690                  * The needed time has passed but for
10691                  * the gain cycle extra rules apply:
10692                  * 1) If we have seen loss, we exit
10693                  * 2) If we have not reached the target
10694                  *    we stay in GAIN (gain-to-target).
10695                  */
10696                 if (google_consider_lost && losses)
10697                         goto change_state;
10698                 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) {
10699                         return;
10700                 }
10701         }
10702 change_state:
10703         /* For gain we must reach our target, all others last 1 rttProp */
10704         bbr_substate_change(bbr, cts, __LINE__, 1);
10705 }
10706
10707 static void
10708 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10709 {
10710         uint32_t flight, bbr_cur_cycle_time;
10711
10712         if (bbr->rc_use_google) {
10713                 bbr_set_probebw_google_gains(bbr, cts, losses);
10714                 return;
10715         }
10716         if (cts == 0) {
10717                 /*
10718                  * Never alow cts to be 0 we
10719                  * do this so we can judge if
10720                  * we have set a timestamp.
10721                  */
10722                 cts = 1;
10723         }
10724         if (bbr_state_is_pkt_epoch)
10725                 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
10726         else
10727                 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP);
10728
10729         if (bbr->r_ctl.rc_bbr_state_atflight == 0) {
10730                 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10731                         flight = ctf_flight_size(bbr->rc_tp,
10732                                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10733                         if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) {
10734                                 /* Keep it slam down */
10735                                 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) {
10736                                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10737                                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10738                                 }
10739                                 if (bbr_sub_drain_app_limit) {
10740                                         /* Go app limited if we are on a long drain */
10741                                         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight);
10742                                 }
10743                         }
10744                         if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) &&
10745                             (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) ||
10746                              (flight >= bbr->r_ctl.flightsize_at_drain))) {
10747                                 /*
10748                                  * Still here after the same time as
10749                                  * the gain. We need to drain harder
10750                                  * for the next srtt. Reduce by a set amount
10751                                  * the gain drop is capped at DRAIN states
10752                                  * value (88).
10753                                  */
10754                                 bbr->r_ctl.flightsize_at_drain = flight;
10755                                 if (bbr_drain_drop_mul &&
10756                                     bbr_drain_drop_div &&
10757                                     (bbr_drain_drop_mul < bbr_drain_drop_div)) {
10758                                         /* Use your specific drop value (def 4/5 = 20%) */
10759                                         bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul;
10760                                         bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div;
10761                                 } else {
10762                                         /* You get drop of 20% */
10763                                         bbr->r_ctl.rc_bbr_hptsi_gain *= 4;
10764                                         bbr->r_ctl.rc_bbr_hptsi_gain /= 5;
10765                                 }
10766                                 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) {
10767                                         /* Reduce our gain again to the bottom  */
10768                                         bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
10769                                 }
10770                                 bbr_log_exit_gain(bbr, cts, 4);
10771                                 /*
10772                                  * Extend out so we wait another
10773                                  * epoch before dropping again.
10774                                  */
10775                                 bbr->r_ctl.gain_epoch = cts;
10776                         }
10777                         if (flight <= bbr->r_ctl.rc_target_at_state) {
10778                                 if (bbr_sub_drain_slam_cwnd &&
10779                                     (bbr->rc_use_google == 0) &&
10780                                     (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10781                                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10782                                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10783                                 }
10784                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10785                                 bbr_log_exit_gain(bbr, cts, 3);
10786                         }
10787                 } else {
10788                         /* Its a gain  */
10789                         if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) {
10790                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10791                                 goto change_state;
10792                         }
10793                         if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) ||
10794                             ((ctf_outstanding(bbr->rc_tp) +  bbr->rc_tp->t_maxseg - 1) >=
10795                              bbr->rc_tp->snd_wnd)) {
10796                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10797                                 bbr_log_exit_gain(bbr, cts, 2);
10798                         }
10799                 }
10800                 /**
10801                  * We fall through and return always one of two things has
10802                  * occured.
10803                  * 1) We are still not at target
10804                  *    <or>
10805                  * 2) We reached the target and set rc_bbr_state_atflight
10806                  *    which means we no longer hit this block
10807                  *    next time we are called.
10808                  */
10809                 return;
10810         }
10811 change_state:
10812         if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time))
10813                 return;
10814         if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) {
10815                 /* Less than a full time-period has passed */
10816                 return;
10817         }
10818         if (bbr->r_ctl.rc_level_state_extra &&
10819             (bbr_state_val(bbr) > BBR_SUB_DRAIN) &&
10820             ((cts - bbr->r_ctl.rc_bbr_state_time) <
10821              (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10822                 /* Less than a full time-period + extra has passed */
10823                 return;
10824         }
10825         if (bbr_gain_gets_extra_too &&
10826             bbr->r_ctl.rc_level_state_extra &&
10827             (bbr_state_val(bbr) == BBR_SUB_GAIN) &&
10828             ((cts - bbr->r_ctl.rc_bbr_state_time) <
10829              (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10830                 /* Less than a full time-period + extra has passed */
10831                 return;
10832         }
10833         bbr_substate_change(bbr, cts, __LINE__, 1);
10834 }
10835
10836 static uint32_t
10837 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain)
10838 {
10839         uint32_t mss, tar;
10840
10841         if (bbr->rc_use_google) {
10842                 /* Google just uses the cwnd target */
10843                 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain);
10844         } else {
10845                 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options),
10846                           bbr->r_ctl.rc_pace_max_segs);
10847                 /* Get the base cwnd with gain rounded to a mss */
10848                 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr),
10849                                                       gain), mss);
10850                 /* Make sure it is within our min */
10851                 if (tar < get_min_cwnd(bbr))
10852                         return (get_min_cwnd(bbr));
10853         }
10854         return (tar);
10855 }
10856
10857 static void
10858 bbr_set_state_target(struct tcp_bbr *bbr, int line)
10859 {
10860         uint32_t tar, meth;
10861
10862         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
10863             ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
10864                 /* Special case using old probe-rtt method */
10865                 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10866                 meth = 1;
10867         } else {
10868                 /* Non-probe-rtt case and reduced probe-rtt  */
10869                 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
10870                     (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) {
10871                         /* For gain cycle we use the hptsi gain */
10872                         tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10873                         meth = 2;
10874                 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) {
10875                         /*
10876                          * If configured, or for google all other states
10877                          * get BBR_UNIT.
10878                          */
10879                         tar = bbr_get_a_state_target(bbr, BBR_UNIT);
10880                         meth = 3;
10881                 } else {
10882                         /*
10883                          * Or we set a target based on the pacing gain
10884                          * for non-google mode and default (non-configured).
10885                          * Note we don't set a target goal below drain (192).
10886                          */
10887                         if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN])  {
10888                                 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]);
10889                                 meth = 4;
10890                         } else {
10891                                 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10892                                 meth = 5;
10893                         }
10894                 }
10895         }
10896         bbr_log_set_of_state_target(bbr, tar, line, meth);
10897         bbr->r_ctl.rc_target_at_state = tar;
10898 }
10899
10900 static void
10901 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
10902 {
10903         /* Change to probe_rtt */
10904         uint32_t time_in;
10905
10906         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10907         bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10908                                              (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10909         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain
10910                                           + bbr->r_ctl.rc_delivered);
10911         /* Setup so we force feed the filter */
10912         if (bbr->rc_use_google || bbr_probertt_sets_rtt)
10913                 bbr->rc_prtt_set_ts = 1;
10914         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10915                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10916                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10917         }
10918         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0);
10919         bbr->r_ctl.rc_rtt_shrinks = cts;
10920         bbr->r_ctl.last_in_probertt = cts;
10921         bbr->r_ctl.rc_probertt_srttchktim = cts;
10922         bbr->r_ctl.rc_bbr_state_time = cts;
10923         bbr->rc_bbr_state = BBR_STATE_PROBE_RTT;
10924         /* We need to force the filter to update */
10925
10926         if ((bbr_sub_drain_slam_cwnd) &&
10927             bbr->rc_hit_state_1 &&
10928             (bbr->rc_use_google == 0) &&
10929             (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10930                 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd)
10931                         bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10932         } else
10933                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10934         /* Update the lost */
10935         bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10936         if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){
10937                 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN)  */
10938                 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10939                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10940                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10941                 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10942                 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6);
10943                 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd;
10944         } else {
10945                 /*
10946                  * We bring it down slowly by using a hptsi gain that is
10947                  * probably 75%. This will slowly float down our outstanding
10948                  * without tampering with the cwnd.
10949                  */
10950                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
10951                 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10952                 bbr_set_state_target(bbr, __LINE__);
10953                 if (bbr_prtt_slam_cwnd &&
10954                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
10955                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10956                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10957                 }
10958         }
10959         if (ctf_flight_size(bbr->rc_tp,
10960                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
10961             bbr->r_ctl.rc_target_at_state) {
10962                 /* We are at target */
10963                 bbr->r_ctl.rc_bbr_enters_probertt = cts;
10964         } else {
10965                 /* We need to come down to reach target before our time begins */
10966                 bbr->r_ctl.rc_bbr_enters_probertt = 0;
10967         }
10968         bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch;
10969         BBR_STAT_INC(bbr_enter_probertt);
10970         bbr_log_exit_gain(bbr, cts, 0);
10971         bbr_log_type_statechange(bbr, cts, line);
10972 }
10973
10974 static void
10975 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts)
10976 {
10977         /*
10978          * Sanity check on probe-rtt intervals.
10979          * In crazy situations where we are competing
10980          * against new-reno flows with huge buffers
10981          * our rtt-prop interval could come to dominate
10982          * things if we can't get through a full set
10983          * of cycles, we need to adjust it.
10984          */
10985         if (bbr_can_adjust_probertt &&
10986             (bbr->rc_use_google == 0)) {
10987                 uint16_t val = 0;
10988                 uint32_t cur_rttp, fval, newval, baseval;
10989
10990                 /* Are we to small and go into probe-rtt to often? */
10991                 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1));
10992                 cur_rttp = roundup(baseval, USECS_IN_SECOND);
10993                 fval = bbr_filter_len_sec * USECS_IN_SECOND;
10994                 if (bbr_is_ratio == 0) {
10995                         if (fval > bbr_rtt_probe_limit)
10996                                 newval = cur_rttp + (fval - bbr_rtt_probe_limit);
10997                         else
10998                                 newval = cur_rttp;
10999                 } else {
11000                         int mul;
11001
11002                         mul = fval / bbr_rtt_probe_limit;
11003                         newval = cur_rttp * mul;
11004                 }
11005                 if (cur_rttp >  bbr->r_ctl.rc_probertt_int) {
11006                         bbr->r_ctl.rc_probertt_int = cur_rttp;
11007                         reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
11008                         val = 1;
11009                 } else {
11010                         /*
11011                          * No adjustments were made
11012                          * do we need to shrink it?
11013                          */
11014                         if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) {
11015                                 if (cur_rttp <= bbr_rtt_probe_limit) {
11016                                         /*
11017                                          * Things have calmed down lets
11018                                          * shrink all the way to default
11019                                          */
11020                                         bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
11021                                         reset_time_small(&bbr->r_ctl.rc_rttprop,
11022                                                          (bbr_filter_len_sec * USECS_IN_SECOND));
11023                                         cur_rttp = bbr_rtt_probe_limit;
11024                                         newval = (bbr_filter_len_sec * USECS_IN_SECOND);
11025                                         val = 2;
11026                                 } else {
11027                                         /*
11028                                          * Well does some adjustment make sense?
11029                                          */
11030                                         if (cur_rttp < bbr->r_ctl.rc_probertt_int) {
11031                                                 /* We can reduce interval time some */
11032                                                 bbr->r_ctl.rc_probertt_int = cur_rttp;
11033                                                 reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
11034                                                 val = 3;
11035                                         }
11036                                 }
11037                         }
11038                 }
11039                 if (val)
11040                         bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val);
11041         }
11042 }
11043
11044 static void
11045 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
11046 {
11047         /* Exit probe-rtt */
11048
11049         if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) {
11050                 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
11051                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11052         }
11053         bbr_log_exit_gain(bbr, cts, 1);
11054         bbr->rc_hit_state_1 = 0;
11055         bbr->r_ctl.rc_rtt_shrinks = cts;
11056         bbr->r_ctl.last_in_probertt = cts;
11057         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0);
11058         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11059         bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp,
11060                                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
11061                                           bbr->r_ctl.rc_delivered);
11062         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11063                 uint32_t time_in;
11064
11065                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11066                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11067         }
11068         if (bbr->rc_filled_pipe) {
11069                 /* Switch to probe_bw */
11070                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11071                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11072                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
11073                 bbr_substate_change(bbr, cts, __LINE__, 0);
11074                 bbr_log_type_statechange(bbr, cts, __LINE__);
11075         } else {
11076                 /* Back to startup */
11077                 bbr->rc_bbr_state = BBR_STATE_STARTUP;
11078                 bbr->r_ctl.rc_bbr_state_time = cts;
11079                 /*
11080                  * We don't want to give a complete free 3
11081                  * measurements until we exit, so we use
11082                  * the number of pe's we were in probe-rtt
11083                  * to add to the startup_epoch. That way
11084                  * we will still retain the old state.
11085                  */
11086                 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt);
11087                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11088                 /* Make sure to use the lower pg when shifting back in */
11089                 if (bbr->r_ctl.rc_lost &&
11090                     bbr_use_lower_gain_in_startup &&
11091                     (bbr->rc_use_google == 0))
11092                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
11093                 else
11094                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
11095                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
11096                 /* Probably not needed but set it anyway */
11097                 bbr_set_state_target(bbr, __LINE__);
11098                 bbr_log_type_statechange(bbr, cts, __LINE__);
11099                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11100                     bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0);
11101         }
11102         bbr_check_probe_rtt_limits(bbr, cts);
11103 }
11104
11105 static int32_t inline
11106 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts)
11107 {
11108         if ((bbr->rc_past_init_win == 1) &&
11109             (bbr->rc_in_persist == 0) &&
11110             (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) {
11111                 return (1);
11112         }
11113         if (bbr_can_force_probertt &&
11114             (bbr->rc_in_persist == 0) &&
11115             (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
11116             ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
11117                 return (1);
11118         }
11119         return (0);
11120 }
11121
11122
11123 static int32_t
11124 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t  pkt_epoch)
11125 {
11126         uint64_t btlbw, gain;
11127         if (pkt_epoch == 0) {
11128                 /*
11129                  * Need to be on a pkt-epoch to continue.
11130                  */
11131                 return (0);
11132         }
11133         btlbw = bbr_get_full_bw(bbr);
11134         gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11135                  (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11136         if (btlbw >= gain) {
11137                 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11138                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11139                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11140                 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
11141         }
11142         if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)
11143                 return (1);
11144         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11145                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
11146         return(0);
11147 }
11148
11149 static int32_t inline
11150 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch)
11151 {
11152         /* Have we gained 25% in the last 3 packet based epoch's? */
11153         uint64_t btlbw, gain;
11154         int do_exit;
11155         int delta, rtt_gain;
11156
11157         if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11158             (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11159                 /*
11160                  * This qualifies as a RTT_PROBE session since we drop the
11161                  * data outstanding to nothing and waited more than
11162                  * bbr_rtt_probe_time.
11163                  */
11164                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11165                 bbr_set_reduced_rtt(bbr, cts, __LINE__);
11166         }
11167         if (bbr_should_enter_probe_rtt(bbr, cts)) {
11168                 bbr_enter_probe_rtt(bbr, cts, __LINE__);
11169                 return (0);
11170         }
11171         if (bbr->rc_use_google)
11172                 return (bbr_google_startup(bbr, cts,  pkt_epoch));
11173
11174         if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11175             (bbr_use_lower_gain_in_startup)) {
11176                 /* Drop to a lower gain 1.5 x since we saw loss */
11177                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
11178         }
11179         if (pkt_epoch == 0) {
11180                 /*
11181                  * Need to be on a pkt-epoch to continue.
11182                  */
11183                 return (0);
11184         }
11185         if (bbr_rtt_gain_thresh) {
11186                 /*
11187                  * Do we allow a flow to stay
11188                  * in startup with no loss and no
11189                  * gain in rtt over a set threshold?
11190                  */
11191                 if (bbr->r_ctl.rc_pkt_epoch_rtt &&
11192                     bbr->r_ctl.startup_last_srtt &&
11193                     (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) {
11194                         delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt;
11195                         rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt;
11196                 } else
11197                         rtt_gain = 0;
11198                 if ((bbr->r_ctl.startup_last_srtt == 0)  ||
11199                     (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt))
11200                         /* First time or new lower value */
11201                         bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
11202
11203                 if ((bbr->r_ctl.rc_lost == 0) &&
11204                     (rtt_gain < bbr_rtt_gain_thresh)) {
11205                         /*
11206                          * No loss, and we are under
11207                          * our gain threhold for
11208                          * increasing RTT.
11209                          */
11210                         if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11211                                 bbr->r_ctl.rc_bbr_last_startup_epoch++;
11212                         bbr_log_startup_event(bbr, cts, rtt_gain,
11213                                               delta, bbr->r_ctl.startup_last_srtt, 10);
11214                         return (0);
11215                 }
11216         }
11217         if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) &&
11218             (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) &&
11219             (!IN_RECOVERY(bbr->rc_tp->t_flags))) {
11220                 /*
11221                  * We only assess if we have a new measurment when
11222                  * we have no loss and are not in recovery.
11223                  * Drag up by one our last_startup epoch so we will hold
11224                  * the number of non-gain we have already accumulated.
11225                  */
11226                 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11227                         bbr->r_ctl.rc_bbr_last_startup_epoch++;
11228                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11229                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9);
11230                 return (0);
11231         }
11232         /* Case where we reduced the lost (bad retransmit) */
11233         if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost)
11234                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11235         bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count;
11236         btlbw = bbr_get_full_bw(bbr);
11237         if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower)
11238                 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11239                          (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11240         else
11241                 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11242                          (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11243         do_exit = 0;
11244         if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw)
11245                 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
11246         if (btlbw >= gain) {
11247                 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11248                 /* Update the lost so we won't exit in next set of tests */
11249                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11250                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11251                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11252         }
11253         if ((bbr->rc_loss_exit &&
11254              (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11255              (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) &&
11256             ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) {
11257                 /*
11258                  * If we had no gain,  we had loss and that loss was above
11259                  * our threshould, the rwnd is not constrained, and we have
11260                  * had at least 3 packet epochs exit. Note that this is
11261                  * switched off by sysctl. Google does not do this by the
11262                  * way.
11263                  */
11264                 if ((ctf_flight_size(bbr->rc_tp,
11265                          (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
11266                      (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) {
11267                         do_exit = 1;
11268                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11269                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4);
11270                 } else {
11271                         /* Just record an updated loss value */
11272                         bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11273                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11274                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5);
11275                 }
11276         } else
11277                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11278         if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) ||
11279             do_exit) {
11280                 /* Return 1 to exit the startup state. */
11281                 return (1);
11282         }
11283         /* Stay in startup */
11284         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11285                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
11286         return (0);
11287 }
11288
11289 static void
11290 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses)
11291 {
11292         /*
11293          * A tick occured in the rtt epoch do we need to do anything?
11294          */
11295 #ifdef BBR_INVARIANTS
11296         if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
11297             (bbr->rc_bbr_state != BBR_STATE_DRAIN) &&
11298             (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) &&
11299             (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
11300             (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) {
11301                 /* Debug code? */
11302                 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state);
11303         }
11304 #endif
11305         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
11306                 /* Do we exit the startup state? */
11307                 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) {
11308                         uint32_t time_in;
11309
11310                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11311                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6);
11312                         bbr->rc_filled_pipe = 1;
11313                         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11314                         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11315
11316                                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11317                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11318                         } else
11319                                 time_in = 0;
11320                         if (bbr->rc_no_pacing)
11321                                 bbr->rc_no_pacing = 0;
11322                         bbr->r_ctl.rc_bbr_state_time = cts;
11323                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg;
11324                         bbr->rc_bbr_state = BBR_STATE_DRAIN;
11325                         bbr_set_state_target(bbr, __LINE__);
11326                         if ((bbr->rc_use_google == 0) &&
11327                             bbr_slam_cwnd_in_main_drain) {
11328                                 /* Here we don't have to worry about probe-rtt */
11329                                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
11330                                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11331                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11332                         }
11333                         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
11334                         bbr_log_type_statechange(bbr, cts, __LINE__);
11335                         if (ctf_flight_size(bbr->rc_tp,
11336                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
11337                             bbr->r_ctl.rc_target_at_state) {
11338                                 /*
11339                                  * Switch to probe_bw if we are already
11340                                  * there
11341                                  */
11342                                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11343                                 bbr_substate_change(bbr, cts, __LINE__, 0);
11344                                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11345                                 bbr_log_type_statechange(bbr, cts, __LINE__);
11346                         }
11347                 }
11348         } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) {
11349                 uint32_t inflight;
11350                 struct tcpcb *tp;
11351
11352                 tp = bbr->rc_tp;
11353                 inflight = ctf_flight_size(tp,
11354                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11355                 if (inflight >= bbr->r_ctl.rc_target_at_state) {
11356                         /* We have reached a flight of the cwnd target */
11357                         bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11358                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11359                         bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
11360                         bbr_set_state_target(bbr, __LINE__);
11361                         /*
11362                          * Rig it so we don't do anything crazy and
11363                          * start fresh with a new randomization.
11364                          */
11365                         bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
11366                         bbr->rc_bbr_substate = BBR_SUB_LEVEL6;
11367                         bbr_substate_change(bbr, cts, __LINE__, 1);
11368                 }
11369         } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) {
11370                 /* Has in-flight reached the bdp (or less)? */
11371                 uint32_t inflight;
11372                 struct tcpcb *tp;
11373
11374                 tp = bbr->rc_tp;
11375                 inflight = ctf_flight_size(tp,
11376                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11377                 if ((bbr->rc_use_google == 0) &&
11378                     bbr_slam_cwnd_in_main_drain &&
11379                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11380                         /*
11381                          * Here we don't have to worry about probe-rtt
11382                          * re-slam it, but keep it slammed down.
11383                          */
11384                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11385                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11386                 }
11387                 if (inflight <= bbr->r_ctl.rc_target_at_state) {
11388                         /* We have drained */
11389                         bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11390                         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11391                         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11392                                 uint32_t time_in;
11393
11394                                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11395                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11396                         }
11397                         if ((bbr->rc_use_google == 0) &&
11398                             bbr_slam_cwnd_in_main_drain &&
11399                             (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
11400                                 /* Restore the cwnd */
11401                                 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
11402                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11403                         }
11404                         /* Setup probe-rtt has being done now RRS-HERE */
11405                         bbr->r_ctl.rc_rtt_shrinks = cts;
11406                         bbr->r_ctl.last_in_probertt = cts;
11407                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0);
11408                         /* Randomly pick a sub-state */
11409                         bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11410                         bbr_substate_change(bbr, cts, __LINE__, 0);
11411                         bbr_log_type_statechange(bbr, cts, __LINE__);
11412                 }
11413         } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) {
11414                 uint32_t flight;
11415
11416                 flight = ctf_flight_size(bbr->rc_tp,
11417                              (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11418                 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered);
11419                 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) &&
11420                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11421                         /*
11422                          * We must keep cwnd at the desired MSS.
11423                          */
11424                         bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
11425                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11426                 } else if ((bbr_prtt_slam_cwnd) &&
11427                            (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11428                         /* Re-slam it */
11429                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11430                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11431                 }
11432                 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) {
11433                         /* Has outstanding reached our target? */
11434                         if (flight <= bbr->r_ctl.rc_target_at_state) {
11435                                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0);
11436                                 bbr->r_ctl.rc_bbr_enters_probertt = cts;
11437                                 /* If time is exactly 0, be 1usec off */
11438                                 if (bbr->r_ctl.rc_bbr_enters_probertt == 0)
11439                                         bbr->r_ctl.rc_bbr_enters_probertt = 1;
11440                                 if (bbr->rc_use_google == 0) {
11441                                         /*
11442                                          * Restore any lowering that as occured to
11443                                          * reach here
11444                                          */
11445                                         if (bbr->r_ctl.bbr_rttprobe_gain_val)
11446                                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
11447                                         else
11448                                                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11449                                 }
11450                         }
11451                         if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) &&
11452                             (bbr->rc_use_google == 0) &&
11453                             bbr->r_ctl.bbr_rttprobe_gain_val &&
11454                             (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) ||
11455                              (flight >= bbr->r_ctl.flightsize_at_drain))) {
11456                                 /*
11457                                  * We have doddled with our current hptsi
11458                                  * gain an srtt and have still not made it
11459                                  * to target, or we have increased our flight.
11460                                  * Lets reduce the gain by xx%
11461                                  * flooring the reduce at DRAIN (based on
11462                                  * mul/div)
11463                                  */
11464                                 int red;
11465
11466                                 bbr->r_ctl.flightsize_at_drain = flight;
11467                                 bbr->r_ctl.rc_probertt_srttchktim = cts;
11468                                 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1);
11469                                 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) {
11470                                         /* Reduce our gain again */
11471                                         bbr->r_ctl.rc_bbr_hptsi_gain -= red;
11472                                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0);
11473                                 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) {
11474                                         /* one more chance before we give up */
11475                                         bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
11476                                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0);
11477                                 } else {
11478                                         /* At the very bottom */
11479                                         bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1);
11480                                 }
11481                         }
11482                 }
11483                 if (bbr->r_ctl.rc_bbr_enters_probertt &&
11484                     (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) &&
11485                     ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) {
11486                         /* Time to exit probe RTT normally */
11487                         bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts);
11488                 }
11489         } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
11490                 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11491                     (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11492                         /*
11493                          * This qualifies as a RTT_PROBE session since we
11494                          * drop the data outstanding to nothing and waited
11495                          * more than bbr_rtt_probe_time.
11496                          */
11497                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11498                         bbr_set_reduced_rtt(bbr, cts, __LINE__);
11499                 }
11500                 if (bbr_should_enter_probe_rtt(bbr, cts)) {
11501                         bbr_enter_probe_rtt(bbr, cts, __LINE__);
11502                 } else {
11503                         bbr_set_probebw_gains(bbr, cts, losses);
11504                 }
11505         }
11506 }
11507
11508 static void
11509 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses)
11510 {
11511         int32_t epoch = 0;
11512
11513         if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
11514                 bbr_set_epoch(bbr, cts, line);
11515                 /* At each epoch doe lt bw sampling */
11516                 epoch = 1;
11517         }
11518         bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses);
11519 }
11520
11521 static int
11522 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
11523     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
11524     int32_t nxt_pkt, struct timeval *tv)
11525 {
11526         int32_t thflags, retval;
11527         uint32_t cts, lcts;
11528         uint32_t tiwin;
11529         struct tcpopt to;
11530         struct tcp_bbr *bbr;
11531         struct bbr_sendmap *rsm;
11532         struct timeval ltv;
11533         int32_t did_out = 0;
11534         int32_t in_recovery;
11535         uint16_t nsegs;
11536         int32_t prev_state;
11537         uint32_t lost;
11538
11539         nsegs = max(1, m->m_pkthdr.lro_nsegs);
11540         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11541         /* add in our stats */
11542         kern_prefetch(bbr, &prev_state);
11543         prev_state = 0;
11544         thflags = th->th_flags;
11545         /*
11546          * If this is either a state-changing packet or current state isn't
11547          * established, we require a write lock on tcbinfo.  Otherwise, we
11548          * allow the tcbinfo to be in either alocked or unlocked, as the
11549          * caller may have unnecessarily acquired a write lock due to a
11550          * race.
11551          */
11552         INP_WLOCK_ASSERT(tp->t_inpcb);
11553         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
11554             __func__));
11555         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
11556             __func__));
11557
11558         tp->t_rcvtime = ticks;
11559         /*
11560          * Unscale the window into a 32-bit value. For the SYN_SENT state
11561          * the scale is zero.
11562          */
11563         tiwin = th->th_win << tp->snd_scale;
11564 #ifdef NETFLIX_STATS
11565         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
11566 #endif
11567         /*
11568          * Parse options on any incoming segment.
11569          */
11570         tcp_dooptions(&to, (u_char *)(th + 1),
11571             (th->th_off << 2) - sizeof(struct tcphdr),
11572             (thflags & TH_SYN) ? TO_SYN : 0);
11573
11574         if (m->m_flags & M_TSTMP) {
11575                 /* Prefer the hardware timestamp if present */
11576                 struct timespec ts;
11577
11578                 mbuf_tstmp2timespec(m, &ts);
11579                 bbr->rc_tv.tv_sec = ts.tv_sec;
11580                 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11581                 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11582         } else if (m->m_flags & M_TSTMP_LRO) {
11583                 /* Next the arrival timestamp */
11584                 struct timespec ts;
11585
11586                 mbuf_tstmp2timespec(m, &ts);
11587                 bbr->rc_tv.tv_sec = ts.tv_sec;
11588                 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11589                 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11590         } else {
11591                 /*
11592                  * Ok just get the current time.
11593                  */
11594                 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv);
11595         }
11596         /*
11597          * If echoed timestamp is later than the current time, fall back to
11598          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
11599          * were used when this connection was established.
11600          */
11601         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
11602                 to.to_tsecr -= tp->ts_offset;
11603                 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv)))
11604                         to.to_tsecr = 0;
11605         }
11606         /*
11607          * If its the first time in we need to take care of options and
11608          * verify we can do SACK for rack!
11609          */
11610         if (bbr->r_state == 0) {
11611                 /*
11612                  * Process options only when we get SYN/ACK back. The SYN
11613                  * case for incoming connections is handled in tcp_syncache.
11614                  * According to RFC1323 the window field in a SYN (i.e., a
11615                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
11616                  * this is traditional behavior, may need to be cleaned up.
11617                  */
11618                 if (bbr->rc_inp == NULL) {
11619                         bbr->rc_inp = tp->t_inpcb;
11620                 }
11621                 /*
11622                  * We need to init rc_inp here since its not init'd when
11623                  * bbr_init is called
11624                  */
11625                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
11626                         if ((to.to_flags & TOF_SCALE) &&
11627                             (tp->t_flags & TF_REQ_SCALE)) {
11628                                 tp->t_flags |= TF_RCVD_SCALE;
11629                                 tp->snd_scale = to.to_wscale;
11630                         }
11631                         /*
11632                          * Initial send window.  It will be updated with the
11633                          * next incoming segment to the scaled value.
11634                          */
11635                         tp->snd_wnd = th->th_win;
11636                         if (to.to_flags & TOF_TS) {
11637                                 tp->t_flags |= TF_RCVD_TSTMP;
11638                                 tp->ts_recent = to.to_tsval;
11639                                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
11640                         }
11641                         if (to.to_flags & TOF_MSS)
11642                                 tcp_mss(tp, to.to_mss);
11643                         if ((tp->t_flags & TF_SACK_PERMIT) &&
11644                             (to.to_flags & TOF_SACKPERM) == 0)
11645                                 tp->t_flags &= ~TF_SACK_PERMIT;
11646                         if (IS_FASTOPEN(tp->t_flags)) {
11647                                 if (to.to_flags & TOF_FASTOPEN) {
11648                                         uint16_t mss;
11649
11650                                         if (to.to_flags & TOF_MSS)
11651                                                 mss = to.to_mss;
11652                                         else
11653                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
11654                                                         mss = TCP6_MSS;
11655                                                 else
11656                                                         mss = TCP_MSS;
11657                                         tcp_fastopen_update_cache(tp, mss,
11658                                             to.to_tfo_len, to.to_tfo_cookie);
11659                                 } else
11660                                         tcp_fastopen_disable_path(tp);
11661                         }
11662                 }
11663                 /*
11664                  * At this point we are at the initial call. Here we decide
11665                  * if we are doing RACK or not. We do this by seeing if
11666                  * TF_SACK_PERMIT is set, if not rack is *not* possible and
11667                  * we switch to the default code.
11668                  */
11669                 if ((tp->t_flags & TF_SACK_PERMIT) == 0) {
11670                         /* Bail */
11671                         tcp_switch_back_to_default(tp);
11672                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
11673                             tlen, iptos);
11674                         return (1);
11675                 }
11676                 /* Set the flag */
11677                 bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
11678                 tcp_set_hpts(tp->t_inpcb);
11679                 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack);
11680         }
11681         if (thflags & TH_ACK) {
11682                 /* Track ack types */
11683                 if (to.to_flags & TOF_SACK)
11684                         BBR_STAT_INC(bbr_acks_with_sacks);
11685                 else
11686                         BBR_STAT_INC(bbr_plain_acks);
11687         }
11688         /*
11689          * This is the one exception case where we set the rack state
11690          * always. All other times (timers etc) we must have a rack-state
11691          * set (so we assure we have done the checks above for SACK).
11692          */
11693         if (bbr->r_state != tp->t_state)
11694                 bbr_set_state(tp, bbr, tiwin);
11695
11696         if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL)
11697                 kern_prefetch(rsm, &prev_state);
11698         prev_state = bbr->r_state;
11699         bbr->rc_ack_was_delayed = 0;
11700         lost = bbr->r_ctl.rc_lost;
11701         bbr->rc_is_pkt_epoch_now = 0;
11702         if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) {
11703                 /* Get the real time into lcts and figure the real delay */
11704                 lcts = tcp_get_usecs(&ltv);
11705                 if (TSTMP_GT(lcts, cts)) {
11706                         bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts;
11707                         bbr->rc_ack_was_delayed = 1;
11708                         if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay,
11709                                      bbr->r_ctl.highest_hdwr_delay))
11710                                 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay;
11711                 } else {
11712                         bbr->r_ctl.rc_ack_hdwr_delay = 0;
11713                         bbr->rc_ack_was_delayed = 0;
11714                 }
11715         } else {
11716                 bbr->r_ctl.rc_ack_hdwr_delay = 0;
11717                 bbr->rc_ack_was_delayed = 0;
11718         }
11719         bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m);
11720         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
11721                 retval = 0;
11722                 m_freem(m);
11723                 goto done_with_input;
11724         }
11725         /*
11726          * If a segment with the ACK-bit set arrives in the SYN-SENT state
11727          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
11728          */
11729         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
11730             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
11731                 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11732                 return (1);
11733         }
11734         in_recovery = IN_RECOVERY(tp->t_flags);
11735         if (tiwin > bbr->r_ctl.rc_high_rwnd)
11736                 bbr->r_ctl.rc_high_rwnd = tiwin;
11737 #ifdef BBR_INVARIANTS
11738         if ((tp->t_inpcb->inp_flags & INP_DROPPED) ||
11739             (tp->t_inpcb->inp_flags2 & INP_FREED)) {
11740                 panic("tp:%p bbr:%p given a dropped inp:%p",
11741                     tp, bbr, tp->t_inpcb);
11742         }
11743 #endif
11744         bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp,
11745                                             (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11746         bbr->rtt_valid = 0;
11747         if (to.to_flags & TOF_TS) {
11748                 bbr->rc_ts_valid = 1;
11749                 bbr->r_ctl.last_inbound_ts = to.to_tsval;
11750         } else {
11751                 bbr->rc_ts_valid = 0;
11752                 bbr->r_ctl.last_inbound_ts = 0;
11753         }
11754         retval = (*bbr->r_substate) (m, th, so,
11755             tp, &to, drop_hdrlen,
11756             tlen, tiwin, thflags, nxt_pkt);
11757 #ifdef BBR_INVARIANTS
11758         if ((retval == 0) &&
11759             (tp->t_inpcb == NULL)) {
11760                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
11761                     retval, tp, prev_state);
11762         }
11763 #endif
11764         if (nxt_pkt == 0)
11765                 BBR_STAT_INC(bbr_rlock_left_ret0);
11766         else
11767                 BBR_STAT_INC(bbr_rlock_left_ret1);
11768         if (retval == 0) {
11769                 /*
11770                  * If retval is 1 the tcb is unlocked and most likely the tp
11771                  * is gone.
11772                  */
11773                 INP_WLOCK_ASSERT(tp->t_inpcb);
11774                 tcp_bbr_xmit_timer_commit(bbr, tp, cts);
11775                 if (bbr->rc_is_pkt_epoch_now)
11776                         bbr_set_pktepoch(bbr, cts, __LINE__);
11777                 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost));
11778                 if (nxt_pkt == 0) {
11779                         if (bbr->r_wanted_output != 0) {
11780                                 bbr->rc_output_starts_timer = 0;
11781                                 did_out = 1;
11782                                 (void)tp->t_fb->tfb_tcp_output(tp);
11783                         } else
11784                                 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0);
11785                 }
11786                 if ((nxt_pkt == 0) &&
11787                     ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
11788                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
11789                      (tp->t_flags & TF_DELACK) ||
11790                      ((tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
11791                       (tp->t_state <= TCPS_CLOSING)))) {
11792                         /*
11793                          * We could not send (probably in the hpts but
11794                          * stopped the timer)?
11795                          */
11796                         if ((tp->snd_max == tp->snd_una) &&
11797                             ((tp->t_flags & TF_DELACK) == 0) &&
11798                             (bbr->rc_inp->inp_in_hpts) &&
11799                             (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11800                                 /*
11801                                  * keep alive not needed if we are hptsi
11802                                  * output yet
11803                                  */
11804                                 ;
11805                         } else {
11806                                 if (bbr->rc_inp->inp_in_hpts) {
11807                                         tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
11808                                         if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11809                                             (TSTMP_GT(lcts, bbr->rc_pacer_started))) {
11810                                                 uint32_t del;
11811
11812                                                 del = lcts - bbr->rc_pacer_started;
11813                                                 if (bbr->r_ctl.rc_last_delay_val > del) {
11814                                                         BBR_STAT_INC(bbr_force_timer_start);
11815                                                         bbr->r_ctl.rc_last_delay_val -= del;
11816                                                         bbr->rc_pacer_started = lcts;
11817                                                 } else {
11818                                                         /* We are late */
11819                                                         bbr->r_ctl.rc_last_delay_val = 0;
11820                                                         BBR_STAT_INC(bbr_force_output);
11821                                                         (void)tp->t_fb->tfb_tcp_output(tp);
11822                                                 }
11823                                         }
11824                                 }
11825                                 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val,
11826                                     0);
11827                         }
11828                 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) {
11829                         /* Do we have the correct timer running? */
11830                         bbr_timer_audit(tp, bbr, lcts, &so->so_snd);
11831                 }
11832                 /* Do we have a new state */
11833                 if (bbr->r_state != tp->t_state)
11834                         bbr_set_state(tp, bbr, tiwin);
11835 done_with_input:
11836                 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out);
11837                 if (did_out)
11838                         bbr->r_wanted_output = 0;
11839 #ifdef BBR_INVARIANTS
11840                 if (tp->t_inpcb == NULL) {
11841                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
11842                             did_out,
11843                             retval, tp, prev_state);
11844                 }
11845 #endif
11846         }
11847         return (retval);
11848 }
11849
11850 static void
11851 bbr_log_type_hrdwtso(struct tcpcb *tp, struct tcp_bbr *bbr, int len, int mod, int what_we_can_send)
11852 {
11853         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
11854                 union tcp_log_stackspecific log;
11855                 struct timeval tv;
11856                 uint32_t cts;
11857
11858                 cts = tcp_get_usecs(&tv);
11859                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
11860                 log.u_bbr.flex1 = bbr->r_ctl.rc_pace_min_segs;
11861                 log.u_bbr.flex2 = what_we_can_send;
11862                 log.u_bbr.flex3 = bbr->r_ctl.rc_pace_max_segs;
11863                 log.u_bbr.flex4 = len;
11864                 log.u_bbr.flex5 = 0;
11865                 log.u_bbr.flex7 = mod;
11866                 log.u_bbr.flex8 = 1;
11867                 TCP_LOG_EVENTP(tp, NULL,
11868                     &tp->t_inpcb->inp_socket->so_rcv,
11869                     &tp->t_inpcb->inp_socket->so_snd,
11870                     TCP_HDWR_TLS, 0,
11871                     0, &log, false, &tv);
11872         }
11873 }
11874
11875 static void
11876 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
11877     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
11878 {
11879         struct timeval tv;
11880         int retval;
11881
11882         /* First lets see if we have old packets */
11883         if (tp->t_in_pkt) {
11884                 if (ctf_do_queued_segments(so, tp, 1)) {
11885                         m_freem(m);
11886                         return;
11887                 }
11888         }
11889         if (m->m_flags & M_TSTMP_LRO) {
11890                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
11891                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
11892         } else {
11893                 /* Should not be should we kassert instead? */
11894                 tcp_get_usecs(&tv);
11895         }
11896         retval = bbr_do_segment_nounlock(m, th, so, tp,
11897                                          drop_hdrlen, tlen, iptos, 0, &tv);
11898         if (retval == 0)
11899                 INP_WUNLOCK(tp->t_inpcb);
11900 }
11901
11902 /*
11903  * Return how much data can be sent without violating the
11904  * cwnd or rwnd.
11905  */
11906
11907 static inline uint32_t
11908 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin,
11909     uint32_t avail, int32_t sb_offset, uint32_t cts)
11910 {
11911         uint32_t len;
11912
11913         if (ctf_outstanding(tp) >= tp->snd_wnd) {
11914                 /* We never want to go over our peers rcv-window */
11915                 len = 0;
11916         } else {
11917                 uint32_t flight;
11918
11919                 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11920                 if (flight >= sendwin) {
11921                         /*
11922                          * We have in flight what we are allowed by cwnd (if
11923                          * it was rwnd blocking it would have hit above out
11924                          * >= tp->snd_wnd).
11925                          */
11926                         return (0);
11927                 }
11928                 len = sendwin - flight;
11929                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
11930                         /* We would send too much (beyond the rwnd) */
11931                         len = tp->snd_wnd - ctf_outstanding(tp);
11932                 }
11933                 if ((len + sb_offset) > avail) {
11934                         /*
11935                          * We don't have that much in the SB, how much is
11936                          * there?
11937                          */
11938                         len = avail - sb_offset;
11939                 }
11940         }
11941         return (len);
11942 }
11943
11944 static inline void
11945 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11946 {
11947 #ifdef NETFLIX_STATS
11948         TCPSTAT_INC(tcps_sndpack_error);
11949         TCPSTAT_ADD(tcps_sndbyte_error, len);
11950 #endif
11951 }
11952
11953 static inline void
11954 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11955 {
11956         if (error) {
11957                 bbr_do_error_accounting(tp, bbr, rsm, len, error);
11958                 return;
11959         }
11960         if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
11961                 /* Window probe */
11962                 TCPSTAT_INC(tcps_sndprobe);
11963 #ifdef NETFLIX_STATS
11964                 stats_voi_update_abs_u32(tp->t_stats,
11965                     VOI_TCP_RETXPB, len);
11966 #endif
11967         } else if (rsm) {
11968                 if (rsm->r_flags & BBR_TLP) {
11969                         /*
11970                          * TLP should not count in retran count, but in its
11971                          * own bin
11972                          */
11973 #ifdef NETFLIX_STATS
11974                         tp->t_sndtlppack++;
11975                         tp->t_sndtlpbyte += len;
11976                         TCPSTAT_INC(tcps_tlpresends);
11977                         TCPSTAT_ADD(tcps_tlpresend_bytes, len);
11978 #endif
11979                 } else {
11980                         /* Retransmit */
11981                         tp->t_sndrexmitpack++;
11982                         TCPSTAT_INC(tcps_sndrexmitpack);
11983                         TCPSTAT_ADD(tcps_sndrexmitbyte, len);
11984 #ifdef NETFLIX_STATS
11985                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
11986                             len);
11987 #endif
11988                 }
11989                 /*
11990                  * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is
11991                  * sub-state
11992                  */
11993                 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len);
11994                 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) {
11995                         /* Non probe_bw log in 1, 2, or 4. */
11996                         counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len);
11997                 } else {
11998                         /*
11999                          * Log our probe state 3, and log also 5-13 to show
12000                          * us the recovery sub-state for the send. This
12001                          * means that 3 == (5+6+7+8+9+10+11+12+13)
12002                          */
12003                         counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len);
12004                         counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len);
12005                 }
12006                 /* Place in both 16's the totals of retransmitted */
12007                 counter_u64_add(bbr_state_lost[16], len);
12008                 counter_u64_add(bbr_state_resend[16], len);
12009                 /* Place in 17's the total sent */
12010                 counter_u64_add(bbr_state_resend[17], len);
12011                 counter_u64_add(bbr_state_lost[17], len);
12012
12013         } else {
12014                 /* New sends */
12015                 TCPSTAT_INC(tcps_sndpack);
12016                 TCPSTAT_ADD(tcps_sndbyte, len);
12017                 /* Place in 17's the total sent */
12018                 counter_u64_add(bbr_state_resend[17], len);
12019                 counter_u64_add(bbr_state_lost[17], len);
12020 #ifdef NETFLIX_STATS
12021                 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
12022                     len);
12023 #endif
12024         }
12025 }
12026
12027 static void
12028 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level)
12029 {
12030         if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) {
12031                 /*
12032                  * Limit the cwnd to not be above N x the target plus whats
12033                  * is outstanding. The target is based on the current b/w
12034                  * estimate.
12035                  */
12036                 uint32_t target;
12037
12038                 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT);
12039                 target += ctf_outstanding(tp);
12040                 target *= bbr_target_cwnd_mult_limit;
12041                 if (tp->snd_cwnd > target)
12042                         tp->snd_cwnd = target;
12043                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__);
12044         }
12045 }
12046
12047 static int
12048 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg)
12049 {
12050         /*
12051          * "adv" is the amount we could increase the window, taking into
12052          * account that we are limited by TCP_MAXWIN << tp->rcv_scale.
12053          */
12054         uint32_t adv;
12055         int32_t oldwin;
12056
12057         adv = min(recwin, TCP_MAXWIN << tp->rcv_scale);
12058         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
12059                 oldwin = (tp->rcv_adv - tp->rcv_nxt);
12060                 adv -= oldwin;
12061         } else
12062                 oldwin = 0;
12063
12064         /*
12065          * If the new window size ends up being the same as the old size
12066          * when it is scaled, then don't force a window update.
12067          */
12068         if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
12069                 return (0);
12070
12071         if (adv >= (2 * maxseg) &&
12072             (adv >= (so->so_rcv.sb_hiwat / 4) ||
12073             recwin <= (so->so_rcv.sb_hiwat / 8) ||
12074             so->so_rcv.sb_hiwat <= 8 * maxseg)) {
12075                 return (1);
12076         }
12077         if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat)
12078                 return (1);
12079         return (0);
12080 }
12081
12082 /*
12083  * Return 0 on success and a errno on failure to send.
12084  * Note that a 0 return may not mean we sent anything
12085  * if the TCB was on the hpts. A non-zero return
12086  * does indicate the error we got from ip[6]_output.
12087  */
12088 static int
12089 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv)
12090 {
12091         struct socket *so;
12092         int32_t len;
12093         uint32_t cts;
12094         uint32_t recwin, sendwin;
12095         int32_t sb_offset;
12096         int32_t flags, abandon, error = 0;
12097         struct tcp_log_buffer *lgb = NULL;
12098         struct mbuf *m;
12099         struct mbuf *mb;
12100         uint32_t if_hw_tsomaxsegcount = 0;
12101         uint32_t if_hw_tsomaxsegsize = 0;
12102         uint32_t if_hw_tsomax = 0;
12103         struct ip *ip = NULL;
12104 #ifdef TCPDEBUG
12105         struct ipovly *ipov = NULL;
12106 #endif
12107         struct tcp_bbr *bbr;
12108         struct tcphdr *th;
12109 #ifdef NETFLIX_TCPOUDP
12110         struct udphdr *udp = NULL;
12111 #endif
12112         u_char opt[TCP_MAXOLEN];
12113         unsigned ipoptlen, optlen, hdrlen;
12114 #ifdef NETFLIX_TCPOUDP
12115         unsigned ulen;
12116 #endif
12117         uint32_t bbr_seq;
12118         uint32_t delay_calc=0;
12119         uint8_t doing_tlp = 0;
12120         uint8_t local_options;
12121 #ifdef BBR_INVARIANTS
12122         uint8_t doing_retran_from = 0;
12123         uint8_t picked_up_retran = 0;
12124 #endif
12125         uint8_t wanted_cookie = 0;
12126         uint8_t more_to_rxt=0;
12127         int32_t prefetch_so_done = 0;
12128         int32_t prefetch_rsm = 0;
12129         uint32_t what_we_can = 0;
12130         uint32_t tot_len = 0;
12131         uint32_t rtr_cnt = 0;
12132         uint32_t maxseg, pace_max_segs, p_maxseg;
12133         int32_t csum_flags;
12134         int32_t hw_tls;
12135 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12136         unsigned ipsec_optlen = 0;
12137
12138 #endif
12139         volatile int32_t sack_rxmit;
12140         struct bbr_sendmap *rsm = NULL;
12141         int32_t tso, mtu;
12142         int force_tso = 0;
12143         struct tcpopt to;
12144         int32_t slot = 0;
12145         struct inpcb *inp;
12146         struct sockbuf *sb;
12147         uint32_t hpts_calling;
12148 #ifdef INET6
12149         struct ip6_hdr *ip6 = NULL;
12150         int32_t isipv6;
12151 #endif
12152         uint8_t app_limited = BBR_JR_SENT_DATA;
12153         uint8_t filled_all = 0;
12154         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
12155         /* We take a cache hit here */
12156         memcpy(&bbr->rc_tv, tv, sizeof(struct timeval));
12157         cts = tcp_tv_to_usectick(&bbr->rc_tv);
12158         inp = bbr->rc_inp;
12159         so = inp->inp_socket;
12160         sb = &so->so_snd;
12161 #ifdef KERN_TLS
12162         if (sb->sb_flags & SB_TLS_IFNET)
12163                 hw_tls = 1;
12164         else
12165 #endif
12166                 hw_tls = 0;
12167         kern_prefetch(sb, &maxseg);
12168         maxseg = tp->t_maxseg - bbr->rc_last_options;
12169         if (bbr_minseg(bbr) < maxseg) {
12170                 tcp_bbr_tso_size_check(bbr, cts);
12171         }
12172         /* Remove any flags that indicate we are pacing on the inp  */
12173         pace_max_segs = bbr->r_ctl.rc_pace_max_segs;
12174         p_maxseg = min(maxseg, pace_max_segs);
12175         INP_WLOCK_ASSERT(inp);
12176 #ifdef TCP_OFFLOAD
12177         if (tp->t_flags & TF_TOE)
12178                 return (tcp_offload_output(tp));
12179 #endif
12180
12181 #ifdef INET6
12182         if (bbr->r_state) {
12183                 /* Use the cache line loaded if possible */
12184                 isipv6 = bbr->r_is_v6;
12185         } else {
12186                 isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
12187         }
12188 #endif
12189         if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
12190             inp->inp_in_hpts) {
12191                 /*
12192                  * We are on the hpts for some timer but not hptsi output.
12193                  * Possibly remove from the hpts so we can send/recv etc.
12194                  */
12195                 if ((tp->t_flags & TF_ACKNOW) == 0) {
12196                         /*
12197                          * No immediate demand right now to send an ack, but
12198                          * the user may have read, making room for new data
12199                          * (a window update). If so we may want to cancel
12200                          * whatever timer is running (KEEP/DEL-ACK?) and
12201                          * continue to send out a window update. Or we may
12202                          * have gotten more data into the socket buffer to
12203                          * send.
12204                          */
12205                         recwin = min(max(sbspace(&so->so_rcv), 0),
12206                             TCP_MAXWIN << tp->rcv_scale);
12207                         if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) &&
12208                             ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <=
12209                             (tp->snd_max - tp->snd_una))) {
12210                                 /*
12211                                  * Nothing new to send and no window update
12212                                  * is needed to send. Lets just return and
12213                                  * let the timer-run off.
12214                                  */
12215                                 return (0);
12216                         }
12217                 }
12218                 tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12219                 bbr_timer_cancel(bbr, __LINE__, cts);
12220         }
12221         if (bbr->r_ctl.rc_last_delay_val) {
12222                 /* Calculate a rough delay for early escape to sending  */
12223                 if (SEQ_GT(cts, bbr->rc_pacer_started))
12224                         delay_calc = cts - bbr->rc_pacer_started;
12225                 if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12226                         delay_calc -= bbr->r_ctl.rc_last_delay_val;
12227                 else
12228                         delay_calc = 0;
12229         }
12230         /* Mark that we have called bbr_output(). */
12231         if ((bbr->r_timer_override) ||
12232             (tp->t_flags & TF_FORCEDATA) ||
12233             (tp->t_state < TCPS_ESTABLISHED)) {
12234                 /* Timeouts or early states are exempt */
12235                 if (inp->inp_in_hpts)
12236                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12237         } else if (inp->inp_in_hpts) {
12238                 if ((bbr->r_ctl.rc_last_delay_val) &&
12239                     (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
12240                     delay_calc) {
12241                         /*
12242                          * We were being paced for output and the delay has
12243                          * already exceeded when we were supposed to be
12244                          * called, lets go ahead and pull out of the hpts
12245                          * and call output.
12246                          */
12247                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1);
12248                         bbr->r_ctl.rc_last_delay_val = 0;
12249                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12250                 } else if (tp->t_state == TCPS_CLOSED) {
12251                         bbr->r_ctl.rc_last_delay_val = 0;
12252                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12253                 } else {
12254                         /*
12255                          * On the hpts, you shall not pass! even if ACKNOW
12256                          * is on, we will when the hpts fires, unless of
12257                          * course we are overdue.
12258                          */
12259                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1);
12260                         return (0);
12261                 }
12262         }
12263         bbr->rc_cwnd_limited = 0;
12264         if (bbr->r_ctl.rc_last_delay_val) {
12265                 /* recalculate the real delay and deal with over/under  */
12266                 if (SEQ_GT(cts, bbr->rc_pacer_started))
12267                         delay_calc = cts - bbr->rc_pacer_started;
12268                 else
12269                         delay_calc = 0;
12270                 if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12271                         /* Setup the delay which will be added in */
12272                         delay_calc -= bbr->r_ctl.rc_last_delay_val;
12273                 else {
12274                         /*
12275                          * We are early setup to adjust
12276                          * our slot time.
12277                          */
12278                         uint64_t merged_val;
12279
12280                         bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc);
12281                         bbr->r_agg_early_set = 1;
12282                         if (bbr->r_ctl.rc_hptsi_agg_delay) {
12283                                 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) {
12284                                         /* Nope our previous late cancels out the early */
12285                                         bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early;
12286                                         bbr->r_agg_early_set = 0;
12287                                         bbr->r_ctl.rc_agg_early = 0;
12288                                 } else {
12289                                         bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay;
12290                                         bbr->r_ctl.rc_hptsi_agg_delay = 0;
12291                                 }
12292                         }
12293                         merged_val = bbr->rc_pacer_started;
12294                         merged_val <<= 32;
12295                         merged_val |= bbr->r_ctl.rc_last_delay_val;
12296                         bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls,
12297                                                  bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val,
12298                                                  bbr->r_agg_early_set, 3);
12299                         bbr->r_ctl.rc_last_delay_val = 0;
12300                         BBR_STAT_INC(bbr_early);
12301                         delay_calc = 0;
12302                 }
12303         } else {
12304                 /* We were not delayed due to hptsi */
12305                 if (bbr->r_agg_early_set)
12306                         bbr->r_ctl.rc_agg_early = 0;
12307                 bbr->r_agg_early_set = 0;
12308                 delay_calc = 0;
12309         }
12310         if (delay_calc) {
12311                 /*
12312                  * We had a hptsi delay which means we are falling behind on
12313                  * sending at the expected rate. Calculate an extra amount
12314                  * of data we can send, if any, to put us back on track.
12315                  */
12316                 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay)
12317                         bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff;
12318                 else
12319                         bbr->r_ctl.rc_hptsi_agg_delay += delay_calc;
12320         }
12321         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12322         if ((tp->snd_una == tp->snd_max) &&
12323             (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
12324             (sbavail(sb))) {
12325                 /*
12326                  * Ok we have been idle with nothing outstanding
12327                  * we possibly need to start fresh with either a new
12328                  * suite of states or a fast-ramp up.
12329                  */
12330                 bbr_restart_after_idle(bbr,
12331                                        cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time));
12332         }
12333         /*
12334          * Now was there a hptsi delay where we are behind? We only count
12335          * being behind if: a) We are not in recovery. b) There was a delay.
12336          * <and> c) We had room to send something.
12337          *
12338          */
12339         hpts_calling = inp->inp_hpts_calls;
12340         inp->inp_hpts_calls = 0;
12341         if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
12342                 if (bbr_process_timers(tp, bbr, cts, hpts_calling)) {
12343                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1);
12344                         return (0);
12345                 }
12346         }
12347         bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12348         if (hpts_calling &&
12349             (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12350                 bbr->r_ctl.rc_last_delay_val = 0;
12351         }
12352         bbr->r_timer_override = 0;
12353         bbr->r_wanted_output = 0;
12354         /*
12355          * For TFO connections in SYN_RECEIVED, only allow the initial
12356          * SYN|ACK and those sent by the retransmit timer.
12357          */
12358         if (IS_FASTOPEN(tp->t_flags) &&
12359             ((tp->t_state == TCPS_SYN_RECEIVED) ||
12360              (tp->t_state == TCPS_SYN_SENT)) &&
12361             SEQ_GT(tp->snd_max, tp->snd_una) && /* inital SYN or SYN|ACK sent */
12362             (tp->t_rxtshift == 0)) {    /* not a retransmit */
12363                 return (0);
12364         }
12365         /*
12366          * Before sending anything check for a state update. For hpts
12367          * calling without input this is important. If its input calling
12368          * then this was already done.
12369          */
12370         if (bbr->rc_use_google == 0)
12371                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12372 again:
12373         /*
12374          * If we've recently taken a timeout, snd_max will be greater than
12375          * snd_max. BBR in general does not pay much attention to snd_nxt
12376          * for historic reasons the persist timer still uses it. This means
12377          * we have to look at it. All retransmissions that are not persits
12378          * use the rsm that needs to be sent so snd_nxt is ignored. At the
12379          * end of this routine we pull snd_nxt always up to snd_max.
12380          */
12381         doing_tlp = 0;
12382 #ifdef BBR_INVARIANTS
12383         doing_retran_from = picked_up_retran = 0;
12384 #endif
12385         error = 0;
12386         tso = 0;
12387         slot = 0;
12388         mtu = 0;
12389         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12390         sb_offset = tp->snd_max - tp->snd_una;
12391         flags = tcp_outflags[tp->t_state];
12392         sack_rxmit = 0;
12393         len = 0;
12394         rsm = NULL;
12395         if (flags & TH_RST) {
12396                 SOCKBUF_LOCK(sb);
12397                 goto send;
12398         }
12399 recheck_resend:
12400         while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
12401                 /* We need to always have one in reserve */
12402                 rsm = bbr_alloc(bbr);
12403                 if (rsm == NULL) {
12404                         error = ENOMEM;
12405                         /* Lie to get on the hpts */
12406                         tot_len = tp->t_maxseg;
12407                         if (hpts_calling)
12408                                 /* Retry in a ms */
12409                                 slot = 1001;
12410                         goto just_return_nolock;
12411                 }
12412                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
12413                 bbr->r_ctl.rc_free_cnt++;
12414                 rsm = NULL;
12415         }
12416         /* What do we send, a resend? */
12417         if (bbr->r_ctl.rc_resend == NULL) {
12418                 /* Check for rack timeout */
12419                 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
12420                 if (bbr->r_ctl.rc_resend) {
12421 #ifdef BBR_INVARIANTS
12422                         picked_up_retran = 1;
12423 #endif
12424                         bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend);
12425                 }
12426         }
12427         if (bbr->r_ctl.rc_resend) {
12428                 rsm = bbr->r_ctl.rc_resend;
12429 #ifdef BBR_INVARIANTS
12430                 doing_retran_from = 1;
12431 #endif
12432                 /* Remove any TLP flags its a RACK or T-O */
12433                 rsm->r_flags &= ~BBR_TLP;
12434                 bbr->r_ctl.rc_resend = NULL;
12435                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
12436 #ifdef BBR_INVARIANTS
12437                         panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n",
12438                             tp, bbr, rsm, rsm->r_start, tp->snd_una);
12439                         goto recheck_resend;
12440 #else
12441                         /* TSNH */
12442                         rsm = NULL;
12443                         goto recheck_resend;
12444 #endif
12445                 }
12446                 rtr_cnt++;
12447                 if (rsm->r_flags & BBR_HAS_SYN) {
12448                         /* Only retransmit a SYN by itself */
12449                         len = 0;
12450                         if ((flags & TH_SYN) == 0) {
12451                                 /* Huh something is wrong */
12452                                 rsm->r_start++;
12453                                 if (rsm->r_start == rsm->r_end) {
12454                                         /* Clean it up, somehow we missed the ack? */
12455                                         bbr_log_syn(tp, NULL);
12456                                 } else {
12457                                         /* TFO with data? */
12458                                         rsm->r_flags &= ~BBR_HAS_SYN;
12459                                         len = rsm->r_end - rsm->r_start;
12460                                 }
12461                         } else {
12462                                 /* Retransmitting SYN */
12463                                 rsm = NULL;
12464                                 SOCKBUF_LOCK(sb);
12465                                 goto send;
12466                         }
12467                 } else
12468                         len = rsm->r_end - rsm->r_start;
12469                 if ((bbr->rc_resends_use_tso == 0) &&
12470 #ifdef KERN_TLS
12471                     ((sb->sb_flags & SB_TLS_IFNET) == 0) &&
12472 #endif
12473                     (len > maxseg)) {
12474                         len = maxseg;
12475                         more_to_rxt = 1;
12476                 }
12477                 sb_offset = rsm->r_start - tp->snd_una;
12478                 if (len > 0) {
12479                         sack_rxmit = 1;
12480                         TCPSTAT_INC(tcps_sack_rexmits);
12481                         TCPSTAT_ADD(tcps_sack_rexmit_bytes,
12482                             min(len, maxseg));
12483                 } else {
12484                         /* I dont think this can happen */
12485                         rsm = NULL;
12486                         goto recheck_resend;
12487                 }
12488                 BBR_STAT_INC(bbr_resends_set);
12489         } else if (bbr->r_ctl.rc_tlp_send) {
12490                 /*
12491                  * Tail loss probe
12492                  */
12493                 doing_tlp = 1;
12494                 rsm = bbr->r_ctl.rc_tlp_send;
12495                 bbr->r_ctl.rc_tlp_send = NULL;
12496                 sack_rxmit = 1;
12497                 len = rsm->r_end - rsm->r_start;
12498                 rtr_cnt++;
12499                 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12500                         len = maxseg;
12501
12502                 if (SEQ_GT(tp->snd_una, rsm->r_start)) {
12503 #ifdef BBR_INVARIANTS
12504                         panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u",
12505                             tp, bbr, tp->snd_una, rsm, rsm->r_start);
12506 #else
12507                         /* TSNH */
12508                         rsm = NULL;
12509                         goto recheck_resend;
12510 #endif
12511                 }
12512                 sb_offset = rsm->r_start - tp->snd_una;
12513                 BBR_STAT_INC(bbr_tlp_set);
12514         }
12515         /*
12516          * Enforce a connection sendmap count limit if set
12517          * as long as we are not retransmiting.
12518          */
12519         if ((rsm == NULL) &&
12520             (bbr_tcp_map_entries_limit > 0) &&
12521             (bbr->r_ctl.rc_num_maps_alloced >= bbr_tcp_map_entries_limit)) {
12522                 BBR_STAT_INC(bbr_alloc_limited);
12523                 if (!bbr->alloc_limit_reported) {
12524                         bbr->alloc_limit_reported = 1;
12525                         BBR_STAT_INC(bbr_alloc_limited_conns);
12526                 }
12527                 goto just_return_nolock;
12528         }
12529 #ifdef BBR_INVARIANTS
12530         if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) {
12531                 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u",
12532                     tp, bbr, rsm, sb_offset, len);
12533         }
12534 #endif
12535         /*
12536          * Get standard flags, and add SYN or FIN if requested by 'hidden'
12537          * state flags.
12538          */
12539         if (tp->t_flags & TF_NEEDFIN && (rsm == NULL))
12540                 flags |= TH_FIN;
12541         if (tp->t_flags & TF_NEEDSYN)
12542                 flags |= TH_SYN;
12543
12544         if (rsm && (rsm->r_flags & BBR_HAS_FIN)) {
12545                 /* we are retransmitting the fin */
12546                 len--;
12547                 if (len) {
12548                         /*
12549                          * When retransmitting data do *not* include the
12550                          * FIN. This could happen from a TLP probe if we
12551                          * allowed data with a FIN.
12552                          */
12553                         flags &= ~TH_FIN;
12554                 }
12555         } else if (rsm) {
12556                 if (flags & TH_FIN)
12557                         flags &= ~TH_FIN;
12558         }
12559         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
12560                 void *end_rsm;
12561
12562                 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
12563                 if (end_rsm)
12564                         kern_prefetch(end_rsm, &prefetch_rsm);
12565                 prefetch_rsm = 1;
12566         }
12567         SOCKBUF_LOCK(sb);
12568         /*
12569          * If in persist timeout with window of 0, send 1 byte. Otherwise,
12570          * if window is small but nonzero and time TF_SENTFIN expired, we
12571          * will send what we can and go to transmit state.
12572          */
12573         if (tp->t_flags & TF_FORCEDATA) {
12574                 if ((sendwin == 0) || (sendwin <= (tp->snd_max - tp->snd_una))) {
12575                         /*
12576                          * If we still have some data to send, then clear
12577                          * the FIN bit.  Usually this would happen below
12578                          * when it realizes that we aren't sending all the
12579                          * data.  However, if we have exactly 1 byte of
12580                          * unsent data, then it won't clear the FIN bit
12581                          * below, and if we are in persist state, we wind up
12582                          * sending the packet without recording that we sent
12583                          * the FIN bit.
12584                          *
12585                          * We can't just blindly clear the FIN bit, because
12586                          * if we don't have any more data to send then the
12587                          * probe will be the FIN itself.
12588                          */
12589                         if (sb_offset < sbused(sb))
12590                                 flags &= ~TH_FIN;
12591                         sendwin = 1;
12592                 } else {
12593                         if ((bbr->rc_in_persist != 0) &&
12594                             (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
12595                                                bbr_minseg(bbr)))) {
12596                                 /* Exit persists if there is space */
12597                                 bbr_exit_persist(tp, bbr, cts, __LINE__);
12598                         }
12599                         if (rsm == NULL) {
12600                                 /*
12601                                  * If we are dropping persist mode then we
12602                                  * need to correct sb_offset if not a
12603                                  * retransmit.
12604                                  */
12605                                 sb_offset = tp->snd_max - tp->snd_una;
12606                         }
12607                 }
12608         }
12609         /*
12610          * If snd_nxt == snd_max and we have transmitted a FIN, the
12611          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
12612          * negative length.  This can also occur when TCP opens up its
12613          * congestion window while receiving additional duplicate acks after
12614          * fast-retransmit because TCP will reset snd_nxt to snd_max after
12615          * the fast-retransmit.
12616          *
12617          * In the normal retransmit-FIN-only case, however, snd_nxt will be
12618          * set to snd_una, the sb_offset will be 0, and the length may wind
12619          * up 0.
12620          *
12621          * If sack_rxmit is true we are retransmitting from the scoreboard
12622          * in which case len is already set.
12623          */
12624         if (sack_rxmit == 0) {
12625                 uint32_t avail;
12626
12627                 avail = sbavail(sb);
12628                 if (SEQ_GT(tp->snd_max, tp->snd_una))
12629                         sb_offset = tp->snd_max - tp->snd_una;
12630                 else
12631                         sb_offset = 0;
12632                 if (bbr->rc_tlp_new_data) {
12633                         /* TLP is forcing out new data */
12634                         uint32_t tlplen;
12635
12636                         doing_tlp = 1;
12637                         tlplen = maxseg;
12638
12639                         if (tlplen > (uint32_t)(avail - sb_offset)) {
12640                                 tlplen = (uint32_t)(avail - sb_offset);
12641                         }
12642                         if (tlplen > tp->snd_wnd) {
12643                                 len = tp->snd_wnd;
12644                         } else {
12645                                 len = tlplen;
12646                         }
12647                         bbr->rc_tlp_new_data = 0;
12648                 } else {
12649                         what_we_can = len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts);
12650                         if ((len < p_maxseg) &&
12651                             (bbr->rc_in_persist == 0) &&
12652                             (ctf_outstanding(tp) >= (2 * p_maxseg)) &&
12653                             ((avail - sb_offset) >= p_maxseg)) {
12654                                 /*
12655                                  * We are not completing whats in the socket
12656                                  * buffer (i.e. there is at least a segment
12657                                  * waiting to send) and we have 2 or more
12658                                  * segments outstanding. There is no sense
12659                                  * of sending a little piece. Lets defer and
12660                                  * and wait until we can send a whole
12661                                  * segment.
12662                                  */
12663                                 len = 0;
12664                         }
12665                         if ((tp->t_flags & TF_FORCEDATA) && (bbr->rc_in_persist)) {
12666                                 /*
12667                                  * We are in persists, figure out if
12668                                  * a retransmit is available (maybe the previous
12669                                  * persists we sent) or if we have to send new
12670                                  * data.
12671                                  */
12672                                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
12673                                 if (rsm) {
12674                                         len = rsm->r_end - rsm->r_start;
12675                                         if (rsm->r_flags & BBR_HAS_FIN)
12676                                                 len--;
12677                                         if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12678                                                 len = maxseg;
12679                                         if (len > 1)
12680                                                 BBR_STAT_INC(bbr_persist_reneg);
12681                                         /*
12682                                          * XXXrrs we could force the len to
12683                                          * 1 byte here to cause the chunk to
12684                                          * split apart.. but that would then
12685                                          * mean we always retransmit it as
12686                                          * one byte even after the window
12687                                          * opens.
12688                                          */
12689                                         sack_rxmit = 1;
12690                                         sb_offset = rsm->r_start - tp->snd_una;
12691                                 } else {
12692                                         /*
12693                                          * First time through in persists or peer
12694                                          * acked our one byte. Though we do have
12695                                          * to have something in the sb.
12696                                          */
12697                                         len = 1;
12698                                         sb_offset = 0;
12699                                         if (avail == 0)
12700                                             len = 0;
12701                                 }
12702                         }
12703                 }
12704         }
12705         if (prefetch_so_done == 0) {
12706                 kern_prefetch(so, &prefetch_so_done);
12707                 prefetch_so_done = 1;
12708         }
12709         /*
12710          * Lop off SYN bit if it has already been sent.  However, if this is
12711          * SYN-SENT state and if segment contains data and if we don't know
12712          * that foreign host supports TAO, suppress sending segment.
12713          */
12714         if ((flags & TH_SYN) && (rsm == NULL) &&
12715             SEQ_GT(tp->snd_max, tp->snd_una)) {
12716                 if (tp->t_state != TCPS_SYN_RECEIVED)
12717                         flags &= ~TH_SYN;
12718                 /*
12719                  * When sending additional segments following a TFO SYN|ACK,
12720                  * do not include the SYN bit.
12721                  */
12722                 if (IS_FASTOPEN(tp->t_flags) &&
12723                     (tp->t_state == TCPS_SYN_RECEIVED))
12724                         flags &= ~TH_SYN;
12725                 sb_offset--, len++;
12726                 if (sbavail(sb) == 0)
12727                         len = 0;
12728         } else if ((flags & TH_SYN) && rsm) {
12729                 /*
12730                  * Subtract one from the len for the SYN being
12731                  * retransmitted.
12732                  */
12733                 len--;
12734         }
12735         /*
12736          * Be careful not to send data and/or FIN on SYN segments. This
12737          * measure is needed to prevent interoperability problems with not
12738          * fully conformant TCP implementations.
12739          */
12740         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
12741                 len = 0;
12742                 flags &= ~TH_FIN;
12743         }
12744         /*
12745          * On TFO sockets, ensure no data is sent in the following cases:
12746          *
12747          *  - When retransmitting SYN|ACK on a passively-created socket
12748          *  - When retransmitting SYN on an actively created socket
12749          *  - When sending a zero-length cookie (cookie request) on an
12750          *    actively created socket
12751          *  - When the socket is in the CLOSED state (RST is being sent)
12752          */
12753         if (IS_FASTOPEN(tp->t_flags) &&
12754             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
12755              ((tp->t_state == TCPS_SYN_SENT) &&
12756               (tp->t_tfo_client_cookie_len == 0)) ||
12757              (flags & TH_RST))) {
12758                 len = 0;
12759                 sack_rxmit = 0;
12760                 rsm = NULL;
12761         }
12762         /* Without fast-open there should never be data sent on a SYN */
12763         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags)))
12764                 len = 0;
12765         if (len <= 0) {
12766                 /*
12767                  * If FIN has been sent but not acked, but we haven't been
12768                  * called to retransmit, len will be < 0.  Otherwise, window
12769                  * shrank after we sent into it.  If window shrank to 0,
12770                  * cancel pending retransmit, pull snd_nxt back to (closed)
12771                  * window, and set the persist timer if it isn't already
12772                  * going.  If the window didn't close completely, just wait
12773                  * for an ACK.
12774                  *
12775                  * We also do a general check here to ensure that we will
12776                  * set the persist timer when we have data to send, but a
12777                  * 0-byte window. This makes sure the persist timer is set
12778                  * even if the packet hits one of the "goto send" lines
12779                  * below.
12780                  */
12781                 len = 0;
12782                 if ((tp->snd_wnd == 0) &&
12783                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12784                     (tp->snd_una == tp->snd_max) &&
12785                     (sb_offset < (int)sbavail(sb))) {
12786                         /*
12787                          * Not enough room in the rwnd to send
12788                          * a paced segment out.
12789                          */
12790                         bbr_enter_persist(tp, bbr, cts, __LINE__);
12791                 }
12792         } else if ((rsm == NULL) &&
12793                    (doing_tlp == 0) &&
12794                    (len < bbr->r_ctl.rc_pace_max_segs)) {
12795                 /*
12796                  * We are not sending a full segment for
12797                  * some reason. Should we not send anything (think
12798                  * sws or persists)?
12799                  */
12800                 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12801                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12802                     (len < (int)(sbavail(sb) - sb_offset))) {
12803                         /*
12804                          * Here the rwnd is less than
12805                          * the pacing size, this is not a retransmit,
12806                          * we are established and
12807                          * the send is not the last in the socket buffer
12808                          * lets not send, and possibly enter persists.
12809                          */
12810                         len = 0;
12811                         if (tp->snd_max == tp->snd_una)
12812                                 bbr_enter_persist(tp, bbr, cts, __LINE__);
12813                 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) &&
12814                            (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12815                                                  bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12816                            (len < (int)(sbavail(sb) - sb_offset)) &&
12817                            (len < bbr_minseg(bbr))) {
12818                         /*
12819                          * Here we are not retransmitting, and
12820                          * the cwnd is not so small that we could
12821                          * not send at least a min size (rxt timer
12822                          * not having gone off), We have 2 segments or
12823                          * more already in flight, its not the tail end
12824                          * of the socket buffer  and the cwnd is blocking
12825                          * us from sending out minimum pacing segment size.
12826                          * Lets not send anything.
12827                          */
12828                         bbr->rc_cwnd_limited = 1;
12829                         len = 0;
12830                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
12831                             min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12832                            (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12833                                                  bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12834                            (len < (int)(sbavail(sb) - sb_offset)) &&
12835                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
12836                         /*
12837                          * Here we have a send window but we have
12838                          * filled it up and we can't send another pacing segment.
12839                          * We also have in flight more than 2 segments
12840                          * and we are not completing the sb i.e. we allow
12841                          * the last bytes of the sb to go out even if
12842                          * its not a full pacing segment.
12843                          */
12844                         len = 0;
12845                 }
12846         }
12847         /* len will be >= 0 after this point. */
12848         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
12849         tcp_sndbuf_autoscale(tp, so, sendwin);
12850         /*
12851          *
12852          */
12853         if (bbr->rc_in_persist &&
12854             len &&
12855             (rsm == NULL) &&
12856             (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) {
12857                 /*
12858                  * We are in persist, not doing a retransmit and don't have enough space
12859                  * yet to send a full TSO. So is it at the end of the sb
12860                  * if so we need to send else nuke to 0 and don't send.
12861                  */
12862                 int sbleft;
12863                 if (sbavail(sb) > sb_offset)
12864                         sbleft = sbavail(sb) - sb_offset;
12865                 else
12866                         sbleft = 0;
12867                 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) {
12868                         /* not at end of sb lets not send */
12869                         len = 0;
12870                 }
12871         }
12872         /*
12873          * Decide if we can use TCP Segmentation Offloading (if supported by
12874          * hardware).
12875          *
12876          * TSO may only be used if we are in a pure bulk sending state.  The
12877          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
12878          * options prevent using TSO.  With TSO the TCP header is the same
12879          * (except for the sequence number) for all generated packets.  This
12880          * makes it impossible to transmit any options which vary per
12881          * generated segment or packet.
12882          *
12883          * IPv4 handling has a clear separation of ip options and ip header
12884          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen()
12885          * does the right thing below to provide length of just ip options
12886          * and thus checking for ipoptlen is enough to decide if ip options
12887          * are present.
12888          */
12889 #ifdef INET6
12890         if (isipv6)
12891                 ipoptlen = ip6_optlen(inp);
12892         else
12893 #endif
12894         if (inp->inp_options)
12895                 ipoptlen = inp->inp_options->m_len -
12896                     offsetof(struct ipoption, ipopt_list);
12897         else
12898                 ipoptlen = 0;
12899 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12900         /*
12901          * Pre-calculate here as we save another lookup into the darknesses
12902          * of IPsec that way and can actually decide if TSO is ok.
12903          */
12904 #ifdef INET6
12905         if (isipv6 && IPSEC_ENABLED(ipv6))
12906                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
12907 #ifdef INET
12908         else
12909 #endif
12910 #endif                          /* INET6 */
12911 #ifdef INET
12912         if (IPSEC_ENABLED(ipv4))
12913                 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
12914 #endif                          /* INET */
12915 #endif                          /* IPSEC */
12916 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12917         ipoptlen += ipsec_optlen;
12918 #endif
12919         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
12920             (len > maxseg) &&
12921             (tp->t_port == 0) &&
12922             ((tp->t_flags & TF_SIGNATURE) == 0) &&
12923             tp->rcv_numsacks == 0 &&
12924             ipoptlen == 0)
12925                 tso = 1;
12926
12927         recwin = min(max(sbspace(&so->so_rcv), 0),
12928             TCP_MAXWIN << tp->rcv_scale);
12929         /*
12930          * Sender silly window avoidance.   We transmit under the following
12931          * conditions when len is non-zero:
12932          *
12933          * - We have a full segment (or more with TSO) - This is the last
12934          * buffer in a write()/send() and we are either idle or running
12935          * NODELAY - we've timed out (e.g. persist timer) - we have more
12936          * then 1/2 the maximum send window's worth of data (receiver may be
12937          * limited the window size) - we need to retransmit
12938          */
12939         if (rsm)
12940                 goto send;
12941         if (len) {
12942                 if (sack_rxmit)
12943                         goto send;
12944                 if (len >= p_maxseg)
12945                         goto send;
12946                 /*
12947                  * NOTE! on localhost connections an 'ack' from the remote
12948                  * end may occur synchronously with the output and cause us
12949                  * to flush a buffer queued with moretocome.  XXX
12950                  *
12951                  */
12952                 if (((tp->t_flags & TF_MORETOCOME) == 0) &&     /* normal case */
12953                     ((tp->t_flags & TF_NODELAY) ||
12954                     ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) &&
12955                     (tp->t_flags & TF_NOPUSH) == 0) {
12956                         goto send;
12957                 }
12958                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
12959                         goto send;
12960                 }
12961                 if (tp->t_flags & TF_FORCEDATA) {       /* typ. timeout case */
12962                         goto send;
12963                 }
12964                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
12965                         goto send;
12966                 }
12967         }
12968         /*
12969          * Sending of standalone window updates.
12970          *
12971          * Window updates are important when we close our window due to a
12972          * full socket buffer and are opening it again after the application
12973          * reads data from it.  Once the window has opened again and the
12974          * remote end starts to send again the ACK clock takes over and
12975          * provides the most current window information.
12976          *
12977          * We must avoid the silly window syndrome whereas every read from
12978          * the receive buffer, no matter how small, causes a window update
12979          * to be sent.  We also should avoid sending a flurry of window
12980          * updates when the socket buffer had queued a lot of data and the
12981          * application is doing small reads.
12982          *
12983          * Prevent a flurry of pointless window updates by only sending an
12984          * update when we can increase the advertized window by more than
12985          * 1/4th of the socket buffer capacity.  When the buffer is getting
12986          * full or is very small be more aggressive and send an update
12987          * whenever we can increase by two mss sized segments. In all other
12988          * situations the ACK's to new incoming data will carry further
12989          * window increases.
12990          *
12991          * Don't send an independent window update if a delayed ACK is
12992          * pending (it will get piggy-backed on it) or the remote side
12993          * already has done a half-close and won't send more data.  Skip
12994          * this if the connection is in T/TCP half-open state.
12995          */
12996         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
12997             !(tp->t_flags & TF_DELACK) &&
12998             !TCPS_HAVERCVDFIN(tp->t_state)) {
12999                 /* Check to see if we should do a window update */
13000                 if (bbr_window_update_needed(tp, so, recwin, maxseg))
13001                         goto send;
13002         }
13003         /*
13004          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
13005          * is also a catch-all for the retransmit timer timeout case.
13006          */
13007         if (tp->t_flags & TF_ACKNOW) {
13008                 goto send;
13009         }
13010         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
13011                 goto send;
13012         }
13013         if (SEQ_GT(tp->snd_up, tp->snd_una)) {
13014                 goto send;
13015         }
13016         /*
13017          * If our state indicates that FIN should be sent and we have not
13018          * yet done so, then we need to send.
13019          */
13020         if (flags & TH_FIN &&
13021             ((tp->t_flags & TF_SENTFIN) == 0)) {
13022                 goto send;
13023         }
13024         /*
13025          * No reason to send a segment, just return.
13026          */
13027 just_return:
13028         SOCKBUF_UNLOCK(sb);
13029 just_return_nolock:
13030         if (tot_len)
13031                 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
13032         if (bbr->rc_no_pacing)
13033                 slot = 0;
13034         if (tot_len == 0) {
13035                 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >=
13036                     tp->snd_wnd) {
13037                         BBR_STAT_INC(bbr_rwnd_limited);
13038                         app_limited = BBR_JR_RWND_LIMITED;
13039                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
13040                         if ((bbr->rc_in_persist == 0) &&
13041                             TCPS_HAVEESTABLISHED(tp->t_state) &&
13042                             (tp->snd_max == tp->snd_una) &&
13043                             sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
13044                                 /* No send window.. we must enter persist */
13045                                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
13046                         }
13047                 } else if (ctf_outstanding(tp) >= sbavail(sb)) {
13048                         BBR_STAT_INC(bbr_app_limited);
13049                         app_limited = BBR_JR_APP_LIMITED;
13050                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
13051                 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13052                                                  bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) {
13053                         BBR_STAT_INC(bbr_cwnd_limited);
13054                         app_limited = BBR_JR_CWND_LIMITED;
13055                         bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13056                                                                         bbr->r_ctl.rc_lost_bytes)));
13057                         bbr->rc_cwnd_limited = 1;
13058                 } else {
13059                         BBR_STAT_INC(bbr_app_limited);
13060                         app_limited = BBR_JR_APP_LIMITED;
13061                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
13062                 }
13063                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
13064                 bbr->r_agg_early_set = 0;
13065                 bbr->r_ctl.rc_agg_early = 0;
13066                 bbr->r_ctl.rc_last_delay_val = 0;
13067         } else if (bbr->rc_use_google == 0)
13068                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
13069         /* Are we app limited? */
13070         if ((app_limited == BBR_JR_APP_LIMITED) ||
13071             (app_limited == BBR_JR_RWND_LIMITED)) {
13072                 /**
13073                  * We are application limited.
13074                  */
13075                 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13076                                                                        bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered);
13077         }
13078         if (tot_len == 0)
13079                 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1);
13080         tp->t_flags &= ~TF_FORCEDATA;
13081         /* Dont update the time if we did not send */
13082         bbr->r_ctl.rc_last_delay_val = 0;
13083         bbr->rc_output_starts_timer = 1;
13084         bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len);
13085         bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len);
13086         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
13087                 /* Make sure snd_nxt is drug up */
13088                 tp->snd_nxt = tp->snd_max;
13089         }
13090         return (error);
13091
13092 send:
13093         if (doing_tlp == 0) {
13094                 /*
13095                  * Data not a TLP, and its not the rxt firing. If it is the
13096                  * rxt firing, we want to leave the tlp_in_progress flag on
13097                  * so we don't send another TLP. It has to be a rack timer
13098                  * or normal send (response to acked data) to clear the tlp
13099                  * in progress flag.
13100                  */
13101                 bbr->rc_tlp_in_progress = 0;
13102                 bbr->rc_tlp_rtx_out = 0;
13103         } else {
13104                 /*
13105                  * Its a TLP.
13106                  */
13107                 bbr->rc_tlp_in_progress = 1;
13108         }
13109         bbr_timer_cancel(bbr, __LINE__, cts);
13110         if (rsm == NULL) {
13111                 if (sbused(sb) > 0) {
13112                         /*
13113                          * This is sub-optimal. We only send a stand alone
13114                          * FIN on its own segment.
13115                          */
13116                         if (flags & TH_FIN) {
13117                                 flags &= ~TH_FIN;
13118                                 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) {
13119                                         /* Lets not send this */
13120                                         slot = 0;
13121                                         goto just_return;
13122                                 }
13123                         }
13124                 }
13125         } else {
13126                 /*
13127                  * We do *not* send a FIN on a retransmit if it has data.
13128                  * The if clause here where len > 1 should never come true.
13129                  */
13130                 if ((len > 0) &&
13131                     (((rsm->r_flags & BBR_HAS_FIN) == 0) &&
13132                     (flags & TH_FIN))) {
13133                         flags &= ~TH_FIN;
13134                         len--;
13135                 }
13136         }
13137         SOCKBUF_LOCK_ASSERT(sb);
13138         if (len > 0) {
13139                 if ((tp->snd_una == tp->snd_max) &&
13140                     (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
13141                         /*
13142                          * This qualifies as a RTT_PROBE session since we
13143                          * drop the data outstanding to nothing and waited
13144                          * more than bbr_rtt_probe_time.
13145                          */
13146                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
13147                         bbr_set_reduced_rtt(bbr, cts, __LINE__);
13148                 }
13149                 if (len >= maxseg)
13150                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
13151                 else
13152                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
13153         }
13154         /*
13155          * Before ESTABLISHED, force sending of initial options unless TCP
13156          * set not to do any options. NOTE: we assume that the IP/TCP header
13157          * plus TCP options always fit in a single mbuf, leaving room for a
13158          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
13159          * + optlen <= MCLBYTES
13160          */
13161         optlen = 0;
13162 #ifdef INET6
13163         if (isipv6)
13164                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
13165         else
13166 #endif
13167                 hdrlen = sizeof(struct tcpiphdr);
13168
13169         /*
13170          * Compute options for segment. We only have to care about SYN and
13171          * established connection segments.  Options for SYN-ACK segments
13172          * are handled in TCP syncache.
13173          */
13174         to.to_flags = 0;
13175         local_options = 0;
13176         if ((tp->t_flags & TF_NOOPT) == 0) {
13177                 /* Maximum segment size. */
13178                 if (flags & TH_SYN) {
13179                         to.to_mss = tcp_mssopt(&inp->inp_inc);
13180 #ifdef NETFLIX_TCPOUDP
13181                         if (tp->t_port)
13182                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
13183 #endif
13184                         to.to_flags |= TOF_MSS;
13185                         /*
13186                          * On SYN or SYN|ACK transmits on TFO connections,
13187                          * only include the TFO option if it is not a
13188                          * retransmit, as the presence of the TFO option may
13189                          * have caused the original SYN or SYN|ACK to have
13190                          * been dropped by a middlebox.
13191                          */
13192                         if (IS_FASTOPEN(tp->t_flags) &&
13193                             (tp->t_rxtshift == 0)) {
13194                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
13195                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
13196                                         to.to_tfo_cookie =
13197                                             (u_int8_t *)&tp->t_tfo_cookie.server;
13198                                         to.to_flags |= TOF_FASTOPEN;
13199                                         wanted_cookie = 1;
13200                                 } else if (tp->t_state == TCPS_SYN_SENT) {
13201                                         to.to_tfo_len =
13202                                             tp->t_tfo_client_cookie_len;
13203                                         to.to_tfo_cookie =
13204                                             tp->t_tfo_cookie.client;
13205                                         to.to_flags |= TOF_FASTOPEN;
13206                                         wanted_cookie = 1;
13207                                 }
13208                         }
13209                 }
13210                 /* Window scaling. */
13211                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
13212                         to.to_wscale = tp->request_r_scale;
13213                         to.to_flags |= TOF_SCALE;
13214                 }
13215                 /* Timestamps. */
13216                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
13217                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
13218                         to.to_tsval =   tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset;
13219                         to.to_tsecr = tp->ts_recent;
13220                         to.to_flags |= TOF_TS;
13221                         local_options += TCPOLEN_TIMESTAMP + 2;
13222                 }
13223                 /* Set receive buffer autosizing timestamp. */
13224                 if (tp->rfbuf_ts == 0 &&
13225                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
13226                         tp->rfbuf_ts =  tcp_tv_to_mssectick(&bbr->rc_tv);
13227                 /* Selective ACK's. */
13228                 if (flags & TH_SYN)
13229                         to.to_flags |= TOF_SACKPERM;
13230                 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13231                     tp->rcv_numsacks > 0) {
13232                         to.to_flags |= TOF_SACK;
13233                         to.to_nsacks = tp->rcv_numsacks;
13234                         to.to_sacks = (u_char *)tp->sackblks;
13235                 }
13236 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13237                 /* TCP-MD5 (RFC2385). */
13238                 if (tp->t_flags & TF_SIGNATURE)
13239                         to.to_flags |= TOF_SIGNATURE;
13240 #endif                          /* TCP_SIGNATURE */
13241
13242                 /* Processing the options. */
13243                 hdrlen += (optlen = tcp_addoptions(&to, opt));
13244                 /*
13245                  * If we wanted a TFO option to be added, but it was unable
13246                  * to fit, ensure no data is sent.
13247                  */
13248                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
13249                     !(to.to_flags & TOF_FASTOPEN))
13250                         len = 0;
13251         }
13252 #ifdef NETFLIX_TCPOUDP
13253         if (tp->t_port) {
13254                 if (V_tcp_udp_tunneling_port == 0) {
13255                         /* The port was removed?? */
13256                         SOCKBUF_UNLOCK(&so->so_snd);
13257                         return (EHOSTUNREACH);
13258                 }
13259
13260                 hdrlen += sizeof(struct udphdr);
13261         }
13262 #endif
13263 #ifdef INET6
13264         if (isipv6)
13265                 ipoptlen = ip6_optlen(tp->t_inpcb);
13266         else
13267 #endif
13268         if (tp->t_inpcb->inp_options)
13269                 ipoptlen = tp->t_inpcb->inp_options->m_len -
13270                     offsetof(struct ipoption, ipopt_list);
13271         else
13272                 ipoptlen = 0;
13273         ipoptlen = 0;
13274 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
13275         ipoptlen += ipsec_optlen;
13276 #endif
13277         if (bbr->rc_last_options != local_options) {
13278                 /*
13279                  * Cache the options length this generally does not change
13280                  * on a connection. We use this to calculate TSO.
13281                  */
13282                 bbr->rc_last_options = local_options;
13283         }
13284         maxseg = tp->t_maxseg - (ipoptlen + optlen);
13285         p_maxseg = min(maxseg, pace_max_segs);
13286         /*
13287          * Adjust data length if insertion of options will bump the packet
13288          * length beyond the t_maxseg length. Clear the FIN bit because we
13289          * cut off the tail of the segment.
13290          */
13291 #ifdef KERN_TLS
13292         /* force TSO for so TLS offload can get mss */
13293         if (sb->sb_flags & SB_TLS_IFNET) {
13294                 force_tso = 1;
13295         }
13296 #endif
13297
13298         if (len > maxseg) {
13299                 if (len != 0 && (flags & TH_FIN)) {
13300                         flags &= ~TH_FIN;
13301                 }
13302                 if (tso) {
13303                         uint32_t moff;
13304                         int32_t max_len;
13305
13306                         /* extract TSO information */
13307                         if_hw_tsomax = tp->t_tsomax;
13308                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
13309                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
13310                         KASSERT(ipoptlen == 0,
13311                             ("%s: TSO can't do IP options", __func__));
13312
13313                         /*
13314                          * Check if we should limit by maximum payload
13315                          * length:
13316                          */
13317                         if (if_hw_tsomax != 0) {
13318                                 /* compute maximum TSO length */
13319                                 max_len = (if_hw_tsomax - hdrlen -
13320                                     max_linkhdr);
13321                                 if (max_len <= 0) {
13322                                         len = 0;
13323                                 } else if (len > max_len) {
13324                                         len = max_len;
13325                                 }
13326                         }
13327                         /*
13328                          * Prevent the last segment from being fractional
13329                          * unless the send sockbuf can be emptied:
13330                          */
13331                         if (((sb_offset + len) < sbavail(sb)) &&
13332                             (hw_tls == 0)) {
13333                                 moff = len % (uint32_t)maxseg;
13334                                 if (moff != 0) {
13335                                         len -= moff;
13336                                 }
13337                         }
13338                         /*
13339                          * In case there are too many small fragments don't
13340                          * use TSO:
13341                          */
13342                         if (len <= maxseg) {
13343                                 len = maxseg;
13344                                 tso = 0;
13345                         }
13346                 } else {
13347                         /* Not doing TSO */
13348                         if (optlen + ipoptlen >= tp->t_maxseg) {
13349                                 /*
13350                                  * Since we don't have enough space to put
13351                                  * the IP header chain and the TCP header in
13352                                  * one packet as required by RFC 7112, don't
13353                                  * send it. Also ensure that at least one
13354                                  * byte of the payload can be put into the
13355                                  * TCP segment.
13356                                  */
13357                                 SOCKBUF_UNLOCK(&so->so_snd);
13358                                 error = EMSGSIZE;
13359                                 sack_rxmit = 0;
13360                                 goto out;
13361                         }
13362                         len = maxseg;
13363                 }
13364         } else {
13365                 /* Not doing TSO */
13366                 if_hw_tsomaxsegcount = 0;
13367                 tso = 0;
13368         }
13369         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
13370             ("%s: len > IP_MAXPACKET", __func__));
13371 #ifdef DIAGNOSTIC
13372 #ifdef INET6
13373         if (max_linkhdr + hdrlen > MCLBYTES)
13374 #else
13375         if (max_linkhdr + hdrlen > MHLEN)
13376 #endif
13377                 panic("tcphdr too big");
13378 #endif
13379         /*
13380          * This KASSERT is here to catch edge cases at a well defined place.
13381          * Before, those had triggered (random) panic conditions further
13382          * down.
13383          */
13384 #ifdef BBR_INVARIANTS
13385         if (sack_rxmit) {
13386                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
13387                         panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u",
13388                             rsm, tp, bbr, rsm->r_start, tp->snd_una);
13389                 }
13390         }
13391 #endif
13392         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
13393         if ((len == 0) &&
13394             (flags & TH_FIN) &&
13395             (sbused(sb))) {
13396                 /*
13397                  * We have outstanding data, don't send a fin by itself!.
13398                  */
13399                 slot = 0;
13400                 goto just_return;
13401         }
13402         /*
13403          * Grab a header mbuf, attaching a copy of data to be transmitted,
13404          * and initialize the header from the template for sends on this
13405          * connection.
13406          */
13407         if (len) {
13408                 uint32_t moff;
13409                 uint32_t orig_len;
13410
13411                 /*
13412                  * We place a limit on sending with hptsi.
13413                  */
13414                 if ((rsm == NULL) && len > pace_max_segs)
13415                         len = pace_max_segs;
13416                 if (len <= maxseg)
13417                         tso = 0;
13418 #ifdef INET6
13419                 if (MHLEN < hdrlen + max_linkhdr)
13420                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
13421                 else
13422 #endif
13423                         m = m_gethdr(M_NOWAIT, MT_DATA);
13424
13425                 if (m == NULL) {
13426                         BBR_STAT_INC(bbr_failed_mbuf_aloc);
13427                         bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13428                         SOCKBUF_UNLOCK(sb);
13429                         error = ENOBUFS;
13430                         sack_rxmit = 0;
13431                         goto out;
13432                 }
13433                 m->m_data += max_linkhdr;
13434                 m->m_len = hdrlen;
13435                 /*
13436                  * Start the m_copy functions from the closest mbuf to the
13437                  * sb_offset in the socket buffer chain.
13438                  */
13439                 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) {
13440 #ifdef BBR_INVARIANTS
13441                         if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0)))
13442                                 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u",
13443                                     tp, bbr, len, sb_offset, sbavail(sb), rsm,
13444                                     doing_retran_from,
13445                                     picked_up_retran,
13446                                     doing_tlp);
13447
13448 #endif
13449                         /*
13450                          * In this messed up situation we have two choices,
13451                          * a) pretend the send worked, and just start timers
13452                          * and what not (not good since that may lead us
13453                          * back here a lot). <or> b) Send the lowest segment
13454                          * in the map. <or> c) Drop the connection. Lets do
13455                          * <b> which if it continues to happen will lead to
13456                          * <c> via timeouts.
13457                          */
13458                         BBR_STAT_INC(bbr_offset_recovery);
13459                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
13460                         sb_offset = 0;
13461                         if (rsm == NULL) {
13462                                 sack_rxmit = 0;
13463                                 len = sbavail(sb);
13464                         } else {
13465                                 sack_rxmit = 1;
13466                                 if (rsm->r_start != tp->snd_una) {
13467                                         /*
13468                                          * Things are really messed up, <c>
13469                                          * is the only thing to do.
13470                                          */
13471                                         BBR_STAT_INC(bbr_offset_drop);
13472                                         tcp_set_inp_to_drop(inp, EFAULT);
13473                                         return (0);
13474                                 }
13475                                 len = rsm->r_end - rsm->r_start;
13476                         }
13477                         if (len > sbavail(sb))
13478                                 len = sbavail(sb);
13479                         if (len > maxseg)
13480                                 len = maxseg;
13481                 }
13482                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
13483                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
13484                         m_copydata(mb, moff, (int)len,
13485                             mtod(m, caddr_t)+hdrlen);
13486                         if (rsm == NULL)
13487                                 sbsndptr_adv(sb, mb, len);
13488                         m->m_len += len;
13489                 } else {
13490                         struct sockbuf *msb;
13491
13492                         if (rsm)
13493                                 msb = NULL;
13494                         else
13495                                 msb = sb;
13496 #ifdef BBR_INVARIANTS
13497                         if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) {
13498                                 if (rsm) {
13499                                         panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u rsm:%p snd_una:%u rsm_start:%u flg:%x %u:%u:%u sr:%d ",
13500                                             tp, bbr, len, moff,
13501                                             sbavail(sb), rsm,
13502                                             tp->snd_una, rsm->r_flags, rsm->r_start,
13503                                             doing_retran_from,
13504                                             picked_up_retran,
13505                                             doing_tlp, sack_rxmit);
13506                                 } else {
13507                                         panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u",
13508                                             tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una);
13509                                 }
13510                         }
13511 #endif
13512                         orig_len = len;
13513                         m->m_next = tcp_m_copym(
13514 #ifdef NETFLIX_COPY_ARGS
13515                                 tp,
13516 #endif
13517                                 mb, moff, &len,
13518                                 if_hw_tsomaxsegcount,
13519                                 if_hw_tsomaxsegsize, msb,
13520                                 ((rsm == NULL) ? hw_tls : 0)
13521 #ifdef NETFLIX_COPY_ARGS
13522                                 , &filled_all
13523 #endif
13524                                 );
13525                         if (len <= maxseg && !force_tso) {
13526                                 /*
13527                                  * Must have ran out of mbufs for the copy
13528                                  * shorten it to no longer need tso. Lets
13529                                  * not put on sendalot since we are low on
13530                                  * mbufs.
13531                                  */
13532                                 tso = 0;
13533                         }
13534                         if (m->m_next == NULL) {
13535                                 SOCKBUF_UNLOCK(sb);
13536                                 (void)m_free(m);
13537                                 error = ENOBUFS;
13538                                 sack_rxmit = 0;
13539                                 goto out;
13540                         }
13541                 }
13542 #ifdef BBR_INVARIANTS
13543                 if (tso && len < maxseg) {
13544                         panic("tp:%p tso on, but len:%d < maxseg:%d",
13545                             tp, len, maxseg);
13546                 }
13547                 if (tso && if_hw_tsomaxsegcount) {
13548                         int32_t seg_cnt = 0;
13549                         struct mbuf *foo;
13550
13551                         foo = m;
13552                         while (foo) {
13553                                 seg_cnt++;
13554                                 foo = foo->m_next;
13555                         }
13556                         if (seg_cnt > if_hw_tsomaxsegcount) {
13557                                 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount);
13558                         }
13559                 }
13560 #endif
13561                 /*
13562                  * If we're sending everything we've got, set PUSH. (This
13563                  * will keep happy those implementations which only give
13564                  * data to the user when a buffer fills or a PUSH comes in.)
13565                  */
13566                 if (sb_offset + len == sbused(sb) &&
13567                     sbused(sb) &&
13568                     !(flags & TH_SYN)) {
13569                         flags |= TH_PUSH;
13570                 }
13571                 SOCKBUF_UNLOCK(sb);
13572         } else {
13573                 SOCKBUF_UNLOCK(sb);
13574                 if (tp->t_flags & TF_ACKNOW)
13575                         TCPSTAT_INC(tcps_sndacks);
13576                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
13577                         TCPSTAT_INC(tcps_sndctrl);
13578                 else if (SEQ_GT(tp->snd_up, tp->snd_una))
13579                         TCPSTAT_INC(tcps_sndurg);
13580                 else
13581                         TCPSTAT_INC(tcps_sndwinup);
13582
13583                 m = m_gethdr(M_NOWAIT, MT_DATA);
13584                 if (m == NULL) {
13585                         BBR_STAT_INC(bbr_failed_mbuf_aloc);
13586                         bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13587                         error = ENOBUFS;
13588                         /* Fudge the send time since we could not send */
13589                         sack_rxmit = 0;
13590                         goto out;
13591                 }
13592 #ifdef INET6
13593                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
13594                     MHLEN >= hdrlen) {
13595                         M_ALIGN(m, hdrlen);
13596                 } else
13597 #endif
13598                         m->m_data += max_linkhdr;
13599                 m->m_len = hdrlen;
13600         }
13601         SOCKBUF_UNLOCK_ASSERT(sb);
13602         m->m_pkthdr.rcvif = (struct ifnet *)0;
13603 #ifdef MAC
13604         mac_inpcb_create_mbuf(inp, m);
13605 #endif
13606 #ifdef INET6
13607         if (isipv6) {
13608                 ip6 = mtod(m, struct ip6_hdr *);
13609 #ifdef NETFLIX_TCPOUDP
13610                 if (tp->t_port) {
13611                         udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr));
13612                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13613                         udp->uh_dport = tp->t_port;
13614                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
13615                         udp->uh_ulen = htons(ulen);
13616                         th = (struct tcphdr *)(udp + 1);
13617                 } else {
13618 #endif
13619                         th = (struct tcphdr *)(ip6 + 1);
13620
13621 #ifdef NETFLIX_TCPOUDP
13622                 }
13623 #endif
13624                 tcpip_fillheaders(inp,
13625 #ifdef NETFLIX_TCPOUDP
13626                                   tp->t_port,
13627 #endif
13628                                   ip6, th);
13629         } else
13630 #endif                          /* INET6 */
13631         {
13632                 ip = mtod(m, struct ip *);
13633 #ifdef TCPDEBUG
13634                 ipov = (struct ipovly *)ip;
13635 #endif
13636 #ifdef NETFLIX_TCPOUDP
13637                 if (tp->t_port) {
13638                         udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip));
13639                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13640                         udp->uh_dport = tp->t_port;
13641                         ulen = hdrlen + len - sizeof(struct ip);
13642                         udp->uh_ulen = htons(ulen);
13643                         th = (struct tcphdr *)(udp + 1);
13644                 } else
13645 #endif
13646                         th = (struct tcphdr *)(ip + 1);
13647                 tcpip_fillheaders(inp,
13648 #ifdef NETFLIX_TCPOUDP
13649                                   tp->t_port,
13650 #endif
13651                                   ip, th);
13652         }
13653         /*
13654          * If we are doing retransmissions, then snd_nxt will not reflect
13655          * the first unsent octet.  For ACK only packets, we do not want the
13656          * sequence number of the retransmitted packet, we want the sequence
13657          * number of the next unsent octet.  So, if there is no data (and no
13658          * SYN or FIN), use snd_max instead of snd_nxt when filling in
13659          * ti_seq.  But if we are in persist state, snd_max might reflect
13660          * one byte beyond the right edge of the window, so use snd_nxt in
13661          * that case, since we know we aren't doing a retransmission.
13662          * (retransmit and persist are mutually exclusive...)
13663          */
13664         if (sack_rxmit == 0) {
13665                 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) {
13666                         /* New data (including new persists) */
13667                         th->th_seq = htonl(tp->snd_max);
13668                         bbr_seq = tp->snd_max;
13669                 } else if (flags & TH_SYN) {
13670                         /* Syn's always send from iss */
13671                         th->th_seq = htonl(tp->iss);
13672                         bbr_seq = tp->iss;
13673                 } else if (flags & TH_FIN) {
13674                         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) {
13675                                 /*
13676                                  * If we sent the fin already its 1 minus
13677                                  * snd_max
13678                                  */
13679                                 th->th_seq = (htonl(tp->snd_max - 1));
13680                                 bbr_seq = (tp->snd_max - 1);
13681                         } else {
13682                                 /* First time FIN use snd_max */
13683                                 th->th_seq = htonl(tp->snd_max);
13684                                 bbr_seq = tp->snd_max;
13685                         }
13686                 } else if (flags & TH_RST) {
13687                         /*
13688                          * For a Reset send the last cum ack in sequence
13689                          * (this like any other choice may still generate a
13690                          * challenge ack, if a ack-update packet is in
13691                          * flight).
13692                          */
13693                         th->th_seq = htonl(tp->snd_una);
13694                         bbr_seq = tp->snd_una;
13695                 } else {
13696                         /*
13697                          * len == 0 and not persist we use snd_max, sending
13698                          * an ack unless we have sent the fin then its 1
13699                          * minus.
13700                          */
13701                         /*
13702                          * XXXRRS Question if we are in persists and we have
13703                          * nothing outstanding to send and we have not sent
13704                          * a FIN, we will send an ACK. In such a case it
13705                          * might be better to send (tp->snd_una - 1) which
13706                          * would force the peer to ack.
13707                          */
13708                         if (tp->t_flags & TF_SENTFIN) {
13709                                 th->th_seq = htonl(tp->snd_max - 1);
13710                                 bbr_seq = (tp->snd_max - 1);
13711                         } else {
13712                                 th->th_seq = htonl(tp->snd_max);
13713                                 bbr_seq = tp->snd_max;
13714                         }
13715                 }
13716         } else {
13717                 /* All retransmits use the rsm to guide the send */
13718                 th->th_seq = htonl(rsm->r_start);
13719                 bbr_seq = rsm->r_start;
13720         }
13721         th->th_ack = htonl(tp->rcv_nxt);
13722         if (optlen) {
13723                 bcopy(opt, th + 1, optlen);
13724                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
13725         }
13726         th->th_flags = flags;
13727         /*
13728          * Calculate receive window.  Don't shrink window, but avoid silly
13729          * window syndrome.
13730          */
13731         if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) &&
13732                                   recwin < maxseg)))
13733                 recwin = 0;
13734         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
13735             recwin < (tp->rcv_adv - tp->rcv_nxt))
13736                 recwin = (tp->rcv_adv - tp->rcv_nxt);
13737         if (recwin > TCP_MAXWIN << tp->rcv_scale)
13738                 recwin = TCP_MAXWIN << tp->rcv_scale;
13739
13740         /*
13741          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
13742          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
13743          * handled in syncache.
13744          */
13745         if (flags & TH_SYN)
13746                 th->th_win = htons((u_short)
13747                     (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
13748         else
13749                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
13750         /*
13751          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
13752          * window.  This may cause the remote transmitter to stall.  This
13753          * flag tells soreceive() to disable delayed acknowledgements when
13754          * draining the buffer.  This can occur if the receiver is
13755          * attempting to read more data than can be buffered prior to
13756          * transmitting on the connection.
13757          */
13758         if (th->th_win == 0) {
13759                 tp->t_sndzerowin++;
13760                 tp->t_flags |= TF_RXWIN0SENT;
13761         } else
13762                 tp->t_flags &= ~TF_RXWIN0SENT;
13763         if (SEQ_GT(tp->snd_up, tp->snd_max)) {
13764                 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_max));
13765                 th->th_flags |= TH_URG;
13766         } else
13767                 /*
13768                  * If no urgent pointer to send, then we pull the urgent
13769                  * pointer to the left edge of the send window so that it
13770                  * doesn't drift into the send window on sequence number
13771                  * wraparound.
13772                  */
13773                 tp->snd_up = tp->snd_una;       /* drag it along */
13774
13775 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13776         if (to.to_flags & TOF_SIGNATURE) {
13777                 /*
13778                  * Calculate MD5 signature and put it into the place
13779                  * determined before. NOTE: since TCP options buffer doesn't
13780                  * point into mbuf's data, calculate offset and use it.
13781                  */
13782                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
13783                     (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
13784                         /*
13785                          * Do not send segment if the calculation of MD5
13786                          * digest has failed.
13787                          */
13788                         goto out;
13789                 }
13790         }
13791 #endif
13792
13793         /*
13794          * Put TCP length in extended header, and then checksum extended
13795          * header and data.
13796          */
13797         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
13798 #ifdef INET6
13799         if (isipv6) {
13800                 /*
13801                  * ip6_plen is not need to be filled now, and will be filled
13802                  * in ip6_output.
13803                  */
13804 #ifdef NETFLIX_TCPOUDP
13805                 if (tp->t_port) {
13806                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
13807                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13808                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
13809                         th->th_sum = htons(0);
13810                         UDPSTAT_INC(udps_opackets);
13811                 } else {
13812 #endif
13813                         csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
13814                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13815                         th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
13816                             optlen + len, IPPROTO_TCP, 0);
13817 #ifdef NETFLIX_TCPOUDP
13818                 }
13819 #endif
13820         }
13821 #endif
13822 #if defined(INET6) && defined(INET)
13823         else
13824 #endif
13825 #ifdef INET
13826         {
13827 #ifdef NETFLIX_TCPOUDP
13828                 if (tp->t_port) {
13829                         m->m_pkthdr.csum_flags = CSUM_UDP;
13830                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13831                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
13832                             ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
13833                         th->th_sum = htons(0);
13834                         UDPSTAT_INC(udps_opackets);
13835                 } else {
13836 #endif
13837                         csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP;
13838                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13839                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
13840                             ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
13841                             IPPROTO_TCP + len + optlen));
13842 #ifdef NETFLIX_TCPOUDP
13843                 }
13844 #endif
13845                 /* IP version must be set here for ipv4/ipv6 checking later */
13846                 KASSERT(ip->ip_v == IPVERSION,
13847                     ("%s: IP version incorrect: %d", __func__, ip->ip_v));
13848         }
13849 #endif
13850
13851         /*
13852          * Enable TSO and specify the size of the segments. The TCP pseudo
13853          * header checksum is always provided. XXX: Fixme: This is currently
13854          * not the case for IPv6.
13855          */
13856         if (tso || force_tso) {
13857                 KASSERT(force_tso || len > maxseg,
13858                     ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg));
13859                 m->m_pkthdr.csum_flags |= CSUM_TSO;
13860                 csum_flags |= CSUM_TSO;
13861                 m->m_pkthdr.tso_segsz = maxseg;
13862         }
13863         KASSERT(len + hdrlen == m_length(m, NULL),
13864             ("%s: mbuf chain different than expected: %d + %u != %u",
13865             __func__, len, hdrlen, m_length(m, NULL)));
13866
13867 #ifdef TCP_HHOOK
13868         /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */
13869         hhook_run_tcp_est_out(tp, th, &to, len, tso);
13870 #endif
13871 #ifdef TCPDEBUG
13872         /*
13873          * Trace.
13874          */
13875         if (so->so_options & SO_DEBUG) {
13876                 u_short save = 0;
13877
13878 #ifdef INET6
13879                 if (!isipv6)
13880 #endif
13881                 {
13882                         save = ipov->ih_len;
13883                         ipov->ih_len = htons(m->m_pkthdr.len    /* - hdrlen +
13884                               * (th->th_off << 2) */ );
13885                 }
13886                 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
13887 #ifdef INET6
13888                 if (!isipv6)
13889 #endif
13890                         ipov->ih_len = save;
13891         }
13892 #endif                          /* TCPDEBUG */
13893
13894         /* Log to the black box */
13895         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13896                 union tcp_log_stackspecific log;
13897
13898                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
13899                 /* Record info on type of transmission */
13900                 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay;
13901                 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3);
13902                 log.u_bbr.flex3 = maxseg;
13903                 log.u_bbr.flex4 = delay_calc;
13904                 /* Encode filled_all into the upper flex5 bit */
13905                 log.u_bbr.flex5 = bbr->rc_past_init_win;
13906                 log.u_bbr.flex5 <<= 1;
13907                 log.u_bbr.flex5 |= bbr->rc_no_pacing;
13908                 log.u_bbr.flex5 <<= 29;
13909                 if (filled_all)
13910                         log.u_bbr.flex5 |= 0x80000000;
13911                 log.u_bbr.flex5 |= tp->t_maxseg;
13912                 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs;
13913                 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr);
13914                 /* lets poke in the low and the high here for debugging */
13915                 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
13916                 if (rsm || sack_rxmit) {
13917                         if (doing_tlp)
13918                                 log.u_bbr.flex8 = 2;
13919                         else
13920                                 log.u_bbr.flex8 = 1;
13921                 } else {
13922                         log.u_bbr.flex8 = 0;
13923                 }
13924                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
13925                     len, &log, false, NULL, NULL, 0, tv);
13926         } else {
13927                 lgb = NULL;
13928         }
13929         /*
13930          * Fill in IP length and desired time to live and send to IP level.
13931          * There should be a better way to handle ttl and tos; we could keep
13932          * them in the template, but need a way to checksum without them.
13933          */
13934         /*
13935          * m->m_pkthdr.len should have been set before cksum calcuration,
13936          * because in6_cksum() need it.
13937          */
13938 #ifdef INET6
13939         if (isipv6) {
13940                 /*
13941                  * we separately set hoplimit for every segment, since the
13942                  * user might want to change the value via setsockopt. Also,
13943                  * desired default hop limit might be changed via Neighbor
13944                  * Discovery.
13945                  */
13946                 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
13947
13948                 /*
13949                  * Set the packet size here for the benefit of DTrace
13950                  * probes. ip6_output() will set it properly; it's supposed
13951                  * to include the option header lengths as well.
13952                  */
13953                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
13954
13955                 if (V_path_mtu_discovery && maxseg > V_tcp_minmss)
13956                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13957                 else
13958                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13959
13960                 if (tp->t_state == TCPS_SYN_SENT)
13961                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
13962
13963                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
13964                 /* TODO: IPv6 IP6TOS_ECT bit on */
13965                 error = ip6_output(m, inp->in6p_outputopts,
13966                     &inp->inp_route6,
13967                     ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
13968                     NULL, NULL, inp);
13969
13970                 if (error == EMSGSIZE && inp->inp_route6.ro_rt != NULL)
13971                         mtu = inp->inp_route6.ro_rt->rt_mtu;
13972         }
13973 #endif                          /* INET6 */
13974 #if defined(INET) && defined(INET6)
13975         else
13976 #endif
13977 #ifdef INET
13978         {
13979                 ip->ip_len = htons(m->m_pkthdr.len);
13980 #ifdef INET6
13981                 if (isipv6)
13982                         ip->ip_ttl = in6_selecthlim(inp, NULL);
13983 #endif                          /* INET6 */
13984                 /*
13985                  * If we do path MTU discovery, then we set DF on every
13986                  * packet. This might not be the best thing to do according
13987                  * to RFC3390 Section 2. However the tcp hostcache migitates
13988                  * the problem so it affects only the first tcp connection
13989                  * with a host.
13990                  *
13991                  * NB: Don't set DF on small MTU/MSS to have a safe
13992                  * fallback.
13993                  */
13994                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
13995                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13996                         if (tp->t_port == 0 || len < V_tcp_minmss) {
13997                                 ip->ip_off |= htons(IP_DF);
13998                         }
13999                 } else {
14000                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
14001                 }
14002
14003                 if (tp->t_state == TCPS_SYN_SENT)
14004                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
14005
14006                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
14007
14008                 error = ip_output(m, inp->inp_options, &inp->inp_route,
14009                     ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
14010                     inp);
14011                 if (error == EMSGSIZE && inp->inp_route.ro_rt != NULL)
14012                         mtu = inp->inp_route.ro_rt->rt_mtu;
14013         }
14014 #endif                          /* INET */
14015 out:
14016
14017         if (lgb) {
14018                 lgb->tlb_errno = error;
14019                 lgb = NULL;
14020         }
14021         /*
14022          * In transmit state, time the transmission and arrange for the
14023          * retransmit.  In persist state, just set snd_max.
14024          */
14025         if (error == 0) {
14026                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
14027                     (tp->t_flags & TF_SACK_PERMIT) &&
14028                     tp->rcv_numsacks > 0)
14029                         tcp_clean_dsack_blocks(tp);
14030                 /* We sent an ack clear the bbr_segs_rcvd count */
14031                 bbr->output_error_seen = 0;
14032                 bbr->oerror_cnt = 0;
14033                 bbr->bbr_segs_rcvd = 0;
14034                 if (len == 0)
14035                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1);
14036                 else if (hw_tls) {
14037                         if (filled_all ||
14038                             (len >= bbr->r_ctl.rc_pace_max_segs))
14039                                 BBR_STAT_INC(bbr_meets_tso_thresh);
14040                         else {
14041                                 if (doing_tlp) {
14042                                         BBR_STAT_INC(bbr_miss_tlp);
14043                                         bbr_log_type_hrdwtso(tp, bbr, len, 1, what_we_can);
14044
14045
14046                                 } else if (rsm) {
14047                                         BBR_STAT_INC(bbr_miss_retran);
14048                                         bbr_log_type_hrdwtso(tp, bbr, len, 2, what_we_can);
14049                                 } else if ((ctf_outstanding(tp) + bbr->r_ctl.rc_pace_max_segs) > sbavail(sb)) {
14050                                         BBR_STAT_INC(bbr_miss_tso_app);
14051                                         bbr_log_type_hrdwtso(tp, bbr, len, 3, what_we_can);
14052                                 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
14053                                                                  bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_pace_max_segs) > tp->snd_cwnd) {
14054                                         BBR_STAT_INC(bbr_miss_tso_cwnd);
14055                                         bbr_log_type_hrdwtso(tp, bbr, len, 4, what_we_can);
14056                                 } else if ((ctf_outstanding(tp) + bbr->r_ctl.rc_pace_max_segs) > tp->snd_wnd) {
14057                                         BBR_STAT_INC(bbr_miss_tso_rwnd);
14058                                         bbr_log_type_hrdwtso(tp, bbr, len, 5, what_we_can);
14059                                 } else {
14060                                         BBR_STAT_INC(bbr_miss_unknown);
14061                                         bbr_log_type_hrdwtso(tp, bbr, len, 6, what_we_can);
14062                                 }
14063                         }
14064                 }
14065                 /* Do accounting for new sends */
14066                 if ((len > 0) && (rsm == NULL)) {
14067                         int idx;
14068                         if (tp->snd_una == tp->snd_max) {
14069                                 /*
14070                                  * Special case to match google, when
14071                                  * nothing is in flight the delivered
14072                                  * time does get updated to the current
14073                                  * time (see tcp_rate_bsd.c).
14074                                  */
14075                                 bbr->r_ctl.rc_del_time = cts;
14076                         }
14077                         if (len >= maxseg) {
14078                                 idx = (len / maxseg) + 3;
14079                                 if (idx >= TCP_MSS_ACCT_ATIMER)
14080                                         counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1);
14081                                 else
14082                                         counter_u64_add(bbr_out_size[idx], 1);
14083                         } else {
14084                                 /* smaller than a MSS */
14085                                 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options);
14086                                 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV)
14087                                         idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1);
14088                                 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1);
14089                         }
14090                 }
14091         }
14092         abandon = 0;
14093         /*
14094          * We must do the send accounting before we log the output,
14095          * otherwise the state of the rsm could change and we account to the
14096          * wrong bucket.
14097          */
14098         if (len > 0) {
14099                 bbr_do_send_accounting(tp, bbr, rsm, len, error);
14100                 if (error == 0) {
14101                         if (tp->snd_una == tp->snd_max)
14102                                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
14103                 }
14104         }
14105         bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error,
14106             cts, mb, &abandon, rsm, 0, sb);
14107         if (abandon) {
14108                 /*
14109                  * If bbr_log_output destroys the TCB or sees a TH_RST being
14110                  * sent we should hit this condition.
14111                  */
14112                 return (0);
14113         }
14114         if (((tp->t_flags & TF_FORCEDATA) == 0) ||
14115             (bbr->rc_in_persist == 0)) {
14116                 /*
14117                  * Advance snd_nxt over sequence space of this segment.
14118                  */
14119                 if (error)
14120                         /* We don't log or do anything with errors */
14121                         goto skip_upd;
14122
14123                 if (tp->snd_una == tp->snd_max &&
14124                     (len || (flags & (TH_SYN | TH_FIN)))) {
14125                         /*
14126                          * Update the time we just added data since none was
14127                          * outstanding.
14128                          */
14129                         bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
14130                         bbr->rc_tp->t_acktime  = ticks;
14131                 }
14132                 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) {
14133                         if (flags & TH_SYN) {
14134                                 tp->snd_max++;
14135                         }
14136                         if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
14137                                 tp->snd_max++;
14138                                 tp->t_flags |= TF_SENTFIN;
14139                         }
14140                 }
14141                 if (sack_rxmit == 0)
14142                         tp->snd_max += len;
14143 skip_upd:
14144                 if ((error == 0) && len)
14145                         tot_len += len;
14146         } else {
14147                 /* Persists case */
14148                 int32_t xlen = len;
14149
14150                 if (error)
14151                         goto nomore;
14152
14153                 if (flags & TH_SYN)
14154                         ++xlen;
14155                 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
14156                         ++xlen;
14157                         tp->t_flags |= TF_SENTFIN;
14158                 }
14159                 if (xlen && (tp->snd_una == tp->snd_max)) {
14160                         /*
14161                          * Update the time we just added data since none was
14162                          * outstanding.
14163                          */
14164                         bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
14165                         bbr->rc_tp->t_acktime = ticks;
14166                 }
14167                 if (sack_rxmit == 0)
14168                         tp->snd_max += xlen;
14169                 tot_len += (len + optlen + ipoptlen);
14170         }
14171 nomore:
14172         if (error) {
14173                 /*
14174                  * Failures do not advance the seq counter above. For the
14175                  * case of ENOBUFS we will fall out and become ack-clocked.
14176                  * capping the cwnd at the current flight.
14177                  * Everything else will just have to retransmit with the timer
14178                  * (no pacer).
14179                  */
14180                 SOCKBUF_UNLOCK_ASSERT(sb);
14181                 BBR_STAT_INC(bbr_saw_oerr);
14182                 /* Clear all delay/early tracks */
14183                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
14184                 bbr->r_ctl.rc_agg_early = 0;
14185                 bbr->r_agg_early_set = 0;
14186                 bbr->output_error_seen = 1;
14187                 if (bbr->oerror_cnt < 0xf)
14188                         bbr->oerror_cnt++;
14189                 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) {
14190                         /* drop the session */
14191                         tcp_set_inp_to_drop(inp, ENETDOWN);
14192                 }
14193                 switch (error) {
14194                 case ENOBUFS:
14195                         /*
14196                          * Make this guy have to get ack's to send
14197                          * more but lets make sure we don't
14198                          * slam him below a T-O (1MSS).
14199                          */
14200                         if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
14201                                 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
14202                                                                     bbr->r_ctl.rc_lost_bytes)) - maxseg;
14203                                 if (tp->snd_cwnd < maxseg)
14204                                         tp->snd_cwnd = maxseg;
14205                         }
14206                         slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt;
14207                         BBR_STAT_INC(bbr_saw_enobuf);
14208                         if (bbr->bbr_hdrw_pacing)
14209                                 counter_u64_add(bbr_hdwr_pacing_enobuf, 1);
14210                         else
14211                                 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1);
14212                         /*
14213                          * Here even in the enobuf's case we want to do our
14214                          * state update. The reason being we may have been
14215                          * called by the input function. If so we have had
14216                          * things change.
14217                          */
14218                         error = 0;
14219                         goto enobufs;
14220                 case EMSGSIZE:
14221                         /*
14222                          * For some reason the interface we used initially
14223                          * to send segments changed to another or lowered
14224                          * its MTU. If TSO was active we either got an
14225                          * interface without TSO capabilits or TSO was
14226                          * turned off. If we obtained mtu from ip_output()
14227                          * then update it and try again.
14228                          */
14229                         /* Turn on tracing (or try to) */
14230                         {
14231                                 int old_maxseg;
14232
14233                                 old_maxseg = tp->t_maxseg;
14234                                 BBR_STAT_INC(bbr_saw_emsgsiz);
14235                                 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts);
14236                                 if (mtu != 0)
14237                                         tcp_mss_update(tp, -1, mtu, NULL, NULL);
14238                                 if (old_maxseg <= tp->t_maxseg) {
14239                                         /* Huh it did not shrink? */
14240                                         tp->t_maxseg = old_maxseg - 40;
14241                                         bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts);
14242                                 }
14243                                 tp->t_flags &= ~TF_FORCEDATA;
14244                                 /*
14245                                  * Nuke all other things that can interfere
14246                                  * with slot
14247                                  */
14248                                 if ((tot_len + len) && (len >= tp->t_maxseg)) {
14249                                         slot = bbr_get_pacing_delay(bbr,
14250                                             bbr->r_ctl.rc_bbr_hptsi_gain,
14251                                             (tot_len + len), cts, 0);
14252                                         if (slot < bbr_error_base_paceout)
14253                                                 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
14254                                 } else
14255                                         slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
14256                                 bbr->rc_output_starts_timer = 1;
14257                                 bbr_start_hpts_timer(bbr, tp, cts, 10, slot,
14258                                     tot_len);
14259                                 return (error);
14260                         }
14261                 case EPERM:
14262                         tp->t_softerror = error;
14263                         /* Fall through */
14264                 case EHOSTDOWN:
14265                 case EHOSTUNREACH:
14266                 case ENETDOWN:
14267                 case ENETUNREACH:
14268                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
14269                                 tp->t_softerror = error;
14270                         }
14271                         /* FALLTHROUGH */
14272                 default:
14273                         tp->t_flags &= ~TF_FORCEDATA;
14274                         slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt;
14275                         bbr->rc_output_starts_timer = 1;
14276                         bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0);
14277                         return (error);
14278                 }
14279 #ifdef NETFLIX_STATS
14280         } else if (((tp->t_flags & TF_GPUTINPROG) == 0) &&
14281                     len &&
14282                     (rsm == NULL) &&
14283             (bbr->rc_in_persist == 0)) {
14284                 tp->gput_seq = bbr_seq;
14285                 tp->gput_ack = bbr_seq +
14286                     min(sbavail(&so->so_snd) - sb_offset, sendwin);
14287                 tp->gput_ts = cts;
14288                 tp->t_flags |= TF_GPUTINPROG;
14289 #endif
14290         }
14291         TCPSTAT_INC(tcps_sndtotal);
14292         if ((bbr->bbr_hdw_pace_ena) &&
14293             (bbr->bbr_attempt_hdwr_pace == 0) &&
14294             (bbr->rc_past_init_win) &&
14295             (bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
14296             (get_filter_value(&bbr->r_ctl.rc_delrate)) &&
14297             (inp->inp_route.ro_rt &&
14298              inp->inp_route.ro_rt->rt_ifp)) {
14299                 /*
14300                  * We are past the initial window and
14301                  * have at least one measurement so we
14302                  * could use hardware pacing if its available.
14303                  * We have an interface and we have not attempted
14304                  * to setup hardware pacing, lets try to now.
14305                  */
14306                 uint64_t rate_wanted;
14307                 int err = 0;
14308
14309                 rate_wanted = bbr_get_hardware_rate(bbr);
14310                 bbr->bbr_attempt_hdwr_pace = 1;
14311                 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp,
14312                                                       inp->inp_route.ro_rt->rt_ifp,
14313                                                       rate_wanted,
14314                                                       (RS_PACING_GEQ|RS_PACING_SUB_OK),
14315                                                       &err);
14316                 if (bbr->r_ctl.crte) {
14317                         bbr_type_log_hdwr_pacing(bbr,
14318                                                  bbr->r_ctl.crte->ptbl->rs_ifp,
14319                                                  rate_wanted,
14320                                                  bbr->r_ctl.crte->rate,
14321                                                  __LINE__, cts, err);
14322                         BBR_STAT_INC(bbr_hdwr_rl_add_ok);
14323                         counter_u64_add(bbr_flows_nohdwr_pacing, -1);
14324                         counter_u64_add(bbr_flows_whdwr_pacing, 1);
14325                         bbr->bbr_hdrw_pacing = 1;
14326                         /* Now what is our gain status? */
14327                         if (bbr->r_ctl.crte->rate < rate_wanted) {
14328                                 /* We have a problem */
14329                                 bbr_setup_less_of_rate(bbr, cts,
14330                                                        bbr->r_ctl.crte->rate, rate_wanted);
14331                         } else {
14332                                 /* We are good */
14333                                 bbr->gain_is_limited = 0;
14334                                 bbr->skip_gain = 0;
14335                         }
14336                         tcp_bbr_tso_size_check(bbr, cts);
14337                 } else {
14338                         bbr_type_log_hdwr_pacing(bbr,
14339                                                  inp->inp_route.ro_rt->rt_ifp,
14340                                                  rate_wanted,
14341                                                  0,
14342                                                  __LINE__, cts, err);
14343                         BBR_STAT_INC(bbr_hdwr_rl_add_fail);
14344                 }
14345         }
14346         if (bbr->bbr_hdrw_pacing) {
14347                 /*
14348                  * Worry about cases where the route
14349                  * changes or something happened that we
14350                  * lost our hardware pacing possibly during
14351                  * the last ip_output call.
14352                  */
14353                 if (inp->inp_snd_tag == NULL) {
14354                         /* A change during ip output disabled hw pacing? */
14355                         bbr->bbr_hdrw_pacing = 0;
14356                 } else if ((inp->inp_route.ro_rt == NULL) ||
14357                     (inp->inp_route.ro_rt->rt_ifp != inp->inp_snd_tag->ifp)) {
14358                         /*
14359                          * We had an interface or route change,
14360                          * detach from the current hdwr pacing
14361                          * and setup to re-attempt next go
14362                          * round.
14363                          */
14364                         bbr->bbr_hdrw_pacing = 0;
14365                         bbr->bbr_attempt_hdwr_pace = 0;
14366                         tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
14367                         tcp_bbr_tso_size_check(bbr, cts);
14368                 }
14369         }
14370         /*
14371          * Data sent (as far as we can tell). If this advertises a larger
14372          * window than any other segment, then remember the size of the
14373          * advertised window. Any pending ACK has now been sent.
14374          */
14375         if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
14376                 tp->rcv_adv = tp->rcv_nxt + recwin;
14377
14378         tp->last_ack_sent = tp->rcv_nxt;
14379         if ((error == 0) &&
14380             (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) &&
14381             (doing_tlp == 0) &&
14382             (tso == 0) &&
14383             (hw_tls == 0) &&
14384             (len > 0) &&
14385             ((flags & TH_RST) == 0) &&
14386             (IN_RECOVERY(tp->t_flags) == 0) &&
14387             (bbr->rc_in_persist == 0) &&
14388             ((tp->t_flags & TF_FORCEDATA) == 0) &&
14389             (tot_len < bbr->r_ctl.rc_pace_max_segs)) {
14390                 /*
14391                  * For non-tso we need to goto again until we have sent out
14392                  * enough data to match what we are hptsi out every hptsi
14393                  * interval.
14394                  */
14395                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14396                         /* Make sure snd_nxt is drug up */
14397                         tp->snd_nxt = tp->snd_max;
14398                 }
14399                 if (rsm != NULL) {
14400                         rsm = NULL;
14401                         goto skip_again;
14402                 }
14403                 rsm = NULL;
14404                 sack_rxmit = 0;
14405                 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK | TF_FORCEDATA);
14406                 goto again;
14407         }
14408 skip_again:
14409         if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) {
14410                 /*
14411                  * Calculate/Re-Calculate the hptsi slot in usecs based on
14412                  * what we have sent so far
14413                  */
14414                 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
14415                 if (bbr->rc_no_pacing)
14416                         slot = 0;
14417         }
14418         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK | TF_FORCEDATA);
14419 enobufs:
14420         if (bbr->rc_use_google == 0)
14421                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
14422         bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
14423                                                         bbr->r_ctl.rc_lost_bytes)));
14424         bbr->rc_output_starts_timer = 1;
14425         if (bbr->bbr_use_rack_cheat &&
14426             (more_to_rxt ||
14427              ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) {
14428                 /* Rack cheats and shotguns out all rxt's 1ms apart */
14429                 if (slot > 1000)
14430                         slot = 1000;
14431         }
14432         if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) {
14433                 /*
14434                  * We don't change the tso size until some number of sends
14435                  * to give the hardware commands time to get down
14436                  * to the interface.
14437                  */
14438                 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++;
14439                 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) {
14440                         bbr->hw_pacing_set = 1;
14441                         tcp_bbr_tso_size_check(bbr, cts);
14442                 }
14443         }
14444         bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len);
14445         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14446                 /* Make sure snd_nxt is drug up */
14447                 tp->snd_nxt = tp->snd_max;
14448         }
14449         return (error);
14450
14451 }
14452
14453 /*
14454  * See bbr_output_wtime() for return values.
14455  */
14456 static int
14457 bbr_output(struct tcpcb *tp)
14458 {
14459         int32_t ret;
14460         struct timeval tv;
14461         struct tcp_bbr *bbr;
14462
14463         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14464         INP_WLOCK_ASSERT(tp->t_inpcb);
14465         (void)tcp_get_usecs(&tv);
14466         ret = bbr_output_wtime(tp, &tv);
14467         return (ret);
14468 }
14469
14470 static void
14471 bbr_mtu_chg(struct tcpcb *tp)
14472 {
14473         struct tcp_bbr *bbr;
14474         struct bbr_sendmap *rsm, *frsm = NULL;
14475         uint32_t maxseg;
14476
14477         /*
14478          * The MTU has changed. a) Clear the sack filter. b) Mark everything
14479          * over the current size as SACK_PASS so a retransmit will occur.
14480          */
14481
14482         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14483         maxseg = tp->t_maxseg - bbr->rc_last_options;
14484         sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
14485         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
14486                 /* Don't mess with ones acked (by sack?) */
14487                 if (rsm->r_flags & BBR_ACKED)
14488                         continue;
14489                 if ((rsm->r_end - rsm->r_start) > maxseg) {
14490                         /*
14491                          * We mark sack-passed on all the previous large
14492                          * sends we did. This will force them to retransmit.
14493                          */
14494                         rsm->r_flags |= BBR_SACK_PASSED;
14495                         if (((rsm->r_flags & BBR_MARKED_LOST) == 0) &&
14496                             bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) {
14497                                 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
14498                                 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
14499                                 rsm->r_flags |= BBR_MARKED_LOST;
14500                         }
14501                         if (frsm == NULL)
14502                                 frsm = rsm;
14503                 }
14504         }
14505         if (frsm) {
14506                 bbr->r_ctl.rc_resend = frsm;
14507         }
14508 }
14509
14510 /*
14511  * bbr_ctloutput() must drop the inpcb lock before performing copyin on
14512  * socket option arguments.  When it re-acquires the lock after the copy, it
14513  * has to revalidate that the connection is still valid for the socket
14514  * option.
14515  */
14516 static int
14517 bbr_set_sockopt(struct socket *so, struct sockopt *sopt,
14518                 struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14519 {
14520         int32_t error = 0, optval;
14521
14522         switch (sopt->sopt_name) {
14523         case TCP_RACK_PACE_MAX_SEG:
14524         case TCP_RACK_MIN_TO:
14525         case TCP_RACK_REORD_THRESH:
14526         case TCP_RACK_REORD_FADE:
14527         case TCP_RACK_TLP_THRESH:
14528         case TCP_RACK_PKT_DELAY:
14529         case TCP_BBR_ALGORITHM:
14530         case TCP_BBR_TSLIMITS:
14531         case TCP_BBR_IWINTSO:
14532         case TCP_BBR_RECFORCE:
14533         case TCP_BBR_STARTUP_PG:
14534         case TCP_BBR_DRAIN_PG:
14535         case TCP_BBR_RWND_IS_APP:
14536         case TCP_BBR_PROBE_RTT_INT:
14537         case TCP_BBR_PROBE_RTT_GAIN:
14538         case TCP_BBR_PROBE_RTT_LEN:
14539         case TCP_BBR_STARTUP_LOSS_EXIT:
14540         case TCP_BBR_USEDEL_RATE:
14541         case TCP_BBR_MIN_RTO:
14542         case TCP_BBR_MAX_RTO:
14543         case TCP_BBR_PACE_PER_SEC:
14544         case TCP_DELACK:
14545         case TCP_BBR_PACE_DEL_TAR:
14546         case TCP_BBR_SEND_IWND_IN_TSO:
14547         case TCP_BBR_EXTRA_STATE:
14548         case TCP_BBR_UTTER_MAX_TSO:
14549         case TCP_BBR_MIN_TOPACEOUT:
14550         case TCP_BBR_FLOOR_MIN_TSO:
14551         case TCP_BBR_TSTMP_RAISES:
14552         case TCP_BBR_POLICER_DETECT:
14553         case TCP_BBR_USE_RACK_CHEAT:
14554         case TCP_DATA_AFTER_CLOSE:
14555         case TCP_BBR_HDWR_PACE:
14556         case TCP_BBR_PACE_SEG_MAX:
14557         case TCP_BBR_PACE_SEG_MIN:
14558         case TCP_BBR_PACE_CROSS:
14559         case TCP_BBR_PACE_OH:
14560 #ifdef NETFLIX_PEAKRATE
14561         case TCP_MAXPEAKRATE:
14562 #endif
14563         case TCP_BBR_TMR_PACE_OH:
14564         case TCP_BBR_RACK_RTT_USE:
14565         case TCP_BBR_RETRAN_WTSO:
14566                 break;
14567         default:
14568                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14569                 break;
14570         }
14571         INP_WUNLOCK(inp);
14572         error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
14573         if (error)
14574                 return (error);
14575         INP_WLOCK(inp);
14576         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
14577                 INP_WUNLOCK(inp);
14578                 return (ECONNRESET);
14579         }
14580         tp = intotcpcb(inp);
14581         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14582         switch (sopt->sopt_name) {
14583         case TCP_BBR_PACE_PER_SEC:
14584                 BBR_OPTS_INC(tcp_bbr_pace_per_sec);
14585                 bbr->r_ctl.bbr_hptsi_per_second = optval;
14586                 break;
14587         case TCP_BBR_PACE_DEL_TAR:
14588                 BBR_OPTS_INC(tcp_bbr_pace_del_tar);
14589                 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval;
14590                 break;
14591         case TCP_BBR_PACE_SEG_MAX:
14592                 BBR_OPTS_INC(tcp_bbr_pace_seg_max);
14593                 bbr->r_ctl.bbr_hptsi_segments_max = optval;
14594                 break;
14595         case TCP_BBR_PACE_SEG_MIN:
14596                 BBR_OPTS_INC(tcp_bbr_pace_seg_min);
14597                 bbr->r_ctl.bbr_hptsi_bytes_min = optval;
14598                 break;
14599         case TCP_BBR_PACE_CROSS:
14600                 BBR_OPTS_INC(tcp_bbr_pace_cross);
14601                 bbr->r_ctl.bbr_cross_over = optval;
14602                 break;
14603         case TCP_BBR_ALGORITHM:
14604                 BBR_OPTS_INC(tcp_bbr_algorithm);
14605                 if (optval && (bbr->rc_use_google == 0)) {
14606                         /* Turn on the google mode */
14607                         bbr_google_mode_on(bbr);
14608                         if ((optval > 3) && (optval < 500)) {
14609                                 /*
14610                                  * Must be at least greater than .3%
14611                                  * and must be less than 50.0%.
14612                                  */
14613                                 bbr->r_ctl.bbr_google_discount = optval;
14614                         }
14615                 } else if ((optval == 0) && (bbr->rc_use_google == 1)) {
14616                         /* Turn off the google mode */
14617                         bbr_google_mode_off(bbr);
14618                 }
14619                 break;
14620         case TCP_BBR_TSLIMITS:
14621                 BBR_OPTS_INC(tcp_bbr_tslimits);
14622                 if (optval == 1)
14623                         bbr->rc_use_ts_limit = 1;
14624                 else if (optval == 0)
14625                         bbr->rc_use_ts_limit = 0;
14626                 else
14627                         error = EINVAL;
14628                 break;
14629
14630         case TCP_BBR_IWINTSO:
14631                 BBR_OPTS_INC(tcp_bbr_iwintso);
14632                 if ((optval >= 0) && (optval < 128)) {
14633                         uint32_t twin;
14634
14635                         bbr->rc_init_win = optval;
14636                         twin = bbr_initial_cwnd(bbr, tp);
14637                         if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd))
14638                                 tp->snd_cwnd = twin;
14639                         else
14640                                 error = EBUSY;
14641                 } else
14642                         error = EINVAL;
14643                 break;
14644         case TCP_BBR_STARTUP_PG:
14645                 BBR_OPTS_INC(tcp_bbr_startup_pg);
14646                 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) {
14647                         bbr->r_ctl.rc_startup_pg = optval;
14648                         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
14649                                 bbr->r_ctl.rc_bbr_hptsi_gain = optval;
14650                         }
14651                 } else
14652                         error = EINVAL;
14653                 break;
14654         case TCP_BBR_DRAIN_PG:
14655                 BBR_OPTS_INC(tcp_bbr_drain_pg);
14656                 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE))
14657                         bbr->r_ctl.rc_drain_pg = optval;
14658                 else
14659                         error = EINVAL;
14660                 break;
14661         case TCP_BBR_PROBE_RTT_LEN:
14662                 BBR_OPTS_INC(tcp_bbr_probertt_len);
14663                 if (optval <= 1)
14664                         reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND));
14665                 else
14666                         error = EINVAL;
14667                 break;
14668         case TCP_BBR_PROBE_RTT_GAIN:
14669                 BBR_OPTS_INC(tcp_bbr_probertt_gain);
14670                 if (optval <= BBR_UNIT)
14671                         bbr->r_ctl.bbr_rttprobe_gain_val = optval;
14672                 else
14673                         error = EINVAL;
14674                 break;
14675         case TCP_BBR_PROBE_RTT_INT:
14676                 BBR_OPTS_INC(tcp_bbr_probe_rtt_int);
14677                 if (optval > 1000)
14678                         bbr->r_ctl.rc_probertt_int = optval;
14679                 else
14680                         error = EINVAL;
14681                 break;
14682         case TCP_BBR_MIN_TOPACEOUT:
14683                 BBR_OPTS_INC(tcp_bbr_topaceout);
14684                 if (optval == 0) {
14685                         bbr->no_pacing_until = 0;
14686                         bbr->rc_no_pacing = 0;
14687                 } else if (optval <= 0x00ff) {
14688                         bbr->no_pacing_until = optval;
14689                         if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) &&
14690                             (bbr->rc_bbr_state == BBR_STATE_STARTUP)){
14691                                 /* Turn on no pacing */
14692                                 bbr->rc_no_pacing = 1;
14693                         }
14694                 } else
14695                         error = EINVAL;
14696                 break;
14697         case TCP_BBR_STARTUP_LOSS_EXIT:
14698                 BBR_OPTS_INC(tcp_bbr_startup_loss_exit);
14699                 bbr->rc_loss_exit = optval;
14700                 break;
14701         case TCP_BBR_USEDEL_RATE:
14702                 error = EINVAL;
14703                 break;
14704         case TCP_BBR_MIN_RTO:
14705                 BBR_OPTS_INC(tcp_bbr_min_rto);
14706                 bbr->r_ctl.rc_min_rto_ms = optval;
14707                 break;
14708         case TCP_BBR_MAX_RTO:
14709                 BBR_OPTS_INC(tcp_bbr_max_rto);
14710                 bbr->rc_max_rto_sec = optval;
14711                 break;
14712         case TCP_RACK_MIN_TO:
14713                 /* Minimum time between rack t-o's in ms */
14714                 BBR_OPTS_INC(tcp_rack_min_to);
14715                 bbr->r_ctl.rc_min_to = optval;
14716                 break;
14717         case TCP_RACK_REORD_THRESH:
14718                 /* RACK reorder threshold (shift amount) */
14719                 BBR_OPTS_INC(tcp_rack_reord_thresh);
14720                 if ((optval > 0) && (optval < 31))
14721                         bbr->r_ctl.rc_reorder_shift = optval;
14722                 else
14723                         error = EINVAL;
14724                 break;
14725         case TCP_RACK_REORD_FADE:
14726                 /* Does reordering fade after ms time */
14727                 BBR_OPTS_INC(tcp_rack_reord_fade);
14728                 bbr->r_ctl.rc_reorder_fade = optval;
14729                 break;
14730         case TCP_RACK_TLP_THRESH:
14731                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
14732                 BBR_OPTS_INC(tcp_rack_tlp_thresh);
14733                 if (optval)
14734                         bbr->rc_tlp_threshold = optval;
14735                 else
14736                         error = EINVAL;
14737                 break;
14738         case TCP_BBR_USE_RACK_CHEAT:
14739                 BBR_OPTS_INC(tcp_use_rackcheat);
14740                 if (bbr->rc_use_google) {
14741                         error = EINVAL;
14742                         break;
14743                 }
14744                 BBR_OPTS_INC(tcp_rack_cheat);
14745                 if (optval)
14746                         bbr->bbr_use_rack_cheat = 1;
14747                 else
14748                         bbr->bbr_use_rack_cheat = 0;
14749                 break;
14750         case TCP_BBR_FLOOR_MIN_TSO:
14751                 BBR_OPTS_INC(tcp_utter_max_tso);
14752                 if ((optval >= 0) && (optval < 40))
14753                         bbr->r_ctl.bbr_hptsi_segments_floor = optval;
14754                 else
14755                         error = EINVAL;
14756                 break;
14757         case TCP_BBR_UTTER_MAX_TSO:
14758                 BBR_OPTS_INC(tcp_utter_max_tso);
14759                 if ((optval >= 0) && (optval < 0xffff))
14760                         bbr->r_ctl.bbr_utter_max = optval;
14761                 else
14762                         error = EINVAL;
14763                 break;
14764
14765         case TCP_BBR_EXTRA_STATE:
14766                 BBR_OPTS_INC(tcp_extra_state);
14767                 if (optval)
14768                         bbr->rc_use_idle_restart = 1;
14769                 else
14770                         bbr->rc_use_idle_restart = 0;
14771                 break;
14772         case TCP_BBR_SEND_IWND_IN_TSO:
14773                 BBR_OPTS_INC(tcp_iwnd_tso);
14774                 if (optval) {
14775                         bbr->bbr_init_win_cheat = 1;
14776                         if (bbr->rc_past_init_win == 0) {
14777                                 uint32_t cts;
14778                                 cts = tcp_get_usecs(&bbr->rc_tv);
14779                                 tcp_bbr_tso_size_check(bbr, cts);
14780                         }
14781                 } else
14782                         bbr->bbr_init_win_cheat = 0;
14783                 break;
14784         case TCP_BBR_HDWR_PACE:
14785                 BBR_OPTS_INC(tcp_hdwr_pacing);
14786                 if (optval){
14787                         bbr->bbr_hdw_pace_ena = 1;
14788                         bbr->bbr_attempt_hdwr_pace = 0;
14789                 } else {
14790                         bbr->bbr_hdw_pace_ena = 0;
14791 #ifdef RATELIMIT
14792                         if (bbr->bbr_hdrw_pacing) {
14793                                 bbr->bbr_hdrw_pacing = 0;
14794                                 in_pcbdetach_txrtlmt(bbr->rc_inp);
14795                         }
14796 #endif
14797                 }
14798                 break;
14799
14800         case TCP_DELACK:
14801                 BBR_OPTS_INC(tcp_delack);
14802                 if (optval < 100) {
14803                         if (optval == 0) /* off */
14804                                 tp->t_delayed_ack = 0;
14805                         else if (optval == 1) /* on which is 2 */
14806                                 tp->t_delayed_ack = 2;
14807                         else /* higher than 2 and less than 100 */
14808                                 tp->t_delayed_ack = optval;
14809                         if (tp->t_flags & TF_DELACK) {
14810                                 tp->t_flags &= ~TF_DELACK;
14811                                 tp->t_flags |= TF_ACKNOW;
14812                                 bbr_output(tp);
14813                         }
14814                 } else
14815                         error = EINVAL;
14816                 break;
14817         case TCP_RACK_PKT_DELAY:
14818                 /* RACK added ms i.e. rack-rtt + reord + N */
14819                 BBR_OPTS_INC(tcp_rack_pkt_delay);
14820                 bbr->r_ctl.rc_pkt_delay = optval;
14821                 break;
14822 #ifdef NETFLIX_PEAKRATE
14823         case TCP_MAXPEAKRATE:
14824                 BBR_OPTS_INC(tcp_maxpeak);
14825                 error = tcp_set_maxpeakrate(tp, optval);
14826                 if (!error)
14827                         tp->t_peakrate_thr = tp->t_maxpeakrate;
14828                 break;
14829 #endif
14830         case TCP_BBR_RETRAN_WTSO:
14831                 BBR_OPTS_INC(tcp_retran_wtso);
14832                 if (optval)
14833                         bbr->rc_resends_use_tso = 1;
14834                 else
14835                         bbr->rc_resends_use_tso = 0;
14836                 break;
14837         case TCP_DATA_AFTER_CLOSE:
14838                 BBR_OPTS_INC(tcp_data_ac);
14839                 if (optval)
14840                         bbr->rc_allow_data_af_clo = 1;
14841                 else
14842                         bbr->rc_allow_data_af_clo = 0;
14843                 break;
14844         case TCP_BBR_POLICER_DETECT:
14845                 BBR_OPTS_INC(tcp_policer_det);
14846                 if (bbr->rc_use_google == 0)
14847                         error = EINVAL;
14848                 else if (optval)
14849                         bbr->r_use_policer = 1;
14850                 else
14851                         bbr->r_use_policer = 0;
14852                 break;
14853
14854         case TCP_BBR_TSTMP_RAISES:
14855                 BBR_OPTS_INC(tcp_ts_raises);
14856                 if (optval)
14857                         bbr->ts_can_raise = 1;
14858                 else
14859                         bbr->ts_can_raise = 0;
14860                 break;
14861         case TCP_BBR_TMR_PACE_OH:
14862                 BBR_OPTS_INC(tcp_pacing_oh_tmr);
14863                 if (bbr->rc_use_google) {
14864                         error = EINVAL;
14865                 } else {
14866                         if (optval)
14867                                 bbr->r_ctl.rc_incr_tmrs = 1;
14868                         else
14869                                 bbr->r_ctl.rc_incr_tmrs = 0;
14870                 }
14871                 break;
14872         case TCP_BBR_PACE_OH:
14873                 BBR_OPTS_INC(tcp_pacing_oh);
14874                 if (bbr->rc_use_google) {
14875                         error = EINVAL;
14876                 } else {
14877                         if (optval > (BBR_INCL_TCP_OH|
14878                                       BBR_INCL_IP_OH|
14879                                       BBR_INCL_ENET_OH)) {
14880                                 error = EINVAL;
14881                                 break;
14882                         }
14883                         if (optval & BBR_INCL_TCP_OH)
14884                                 bbr->r_ctl.rc_inc_tcp_oh = 1;
14885                         else
14886                                 bbr->r_ctl.rc_inc_tcp_oh = 0;
14887                         if (optval & BBR_INCL_IP_OH)
14888                                 bbr->r_ctl.rc_inc_ip_oh = 1;
14889                         else
14890                                 bbr->r_ctl.rc_inc_ip_oh = 0;
14891                         if (optval & BBR_INCL_ENET_OH)
14892                                 bbr->r_ctl.rc_inc_enet_oh = 1;
14893                         else
14894                                 bbr->r_ctl.rc_inc_enet_oh = 0;
14895                 }
14896                 break;
14897         default:
14898                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14899                 break;
14900         }
14901 #ifdef NETFLIX_STATS
14902         tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
14903 #endif
14904         INP_WUNLOCK(inp);
14905         return (error);
14906 }
14907
14908 /*
14909  * return 0 on success, error-num on failure
14910  */
14911 static int
14912 bbr_get_sockopt(struct socket *so, struct sockopt *sopt,
14913     struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14914 {
14915         int32_t error, optval;
14916
14917         /*
14918          * Because all our options are either boolean or an int, we can just
14919          * pull everything into optval and then unlock and copy. If we ever
14920          * add a option that is not a int, then this will have quite an
14921          * impact to this routine.
14922          */
14923         switch (sopt->sopt_name) {
14924         case TCP_BBR_PACE_PER_SEC:
14925                 optval = bbr->r_ctl.bbr_hptsi_per_second;
14926                 break;
14927         case TCP_BBR_PACE_DEL_TAR:
14928                 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar;
14929                 break;
14930         case TCP_BBR_PACE_SEG_MAX:
14931                 optval = bbr->r_ctl.bbr_hptsi_segments_max;
14932                 break;
14933         case TCP_BBR_MIN_TOPACEOUT:
14934                 optval = bbr->no_pacing_until;
14935                 break;
14936         case TCP_BBR_PACE_SEG_MIN:
14937                 optval = bbr->r_ctl.bbr_hptsi_bytes_min;
14938                 break;
14939         case TCP_BBR_PACE_CROSS:
14940                 optval = bbr->r_ctl.bbr_cross_over;
14941                 break;
14942         case TCP_BBR_ALGORITHM:
14943                 optval = bbr->rc_use_google;
14944                 break;
14945         case TCP_BBR_TSLIMITS:
14946                 optval = bbr->rc_use_ts_limit;
14947                 break;
14948         case TCP_BBR_IWINTSO:
14949                 optval = bbr->rc_init_win;
14950                 break;
14951         case TCP_BBR_STARTUP_PG:
14952                 optval = bbr->r_ctl.rc_startup_pg;
14953                 break;
14954         case TCP_BBR_DRAIN_PG:
14955                 optval = bbr->r_ctl.rc_drain_pg;
14956                 break;
14957         case TCP_BBR_PROBE_RTT_INT:
14958                 optval = bbr->r_ctl.rc_probertt_int;
14959                 break;
14960         case TCP_BBR_PROBE_RTT_LEN:
14961                 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND);
14962                 break;
14963         case TCP_BBR_PROBE_RTT_GAIN:
14964                 optval = bbr->r_ctl.bbr_rttprobe_gain_val;
14965                 break;
14966         case TCP_BBR_STARTUP_LOSS_EXIT:
14967                 optval = bbr->rc_loss_exit;
14968                 break;
14969         case TCP_BBR_USEDEL_RATE:
14970                 error = EINVAL;
14971                 break;
14972         case TCP_BBR_MIN_RTO:
14973                 optval = bbr->r_ctl.rc_min_rto_ms;
14974                 break;
14975         case TCP_BBR_MAX_RTO:
14976                 optval = bbr->rc_max_rto_sec;
14977                 break;
14978         case TCP_RACK_PACE_MAX_SEG:
14979                 /* Max segments in a pace */
14980                 optval = bbr->r_ctl.rc_pace_max_segs;
14981                 break;
14982         case TCP_RACK_MIN_TO:
14983                 /* Minimum time between rack t-o's in ms */
14984                 optval = bbr->r_ctl.rc_min_to;
14985                 break;
14986         case TCP_RACK_REORD_THRESH:
14987                 /* RACK reorder threshold (shift amount) */
14988                 optval = bbr->r_ctl.rc_reorder_shift;
14989                 break;
14990         case TCP_RACK_REORD_FADE:
14991                 /* Does reordering fade after ms time */
14992                 optval = bbr->r_ctl.rc_reorder_fade;
14993                 break;
14994         case TCP_BBR_USE_RACK_CHEAT:
14995                 /* Do we use the rack cheat for rxt */
14996                 optval = bbr->bbr_use_rack_cheat;
14997                 break;
14998         case TCP_BBR_FLOOR_MIN_TSO:
14999                 optval = bbr->r_ctl.bbr_hptsi_segments_floor;
15000                 break;
15001         case TCP_BBR_UTTER_MAX_TSO:
15002                 optval = bbr->r_ctl.bbr_utter_max;
15003                 break;
15004         case TCP_BBR_SEND_IWND_IN_TSO:
15005                 /* Do we send TSO size segments initially */
15006                 optval = bbr->bbr_init_win_cheat;
15007                 break;
15008         case TCP_BBR_EXTRA_STATE:
15009                 optval = bbr->rc_use_idle_restart;
15010                 break;
15011         case TCP_RACK_TLP_THRESH:
15012                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
15013                 optval = bbr->rc_tlp_threshold;
15014                 break;
15015         case TCP_RACK_PKT_DELAY:
15016                 /* RACK added ms i.e. rack-rtt + reord + N */
15017                 optval = bbr->r_ctl.rc_pkt_delay;
15018                 break;
15019         case TCP_BBR_RETRAN_WTSO:
15020                 optval = bbr->rc_resends_use_tso;
15021                 break;
15022         case TCP_DATA_AFTER_CLOSE:
15023                 optval = bbr->rc_allow_data_af_clo;
15024                 break;
15025         case TCP_DELACK:
15026                 optval = tp->t_delayed_ack;
15027                 break;
15028         case TCP_BBR_HDWR_PACE:
15029                 optval = bbr->bbr_hdw_pace_ena;
15030                 break;
15031         case TCP_BBR_POLICER_DETECT:
15032                 optval = bbr->r_use_policer;
15033                 break;
15034         case TCP_BBR_TSTMP_RAISES:
15035                 optval = bbr->ts_can_raise;
15036                 break;
15037         case TCP_BBR_TMR_PACE_OH:
15038                 optval = bbr->r_ctl.rc_incr_tmrs;
15039                 break;
15040         case TCP_BBR_PACE_OH:
15041                 optval = 0;
15042                 if (bbr->r_ctl.rc_inc_tcp_oh)
15043                         optval |= BBR_INCL_TCP_OH;
15044                 if (bbr->r_ctl.rc_inc_ip_oh)
15045                         optval |= BBR_INCL_IP_OH;
15046                 if (bbr->r_ctl.rc_inc_enet_oh)
15047                         optval |= BBR_INCL_ENET_OH;
15048                 break;
15049         default:
15050                 return (tcp_default_ctloutput(so, sopt, inp, tp));
15051                 break;
15052         }
15053         INP_WUNLOCK(inp);
15054         error = sooptcopyout(sopt, &optval, sizeof optval);
15055         return (error);
15056 }
15057
15058 /*
15059  * return 0 on success, error-num on failure
15060  */
15061 static int
15062 bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
15063 {
15064         int32_t error = EINVAL;
15065         struct tcp_bbr *bbr;
15066
15067         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
15068         if (bbr == NULL) {
15069                 /* Huh? */
15070                 goto out;
15071         }
15072         if (sopt->sopt_dir == SOPT_SET) {
15073                 return (bbr_set_sockopt(so, sopt, inp, tp, bbr));
15074         } else if (sopt->sopt_dir == SOPT_GET) {
15075                 return (bbr_get_sockopt(so, sopt, inp, tp, bbr));
15076         }
15077 out:
15078         INP_WUNLOCK(inp);
15079         return (error);
15080 }
15081
15082
15083 struct tcp_function_block __tcp_bbr = {
15084         .tfb_tcp_block_name = __XSTRING(STACKNAME),
15085         .tfb_tcp_output = bbr_output,
15086         .tfb_do_queued_segments = ctf_do_queued_segments,
15087         .tfb_do_segment_nounlock = bbr_do_segment_nounlock,
15088         .tfb_tcp_do_segment = bbr_do_segment,
15089         .tfb_tcp_ctloutput = bbr_ctloutput,
15090         .tfb_tcp_fb_init = bbr_init,
15091         .tfb_tcp_fb_fini = bbr_fini,
15092         .tfb_tcp_timer_stop_all = bbr_stopall,
15093         .tfb_tcp_timer_activate = bbr_timer_activate,
15094         .tfb_tcp_timer_active = bbr_timer_active,
15095         .tfb_tcp_timer_stop = bbr_timer_stop,
15096         .tfb_tcp_rexmit_tmr = bbr_remxt_tmr,
15097         .tfb_tcp_handoff_ok = bbr_handoff_ok,
15098         .tfb_tcp_mtu_chg = bbr_mtu_chg
15099 };
15100
15101 static const char *bbr_stack_names[] = {
15102         __XSTRING(STACKNAME),
15103 #ifdef STACKALIAS
15104         __XSTRING(STACKALIAS),
15105 #endif
15106 };
15107
15108 static bool bbr_mod_inited = false;
15109
15110 static int
15111 tcp_addbbr(module_t mod, int32_t type, void *data)
15112 {
15113         int32_t err = 0;
15114         int num_stacks;
15115
15116         switch (type) {
15117         case MOD_LOAD:
15118                 printf("Attempting to load " __XSTRING(MODNAME) "\n");
15119                 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
15120                     sizeof(struct bbr_sendmap),
15121                     NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
15122                 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
15123                     sizeof(struct tcp_bbr),
15124                     NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
15125                 sysctl_ctx_init(&bbr_sysctl_ctx);
15126                 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
15127                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
15128                     OID_AUTO,
15129 #ifdef STACKALIAS
15130                     __XSTRING(STACKALIAS),
15131 #else
15132                     __XSTRING(STACKNAME),
15133 #endif
15134                     CTLFLAG_RW, 0,
15135                     "");
15136                 if (bbr_sysctl_root == NULL) {
15137                         printf("Failed to add sysctl node\n");
15138                         err = EFAULT;
15139                         goto free_uma;
15140                 }
15141                 bbr_init_sysctls();
15142                 num_stacks = nitems(bbr_stack_names);
15143                 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK,
15144                     bbr_stack_names, &num_stacks);
15145                 if (err) {
15146                         printf("Failed to register %s stack name for "
15147                             "%s module\n", bbr_stack_names[num_stacks],
15148                             __XSTRING(MODNAME));
15149                         sysctl_ctx_free(&bbr_sysctl_ctx);
15150         free_uma:
15151                         uma_zdestroy(bbr_zone);
15152                         uma_zdestroy(bbr_pcb_zone);
15153                         bbr_counter_destroy();
15154                         printf("Failed to register " __XSTRING(MODNAME)
15155                             " module err:%d\n", err);
15156                         return (err);
15157                 }
15158                 tcp_lro_reg_mbufq();
15159                 bbr_mod_inited = true;
15160                 printf(__XSTRING(MODNAME) " is now available\n");
15161                 break;
15162         case MOD_QUIESCE:
15163                 err = deregister_tcp_functions(&__tcp_bbr, true, false);
15164                 break;
15165         case MOD_UNLOAD:
15166                 err = deregister_tcp_functions(&__tcp_bbr, false, true);
15167                 if (err == EBUSY)
15168                         break;
15169                 if (bbr_mod_inited) {
15170                         uma_zdestroy(bbr_zone);
15171                         uma_zdestroy(bbr_pcb_zone);
15172                         sysctl_ctx_free(&bbr_sysctl_ctx);
15173                         bbr_counter_destroy();
15174                         printf(__XSTRING(MODNAME)
15175                             " is now no longer available\n");
15176                         bbr_mod_inited = false;
15177                 }
15178                 tcp_lro_dereg_mbufq();
15179                 err = 0;
15180                 break;
15181         default:
15182                 return (EOPNOTSUPP);
15183         }
15184         return (err);
15185 }
15186
15187 static moduledata_t tcp_bbr = {
15188         .name = __XSTRING(MODNAME),
15189             .evhand = tcp_addbbr,
15190             .priv = 0
15191 };
15192
15193 MODULE_VERSION(MODNAME, 1);
15194 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
15195 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);