sys/netinet/tcp_stacks/bbr.c

   1 /*-
   2  * Copyright (c) 2016-2020 Netflix, Inc.
   3  *
   4  * Redistribution and use in source and binary forms, with or without
   5  * modification, are permitted provided that the following conditions
   6  * are met:
   7  * 1. Redistributions of source code must retain the above copyright
   8  *    notice, this list of conditions and the following disclaimer.
   9  * 2. Redistributions in binary form must reproduce the above copyright
  10  *    notice, this list of conditions and the following disclaimer in the
  11  *    documentation and/or other materials provided with the distribution.
  12  *
  13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  23  * SUCH DAMAGE.
  24  *
  25  */
  26 /**
  27  * Author: Randall Stewart <rrs@netflix.com>
  28  * This work is based on the ACM Queue paper
  29  * BBR - Congestion Based Congestion Control
  30  * and also numerous discussions with Neal, Yuchung and Van.
  31  */
  32
  33 #include <sys/cdefs.h>
  34 __FBSDID("$FreeBSD$");
  35
  36 #include "opt_inet.h"
  37 #include "opt_inet6.h"
  38 #include "opt_ipsec.h"
  39 #include "opt_tcpdebug.h"
  40 #include "opt_ratelimit.h"
  41 #include "opt_kern_tls.h"
  42 #include <sys/param.h>
  43 #include <sys/arb.h>
  44 #include <sys/module.h>
  45 #include <sys/kernel.h>
  46 #include <sys/libkern.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 #ifdef STATS
  61 #include <sys/qmath.h>
  62 #include <sys/tree.h>
  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 <sys/protosw.h>
  75 #include <vm/uma.h>
  76 #include <sys/kern_prefetch.h>
  77
  78 #include <net/route.h>
  79 #include <net/route/nhop.h>
  80 #include <net/vnet.h>
  81
  82 #define TCPSTATES               /* for logging */
  83
  84 #include <netinet/in.h>
  85 #include <netinet/in_kdtrace.h>
  86 #include <netinet/in_pcb.h>
  87 #include <netinet/ip.h>
  88 #include <netinet/ip_icmp.h>    /* required for icmp_var.h */
  89 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
  90 #include <netinet/ip_var.h>
  91 #include <netinet/ip6.h>
  92 #include <netinet6/in6_pcb.h>
  93 #include <netinet6/ip6_var.h>
  94 #define TCPOUTFLAGS
  95 #include <netinet/tcp.h>
  96 #include <netinet/tcp_fsm.h>
  97 #include <netinet/tcp_seq.h>
  98 #include <netinet/tcp_timer.h>
  99 #include <netinet/tcp_var.h>
 100 #include <netinet/tcpip.h>
 101 #include <netinet/tcp_hpts.h>
 102 #include <netinet/cc/cc.h>
 103 #include <netinet/tcp_log_buf.h>
 104 #include <netinet/tcp_ratelimit.h>
 105 #include <netinet/tcp_lro.h>
 106 #ifdef TCPDEBUG
 107 #include <netinet/tcp_debug.h>
 108 #endif                          /* TCPDEBUG */
 109 #ifdef TCP_OFFLOAD
 110 #include <netinet/tcp_offload.h>
 111 #endif
 112 #ifdef INET6
 113 #include <netinet6/tcp6_var.h>
 114 #endif
 115 #include <netinet/tcp_fastopen.h>
 116
 117 #include <netipsec/ipsec_support.h>
 118 #include <net/if.h>
 119 #include <net/if_var.h>
 120 #include <net/ethernet.h>
 121
 122 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 123 #include <netipsec/ipsec.h>
 124 #include <netipsec/ipsec6.h>
 125 #endif                          /* IPSEC */
 126
 127 #include <netinet/udp.h>
 128 #include <netinet/udp_var.h>
 129 #include <machine/in_cksum.h>
 130
 131 #ifdef MAC
 132 #include <security/mac/mac_framework.h>
 133 #endif
 134
 135 #include "sack_filter.h"
 136 #include "tcp_bbr.h"
 137 #include "rack_bbr_common.h"
 138 uma_zone_t bbr_zone;
 139 uma_zone_t bbr_pcb_zone;
 140
 141 struct sysctl_ctx_list bbr_sysctl_ctx;
 142 struct sysctl_oid *bbr_sysctl_root;
 143
 144 #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \
 145         (tv) = (value); \
 146         if ((u_long)(tv) < (u_long)(tvmin)) \
 147                 (tv) = (tvmin); \
 148         if ((u_long)(tv) > (u_long)(tvmax)) \
 149                 (tv) = (tvmax); \
 150 } while(0)
 151
 152 /*#define BBR_INVARIANT 1*/
 153
 154 /*
 155  * initial window
 156  */
 157 static uint32_t bbr_def_init_win = 10;
 158 static int32_t bbr_persist_min = 250000;        /* 250ms */
 159 static int32_t bbr_persist_max = 1000000;       /* 1 Second */
 160 static int32_t bbr_cwnd_may_shrink = 0;
 161 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP;
 162 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT;
 163 static int32_t bbr_hardware_pacing_limit = 8000;
 164 static int32_t bbr_quanta = 3;  /* How much extra quanta do we get? */
 165 static int32_t bbr_no_retran = 0;
 166
 167
 168 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */
 169 static int32_t bbr_max_net_error_cnt = 10;
 170 /* Should the following be dynamic too -- loss wise */
 171 static int32_t bbr_rtt_gain_thresh = 0;
 172 /* Measurement controls */
 173 static int32_t bbr_use_google_algo = 1;
 174 static int32_t bbr_ts_limiting = 1;
 175 static int32_t bbr_ts_can_raise = 0;
 176 static int32_t bbr_do_red = 600;
 177 static int32_t bbr_red_scale = 20000;
 178 static int32_t bbr_red_mul = 1;
 179 static int32_t bbr_red_div = 2;
 180 static int32_t bbr_red_growth_restrict = 1;
 181 static int32_t  bbr_target_is_bbunit = 0;
 182 static int32_t bbr_drop_limit = 0;
 183 /*
 184  * How much gain do we need to see to
 185  * stay in startup?
 186  */
 187 static int32_t bbr_marks_rxt_sack_passed = 0;
 188 static int32_t bbr_start_exit = 25;
 189 static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */
 190 static int32_t bbr_startup_loss_thresh = 2000;  /* 20.00% loss */
 191 static int32_t bbr_hptsi_max_mul = 1;   /* These two mul/div assure a min pacing */
 192 static int32_t bbr_hptsi_max_div = 2;   /* time, 0 means turned off. We need this
 193                                          * if we go back ever to where the pacer
 194                                          * has priority over timers.
 195                                          */
 196 static int32_t bbr_policer_call_from_rack_to = 0;
 197 static int32_t bbr_policer_detection_enabled = 1;
 198 static int32_t bbr_min_measurements_req = 1;    /* We need at least 2
 199                                                  * measurments before we are
 200                                                  * "good" note that 2 == 1.
 201                                                  * This is because we use a >
 202                                                  * comparison. This means if
 203                                                  * min_measure was 0, it takes
 204                                                  * num-measures > min(0) and
 205                                                  * you get 1 measurement and
 206                                                  * you are good. Set to 1, you
 207                                                  * have to have two
 208                                                  * measurements (this is done
 209                                                  * to prevent it from being ok
 210                                                  * to have no measurements). */
 211 static int32_t bbr_no_pacing_until = 4;
 212
 213 static int32_t bbr_min_usec_delta = 20000;      /* 20,000 usecs */
 214 static int32_t bbr_min_peer_delta = 20;         /* 20 units */
 215 static int32_t bbr_delta_percent = 150;         /* 15.0 % */
 216
 217 static int32_t bbr_target_cwnd_mult_limit = 8;
 218 /*
 219  * bbr_cwnd_min_val is the number of
 220  * segments we hold to in the RTT probe
 221  * state typically 4.
 222  */
 223 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS;
 224
 225
 226 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS;
 227
 228 static int32_t bbr_gain_to_target = 1;
 229 static int32_t bbr_gain_gets_extra_too = 1;
 230 /*
 231  * bbr_high_gain is the 2/ln(2) value we need
 232  * to double the sending rate in startup. This
 233  * is used for both cwnd and hptsi gain's.
 234  */
 235 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1;
 236 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1;
 237 static int32_t bbr_use_lower_gain_in_startup = 1;
 238
 239 /* thresholds for reduction on drain in sub-states/drain */
 240 static int32_t bbr_drain_rtt = BBR_SRTT;
 241 static int32_t bbr_drain_floor = 88;
 242 static int32_t google_allow_early_out = 1;
 243 static int32_t google_consider_lost = 1;
 244 static int32_t bbr_drain_drop_mul = 4;
 245 static int32_t bbr_drain_drop_div = 5;
 246 static int32_t bbr_rand_ot = 50;
 247 static int32_t bbr_can_force_probertt = 0;
 248 static int32_t bbr_can_adjust_probertt = 1;
 249 static int32_t bbr_probertt_sets_rtt = 0;
 250 static int32_t bbr_can_use_ts_for_rtt = 1;
 251 static int32_t bbr_is_ratio = 0;
 252 static int32_t bbr_sub_drain_app_limit = 1;
 253 static int32_t bbr_prtt_slam_cwnd = 1;
 254 static int32_t bbr_sub_drain_slam_cwnd = 1;
 255 static int32_t bbr_slam_cwnd_in_main_drain = 1;
 256 static int32_t bbr_filter_len_sec = 6;  /* How long does the rttProp filter
 257                                          * hold */
 258 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4);
 259 /*
 260  * bbr_drain_gain is the reverse of the high_gain
 261  * designed to drain back out the standing queue
 262  * that is formed in startup by causing a larger
 263  * hptsi gain and thus drainging the packets
 264  * in flight.
 265  */
 266 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885;
 267 static int32_t bbr_rttprobe_gain = 192;
 268
 269 /*
 270  * The cwnd_gain is the default cwnd gain applied when
 271  * calculating a target cwnd. Note that the cwnd is
 272  * a secondary factor in the way BBR works (see the
 273  * paper and think about it, it will take some time).
 274  * Basically the hptsi_gain spreads the packets out
 275  * so you never get more than BDP to the peer even
 276  * if the cwnd is high. In our implemenation that
 277  * means in non-recovery/retransmission scenarios
 278  * cwnd will never be reached by the flight-size.
 279  */
 280 static int32_t bbr_cwnd_gain = BBR_UNIT * 2;
 281 static int32_t bbr_tlp_type_to_use = BBR_SRTT;
 282 static int32_t bbr_delack_time = 100000;        /* 100ms in useconds */
 283 static int32_t bbr_sack_not_required = 0;       /* set to one to allow non-sack to use bbr */
 284 static int32_t bbr_initial_bw_bps = 62500;      /* 500kbps in bytes ps */
 285 static int32_t bbr_ignore_data_after_close = 1;
 286 static int16_t bbr_hptsi_gain[] = {
 287         (BBR_UNIT *5 / 4),
 288         (BBR_UNIT * 3 / 4),
 289         BBR_UNIT,
 290         BBR_UNIT,
 291         BBR_UNIT,
 292         BBR_UNIT,
 293         BBR_UNIT,
 294         BBR_UNIT
 295 };
 296 int32_t bbr_use_rack_resend_cheat = 1;
 297 int32_t bbr_sends_full_iwnd = 1;
 298
 299 #define BBR_HPTSI_GAIN_MAX 8
 300 /*
 301  * The BBR module incorporates a number of
 302  * TCP ideas that have been put out into the IETF
 303  * over the last few years:
 304  * - Yuchung Cheng's RACK TCP (for which its named) that
 305  *    will stop us using the number of dup acks and instead
 306  *    use time as the gage of when we retransmit.
 307  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
 308  *    of Dukkipati et.al.
 309  * - Van Jacobson's et.al BBR.
 310  *
 311  * RACK depends on SACK, so if an endpoint arrives that
 312  * cannot do SACK the state machine below will shuttle the
 313  * connection back to using the "default" TCP stack that is
 314  * in FreeBSD.
 315  *
 316  * To implement BBR and RACK the original TCP stack was first decomposed
 317  * into a functional state machine with individual states
 318  * for each of the possible TCP connection states. The do_segement
 319  * functions role in life is to mandate the connection supports SACK
 320  * initially and then assure that the RACK state matches the conenction
 321  * state before calling the states do_segment function. Data processing
 322  * of inbound segments also now happens in the hpts_do_segment in general
 323  * with only one exception. This is so we can keep the connection on
 324  * a single CPU.
 325  *
 326  * Each state is simplified due to the fact that the original do_segment
 327  * has been decomposed and we *know* what state we are in (no
 328  * switches on the state) and all tests for SACK are gone. This
 329  * greatly simplifies what each state does.
 330  *
 331  * TCP output is also over-written with a new version since it
 332  * must maintain the new rack scoreboard and has had hptsi
 333  * integrated as a requirment. Still todo is to eliminate the
 334  * use of the callout_() system and use the hpts for all
 335  * timers as well.
 336  */
 337 static uint32_t bbr_rtt_probe_time = 200000;    /* 200ms in micro seconds */
 338 static uint32_t bbr_rtt_probe_cwndtarg = 4;     /* How many mss's outstanding */
 339 static const int32_t bbr_min_req_free = 2;      /* The min we must have on the
 340                                                  * free list */
 341 static int32_t bbr_tlp_thresh = 1;
 342 static int32_t bbr_reorder_thresh = 2;
 343 static int32_t bbr_reorder_fade = 60000000;     /* 0 - never fade, def
 344                                                  * 60,000,000 - 60 seconds */
 345 static int32_t bbr_pkt_delay = 1000;
 346 static int32_t bbr_min_to = 1000;       /* Number of usec's minimum timeout */
 347 static int32_t bbr_incr_timers = 1;
 348
 349 static int32_t bbr_tlp_min = 10000;     /* 10ms in usecs */
 350 static int32_t bbr_delayed_ack_time = 200000;   /* 200ms in usecs */
 351 static int32_t bbr_exit_startup_at_loss = 1;
 352
 353 /*
 354  * bbr_lt_bw_ratio is 1/8th
 355  * bbr_lt_bw_diff is  < 4 Kbit/sec
 356  */
 357 static uint64_t bbr_lt_bw_diff = 4000 / 8;      /* In bytes per second */
 358 static uint64_t bbr_lt_bw_ratio = 8;    /* For 1/8th */
 359 static uint32_t bbr_lt_bw_max_rtts = 48;        /* How many rtt's do we use
 360                                                  * the lt_bw for */
 361 static uint32_t bbr_lt_intvl_min_rtts = 4;      /* Min num of RTT's to measure
 362                                                  * lt_bw */
 363 static int32_t bbr_lt_intvl_fp = 0;             /* False positive epoch diff */
 364 static int32_t bbr_lt_loss_thresh = 196;        /* Lost vs delivered % */
 365 static int32_t bbr_lt_fd_thresh = 100;          /* false detection % */
 366
 367 static int32_t bbr_verbose_logging = 0;
 368 /*
 369  * Currently regular tcp has a rto_min of 30ms
 370  * the backoff goes 12 times so that ends up
 371  * being a total of 122.850 seconds before a
 372  * connection is killed.
 373  */
 374 static int32_t bbr_rto_min_ms = 30;     /* 30ms same as main freebsd */
 375 static int32_t bbr_rto_max_sec = 4;     /* 4 seconds */
 376
 377 /****************************************************/
 378 /* DEFAULT TSO SIZING  (cpu performance impacting)  */
 379 /****************************************************/
 380 /* What amount is our formula using to get TSO size */
 381 static int32_t bbr_hptsi_per_second = 1000;
 382
 383 /*
 384  * For hptsi under bbr_cross_over connections what is delay
 385  * target 7ms (in usec) combined with a seg_max of 2
 386  * gets us close to identical google behavior in
 387  * TSO size selection (possibly more 1MSS sends).
 388  */
 389 static int32_t bbr_hptsi_segments_delay_tar = 7000;
 390
 391 /* Does pacing delay include overhead's in its time calculations? */
 392 static int32_t bbr_include_enet_oh = 0;
 393 static int32_t bbr_include_ip_oh = 1;
 394 static int32_t bbr_include_tcp_oh = 1;
 395 static int32_t bbr_google_discount = 10;
 396
 397 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */
 398 static int32_t bbr_state_is_pkt_epoch = 0;
 399 static int32_t bbr_state_drain_2_tar = 1;
 400 /* What is the max the 0 - bbr_cross_over MBPS TSO target
 401  * can reach using our delay target. Note that this
 402  * value becomes the floor for the cross over
 403  * algorithm.
 404  */
 405 static int32_t bbr_hptsi_segments_max = 2;
 406 static int32_t bbr_hptsi_segments_floor = 1;
 407 static int32_t bbr_hptsi_utter_max = 0;
 408
 409 /* What is the min the 0 - bbr_cross-over MBPS  TSO target can be */
 410 static int32_t bbr_hptsi_bytes_min = 1460;
 411 static int32_t bbr_all_get_min = 0;
 412
 413 /* Cross over point from algo-a to algo-b */
 414 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS;
 415
 416 /* Do we deal with our restart state? */
 417 static int32_t bbr_uses_idle_restart = 0;
 418 static int32_t bbr_idle_restart_threshold = 100000;     /* 100ms in useconds */
 419
 420 /* Do we allow hardware pacing? */
 421 static int32_t bbr_allow_hdwr_pacing = 0;
 422 static int32_t bbr_hdwr_pace_adjust = 2;        /* multipler when we calc the tso size */
 423 static int32_t bbr_hdwr_pace_floor = 1;
 424 static int32_t bbr_hdwr_pacing_delay_cnt = 10;
 425
 426 /****************************************************/
 427 static int32_t bbr_resends_use_tso = 0;
 428 static int32_t bbr_tlp_max_resend = 2;
 429 static int32_t bbr_sack_block_limit = 128;
 430
 431 #define  BBR_MAX_STAT 19
 432 counter_u64_t bbr_state_time[BBR_MAX_STAT];
 433 counter_u64_t bbr_state_lost[BBR_MAX_STAT];
 434 counter_u64_t bbr_state_resend[BBR_MAX_STAT];
 435 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE];
 436 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE];
 437 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE];
 438 counter_u64_t bbr_flows_whdwr_pacing;
 439 counter_u64_t bbr_flows_nohdwr_pacing;
 440
 441 counter_u64_t bbr_nohdwr_pacing_enobuf;
 442 counter_u64_t bbr_hdwr_pacing_enobuf;
 443
 444 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr);
 445
 446 /*
 447  * Static defintions we need for forward declarations.
 448  */
 449 static uint32_t
 450 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain,
 451     uint32_t useconds_time, uint64_t bw);
 452 static uint32_t
 453 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain);
 454 static void
 455      bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win);
 456 static void
 457 bbr_set_probebw_gains(struct tcp_bbr *bbr,  uint32_t cts, uint32_t losses);
 458 static void
 459 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line,
 460                     int dolog);
 461 static uint32_t
 462 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain);
 463 static void
 464 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch,
 465                  int32_t pkt_epoch, uint32_t losses);
 466 static uint32_t
 467 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm);
 468 static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp);
 469 static uint32_t
 470 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
 471     struct bbr_sendmap *rsm, uint32_t srtt,
 472     uint32_t cts);
 473 static void
 474 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts,
 475     int32_t line);
 476 static void
 477      bbr_set_state_target(struct tcp_bbr *bbr, int line);
 478 static void
 479      bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line);
 480
 481 static void
 482      bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line);
 483
 484 static void
 485      tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts);
 486
 487 static void
 488      bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts);
 489
 490 static void
 491      bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt,
 492                          uint32_t line, uint8_t is_start, uint16_t set);
 493
 494 static struct bbr_sendmap *
 495             bbr_find_lowest_rsm(struct tcp_bbr *bbr);
 496 static __inline uint32_t
 497 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type);
 498 static void
 499      bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which);
 500
 501 static void
 502 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
 503     uint32_t thresh, uint32_t to);
 504 static void
 505      bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag);
 506
 507 static void
 508 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot,
 509     uint32_t del_by, uint32_t cts, uint32_t sloton, uint32_t prev_delay);
 510
 511 static void
 512 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr,
 513     uint32_t cts, int32_t line);
 514 static void
 515      bbr_stop_all_timers(struct tcpcb *tp);
 516 static void
 517      bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts);
 518 static void
 519      bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts);
 520 static void
 521      bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts);
 522
 523
 524 static void
 525 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
 526     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod);
 527
 528 static inline uint8_t
 529 bbr_state_val(struct tcp_bbr *bbr)
 530 {
 531         return(bbr->rc_bbr_substate);
 532 }
 533
 534 static inline uint32_t
 535 get_min_cwnd(struct tcp_bbr *bbr)
 536 {
 537         int mss;
 538
 539         mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
 540         if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED)
 541                 return (bbr_cwnd_min_val_hs * mss);
 542         else
 543                 return (bbr_cwnd_min_val * mss);
 544 }
 545
 546 static uint32_t
 547 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr)
 548 {
 549         uint64_t srtt, var;
 550         uint64_t ret_val;
 551
 552         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
 553         if (tp->t_srtt == 0) {
 554                 srtt = (uint64_t)BBR_INITIAL_RTO;
 555                 var = 0;
 556         } else {
 557                 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
 558                 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT);
 559         }
 560         TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]),
 561             bbr_persist_min, bbr_persist_max);
 562         return ((uint32_t)ret_val);
 563 }
 564
 565 static uint32_t
 566 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
 567 {
 568         /*
 569          * Start the FR timer, we do this based on getting the first one in
 570          * the rc_tmap. Note that if its NULL we must stop the timer. in all
 571          * events we need to stop the running timer (if its running) before
 572          * starting the new one.
 573          */
 574         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
 575         int32_t idx;
 576         int32_t is_tlp_timer = 0;
 577         struct bbr_sendmap *rsm;
 578
 579         if (bbr->rc_all_timers_stopped) {
 580                 /* All timers have been stopped none are to run */
 581                 return (0);
 582         }
 583         if (bbr->rc_in_persist) {
 584                 /* We can't start any timer in persists */
 585                 return (bbr_get_persists_timer_val(tp, bbr));
 586         }
 587         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 588         if ((rsm == NULL) ||
 589             ((tp->t_flags & TF_SACK_PERMIT) == 0) ||
 590             (tp->t_state < TCPS_ESTABLISHED)) {
 591                 /* Nothing on the send map */
 592 activate_rxt:
 593                 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
 594                         uint64_t tov;
 595
 596                         time_since_sent = 0;
 597                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 598                         if (rsm) {
 599                                 idx = rsm->r_rtr_cnt - 1;
 600                                 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
 601                                         tstmp_touse = rsm->r_tim_lastsent[idx];
 602                                 else
 603                                         tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
 604                                 if (TSTMP_GT(tstmp_touse, cts))
 605                                     time_since_sent = cts - tstmp_touse;
 606                         }
 607                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
 608                         if (tp->t_srtt == 0)
 609                                 tov = BBR_INITIAL_RTO;
 610                         else
 611                                 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) +
 612                                     ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT);
 613                         if (tp->t_rxtshift)
 614                                 tov *= tcp_backoff[tp->t_rxtshift];
 615                         if (tov > time_since_sent)
 616                                 tov -= time_since_sent;
 617                         else
 618                                 tov = bbr->r_ctl.rc_min_to;
 619                         TCPT_RANGESET_NOSLOP(to, tov,
 620                             (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC),
 621                             (bbr->rc_max_rto_sec * USECS_IN_SECOND));
 622                         bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to);
 623                         return (to);
 624                 }
 625                 return (0);
 626         }
 627         if (rsm->r_flags & BBR_ACKED) {
 628                 rsm = bbr_find_lowest_rsm(bbr);
 629                 if (rsm == NULL) {
 630                         /* No lowest? */
 631                         goto activate_rxt;
 632                 }
 633         }
 634         /* Convert from ms to usecs */
 635         if (rsm->r_flags & BBR_SACK_PASSED) {
 636                 if ((tp->t_flags & TF_SENTFIN) &&
 637                     ((tp->snd_max - tp->snd_una) == 1) &&
 638                     (rsm->r_flags & BBR_HAS_FIN)) {
 639                         /*
 640                          * We don't start a bbr rack timer if all we have is
 641                          * a FIN outstanding.
 642                          */
 643                         goto activate_rxt;
 644                 }
 645                 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK);
 646                 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm);
 647                 idx = rsm->r_rtr_cnt - 1;
 648                 exp = rsm->r_tim_lastsent[idx] + thresh;
 649                 if (SEQ_GEQ(exp, cts)) {
 650                         to = exp - cts;
 651                         if (to < bbr->r_ctl.rc_min_to) {
 652                                 to = bbr->r_ctl.rc_min_to;
 653                         }
 654                 } else {
 655                         to = bbr->r_ctl.rc_min_to;
 656                 }
 657         } else {
 658                 /* Ok we need to do a TLP not RACK */
 659                 if (bbr->rc_tlp_in_progress != 0) {
 660                         /*
 661                          * The previous send was a TLP.
 662                          */
 663                         goto activate_rxt;
 664                 }
 665                 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
 666                 if (rsm == NULL) {
 667                         /* We found no rsm to TLP with. */
 668                         goto activate_rxt;
 669                 }
 670                 if (rsm->r_flags & BBR_HAS_FIN) {
 671                         /* If its a FIN we don't do TLP */
 672                         rsm = NULL;
 673                         goto activate_rxt;
 674                 }
 675                 time_since_sent = 0;
 676                 idx = rsm->r_rtr_cnt - 1;
 677                 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
 678                         tstmp_touse = rsm->r_tim_lastsent[idx];
 679                 else
 680                         tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
 681                 if (TSTMP_GT(tstmp_touse, cts))
 682                     time_since_sent = cts - tstmp_touse;
 683                 is_tlp_timer = 1;
 684                 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use);
 685                 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts);
 686                 if (thresh > time_since_sent)
 687                         to = thresh - time_since_sent;
 688                 else
 689                         to = bbr->r_ctl.rc_min_to;
 690                 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
 691                         /*
 692                          * If the TLP time works out to larger than the max
 693                          * RTO lets not do TLP.. just RTO.
 694                          */
 695                         goto activate_rxt;
 696                 }
 697                 if ((bbr->rc_tlp_rtx_out == 1) &&
 698                     (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) {
 699                         /*
 700                          * Second retransmit of the same TLP
 701                          * lets not.
 702                          */
 703                         bbr->rc_tlp_rtx_out = 0;
 704                         goto activate_rxt;
 705                 }
 706                 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) {
 707                         /*
 708                          * The tail is no longer the last one I did a probe
 709                          * on
 710                          */
 711                         bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
 712                         bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
 713                 }
 714         }
 715         if (is_tlp_timer == 0) {
 716                 BBR_STAT_INC(bbr_to_arm_rack);
 717                 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
 718         } else {
 719                 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to);
 720                 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
 721                         /*
 722                          * We have exceeded how many times we can retran the
 723                          * current TLP timer, switch to the RTO timer.
 724                          */
 725                         goto activate_rxt;
 726                 } else {
 727                         BBR_STAT_INC(bbr_to_arm_tlp);
 728                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
 729                 }
 730         }
 731         return (to);
 732 }
 733
 734 static inline int32_t
 735 bbr_minseg(struct tcp_bbr *bbr)
 736 {
 737         return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options);
 738 }
 739
 740 static void
 741 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len)
 742 {
 743         struct inpcb *inp;
 744         struct hpts_diag diag;
 745         uint32_t delayed_ack = 0;
 746         uint32_t left = 0;
 747         uint32_t hpts_timeout;
 748         uint8_t stopped;
 749         int32_t delay_calc = 0;
 750         uint32_t prev_delay = 0;
 751
 752         inp = tp->t_inpcb;
 753         if (inp->inp_in_hpts) {
 754                 /* A previous call is already set up */
 755                 return;
 756         }
 757         if ((tp->t_state == TCPS_CLOSED) ||
 758             (tp->t_state == TCPS_LISTEN)) {
 759                 return;
 760         }
 761         stopped = bbr->rc_tmr_stopped;
 762         if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) {
 763                 left = bbr->r_ctl.rc_timer_exp - cts;
 764         }
 765         bbr->r_ctl.rc_hpts_flags = 0;
 766         bbr->r_ctl.rc_timer_exp = 0;
 767         prev_delay = bbr->r_ctl.rc_last_delay_val;
 768         if (bbr->r_ctl.rc_last_delay_val &&
 769             (slot == 0)) {
 770                 /*
 771                  * If a previous pacer delay was in place we
 772                  * are not coming from the output side (where
 773                  * we calculate a delay, more likely a timer).
 774                  */
 775                 slot = bbr->r_ctl.rc_last_delay_val;
 776                 if (TSTMP_GT(cts, bbr->rc_pacer_started)) {
 777                         /* Compensate for time passed  */
 778                         delay_calc = cts - bbr->rc_pacer_started;
 779                         if (delay_calc <= slot)
 780                                 slot -= delay_calc;
 781                 }
 782         }
 783         /* Do we have early to make up for by pushing out the pacing time? */
 784         if (bbr->r_agg_early_set) {
 785                 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2);
 786                 slot += bbr->r_ctl.rc_agg_early;
 787                 bbr->r_ctl.rc_agg_early = 0;
 788                 bbr->r_agg_early_set = 0;
 789         }
 790         /* Are we running a total debt that needs to be compensated for? */
 791         if (bbr->r_ctl.rc_hptsi_agg_delay) {
 792                 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) {
 793                         /* We nuke the delay */
 794                         slot -= bbr->r_ctl.rc_hptsi_agg_delay;
 795                         bbr->r_ctl.rc_hptsi_agg_delay = 0;
 796                 } else {
 797                         /* We nuke some of the delay, put in a minimal 100usecs  */
 798                         bbr->r_ctl.rc_hptsi_agg_delay -= slot;
 799                         bbr->r_ctl.rc_last_delay_val = slot = 100;
 800                 }
 801         }
 802         bbr->r_ctl.rc_last_delay_val = slot;
 803         hpts_timeout = bbr_timer_start(tp, bbr, cts);
 804         if (tp->t_flags & TF_DELACK) {
 805                 if (bbr->rc_in_persist == 0) {
 806                         delayed_ack = bbr_delack_time;
 807                 } else {
 808                         /*
 809                          * We are in persists and have
 810                          * gotten a new data element.
 811                          */
 812                         if (hpts_timeout > bbr_delack_time) {
 813                                 /*
 814                                  * Lets make the persists timer (which acks)
 815                                  * be the smaller of hpts_timeout and bbr_delack_time.
 816                                  */
 817                                 hpts_timeout = bbr_delack_time;
 818                         }
 819                 }
 820         }
 821         if (delayed_ack &&
 822             ((hpts_timeout == 0) ||
 823              (delayed_ack < hpts_timeout))) {
 824                 /* We need a Delayed ack timer */
 825                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
 826                 hpts_timeout = delayed_ack;
 827         }
 828         if (slot) {
 829                 /* Mark that we have a pacing timer up */
 830                 BBR_STAT_INC(bbr_paced_segments);
 831                 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
 832         }
 833         /*
 834          * If no timers are going to run and we will fall off thfe hptsi
 835          * wheel, we resort to a keep-alive timer if its configured.
 836          */
 837         if ((hpts_timeout == 0) &&
 838             (slot == 0)) {
 839                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
 840                     (tp->t_state <= TCPS_CLOSING)) {
 841                         /*
 842                          * Ok we have no timer (persists, rack, tlp, rxt  or
 843                          * del-ack), we don't have segments being paced. So
 844                          * all that is left is the keepalive timer.
 845                          */
 846                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
 847                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
 848                         } else {
 849                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
 850                         }
 851                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
 852                 }
 853         }
 854         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
 855             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
 856                 /*
 857                  * RACK, TLP, persists and RXT timers all are restartable
 858                  * based on actions input .. i.e we received a packet (ack
 859                  * or sack) and that changes things (rw, or snd_una etc).
 860                  * Thus we can restart them with a new value. For
 861                  * keep-alive, delayed_ack we keep track of what was left
 862                  * and restart the timer with a smaller value.
 863                  */
 864                 if (left < hpts_timeout)
 865                         hpts_timeout = left;
 866         }
 867         if (bbr->r_ctl.rc_incr_tmrs && slot &&
 868             (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
 869                 /*
 870                  * If configured to do so, and the timer is either
 871                  * the TLP or RXT timer, we need to increase the timeout
 872                  * by the pacing time. Consider the bottleneck at my
 873                  * machine as an example, we are sending something
 874                  * to start a TLP on. The last packet won't be emitted
 875                  * fully until the pacing time (the bottleneck will hold
 876                  * the data in place). Once the packet is emitted that
 877                  * is when we want to start waiting for the TLP. This
 878                  * is most evident with hardware pacing (where the nic
 879                  * is holding the packet(s) before emitting). But it
 880                  * can also show up in the network so we do it for all
 881                  * cases. Technically we would take off one packet from
 882                  * this extra delay but this is easier and being more
 883                  * conservative is probably better.
 884                  */
 885                 hpts_timeout += slot;
 886         }
 887         if (hpts_timeout) {
 888                 /*
 889                  * Hack alert for now we can't time-out over 2147 seconds (a
 890                  * bit more than 35min)
 891                  */
 892                 if (hpts_timeout > 0x7ffffffe)
 893                         hpts_timeout = 0x7ffffffe;
 894                 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
 895         } else
 896                 bbr->r_ctl.rc_timer_exp = 0;
 897         if ((slot) &&
 898             (bbr->rc_use_google ||
 899              bbr->output_error_seen ||
 900              (slot <= hpts_timeout))  ) {
 901                 /*
 902                  * Tell LRO that it can queue packets while
 903                  * we pace.
 904                  */
 905                 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
 906                 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
 907                     (bbr->rc_cwnd_limited == 0)) {
 908                         /*
 909                          * If we are not cwnd limited and we
 910                          * are running a rack timer we put on
 911                          * the do not disturbe even for sack.
 912                          */
 913                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
 914                 } else
 915                         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
 916                 bbr->rc_pacer_started = cts;
 917
 918                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
 919                                            __LINE__, &diag);
 920                 bbr->rc_timer_first = 0;
 921                 bbr->bbr_timer_src = frm;
 922                 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1);
 923                 bbr_log_hpts_diag(bbr, cts, &diag);
 924         } else if (hpts_timeout) {
 925                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
 926                                            __LINE__, &diag);
 927                 /*
 928                  * We add the flag here as well if the slot is set,
 929                  * since hpts will call in to clear the queue first before
 930                  * calling the output routine (which does our timers).
 931                  * We don't want to set the flag if its just a timer
 932                  * else the arrival of data might (that causes us
 933                  * to send more) might get delayed. Imagine being
 934                  * on a keep-alive timer and a request comes in for
 935                  * more data.
 936                  */
 937                 if (slot)
 938                         bbr->rc_pacer_started = cts;
 939                 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
 940                     (bbr->rc_cwnd_limited == 0)) {
 941                         /*
 942                          * For a rack timer, don't wake us even
 943                          * if a sack arrives as long as we are
 944                          * not cwnd limited.
 945                          */
 946                         bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
 947                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
 948                 } else {
 949                         /* All other timers wake us up */
 950                         bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
 951                         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
 952                 }
 953                 bbr->bbr_timer_src = frm;
 954                 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0);
 955                 bbr_log_hpts_diag(bbr, cts, &diag);
 956                 bbr->rc_timer_first = 1;
 957         }
 958         bbr->rc_tmr_stopped = 0;
 959         bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay);
 960 }
 961
 962 static void
 963 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb)
 964 {
 965         /*
 966          * We received an ack, and then did not call send or were bounced
 967          * out due to the hpts was running. Now a timer is up as well, is it
 968          * the right timer?
 969          */
 970         struct inpcb *inp;
 971         struct bbr_sendmap *rsm;
 972         uint32_t hpts_timeout;
 973         int tmr_up;
 974
 975         tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
 976         if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
 977                 return;
 978         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 979         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
 980             (tmr_up == PACE_TMR_RXT)) {
 981                 /* Should be an RXT */
 982                 return;
 983         }
 984         inp = bbr->rc_inp;
 985         if (rsm == NULL) {
 986                 /* Nothing outstanding? */
 987                 if (tp->t_flags & TF_DELACK) {
 988                         if (tmr_up == PACE_TMR_DELACK)
 989                                 /*
 990                                  * We are supposed to have delayed ack up
 991                                  * and we do
 992                                  */
 993                                 return;
 994                 } else if (sbavail(&inp->inp_socket->so_snd) &&
 995                     (tmr_up == PACE_TMR_RXT)) {
 996                         /*
 997                          * if we hit enobufs then we would expect the
 998                          * possiblity of nothing outstanding and the RXT up
 999                          * (and the hptsi timer).
1000                          */
1001                         return;
1002                 } else if (((V_tcp_always_keepalive ||
1003                             inp->inp_socket->so_options & SO_KEEPALIVE) &&
1004                             (tp->t_state <= TCPS_CLOSING)) &&
1005                             (tmr_up == PACE_TMR_KEEP) &&
1006                     (tp->snd_max == tp->snd_una)) {
1007                         /* We should have keep alive up and we do */
1008                         return;
1009                 }
1010         }
1011         if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) {
1012                 if ((tp->t_flags & TF_SENTFIN) &&
1013                     ((tp->snd_max - tp->snd_una) == 1) &&
1014                     (rsm->r_flags & BBR_HAS_FIN)) {
1015                         /* needs to be a RXT */
1016                         if (tmr_up == PACE_TMR_RXT)
1017                                 return;
1018                         else
1019                                 goto wrong_timer;
1020                 } else if (tmr_up == PACE_TMR_RACK)
1021                         return;
1022                 else
1023                         goto wrong_timer;
1024         } else if (rsm && (tmr_up == PACE_TMR_RACK)) {
1025                 /* Rack timer has priority if we have data out */
1026                 return;
1027         } else if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1028                     ((tmr_up == PACE_TMR_TLP) ||
1029             (tmr_up == PACE_TMR_RXT))) {
1030                 /*
1031                  * Either a TLP or RXT is fine if no sack-passed is in place
1032                  * and data is outstanding.
1033                  */
1034                 return;
1035         } else if (tmr_up == PACE_TMR_DELACK) {
1036                 /*
1037                  * If the delayed ack was going to go off before the
1038                  * rtx/tlp/rack timer were going to expire, then that would
1039                  * be the timer in control. Note we don't check the time
1040                  * here trusting the code is correct.
1041                  */
1042                 return;
1043         }
1044         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1045             ((tmr_up == PACE_TMR_RXT) ||
1046              (tmr_up == PACE_TMR_TLP) ||
1047              (tmr_up == PACE_TMR_RACK))) {
1048                 /*
1049                  * We have outstanding data and
1050                  * we *do* have a RACK, TLP or RXT
1051                  * timer running. We won't restart
1052                  * anything here since thats probably ok we
1053                  * will get called with some timer here shortly.
1054                  */
1055                 return;
1056         }
1057         /*
1058          * Ok the timer originally started is not what we want now. We will
1059          * force the hpts to be stopped if any, and restart with the slot
1060          * set to what was in the saved slot.
1061          */
1062 wrong_timer:
1063         if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) {
1064                 if (inp->inp_in_hpts)
1065                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
1066                 bbr_timer_cancel(bbr, __LINE__, cts);
1067                 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val,
1068                     0);
1069         } else {
1070                 /*
1071                  * Output is hptsi so we just need to switch the type of
1072                  * timer. We don't bother with keep-alive, since when we
1073                  * jump through the output, it will start the keep-alive if
1074                  * nothing is sent.
1075                  *
1076                  * We only need a delayed-ack added and or the hpts_timeout.
1077                  */
1078                 hpts_timeout = bbr_timer_start(tp, bbr, cts);
1079                 if (tp->t_flags & TF_DELACK) {
1080                         if (hpts_timeout == 0) {
1081                                 hpts_timeout = bbr_delack_time;
1082                                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1083                         }
1084                         else if (hpts_timeout > bbr_delack_time) {
1085                                 hpts_timeout = bbr_delack_time;
1086                                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1087                         }
1088                 }
1089                 if (hpts_timeout) {
1090                         if (hpts_timeout > 0x7ffffffe)
1091                                 hpts_timeout = 0x7ffffffe;
1092                         bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
1093                 }
1094         }
1095 }
1096
1097 int32_t bbr_clear_lost = 0;
1098
1099 /*
1100  * Considers the two time values now (cts) and earlier.
1101  * If cts is smaller than earlier, we could have
1102  * had a sequence wrap (our counter wraps every
1103  * 70 min or so) or it could be just clock skew
1104  * getting us two differnt time values. Clock skew
1105  * will show up within 10ms or so. So in such
1106  * a case (where cts is behind earlier time by
1107  * less than 10ms) we return 0. Otherwise we
1108  * return the true difference between them.
1109  */
1110 static inline uint32_t
1111 bbr_calc_time(uint32_t cts, uint32_t earlier_time) {
1112         /*
1113          * Given two timestamps, the current time stamp cts, and some other
1114          * time-stamp taken in theory earlier return the difference. The
1115          * trick is here sometimes locking will get the other timestamp
1116          * after the cts. If this occurs we need to return 0.
1117          */
1118         if (TSTMP_GEQ(cts, earlier_time))
1119                 return (cts - earlier_time);
1120         /*
1121          * cts is behind earlier_time if its less than 10ms consider it 0.
1122          * If its more than 10ms difference then we had a time wrap. Else
1123          * its just the normal locking foo. I wonder if we should not go to
1124          * 64bit TS and get rid of this issue.
1125          */
1126         if (TSTMP_GEQ((cts + 10000), earlier_time))
1127                 return (0);
1128         /*
1129          * Ok the time must have wrapped. So we need to answer a large
1130          * amount of time, which the normal subtraction should do.
1131          */
1132         return (cts - earlier_time);
1133 }
1134
1135
1136
1137 static int
1138 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS)
1139 {
1140         uint32_t stat;
1141         int32_t error;
1142
1143         error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t));
1144         if (error || req->newptr == NULL)
1145                 return error;
1146
1147         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
1148         if (error)
1149                 return (error);
1150         if (stat == 1) {
1151 #ifdef BBR_INVARIANTS
1152                 printf("Clearing BBR lost counters\n");
1153 #endif
1154                 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT);
1155                 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT);
1156                 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT);
1157         } else if (stat == 2) {
1158 #ifdef BBR_INVARIANTS
1159                 printf("Clearing BBR option counters\n");
1160 #endif
1161                 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE);
1162         } else if (stat == 3) {
1163 #ifdef BBR_INVARIANTS
1164                 printf("Clearing BBR stats counters\n");
1165 #endif
1166                 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE);
1167         } else if (stat == 4) {
1168 #ifdef BBR_INVARIANTS
1169                 printf("Clearing BBR out-size counters\n");
1170 #endif
1171                 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE);
1172         }
1173         bbr_clear_lost = 0;
1174         return (0);
1175 }
1176
1177 static void
1178 bbr_init_sysctls(void)
1179 {
1180         struct sysctl_oid *bbr_probertt;
1181         struct sysctl_oid *bbr_hptsi;
1182         struct sysctl_oid *bbr_measure;
1183         struct sysctl_oid *bbr_cwnd;
1184         struct sysctl_oid *bbr_timeout;
1185         struct sysctl_oid *bbr_states;
1186         struct sysctl_oid *bbr_startup;
1187         struct sysctl_oid *bbr_policer;
1188
1189         /* Probe rtt controls */
1190         bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1191             SYSCTL_CHILDREN(bbr_sysctl_root),
1192             OID_AUTO,
1193             "probertt",
1194             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1195             "");
1196         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1197             SYSCTL_CHILDREN(bbr_probertt),
1198             OID_AUTO, "gain", CTLFLAG_RW,
1199             &bbr_rttprobe_gain, 192,
1200             "What is the filter gain drop in probe_rtt (0=disable)?");
1201         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1202             SYSCTL_CHILDREN(bbr_probertt),
1203             OID_AUTO, "cwnd", CTLFLAG_RW,
1204             &bbr_rtt_probe_cwndtarg, 4,
1205             "How many mss's are outstanding during probe-rtt");
1206         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1207             SYSCTL_CHILDREN(bbr_probertt),
1208             OID_AUTO, "int", CTLFLAG_RW,
1209             &bbr_rtt_probe_limit, 4000000,
1210             "If RTT has not shrank in this many micro-seconds enter probe-rtt");
1211         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1212             SYSCTL_CHILDREN(bbr_probertt),
1213             OID_AUTO, "mintime", CTLFLAG_RW,
1214             &bbr_rtt_probe_time, 200000,
1215             "How many microseconds in probe-rtt");
1216         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1217             SYSCTL_CHILDREN(bbr_probertt),
1218             OID_AUTO, "filter_len_sec", CTLFLAG_RW,
1219             &bbr_filter_len_sec, 6,
1220             "How long in seconds does the rttProp filter run?");
1221         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1222             SYSCTL_CHILDREN(bbr_probertt),
1223             OID_AUTO, "drain_rtt", CTLFLAG_RW,
1224             &bbr_drain_rtt, BBR_SRTT,
1225             "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?");
1226         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1227             SYSCTL_CHILDREN(bbr_probertt),
1228             OID_AUTO, "can_force", CTLFLAG_RW,
1229             &bbr_can_force_probertt, 0,
1230             "If we keep setting new low rtt's but delay going in probe-rtt can we force in??");
1231         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1232             SYSCTL_CHILDREN(bbr_probertt),
1233             OID_AUTO, "enter_sets_force", CTLFLAG_RW,
1234             &bbr_probertt_sets_rtt, 0,
1235             "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?");
1236         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1237             SYSCTL_CHILDREN(bbr_probertt),
1238             OID_AUTO, "can_adjust", CTLFLAG_RW,
1239             &bbr_can_adjust_probertt, 1,
1240             "Can we dynamically adjust the probe-rtt limits and times?");
1241         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1242             SYSCTL_CHILDREN(bbr_probertt),
1243             OID_AUTO, "is_ratio", CTLFLAG_RW,
1244             &bbr_is_ratio, 0,
1245             "is the limit to filter a ratio?");
1246         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1247             SYSCTL_CHILDREN(bbr_probertt),
1248             OID_AUTO, "use_cwnd", CTLFLAG_RW,
1249             &bbr_prtt_slam_cwnd, 0,
1250             "Should we set/recover cwnd?");
1251         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1252             SYSCTL_CHILDREN(bbr_probertt),
1253             OID_AUTO, "can_use_ts", CTLFLAG_RW,
1254             &bbr_can_use_ts_for_rtt, 1,
1255             "Can we use the ms timestamp if available for retransmistted rtt calculations?");
1256
1257         /* Pacing controls */
1258         bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1259             SYSCTL_CHILDREN(bbr_sysctl_root),
1260             OID_AUTO,
1261             "pacing",
1262             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1263             "");
1264         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1265             SYSCTL_CHILDREN(bbr_hptsi),
1266             OID_AUTO, "hw_pacing", CTLFLAG_RW,
1267             &bbr_allow_hdwr_pacing, 1,
1268             "Do we allow hardware pacing?");
1269         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1270             SYSCTL_CHILDREN(bbr_hptsi),
1271             OID_AUTO, "hw_pacing_limit", CTLFLAG_RW,
1272             &bbr_hardware_pacing_limit, 4000,
1273             "Do we have a limited number of connections for pacing chelsio (0=no limit)?");
1274         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1275             SYSCTL_CHILDREN(bbr_hptsi),
1276             OID_AUTO, "hw_pacing_adj", CTLFLAG_RW,
1277             &bbr_hdwr_pace_adjust, 2,
1278             "Multiplier to calculated tso size?");
1279         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1280             SYSCTL_CHILDREN(bbr_hptsi),
1281             OID_AUTO, "hw_pacing_floor", CTLFLAG_RW,
1282             &bbr_hdwr_pace_floor, 1,
1283             "Do we invoke the hardware pacing floor?");
1284         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1285             SYSCTL_CHILDREN(bbr_hptsi),
1286             OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW,
1287             &bbr_hdwr_pacing_delay_cnt, 10,
1288             "How many packets must be sent after hdwr pacing is enabled");
1289         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1290             SYSCTL_CHILDREN(bbr_hptsi),
1291             OID_AUTO, "bw_cross", CTLFLAG_RW,
1292             &bbr_cross_over, 3000000,
1293             "What is the point where we cross over to linux like TSO size set");
1294         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1295             SYSCTL_CHILDREN(bbr_hptsi),
1296             OID_AUTO, "seg_deltarg", CTLFLAG_RW,
1297             &bbr_hptsi_segments_delay_tar, 7000,
1298             "What is the worse case delay target for hptsi < 48Mbp connections");
1299         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1300             SYSCTL_CHILDREN(bbr_hptsi),
1301             OID_AUTO, "enet_oh", CTLFLAG_RW,
1302             &bbr_include_enet_oh, 0,
1303             "Do we include the ethernet overhead in calculating pacing delay?");
1304         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1305             SYSCTL_CHILDREN(bbr_hptsi),
1306             OID_AUTO, "ip_oh", CTLFLAG_RW,
1307             &bbr_include_ip_oh, 1,
1308             "Do we include the IP overhead in calculating pacing delay?");
1309         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1310             SYSCTL_CHILDREN(bbr_hptsi),
1311             OID_AUTO, "tcp_oh", CTLFLAG_RW,
1312             &bbr_include_tcp_oh, 0,
1313             "Do we include the TCP overhead in calculating pacing delay?");
1314         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1315             SYSCTL_CHILDREN(bbr_hptsi),
1316             OID_AUTO, "google_discount", CTLFLAG_RW,
1317             &bbr_google_discount, 10,
1318             "What is the default google discount percentage wise for pacing (11 = 1.1%%)?");
1319         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1320             SYSCTL_CHILDREN(bbr_hptsi),
1321             OID_AUTO, "all_get_min", CTLFLAG_RW,
1322             &bbr_all_get_min, 0,
1323             "If you are less than a MSS do you just get the min?");
1324         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1325             SYSCTL_CHILDREN(bbr_hptsi),
1326             OID_AUTO, "tso_min", CTLFLAG_RW,
1327             &bbr_hptsi_bytes_min, 1460,
1328             "For 0 -> 24Mbps what is floor number of segments for TSO");
1329         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1330             SYSCTL_CHILDREN(bbr_hptsi),
1331             OID_AUTO, "seg_tso_max", CTLFLAG_RW,
1332             &bbr_hptsi_segments_max, 6,
1333             "For 0 -> 24Mbps what is top number of segments for TSO");
1334         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1335             SYSCTL_CHILDREN(bbr_hptsi),
1336             OID_AUTO, "seg_floor", CTLFLAG_RW,
1337             &bbr_hptsi_segments_floor, 1,
1338             "Minimum TSO size we will fall too in segments");
1339         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1340             SYSCTL_CHILDREN(bbr_hptsi),
1341             OID_AUTO, "utter_max", CTLFLAG_RW,
1342             &bbr_hptsi_utter_max, 0,
1343             "The absolute maximum that any pacing (outside of hardware) can be");
1344         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1345             SYSCTL_CHILDREN(bbr_hptsi),
1346             OID_AUTO, "seg_divisor", CTLFLAG_RW,
1347             &bbr_hptsi_per_second, 100,
1348             "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps ");
1349         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1350             SYSCTL_CHILDREN(bbr_hptsi),
1351             OID_AUTO, "srtt_mul", CTLFLAG_RW,
1352             &bbr_hptsi_max_mul, 1,
1353             "The multiplier for pace len max");
1354         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1355             SYSCTL_CHILDREN(bbr_hptsi),
1356             OID_AUTO, "srtt_div", CTLFLAG_RW,
1357             &bbr_hptsi_max_div, 2,
1358             "The divisor for pace len max");
1359         /* Measurement controls */
1360         bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1361             SYSCTL_CHILDREN(bbr_sysctl_root),
1362             OID_AUTO,
1363             "measure",
1364             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1365             "Measurement controls");
1366         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1367             SYSCTL_CHILDREN(bbr_measure),
1368             OID_AUTO, "min_i_bw", CTLFLAG_RW,
1369             &bbr_initial_bw_bps, 62500,
1370             "Minimum initial b/w in bytes per second");
1371         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1372             SYSCTL_CHILDREN(bbr_measure),
1373             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1374             &bbr_sack_not_required, 0,
1375             "Do we allow bbr to run on connections not supporting SACK?");
1376         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1377             SYSCTL_CHILDREN(bbr_measure),
1378             OID_AUTO, "use_google", CTLFLAG_RW,
1379             &bbr_use_google_algo, 0,
1380             "Use has close to google V1.0 has possible?");
1381         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1382             SYSCTL_CHILDREN(bbr_measure),
1383             OID_AUTO, "ts_limiting", CTLFLAG_RW,
1384             &bbr_ts_limiting, 1,
1385             "Do we attempt to use the peers timestamp to limit b/w caculations?");
1386         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1387             SYSCTL_CHILDREN(bbr_measure),
1388             OID_AUTO, "ts_can_raise", CTLFLAG_RW,
1389             &bbr_ts_can_raise, 0,
1390             "Can we raise the b/w via timestamp b/w calculation?");
1391         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1392             SYSCTL_CHILDREN(bbr_measure),
1393             OID_AUTO, "ts_delta", CTLFLAG_RW,
1394             &bbr_min_usec_delta, 20000,
1395             "How long in usec between ts of our sends in ts validation code?");
1396         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1397             SYSCTL_CHILDREN(bbr_measure),
1398             OID_AUTO, "ts_peer_delta", CTLFLAG_RW,
1399             &bbr_min_peer_delta, 20,
1400             "What min numerical value should be between the peer deltas?");
1401         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1402             SYSCTL_CHILDREN(bbr_measure),
1403             OID_AUTO, "ts_delta_percent", CTLFLAG_RW,
1404             &bbr_delta_percent, 150,
1405             "What percentage (150 = 15.0) do we allow variance for?");
1406         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1407             SYSCTL_CHILDREN(bbr_measure),
1408             OID_AUTO, "min_measure_good_bw", CTLFLAG_RW,
1409             &bbr_min_measurements_req, 1,
1410             "What is the minimum measurment count we need before we switch to our b/w estimate");
1411         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1412             SYSCTL_CHILDREN(bbr_measure),
1413             OID_AUTO, "min_measure_before_pace", CTLFLAG_RW,
1414             &bbr_no_pacing_until, 4,
1415             "How many pkt-epoch's (0 is off) do we need before pacing is on?");
1416         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1417             SYSCTL_CHILDREN(bbr_measure),
1418             OID_AUTO, "quanta", CTLFLAG_RW,
1419             &bbr_quanta, 2,
1420             "Extra quanta to add when calculating the target (ID section 4.2.3.2).");
1421         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1422             SYSCTL_CHILDREN(bbr_measure),
1423             OID_AUTO, "noretran", CTLFLAG_RW,
1424             &bbr_no_retran, 0,
1425             "Should google mode not use retransmission measurements for the b/w estimation?");
1426         /* State controls */
1427         bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1428             SYSCTL_CHILDREN(bbr_sysctl_root),
1429             OID_AUTO,
1430             "states",
1431             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1432             "State controls");
1433         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1434             SYSCTL_CHILDREN(bbr_states),
1435             OID_AUTO, "idle_restart", CTLFLAG_RW,
1436             &bbr_uses_idle_restart, 0,
1437             "Do we use a new special idle_restart state to ramp back up quickly?");
1438         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1439             SYSCTL_CHILDREN(bbr_states),
1440             OID_AUTO, "idle_restart_threshold", CTLFLAG_RW,
1441             &bbr_idle_restart_threshold, 100000,
1442             "How long must we be idle before we restart??");
1443         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1444             SYSCTL_CHILDREN(bbr_states),
1445             OID_AUTO, "use_pkt_epoch", CTLFLAG_RW,
1446             &bbr_state_is_pkt_epoch, 0,
1447             "Do we use a pkt-epoch for substate if 0 rttProp?");
1448         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1449             SYSCTL_CHILDREN(bbr_states),
1450             OID_AUTO, "startup_rtt_gain", CTLFLAG_RW,
1451             &bbr_rtt_gain_thresh, 0,
1452             "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?");
1453         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1454             SYSCTL_CHILDREN(bbr_states),
1455             OID_AUTO, "drain_floor", CTLFLAG_RW,
1456             &bbr_drain_floor, 88,
1457             "What is the lowest we can drain (pg) too?");
1458         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1459             SYSCTL_CHILDREN(bbr_states),
1460             OID_AUTO, "drain_2_target", CTLFLAG_RW,
1461             &bbr_state_drain_2_tar, 1,
1462             "Do we drain to target in drain substate?");
1463         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1464             SYSCTL_CHILDREN(bbr_states),
1465             OID_AUTO, "gain_2_target", CTLFLAG_RW,
1466             &bbr_gain_to_target, 1,
1467             "Does probe bw gain to target??");
1468         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1469             SYSCTL_CHILDREN(bbr_states),
1470             OID_AUTO, "gain_extra_time", CTLFLAG_RW,
1471             &bbr_gain_gets_extra_too, 1,
1472             "Does probe bw gain get the extra time too?");
1473         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1474             SYSCTL_CHILDREN(bbr_states),
1475             OID_AUTO, "ld_div", CTLFLAG_RW,
1476             &bbr_drain_drop_div, 5,
1477             "Long drain drop divider?");
1478         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1479             SYSCTL_CHILDREN(bbr_states),
1480             OID_AUTO, "ld_mul", CTLFLAG_RW,
1481             &bbr_drain_drop_mul, 4,
1482             "Long drain drop multiplier?");
1483         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1484             SYSCTL_CHILDREN(bbr_states),
1485             OID_AUTO, "rand_ot_disc", CTLFLAG_RW,
1486             &bbr_rand_ot, 50,
1487             "Random discount of the ot?");
1488         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1489             SYSCTL_CHILDREN(bbr_states),
1490             OID_AUTO, "dr_filter_life", CTLFLAG_RW,
1491             &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT,
1492             "How many packet-epochs does the b/w delivery rate last?");
1493         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1494             SYSCTL_CHILDREN(bbr_states),
1495             OID_AUTO, "subdrain_applimited", CTLFLAG_RW,
1496             &bbr_sub_drain_app_limit, 0,
1497             "Does our sub-state drain invoke app limited if its long?");
1498         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1499             SYSCTL_CHILDREN(bbr_states),
1500             OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW,
1501             &bbr_sub_drain_slam_cwnd, 0,
1502             "Should we set/recover cwnd for sub-state drain?");
1503         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1504             SYSCTL_CHILDREN(bbr_states),
1505             OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW,
1506             &bbr_slam_cwnd_in_main_drain, 0,
1507             "Should we set/recover cwnd for main-state drain?");
1508         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1509             SYSCTL_CHILDREN(bbr_states),
1510             OID_AUTO, "google_gets_earlyout", CTLFLAG_RW,
1511             &google_allow_early_out, 1,
1512             "Should we allow google probe-bw/drain to exit early at flight target?");
1513         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1514             SYSCTL_CHILDREN(bbr_states),
1515             OID_AUTO, "google_exit_loss", CTLFLAG_RW,
1516             &google_consider_lost, 1,
1517             "Should we have losses exit gain of probebw in google mode??");
1518         /* Startup controls */
1519         bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1520             SYSCTL_CHILDREN(bbr_sysctl_root),
1521             OID_AUTO,
1522             "startup",
1523             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1524             "Startup controls");
1525         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1526             SYSCTL_CHILDREN(bbr_startup),
1527             OID_AUTO, "cheat_iwnd", CTLFLAG_RW,
1528             &bbr_sends_full_iwnd, 1,
1529             "Do we not pace but burst out initial windows has our TSO size?");
1530         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1531             SYSCTL_CHILDREN(bbr_startup),
1532             OID_AUTO, "loss_threshold", CTLFLAG_RW,
1533             &bbr_startup_loss_thresh, 2000,
1534             "In startup what is the loss threshold in a pe that will exit us from startup?");
1535         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1536             SYSCTL_CHILDREN(bbr_startup),
1537             OID_AUTO, "use_lowerpg", CTLFLAG_RW,
1538             &bbr_use_lower_gain_in_startup, 1,
1539             "Should we use a lower hptsi gain if we see loss in startup?");
1540         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1541             SYSCTL_CHILDREN(bbr_startup),
1542             OID_AUTO, "gain", CTLFLAG_RW,
1543             &bbr_start_exit, 25,
1544             "What gain percent do we need to see to stay in startup??");
1545         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1546             SYSCTL_CHILDREN(bbr_startup),
1547             OID_AUTO, "low_gain", CTLFLAG_RW,
1548             &bbr_low_start_exit, 15,
1549             "What gain percent do we need to see to stay in the lower gain startup??");
1550         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1551             SYSCTL_CHILDREN(bbr_startup),
1552             OID_AUTO, "loss_exit", CTLFLAG_RW,
1553             &bbr_exit_startup_at_loss, 1,
1554             "Should we exit startup at loss in an epoch if we are not gaining?");
1555         /* CWND controls */
1556         bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1557             SYSCTL_CHILDREN(bbr_sysctl_root),
1558             OID_AUTO,
1559             "cwnd",
1560             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1561             "Cwnd controls");
1562         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1563             SYSCTL_CHILDREN(bbr_cwnd),
1564             OID_AUTO, "tar_rtt", CTLFLAG_RW,
1565             &bbr_cwndtarget_rtt_touse, 0,
1566             "Target cwnd rtt measurment to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?");
1567         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1568             SYSCTL_CHILDREN(bbr_cwnd),
1569             OID_AUTO, "may_shrink", CTLFLAG_RW,
1570             &bbr_cwnd_may_shrink, 0,
1571             "Can the cwnd shrink if it would grow to more than the target?");
1572         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1573             SYSCTL_CHILDREN(bbr_cwnd),
1574             OID_AUTO, "max_target_limit", CTLFLAG_RW,
1575             &bbr_target_cwnd_mult_limit, 8,
1576             "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?");
1577         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1578             SYSCTL_CHILDREN(bbr_cwnd),
1579             OID_AUTO, "highspeed_min", CTLFLAG_RW,
1580             &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS,
1581             "What is the high-speed min cwnd (rttProp under 1ms)");
1582         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1583             SYSCTL_CHILDREN(bbr_cwnd),
1584             OID_AUTO, "lowspeed_min", CTLFLAG_RW,
1585             &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS,
1586             "What is the min cwnd (rttProp > 1ms)");
1587         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1588             SYSCTL_CHILDREN(bbr_cwnd),
1589             OID_AUTO, "initwin", CTLFLAG_RW,
1590             &bbr_def_init_win, 10,
1591             "What is the BBR initial window, if 0 use tcp version");
1592         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1593             SYSCTL_CHILDREN(bbr_cwnd),
1594             OID_AUTO, "do_loss_red", CTLFLAG_RW,
1595             &bbr_do_red, 600,
1596             "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?");
1597         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1598             SYSCTL_CHILDREN(bbr_cwnd),
1599             OID_AUTO, "red_scale", CTLFLAG_RW,
1600             &bbr_red_scale, 20000,
1601             "What RTT do we scale with?");
1602         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1603             SYSCTL_CHILDREN(bbr_cwnd),
1604             OID_AUTO, "red_growslow", CTLFLAG_RW,
1605             &bbr_red_growth_restrict, 1,
1606             "Do we restrict cwnd growth for whats in flight?");
1607         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1608             SYSCTL_CHILDREN(bbr_cwnd),
1609             OID_AUTO, "red_div", CTLFLAG_RW,
1610             &bbr_red_div, 2,
1611             "If we reduce whats the divisor?");
1612         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1613             SYSCTL_CHILDREN(bbr_cwnd),
1614             OID_AUTO, "red_mul", CTLFLAG_RW,
1615             &bbr_red_mul, 1,
1616             "If we reduce whats the mulitiplier?");
1617         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1618             SYSCTL_CHILDREN(bbr_cwnd),
1619             OID_AUTO, "target_is_unit", CTLFLAG_RW,
1620             &bbr_target_is_bbunit, 0,
1621             "Is the state target the pacing_gain or BBR_UNIT?");
1622         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1623             SYSCTL_CHILDREN(bbr_cwnd),
1624             OID_AUTO, "drop_limit", CTLFLAG_RW,
1625             &bbr_drop_limit, 0,
1626             "Number of segments limit for drop (0=use min_cwnd w/flight)?");
1627
1628         /* Timeout controls */
1629         bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1630             SYSCTL_CHILDREN(bbr_sysctl_root),
1631             OID_AUTO,
1632             "timeout",
1633             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1634             "Time out controls");
1635         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1636             SYSCTL_CHILDREN(bbr_timeout),
1637             OID_AUTO, "delack", CTLFLAG_RW,
1638             &bbr_delack_time, 100000,
1639             "BBR's delayed ack time");
1640         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1641             SYSCTL_CHILDREN(bbr_timeout),
1642             OID_AUTO, "tlp_uses", CTLFLAG_RW,
1643             &bbr_tlp_type_to_use, 3,
1644             "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt");
1645         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1646             SYSCTL_CHILDREN(bbr_timeout),
1647             OID_AUTO, "persmin", CTLFLAG_RW,
1648             &bbr_persist_min, 250000,
1649             "What is the minimum time in microseconds between persists");
1650         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1651             SYSCTL_CHILDREN(bbr_timeout),
1652             OID_AUTO, "persmax", CTLFLAG_RW,
1653             &bbr_persist_max, 1000000,
1654             "What is the largest delay in microseconds between persists");
1655         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1656             SYSCTL_CHILDREN(bbr_timeout),
1657             OID_AUTO, "tlp_minto", CTLFLAG_RW,
1658             &bbr_tlp_min, 10000,
1659             "TLP Min timeout in usecs");
1660         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1661             SYSCTL_CHILDREN(bbr_timeout),
1662             OID_AUTO, "tlp_dack_time", CTLFLAG_RW,
1663             &bbr_delayed_ack_time, 200000,
1664             "TLP delayed ack compensation value");
1665         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1666             SYSCTL_CHILDREN(bbr_sysctl_root),
1667             OID_AUTO, "minrto", CTLFLAG_RW,
1668             &bbr_rto_min_ms, 30,
1669             "Minimum RTO in ms");
1670         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1671             SYSCTL_CHILDREN(bbr_timeout),
1672             OID_AUTO, "maxrto", CTLFLAG_RW,
1673             &bbr_rto_max_sec, 4,
1674             "Maxiumum RTO in seconds -- should be at least as large as min_rto");
1675         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1676             SYSCTL_CHILDREN(bbr_timeout),
1677             OID_AUTO, "tlp_retry", CTLFLAG_RW,
1678             &bbr_tlp_max_resend, 2,
1679             "How many times does TLP retry a single segment or multiple with no ACK");
1680         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1681             SYSCTL_CHILDREN(bbr_timeout),
1682             OID_AUTO, "minto", CTLFLAG_RW,
1683             &bbr_min_to, 1000,
1684             "Minimum rack timeout in useconds");
1685         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1686             SYSCTL_CHILDREN(bbr_timeout),
1687             OID_AUTO, "pktdelay", CTLFLAG_RW,
1688             &bbr_pkt_delay, 1000,
1689             "Extra RACK time (in useconds) besides reordering thresh");
1690         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1691             SYSCTL_CHILDREN(bbr_timeout),
1692             OID_AUTO, "incr_tmrs", CTLFLAG_RW,
1693             &bbr_incr_timers, 1,
1694             "Increase the RXT/TLP timer by the pacing time used?");
1695         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1696             SYSCTL_CHILDREN(bbr_timeout),
1697             OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW,
1698             &bbr_marks_rxt_sack_passed, 0,
1699             "Mark sack passed on all those not ack'd when a RXT hits?");
1700         /* Policer controls */
1701         bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1702             SYSCTL_CHILDREN(bbr_sysctl_root),
1703             OID_AUTO,
1704             "policer",
1705             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1706             "Policer controls");
1707         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1708             SYSCTL_CHILDREN(bbr_policer),
1709             OID_AUTO, "detect_enable", CTLFLAG_RW,
1710             &bbr_policer_detection_enabled, 1,
1711             "Is policer detection enabled??");
1712         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1713             SYSCTL_CHILDREN(bbr_policer),
1714             OID_AUTO, "min_pes", CTLFLAG_RW,
1715             &bbr_lt_intvl_min_rtts, 4,
1716             "Minimum number of PE's?");
1717         SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1718             SYSCTL_CHILDREN(bbr_policer),
1719             OID_AUTO, "bwdiff", CTLFLAG_RW,
1720             &bbr_lt_bw_diff, (4000/8),
1721             "Minimal bw diff?");
1722         SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1723             SYSCTL_CHILDREN(bbr_policer),
1724             OID_AUTO, "bwratio", CTLFLAG_RW,
1725             &bbr_lt_bw_ratio, 8,
1726             "Minimal bw diff?");
1727         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1728             SYSCTL_CHILDREN(bbr_policer),
1729             OID_AUTO, "from_rack_rxt", CTLFLAG_RW,
1730             &bbr_policer_call_from_rack_to, 0,
1731             "Do we call the policer detection code from a rack-timeout?");
1732         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1733             SYSCTL_CHILDREN(bbr_policer),
1734             OID_AUTO, "false_postive", CTLFLAG_RW,
1735             &bbr_lt_intvl_fp, 0,
1736             "What packet epoch do we do false-postive detection at (0=no)?");
1737         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1738             SYSCTL_CHILDREN(bbr_policer),
1739             OID_AUTO, "loss_thresh", CTLFLAG_RW,
1740             &bbr_lt_loss_thresh, 196,
1741             "Loss threshold 196 = 19.6%?");
1742         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1743             SYSCTL_CHILDREN(bbr_policer),
1744             OID_AUTO, "false_postive_thresh", CTLFLAG_RW,
1745             &bbr_lt_fd_thresh, 100,
1746             "What percentage is the false detection threshold (150=15.0)?");
1747         /* All the rest */
1748         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1749             SYSCTL_CHILDREN(bbr_sysctl_root),
1750             OID_AUTO, "cheat_rxt", CTLFLAG_RW,
1751             &bbr_use_rack_resend_cheat, 0,
1752             "Do we burst 1ms between sends on retransmissions (like rack)?");
1753         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1754             SYSCTL_CHILDREN(bbr_sysctl_root),
1755             OID_AUTO, "error_paceout", CTLFLAG_RW,
1756             &bbr_error_base_paceout, 10000,
1757             "When we hit an error what is the min to pace out in usec's?");
1758         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1759             SYSCTL_CHILDREN(bbr_sysctl_root),
1760             OID_AUTO, "kill_paceout", CTLFLAG_RW,
1761             &bbr_max_net_error_cnt, 10,
1762             "When we hit this many errors in a row, kill the session?");
1763         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1764             SYSCTL_CHILDREN(bbr_sysctl_root),
1765             OID_AUTO, "data_after_close", CTLFLAG_RW,
1766             &bbr_ignore_data_after_close, 1,
1767             "Do we hold off sending a RST until all pending data is ack'd");
1768         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1769             SYSCTL_CHILDREN(bbr_sysctl_root),
1770             OID_AUTO, "resend_use_tso", CTLFLAG_RW,
1771             &bbr_resends_use_tso, 0,
1772             "Can resends use TSO?");
1773         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1774             SYSCTL_CHILDREN(bbr_sysctl_root),
1775             OID_AUTO, "sblklimit", CTLFLAG_RW,
1776             &bbr_sack_block_limit, 128,
1777             "When do we start ignoring small sack blocks");
1778         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1779             SYSCTL_CHILDREN(bbr_sysctl_root),
1780             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1781             &bbr_verbose_logging, 0,
1782             "Should BBR black box logging be verbose");
1783         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1784             SYSCTL_CHILDREN(bbr_sysctl_root),
1785             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1786             &bbr_reorder_thresh, 2,
1787             "What factor for rack will be added when seeing reordering (shift right)");
1788         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1789             SYSCTL_CHILDREN(bbr_sysctl_root),
1790             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1791             &bbr_reorder_fade, 0,
1792             "Does reorder detection fade, if so how many ms (0 means never)");
1793         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1794             SYSCTL_CHILDREN(bbr_sysctl_root),
1795             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1796             &bbr_tlp_thresh, 1,
1797             "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1798         /* Stats and counters */
1799         /* The pacing counters for hdwr/software can't be in the array */
1800         bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1801         bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1802         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1803             SYSCTL_CHILDREN(bbr_sysctl_root),
1804             OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD,
1805             &bbr_hdwr_pacing_enobuf,
1806             "Total number of enobufs for hardware paced flows");
1807         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1808             SYSCTL_CHILDREN(bbr_sysctl_root),
1809             OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD,
1810             &bbr_nohdwr_pacing_enobuf,
1811             "Total number of enobufs for non-hardware paced flows");
1812
1813
1814         bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK);
1815         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1816             SYSCTL_CHILDREN(bbr_sysctl_root),
1817             OID_AUTO, "hdwr_pacing", CTLFLAG_RD,
1818             &bbr_flows_whdwr_pacing,
1819             "Total number of hardware paced flows");
1820         bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK);
1821         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1822             SYSCTL_CHILDREN(bbr_sysctl_root),
1823             OID_AUTO, "software_pacing", CTLFLAG_RD,
1824             &bbr_flows_nohdwr_pacing,
1825             "Total number of software paced flows");
1826         COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK);
1827         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1828             OID_AUTO, "stats", CTLFLAG_RD,
1829             bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats");
1830         COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK);
1831         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1832             OID_AUTO, "opts", CTLFLAG_RD,
1833             bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats");
1834         COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK);
1835         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1836             OID_AUTO, "lost", CTLFLAG_RD,
1837             bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur");
1838         COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK);
1839         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1840             OID_AUTO, "stateresend", CTLFLAG_RD,
1841             bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend");
1842         COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK);
1843         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1844             OID_AUTO, "statetime", CTLFLAG_RD,
1845             bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states");
1846         COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1847         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1848             OID_AUTO, "outsize", CTLFLAG_RD,
1849             bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls");
1850         SYSCTL_ADD_PROC(&bbr_sysctl_ctx,
1851             SYSCTL_CHILDREN(bbr_sysctl_root),
1852             OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1853             &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters");
1854 }
1855
1856 static void
1857 bbr_counter_destroy(void)
1858 {
1859         COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE);
1860         COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE);
1861         COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE);
1862         COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT);
1863         COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT);
1864         COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT);
1865         counter_u64_free(bbr_nohdwr_pacing_enobuf);
1866         counter_u64_free(bbr_hdwr_pacing_enobuf);
1867         counter_u64_free(bbr_flows_whdwr_pacing);
1868         counter_u64_free(bbr_flows_nohdwr_pacing);
1869
1870 }
1871
1872 static __inline void
1873 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts)
1874 {
1875         memset(l, 0, sizeof(union tcp_log_stackspecific));
1876         l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate;
1877         l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate);
1878         l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
1879         l->bw_inuse = bbr_get_bw(bbr);
1880         l->inflight = ctf_flight_size(bbr->rc_tp,
1881                           (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
1882         l->applimited = bbr->r_ctl.r_app_limited_until;
1883         l->delivered = bbr->r_ctl.rc_delivered;
1884         l->timeStamp = cts;
1885         l->lost = bbr->r_ctl.rc_lost;
1886         l->bbr_state = bbr->rc_bbr_state;
1887         l->bbr_substate = bbr_state_val(bbr);
1888         l->epoch = bbr->r_ctl.rc_rtt_epoch;
1889         l->lt_epoch = bbr->r_ctl.rc_lt_epoch;
1890         l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
1891         l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain;
1892         l->inhpts = bbr->rc_inp->inp_in_hpts;
1893         l->ininput = bbr->rc_inp->inp_in_input;
1894         l->use_lt_bw = bbr->rc_lt_use_bw;
1895         l->pkts_out = bbr->r_ctl.rc_flight_at_input;
1896         l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch;
1897 }
1898
1899 static void
1900 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason)
1901 {
1902         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1903                 union tcp_log_stackspecific log;
1904
1905                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1906                 log.u_bbr.flex1 = 0;
1907                 log.u_bbr.flex2 = 0;
1908                 log.u_bbr.flex5 = 0;
1909                 log.u_bbr.flex3 = 0;
1910                 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate;
1911                 log.u_bbr.flex7 = reason;
1912                 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt;
1913                 log.u_bbr.flex8 = 0;
1914                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1915                     &bbr->rc_inp->inp_socket->so_rcv,
1916                     &bbr->rc_inp->inp_socket->so_snd,
1917                     BBR_LOG_BW_RED_EV, 0,
1918                     0, &log, false, &bbr->rc_tv);
1919         }
1920 }
1921
1922 static void
1923 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count)
1924 {
1925         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1926                 union tcp_log_stackspecific log;
1927
1928                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1929                 log.u_bbr.flex1 = seq;
1930                 log.u_bbr.flex2 = count;
1931                 log.u_bbr.flex8 = mode;
1932                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1933                     &bbr->rc_inp->inp_socket->so_rcv,
1934                     &bbr->rc_inp->inp_socket->so_snd,
1935                     BBR_LOG_LOWGAIN, 0,
1936                     0, &log, false, &bbr->rc_tv);
1937         }
1938 }
1939
1940
1941
1942 static void
1943 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling,
1944     uint8_t reason, uint32_t p_maxseg, int len)
1945 {
1946         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1947                 union tcp_log_stackspecific log;
1948
1949                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1950                 log.u_bbr.flex1 = p_maxseg;
1951                 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags;
1952                 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
1953                 log.u_bbr.flex4 = reason;
1954                 log.u_bbr.flex5 = bbr->rc_in_persist;
1955                 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val;
1956                 log.u_bbr.flex7 = p_maxseg;
1957                 log.u_bbr.flex8 = bbr->rc_in_persist;
1958                 log.u_bbr.pkts_out = 0;
1959                 log.u_bbr.applimited = len;
1960                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1961                     &bbr->rc_inp->inp_socket->so_rcv,
1962                     &bbr->rc_inp->inp_socket->so_snd,
1963                     BBR_LOG_JUSTRET, 0,
1964                     tlen, &log, false, &bbr->rc_tv);
1965         }
1966 }
1967
1968
1969 static void
1970 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq)
1971 {
1972         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1973                 union tcp_log_stackspecific log;
1974
1975                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1976                 log.u_bbr.flex1 = seq;
1977                 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
1978                 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start;
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_ENTREC, 0,
1983                     0, &log, false, &bbr->rc_tv);
1984         }
1985 }
1986
1987 static void
1988 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)
1989 {
1990         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1991                 union tcp_log_stackspecific log;
1992
1993                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1994                 log.u_bbr.flex1 = tso;
1995                 log.u_bbr.flex2 = maxseg;
1996                 log.u_bbr.flex3 = mtu;
1997                 log.u_bbr.flex4 = csum_flags;
1998                 TCP_LOG_EVENTP(tp, NULL,
1999                     &bbr->rc_inp->inp_socket->so_rcv,
2000                     &bbr->rc_inp->inp_socket->so_snd,
2001                     BBR_LOG_MSGSIZE, 0,
2002                     0, &log, false, &bbr->rc_tv);
2003         }
2004 }
2005
2006 static void
2007 bbr_log_flowend(struct tcp_bbr *bbr)
2008 {
2009         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2010                 union tcp_log_stackspecific log;
2011                 struct sockbuf *r, *s;
2012                 struct timeval tv;
2013
2014                 if (bbr->rc_inp->inp_socket) {
2015                         r = &bbr->rc_inp->inp_socket->so_rcv;
2016                         s = &bbr->rc_inp->inp_socket->so_snd;
2017                 } else {
2018                         r = s = NULL;
2019                 }
2020                 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv));
2021                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2022                     r, s,
2023                     TCP_LOG_FLOWEND, 0,
2024                     0, &log, false, &tv);
2025         }
2026 }
2027
2028 static void
2029 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line,
2030     uint32_t lost, uint32_t del)
2031 {
2032         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2033                 union tcp_log_stackspecific log;
2034
2035                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2036                 log.u_bbr.flex1 = lost;
2037                 log.u_bbr.flex2 = del;
2038                 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw;
2039                 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt;
2040                 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2041                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2042                 log.u_bbr.flex7 = line;
2043                 log.u_bbr.flex8 = 0;
2044                 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count;
2045                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2046                     &bbr->rc_inp->inp_socket->so_rcv,
2047                     &bbr->rc_inp->inp_socket->so_snd,
2048                     BBR_LOG_PKT_EPOCH, 0,
2049                     0, &log, false, &bbr->rc_tv);
2050         }
2051 }
2052
2053 static void
2054 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time)
2055 {
2056         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2057                 union tcp_log_stackspecific log;
2058
2059                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2060                 log.u_bbr.flex1 = bbr->r_ctl.rc_lost;
2061                 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat;
2062                 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat;
2063                 log.u_bbr.flex7 = line;
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_TIME_EPOCH, 0,
2068                     0, &log, false, &bbr->rc_tv);
2069         }
2070 }
2071
2072 static void
2073 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth)
2074 {
2075         if (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, bbr->r_ctl.rc_rcvtime);
2079                 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2080                 log.u_bbr.flex2 = new_tar;
2081                 log.u_bbr.flex3 = line;
2082                 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2083                 log.u_bbr.flex5 = bbr_quanta;
2084                 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs;
2085                 log.u_bbr.flex7 = bbr->rc_last_options;
2086                 log.u_bbr.flex8 = meth;
2087                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2088                     &bbr->rc_inp->inp_socket->so_rcv,
2089                     &bbr->rc_inp->inp_socket->so_snd,
2090                     BBR_LOG_STATE_TARGET, 0,
2091                     0, &log, false, &bbr->rc_tv);
2092         }
2093
2094 }
2095
2096 static void
2097 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2098 {
2099         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2100                 union tcp_log_stackspecific log;
2101
2102                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2103                 log.u_bbr.flex1 = line;
2104                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2105                 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2106                 if (bbr_state_is_pkt_epoch)
2107                         log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
2108                 else
2109                         log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP);
2110                 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2111                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2112                 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000);
2113                 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra;
2114                 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state;
2115                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2116                     &bbr->rc_inp->inp_socket->so_rcv,
2117                     &bbr->rc_inp->inp_socket->so_snd,
2118                     BBR_LOG_STATE, 0,
2119                     0, &log, false, &bbr->rc_tv);
2120         }
2121 }
2122
2123 static void
2124 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied,
2125                     uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond)
2126 {
2127         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2128                 union tcp_log_stackspecific log;
2129
2130                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2131                 log.u_bbr.flex1 = line;
2132                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2133                 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt;
2134                 log.u_bbr.flex4 = applied;
2135                 log.u_bbr.flex5 = rtt;
2136                 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2137                 log.u_bbr.flex7 = cond;
2138                 log.u_bbr.flex8 = reas;
2139                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2140                     &bbr->rc_inp->inp_socket->so_rcv,
2141                     &bbr->rc_inp->inp_socket->so_snd,
2142                     BBR_LOG_RTT_SHRINKS, 0,
2143                     0, &log, false, &bbr->rc_tv);
2144         }
2145 }
2146
2147 static void
2148 bbr_log_type_exit_rec(struct tcp_bbr *bbr)
2149 {
2150         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2151                 union tcp_log_stackspecific log;
2152
2153                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2154                 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start;
2155                 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
2156                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2157                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2158                     &bbr->rc_inp->inp_socket->so_rcv,
2159                     &bbr->rc_inp->inp_socket->so_snd,
2160                     BBR_LOG_EXITREC, 0,
2161                     0, &log, false, &bbr->rc_tv);
2162         }
2163 }
2164
2165 static void
2166 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg,
2167     uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line)
2168 {
2169         if (bbr_verbose_logging && (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 = line;
2174                 log.u_bbr.flex2 = prev_acked;
2175                 log.u_bbr.flex3 = bytes_this_ack;
2176                 log.u_bbr.flex4 = chg;
2177                 log.u_bbr.flex5 = th_ack;
2178                 log.u_bbr.flex6 = target;
2179                 log.u_bbr.flex8 = meth;
2180                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2181                     &bbr->rc_inp->inp_socket->so_rcv,
2182                     &bbr->rc_inp->inp_socket->so_snd,
2183                     BBR_LOG_CWND, 0,
2184                     0, &log, false, &bbr->rc_tv);
2185         }
2186 }
2187
2188 static void
2189 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin)
2190 {
2191         /*
2192          * Log the rtt sample we are applying to the srtt algorithm in
2193          * useconds.
2194          */
2195         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2196                 union tcp_log_stackspecific log;
2197
2198                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2199                 log.u_bbr.flex1 = rtt;
2200                 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time;
2201                 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay;
2202                 log.u_bbr.flex4 = bbr->rc_tp->ts_offset;
2203                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2204                 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv);
2205                 log.u_bbr.flex6 = tsin;
2206                 log.u_bbr.flex7 = 0;
2207                 log.u_bbr.flex8 = bbr->rc_ack_was_delayed;
2208                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2209                     &bbr->rc_inp->inp_socket->so_rcv,
2210                     &bbr->rc_inp->inp_socket->so_snd,
2211                     TCP_LOG_RTT, 0,
2212                     0, &log, false, &bbr->rc_tv);
2213         }
2214 }
2215
2216 static void
2217 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit)
2218 {
2219         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2220                 union tcp_log_stackspecific log;
2221
2222                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2223                 log.u_bbr.flex1 = time_in;
2224                 log.u_bbr.flex2 = line;
2225                 log.u_bbr.flex8 = enter_exit;
2226                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2227                     &bbr->rc_inp->inp_socket->so_rcv,
2228                     &bbr->rc_inp->inp_socket->so_snd,
2229                     BBR_LOG_PERSIST, 0,
2230                     0, &log, false, &bbr->rc_tv);
2231         }
2232 }
2233 static void
2234 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts)
2235 {
2236         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2237                 union tcp_log_stackspecific log;
2238
2239                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2240                 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age;
2241                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2242                 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2243                 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time;
2244                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
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_ACKCLEAR, 0,
2249                     0, &log, false, &bbr->rc_tv);
2250         }
2251 }
2252
2253 static void
2254 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen,
2255                   uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m)
2256 {
2257         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2258                 union tcp_log_stackspecific log;
2259                 struct timeval tv;
2260
2261                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2262                 log.u_bbr.flex1 = nsegs;
2263                 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes;
2264                 if (m) {
2265                         struct timespec ts;
2266
2267                         log.u_bbr.flex3 = m->m_flags;
2268                         if (m->m_flags & M_TSTMP) {
2269                                 mbuf_tstmp2timespec(m, &ts);
2270                                 tv.tv_sec = ts.tv_sec;
2271                                 tv.tv_usec = ts.tv_nsec / 1000;
2272                                 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv);
2273                         } else {
2274                                 log.u_bbr.lt_epoch = 0;
2275                         }
2276                         if (m->m_flags & M_TSTMP_LRO) {
2277                                 tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000;
2278                                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000;
2279                                 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv);
2280                         } else {
2281                                 /* No arrival timestamp */
2282                                 log.u_bbr.flex5 = 0;
2283                         }
2284
2285                         log.u_bbr.pkts_out = tcp_get_usecs(&tv);
2286                 } else {
2287                         log.u_bbr.flex3 = 0;
2288                         log.u_bbr.flex5 = 0;
2289                         log.u_bbr.flex6 = 0;
2290                         log.u_bbr.pkts_out = 0;
2291                 }
2292                 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state;
2293                 log.u_bbr.flex7 = bbr->r_wanted_output;
2294                 log.u_bbr.flex8 = bbr->rc_in_persist;
2295                 TCP_LOG_EVENTP(bbr->rc_tp, th,
2296                     &bbr->rc_inp->inp_socket->so_rcv,
2297                     &bbr->rc_inp->inp_socket->so_snd,
2298                     TCP_LOG_IN, 0,
2299                     tlen, &log, true, &bbr->rc_tv);
2300         }
2301 }
2302
2303 static void
2304 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out)
2305 {
2306         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2307                 union tcp_log_stackspecific log;
2308
2309                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2310                 log.u_bbr.flex1 = did_out;
2311                 log.u_bbr.flex2 = nxt_pkt;
2312                 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val;
2313                 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2314                 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp;
2315                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes;
2316                 log.u_bbr.flex7 = bbr->r_wanted_output;
2317                 log.u_bbr.flex8 = bbr->rc_in_persist;
2318                 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay;
2319                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2320                     &bbr->rc_inp->inp_socket->so_rcv,
2321                     &bbr->rc_inp->inp_socket->so_snd,
2322                     BBR_LOG_DOSEG_DONE, 0,
2323                     0, &log, true, &bbr->rc_tv);
2324         }
2325 }
2326
2327 static void
2328 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts,
2329     int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz)
2330 {
2331         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2332                 union tcp_log_stackspecific log;
2333
2334                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2335                 log.u_bbr.flex1 = line;
2336                 log.u_bbr.flex2 = o_len;
2337                 log.u_bbr.flex3 = segcnt;
2338                 log.u_bbr.flex4 = segsiz;
2339                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2340                     &bbr->rc_inp->inp_socket->so_rcv,
2341                     &bbr->rc_inp->inp_socket->so_snd,
2342                     BBR_LOG_ENOBUF_JMP, ENOBUFS,
2343                     len, &log, true, &bbr->rc_tv);
2344         }
2345 }
2346
2347 static void
2348 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling)
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 = timers;
2355                 log.u_bbr.flex2 = ret;
2356                 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
2357                 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2358                 log.u_bbr.flex5 = cts;
2359                 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2360                 log.u_bbr.flex8 = hpts_calling;
2361                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2362                     &bbr->rc_inp->inp_socket->so_rcv,
2363                     &bbr->rc_inp->inp_socket->so_snd,
2364                     BBR_LOG_TO_PROCESS, 0,
2365                     0, &log, false, &bbr->rc_tv);
2366         }
2367 }
2368
2369 static void
2370 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num)
2371 {
2372         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2373                 union tcp_log_stackspecific log;
2374                 uint64_t ar;
2375
2376                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2377                 log.u_bbr.flex1 = bbr->bbr_timer_src;
2378                 log.u_bbr.flex2 = 0;
2379                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2380                 ar = (uint64_t)(bbr->r_ctl.rc_resend);
2381                 ar >>= 32;
2382                 ar &= 0x00000000ffffffff;
2383                 log.u_bbr.flex4 = (uint32_t)ar;
2384                 ar = (uint64_t)bbr->r_ctl.rc_resend;
2385                 ar &= 0x00000000ffffffff;
2386                 log.u_bbr.flex5 = (uint32_t)ar;
2387                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2388                 log.u_bbr.flex8 = to_num;
2389                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2390                     &bbr->rc_inp->inp_socket->so_rcv,
2391                     &bbr->rc_inp->inp_socket->so_snd,
2392                     BBR_LOG_RTO, 0,
2393                     0, &log, false, &bbr->rc_tv);
2394         }
2395 }
2396
2397 static void
2398 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason)
2399 {
2400         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2401                 union tcp_log_stackspecific log;
2402
2403                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2404                 log.u_bbr.flex1 = flex1;
2405                 log.u_bbr.flex2 = flex2;
2406                 log.u_bbr.flex3 = flex3;
2407                 log.u_bbr.flex4 = 0;
2408                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2409                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2410                 log.u_bbr.flex8 = reason;
2411                 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw;
2412                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2413                     &bbr->rc_inp->inp_socket->so_rcv,
2414                     &bbr->rc_inp->inp_socket->so_snd,
2415                     BBR_LOG_REDUCE, 0,
2416                     0, &log, false, &bbr->rc_tv);
2417         }
2418 }
2419
2420 static void
2421 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag)
2422 {
2423         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2424                 union tcp_log_stackspecific log;
2425
2426                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2427                 log.u_bbr.flex1 = diag->p_nxt_slot;
2428                 log.u_bbr.flex2 = diag->p_cur_slot;
2429                 log.u_bbr.flex3 = diag->slot_req;
2430                 log.u_bbr.flex4 = diag->inp_hptsslot;
2431                 log.u_bbr.flex5 = diag->slot_remaining;
2432                 log.u_bbr.flex6 = diag->need_new_to;
2433                 log.u_bbr.flex7 = diag->p_hpts_active;
2434                 log.u_bbr.flex8 = diag->p_on_min_sleep;
2435                 /* Hijack other fields as needed  */
2436                 log.u_bbr.epoch = diag->have_slept;
2437                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
2438                 log.u_bbr.pkts_out = diag->co_ret;
2439                 log.u_bbr.applimited = diag->hpts_sleep_time;
2440                 log.u_bbr.delivered = diag->p_prev_slot;
2441                 log.u_bbr.inflight = diag->p_runningtick;
2442                 log.u_bbr.bw_inuse = diag->wheel_tick;
2443                 log.u_bbr.rttProp = diag->wheel_cts;
2444                 log.u_bbr.delRate = diag->maxticks;
2445                 log.u_bbr.cur_del_rate = diag->p_curtick;
2446                 log.u_bbr.cur_del_rate <<= 32;
2447                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
2448                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2449                     &bbr->rc_inp->inp_socket->so_rcv,
2450                     &bbr->rc_inp->inp_socket->so_snd,
2451                     BBR_LOG_HPTSDIAG, 0,
2452                     0, &log, false, &bbr->rc_tv);
2453         }
2454 }
2455
2456 static void
2457 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
2458     uint32_t thresh, uint32_t to)
2459 {
2460         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2461                 union tcp_log_stackspecific log;
2462
2463                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2464                 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar;
2465                 log.u_bbr.flex2 = time_since_sent;
2466                 log.u_bbr.flex3 = srtt;
2467                 log.u_bbr.flex4 = thresh;
2468                 log.u_bbr.flex5 = to;
2469                 log.u_bbr.flex6 = bbr->rc_tp->t_srtt;
2470                 log.u_bbr.flex8 = mode;
2471                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2472                     &bbr->rc_inp->inp_socket->so_rcv,
2473                     &bbr->rc_inp->inp_socket->so_snd,
2474                     BBR_LOG_TIMERPREP, 0,
2475                     0, &log, false, &bbr->rc_tv);
2476         }
2477 }
2478
2479 static void
2480 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
2481     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod)
2482 {
2483         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2484                 union tcp_log_stackspecific log;
2485
2486                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2487                 log.u_bbr.flex1 = usecs;
2488                 log.u_bbr.flex2 = len;
2489                 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff);
2490                 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff);
2491                 if (override)
2492                         log.u_bbr.flex5 = (1 << 2);
2493                 else
2494                         log.u_bbr.flex5 = 0;
2495                 log.u_bbr.flex6 = override;
2496                 log.u_bbr.flex7 = gain;
2497                 log.u_bbr.flex8 = mod;
2498                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2499                     &bbr->rc_inp->inp_socket->so_rcv,
2500                     &bbr->rc_inp->inp_socket->so_snd,
2501                     BBR_LOG_HPTSI_CALC, 0,
2502                     len, &log, false, &bbr->rc_tv);
2503         }
2504 }
2505
2506 static void
2507 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2508 {
2509         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2510                 union tcp_log_stackspecific log;
2511
2512                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2513
2514                 log.u_bbr.flex1 = bbr->bbr_timer_src;
2515                 log.u_bbr.flex2 = to;
2516                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2517                 log.u_bbr.flex4 = slot;
2518                 log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot;
2519                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2520                 log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2;
2521                 log.u_bbr.flex8 = which;
2522                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2523                     &bbr->rc_inp->inp_socket->so_rcv,
2524                     &bbr->rc_inp->inp_socket->so_snd,
2525                     BBR_LOG_TIMERSTAR, 0,
2526                     0, &log, false, &bbr->rc_tv);
2527         }
2528 }
2529
2530 static void
2531 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)
2532 {
2533         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2534                 union tcp_log_stackspecific log;
2535
2536                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2537                 log.u_bbr.flex1 = thresh;
2538                 log.u_bbr.flex2 = lro;
2539                 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts;
2540                 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
2541                 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2542                 log.u_bbr.flex6 = srtt;
2543                 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift;
2544                 log.u_bbr.flex8 = frm;
2545                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2546                     &bbr->rc_inp->inp_socket->so_rcv,
2547                     &bbr->rc_inp->inp_socket->so_snd,
2548                     BBR_LOG_THRESH_CALC, 0,
2549                     0, &log, false, &bbr->rc_tv);
2550         }
2551 }
2552
2553 static void
2554 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed)
2555 {
2556         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2557                 union tcp_log_stackspecific log;
2558
2559                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2560                 log.u_bbr.flex1 = line;
2561                 log.u_bbr.flex2 = bbr->bbr_timer_src;
2562                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2563                 log.u_bbr.flex4 = bbr->rc_in_persist;
2564                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2565                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2566                 log.u_bbr.flex8 = hpts_removed;
2567                 log.u_bbr.pkts_out = bbr->rc_pacer_started;
2568                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2569                     &bbr->rc_inp->inp_socket->so_rcv,
2570                     &bbr->rc_inp->inp_socket->so_snd,
2571                     BBR_LOG_TIMERCANC, 0,
2572                     0, &log, false, &bbr->rc_tv);
2573         }
2574 }
2575
2576
2577 static void
2578 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta)
2579 {
2580         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2581                 union tcp_log_stackspecific log;
2582
2583                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2584                 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio;
2585                 log.u_bbr.flex2 = (peer_delta >> 32);
2586                 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff);
2587                 log.u_bbr.flex4 = (delta >> 32);
2588                 log.u_bbr.flex5 = (delta & 0x00000000ffffffff);
2589                 log.u_bbr.flex7 = bbr->rc_ts_clock_set;
2590                 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used;
2591                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2592                     &bbr->rc_inp->inp_socket->so_rcv,
2593                     &bbr->rc_inp->inp_socket->so_snd,
2594                     BBR_LOG_TSTMP_VAL, 0,
2595                     0, &log, false, &bbr->rc_tv);
2596
2597         }
2598 }
2599
2600 static void
2601 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)
2602 {
2603         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2604                 union tcp_log_stackspecific log;
2605
2606                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2607                 log.u_bbr.flex1 = tsosz;
2608                 log.u_bbr.flex2 = tls;
2609                 log.u_bbr.flex3 = tcp_min_hptsi_time;
2610                 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min;
2611                 log.u_bbr.flex5 = old_val;
2612                 log.u_bbr.flex6 = maxseg;
2613                 log.u_bbr.flex7 = bbr->rc_no_pacing;
2614                 log.u_bbr.flex7 <<= 1;
2615                 log.u_bbr.flex7 |= bbr->rc_past_init_win;
2616                 if (hdwr)
2617                         log.u_bbr.flex8 = 0x80 | bbr->rc_use_google;
2618                 else
2619                         log.u_bbr.flex8 = bbr->rc_use_google;
2620                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2621                     &bbr->rc_inp->inp_socket->so_rcv,
2622                     &bbr->rc_inp->inp_socket->so_snd,
2623                     BBR_LOG_BBRTSO, 0,
2624                     0, &log, false, &bbr->rc_tv);
2625         }
2626 }
2627
2628 static void
2629 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm,
2630                       uint32_t flags, uint32_t line)
2631 {
2632         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2633                 union tcp_log_stackspecific log;
2634
2635                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2636                 log.u_bbr.flex1 = line;
2637                 log.u_bbr.flex2 = rsm->r_start;
2638                 log.u_bbr.flex3 = rsm->r_end;
2639                 log.u_bbr.flex4 = rsm->r_delivered;
2640                 log.u_bbr.flex5 = rsm->r_rtr_cnt;
2641                 log.u_bbr.flex6 = rsm->r_dupack;
2642                 log.u_bbr.flex7 = rsm->r_tim_lastsent[0];
2643                 log.u_bbr.flex8 = rsm->r_flags;
2644                 /* Hijack the pkts_out fids */
2645                 log.u_bbr.applimited = flags;
2646                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2647                     &bbr->rc_inp->inp_socket->so_rcv,
2648                     &bbr->rc_inp->inp_socket->so_snd,
2649                     BBR_RSM_CLEARED, 0,
2650                     0, &log, false, &bbr->rc_tv);
2651         }
2652 }
2653
2654 static void
2655 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts,
2656     uint32_t flex3, uint32_t flex2, uint32_t flex5,
2657     uint32_t flex6, uint32_t pkts_out, int flex7,
2658     uint32_t flex4, uint32_t flex1)
2659 {
2660
2661         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2662                 union tcp_log_stackspecific log;
2663
2664                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2665                 log.u_bbr.flex1 = flex1;
2666                 log.u_bbr.flex2 = flex2;
2667                 log.u_bbr.flex3 = flex3;
2668                 log.u_bbr.flex4 = flex4;
2669                 log.u_bbr.flex5 = flex5;
2670                 log.u_bbr.flex6 = flex6;
2671                 log.u_bbr.flex7 = flex7;
2672                 /* Hijack the pkts_out fids */
2673                 log.u_bbr.pkts_out = pkts_out;
2674                 log.u_bbr.flex8 = flex8;
2675                 if (bbr->rc_ack_was_delayed)
2676                         log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay;
2677                 else
2678                         log.u_bbr.epoch = 0;
2679                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2680                     &bbr->rc_inp->inp_socket->so_rcv,
2681                     &bbr->rc_inp->inp_socket->so_snd,
2682                     BBR_LOG_BBRUPD, 0,
2683                     flex2, &log, false, &bbr->rc_tv);
2684         }
2685 }
2686
2687
2688 static void
2689 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason,
2690         uint32_t newbw, uint32_t obw, uint32_t diff,
2691         uint32_t tim)
2692 {
2693         if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2694                 union tcp_log_stackspecific log;
2695
2696                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2697                 log.u_bbr.flex1 = reason;
2698                 log.u_bbr.flex2 = newbw;
2699                 log.u_bbr.flex3 = obw;
2700                 log.u_bbr.flex4 = diff;
2701                 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost;
2702                 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del;
2703                 log.u_bbr.flex7 = bbr->rc_lt_is_sampling;
2704                 log.u_bbr.pkts_out = tim;
2705                 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw;
2706                 if (bbr->rc_lt_use_bw == 0)
2707                         log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
2708                 else
2709                         log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
2710                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2711                     &bbr->rc_inp->inp_socket->so_rcv,
2712                     &bbr->rc_inp->inp_socket->so_snd,
2713                     BBR_LOG_BWSAMP, 0,
2714                     0, &log, false, &bbr->rc_tv);
2715         }
2716 }
2717
2718 static inline void
2719 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line)
2720 {
2721         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2722                 union tcp_log_stackspecific log;
2723
2724                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2725                 log.u_bbr.flex1 = line;
2726                 log.u_bbr.flex2 = tick;
2727                 log.u_bbr.flex3 = tp->t_maxunacktime;
2728                 log.u_bbr.flex4 = tp->t_acktime;
2729                 log.u_bbr.flex8 = event;
2730                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2731                     &bbr->rc_inp->inp_socket->so_rcv,
2732                     &bbr->rc_inp->inp_socket->so_snd,
2733                     BBR_LOG_PROGRESS, 0,
2734                     0, &log, false, &bbr->rc_tv);
2735         }
2736 }
2737
2738 static void
2739 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp,
2740                          uint64_t rate, uint64_t hw_rate, int line, uint32_t cts,
2741                          int error)
2742 {
2743         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2744                 union tcp_log_stackspecific log;
2745
2746                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2747                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2748                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2749                 log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2750                 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2751                 log.u_bbr.bw_inuse = rate;
2752                 log.u_bbr.flex5 = line;
2753                 log.u_bbr.flex6 = error;
2754                 log.u_bbr.flex8 = bbr->skip_gain;
2755                 log.u_bbr.flex8 <<= 1;
2756                 log.u_bbr.flex8 |= bbr->gain_is_limited;
2757                 log.u_bbr.flex8 <<= 1;
2758                 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing;
2759                 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
2760                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2761                     &bbr->rc_inp->inp_socket->so_rcv,
2762                     &bbr->rc_inp->inp_socket->so_snd,
2763                     BBR_LOG_HDWR_PACE, 0,
2764                     0, &log, false, &bbr->rc_tv);
2765         }
2766 }
2767
2768 static void
2769 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)
2770 {
2771         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2772                 union tcp_log_stackspecific log;
2773
2774                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2775                 log.u_bbr.flex1 = slot;
2776                 log.u_bbr.flex2 = del_by;
2777                 log.u_bbr.flex3 = prev_delay;
2778                 log.u_bbr.flex4 = line;
2779                 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val;
2780                 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay;
2781                 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags);
2782                 log.u_bbr.flex8 = bbr->rc_in_persist;
2783                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2784                     &bbr->rc_inp->inp_socket->so_rcv,
2785                     &bbr->rc_inp->inp_socket->so_snd,
2786                     BBR_LOG_BBRSND, 0,
2787                     len, &log, false, &bbr->rc_tv);
2788         }
2789 }
2790
2791 static void
2792 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)
2793 {
2794         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2795                 union tcp_log_stackspecific log;
2796
2797                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2798                 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered;
2799                 log.u_bbr.flex2 = 0;
2800                 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt;
2801                 log.u_bbr.flex4 = end;
2802                 log.u_bbr.flex5 = seq;
2803                 log.u_bbr.flex6 = t;
2804                 log.u_bbr.flex7 = match;
2805                 log.u_bbr.flex8 = flags;
2806                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2807                     &bbr->rc_inp->inp_socket->so_rcv,
2808                     &bbr->rc_inp->inp_socket->so_snd,
2809                     BBR_LOG_BBRRTT, 0,
2810                     0, &log, false, &bbr->rc_tv);
2811         }
2812 }
2813
2814 static void
2815 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method)
2816 {
2817         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2818                 union tcp_log_stackspecific log;
2819
2820                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2821                 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2822                 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
2823                 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch;
2824                 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2825                 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs;
2826                 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight;
2827                 log.u_bbr.flex7 = 0;
2828                 log.u_bbr.flex8 = entry_method;
2829                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2830                     &bbr->rc_inp->inp_socket->so_rcv,
2831                     &bbr->rc_inp->inp_socket->so_snd,
2832                     BBR_LOG_EXIT_GAIN, 0,
2833                     0, &log, false, &bbr->rc_tv);
2834         }
2835 }
2836
2837 static void
2838 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired)
2839 {
2840         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2841                 union tcp_log_stackspecific log;
2842
2843                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2844                 /* R-HU */
2845                 log.u_bbr.flex1 = 0;
2846                 log.u_bbr.flex2 = 0;
2847                 log.u_bbr.flex3 = 0;
2848                 log.u_bbr.flex4 = 0;
2849                 log.u_bbr.flex7 = 0;
2850                 log.u_bbr.flex8 = settings_desired;
2851
2852                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2853                     &bbr->rc_inp->inp_socket->so_rcv,
2854                     &bbr->rc_inp->inp_socket->so_snd,
2855                     BBR_LOG_SETTINGS_CHG, 0,
2856                     0, &log, false, &bbr->rc_tv);
2857         }
2858 }
2859
2860 /*
2861  * Returns the bw from the our filter.
2862  */
2863 static inline uint64_t
2864 bbr_get_full_bw(struct tcp_bbr *bbr)
2865 {
2866         uint64_t bw;
2867
2868         bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2869
2870         return (bw);
2871 }
2872
2873 static inline void
2874 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2875 {
2876         uint64_t calclr;
2877         uint32_t lost, del;
2878
2879         if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch)
2880                 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch;
2881         else
2882                 lost = 0;
2883         del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del;
2884         if (lost == 0)  {
2885                 calclr = 0;
2886         } else if (del) {
2887                 calclr = lost;
2888                 calclr *= (uint64_t)1000;
2889                 calclr /= (uint64_t)del;
2890         } else {
2891                 /* Nothing delivered? 100.0% loss */
2892                 calclr = 1000;
2893         }
2894         bbr->r_ctl.rc_pkt_epoch_loss_rate =  (uint32_t)calclr;
2895         if (IN_RECOVERY(bbr->rc_tp->t_flags))
2896                 bbr->r_ctl.recovery_lr += (uint32_t)calclr;
2897         bbr->r_ctl.rc_pkt_epoch++;
2898         if (bbr->rc_no_pacing &&
2899             (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) {
2900                 bbr->rc_no_pacing = 0;
2901                 tcp_bbr_tso_size_check(bbr, cts);
2902         }
2903         bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time);
2904         bbr->r_ctl.rc_pkt_epoch_time = cts;
2905         /* What was our loss rate */
2906         bbr_log_pkt_epoch(bbr, cts, line, lost, del);
2907         bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered;
2908         bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost;
2909 }
2910
2911 static inline void
2912 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2913 {
2914         uint32_t epoch_time;
2915
2916         /* Tick the RTT clock */
2917         bbr->r_ctl.rc_rtt_epoch++;
2918         epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start;
2919         bbr_log_time_epoch(bbr, cts, line, epoch_time);
2920         bbr->r_ctl.rc_rcv_epoch_start = cts;
2921 }
2922
2923
2924 static inline void
2925 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked)
2926 {
2927         if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) {
2928                 bbr->rc_is_pkt_epoch_now = 1;
2929         }
2930 }
2931
2932 /*
2933  * Returns the bw from either the b/w filter
2934  * or from the lt_bw (if the connection is being
2935  * policed).
2936  */
2937 static inline uint64_t
2938 __bbr_get_bw(struct tcp_bbr *bbr)
2939 {
2940         uint64_t bw, min_bw;
2941         uint64_t rtt;
2942         int gm_measure_cnt = 1;
2943
2944         /*
2945          * For startup we make, like google, a
2946          * minimum b/w. This is generated from the
2947          * IW and the rttProp. We do fall back to srtt
2948          * if for some reason (initial handshake) we don't
2949          * have a rttProp. We, in the worst case, fall back
2950          * to the configured min_bw (rc_initial_hptsi_bw).
2951          */
2952         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
2953                 /* Attempt first to use rttProp */
2954                 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2955                 if (rtt && (rtt < 0xffffffff)) {
2956 measure:
2957                         min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2958                                 ((uint64_t)1000000);
2959                         min_bw /= rtt;
2960                         if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2961                                 min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2962                         }
2963
2964                 } else if (bbr->rc_tp->t_srtt != 0) {
2965                         /* No rttProp, use srtt? */
2966                         rtt = bbr_get_rtt(bbr, BBR_SRTT);
2967                         goto measure;
2968                 } else {
2969                         min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2970                 }
2971         } else
2972                 min_bw = 0;
2973
2974         if ((bbr->rc_past_init_win == 0) &&
2975             (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp)))
2976                 bbr->rc_past_init_win = 1;
2977         if ((bbr->rc_use_google)  && (bbr->r_ctl.r_measurement_count >= 1))
2978                 gm_measure_cnt = 0;
2979         if (gm_measure_cnt &&
2980             ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) ||
2981              (bbr->rc_past_init_win == 0))) {
2982                 /* For google we use our guess rate until we get 1 measurement */
2983
2984 use_initial_window:
2985                 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2986                 if (rtt && (rtt < 0xffffffff)) {
2987                         /*
2988                          * We have an RTT measurment. Use that in
2989                          * combination with our initial window to calculate
2990                          * a b/w.
2991                          */
2992                         bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2993                                 ((uint64_t)1000000);
2994                         bw /= rtt;
2995                         if (bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2996                                 bw = bbr->r_ctl.rc_initial_hptsi_bw;
2997                         }
2998                 } else {
2999                         /* Drop back to the 40 and punt to a default */
3000                         bw = bbr->r_ctl.rc_initial_hptsi_bw;
3001                 }
3002                 if (bw < 1)
3003                         /* Probably should panic */
3004                         bw = 1;
3005                 if (bw > min_bw)
3006                         return (bw);
3007                 else
3008                         return (min_bw);
3009         }
3010         if (bbr->rc_lt_use_bw)
3011                 bw = bbr->r_ctl.rc_lt_bw;
3012         else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0))
3013                 bw = bbr->r_ctl.red_bw;
3014         else
3015                 bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3016         if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) {
3017                 /*
3018                  * Enforce user set rate limit, keep in mind that
3019                  * t_peakrate_thr is in B/s already
3020                  */
3021                 bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw);
3022         }
3023         if (bw == 0) {
3024                 /* We should not be at 0, go to the initial window then  */
3025                 goto use_initial_window;
3026         }
3027         if (bw < 1)
3028                 /* Probably should panic */
3029                 bw = 1;
3030         if (bw < min_bw)
3031                 bw = min_bw;
3032         return (bw);
3033 }
3034
3035 static inline uint64_t
3036 bbr_get_bw(struct tcp_bbr *bbr)
3037 {
3038         uint64_t bw;
3039
3040         bw = __bbr_get_bw(bbr);
3041         return (bw);
3042 }
3043
3044 static inline void
3045 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts)
3046 {
3047         bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch;
3048         bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time;
3049         bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3050         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3051 }
3052
3053 static inline void
3054 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts)
3055 {
3056         bbr->rc_lt_is_sampling = 0;
3057         bbr->rc_lt_use_bw = 0;
3058         bbr->r_ctl.rc_lt_bw = 0;
3059         bbr_reset_lt_bw_interval(bbr, cts);
3060 }
3061
3062 static inline void
3063 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin)
3064 {
3065         uint64_t diff;
3066
3067         /* Do we have a previous sample? */
3068         if (bbr->r_ctl.rc_lt_bw) {
3069                 /* Get the diff in bytes per second */
3070                 if (bbr->r_ctl.rc_lt_bw > bw)
3071                         diff = bbr->r_ctl.rc_lt_bw - bw;
3072                 else
3073                         diff = bw - bbr->r_ctl.rc_lt_bw;
3074                 if ((diff <= bbr_lt_bw_diff) ||
3075                     (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) {
3076                         /* Consider us policed */
3077                         uint32_t saved_bw;
3078
3079                         saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw;
3080                         bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2;   /* average of two */
3081                         bbr->rc_lt_use_bw = 1;
3082                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
3083                         /*
3084                          * Use pkt based epoch for measuring length of
3085                          * policer up
3086                          */
3087                         bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch;
3088                         /*
3089                          * reason 4 is we need to start consider being
3090                          * policed
3091                          */
3092                         bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin);
3093                         return;
3094                 }
3095         }
3096         bbr->r_ctl.rc_lt_bw = bw;
3097         bbr_reset_lt_bw_interval(bbr, cts);
3098         bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin);
3099 }
3100
3101 static void
3102 bbr_randomize_extra_state_time(struct tcp_bbr *bbr)
3103 {
3104         uint32_t ran, deduct;
3105
3106         ran = arc4random_uniform(bbr_rand_ot);
3107         if (ran) {
3108                 deduct = bbr->r_ctl.rc_level_state_extra / ran;
3109                 bbr->r_ctl.rc_level_state_extra -= deduct;
3110         }
3111 }
3112 /*
3113  * Return randomly the starting state
3114  * to use in probebw.
3115  */
3116 static uint8_t
3117 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts)
3118 {
3119         uint32_t ran;
3120         uint8_t ret_val;
3121
3122         /* Initialize the offset to 0 */
3123         bbr->r_ctl.rc_exta_time_gd = 0;
3124         bbr->rc_hit_state_1 = 0;
3125         bbr->r_ctl.rc_level_state_extra = 0;
3126         ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1));
3127         /*
3128          * The math works funny here :) the return value is used to set the
3129          * substate and then the state change is called which increments by
3130          * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when
3131          * we fully enter the state. Note that the (8 - 1 - ran) assures that
3132          * we return 1 - 7, so we dont return 0 and end up starting in
3133          * state 1 (DRAIN).
3134          */
3135         ret_val = BBR_SUBSTATE_COUNT - 1 - ran;
3136         /* Set an epoch */
3137         if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP))
3138                 bbr_set_epoch(bbr, cts, __LINE__);
3139
3140         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
3141         return (ret_val);
3142 }
3143
3144 static void
3145 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected)
3146 {
3147         uint32_t diff, d_time;
3148         uint64_t del_time, bw, lost, delivered;
3149
3150         if (bbr->r_use_policer == 0)
3151                 return;
3152         if (bbr->rc_lt_use_bw) {
3153                 /* We are using lt bw do we stop yet? */
3154                 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
3155                 if (diff > bbr_lt_bw_max_rtts) {
3156                         /* Reset it all */
3157 reset_all:
3158                         bbr_reset_lt_bw_sampling(bbr, cts);
3159                         if (bbr->rc_filled_pipe) {
3160                                 bbr_set_epoch(bbr, cts, __LINE__);
3161                                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
3162                                 bbr_substate_change(bbr, cts, __LINE__, 0);
3163                                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
3164                                 bbr_log_type_statechange(bbr, cts, __LINE__);
3165                         } else {
3166                                 /*
3167                                  * This should not happen really
3168                                  * unless we remove the startup/drain
3169                                  * restrictions above.
3170                                  */
3171                                 bbr->rc_bbr_state = BBR_STATE_STARTUP;
3172                                 bbr_set_epoch(bbr, cts, __LINE__);
3173                                 bbr->r_ctl.rc_bbr_state_time = cts;
3174                                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
3175                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
3176                                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
3177                                 bbr_set_state_target(bbr, __LINE__);
3178                                 bbr_log_type_statechange(bbr, cts, __LINE__);
3179                         }
3180                         /* reason 0 is to stop using lt-bw */
3181                         bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0);
3182                         return;
3183                 }
3184                 if (bbr_lt_intvl_fp == 0) {
3185                         /* Not doing false-postive detection */
3186                         return;
3187                 }
3188                 /* False positive detection */
3189                 if (diff == bbr_lt_intvl_fp) {
3190                         /* At bbr_lt_intvl_fp we record the lost */
3191                         bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3192                         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3193                 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) {
3194                         /* Now is our loss rate still high? */
3195                         lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3196                         delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3197                         if ((delivered == 0) ||
3198                             (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) {
3199                                 /* No still below our threshold */
3200                                 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0);
3201                         } else {
3202                                 /* Yikes its still high, it must be a false positive */
3203                                 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0);
3204                                 goto reset_all;
3205                         }
3206                 }
3207                 return;
3208         }
3209         /*
3210          * Wait for the first loss before sampling, to let the policer
3211          * exhaust its tokens and estimate the steady-state rate allowed by
3212          * the policer. Starting samples earlier includes bursts that
3213          * over-estimate the bw.
3214          */
3215         if (bbr->rc_lt_is_sampling == 0) {
3216                 /* reason 1 is to begin doing the sampling  */
3217                 if (loss_detected == 0)
3218                         return;
3219                 bbr_reset_lt_bw_interval(bbr, cts);
3220                 bbr->rc_lt_is_sampling = 1;
3221                 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0);
3222                 return;
3223         }
3224         /* Now how long were we delivering long term last> */
3225         if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time))
3226                 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time;
3227         else
3228                 d_time = 0;
3229
3230         /* To avoid underestimates, reset sampling if we run out of data. */
3231         if (bbr->r_ctl.r_app_limited_until) {
3232                 /* Can not measure in app-limited state */
3233                 bbr_reset_lt_bw_sampling(bbr, cts);
3234                 /* reason 2 is to reset sampling due to app limits  */
3235                 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time);
3236                 return;
3237         }
3238         diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
3239         if (diff < bbr_lt_intvl_min_rtts) {
3240                 /*
3241                  * need more samples (we don't
3242                  * start on a round like linux so
3243                  * we need 1 more).
3244                  */
3245                 /* 6 is not_enough time or no-loss */
3246                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3247                 return;
3248         }
3249         if (diff > (4 * bbr_lt_intvl_min_rtts)) {
3250                 /*
3251                  * For now if we wait too long, reset all sampling. We need
3252                  * to do some research here, its possible that we should
3253                  * base this on how much loss as occurred.. something like
3254                  * if its under 10% (or some thresh) reset all otherwise
3255                  * don't.  Thats for phase II I guess.
3256                  */
3257                 bbr_reset_lt_bw_sampling(bbr, cts);
3258                 /* reason 3 is to reset sampling due too long of sampling */
3259                 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3260                 return;
3261         }
3262         /*
3263          * End sampling interval when a packet is lost, so we estimate the
3264          * policer tokens were exhausted. Stopping the sampling before the
3265          * tokens are exhausted under-estimates the policed rate.
3266          */
3267         if (loss_detected == 0) {
3268                 /* 6 is not_enough time or no-loss */
3269                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3270                 return;
3271         }
3272         /* Calculate packets lost and delivered in sampling interval. */
3273         lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3274         delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3275         if ((delivered == 0) ||
3276             (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) {
3277                 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time);
3278                 return;
3279         }
3280         if (d_time < 1000) {
3281                 /* Not enough time. wait */
3282                 /* 6 is not_enough time or no-loss */
3283                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3284                 return;
3285         }
3286         if (d_time >= (0xffffffff / USECS_IN_MSEC)) {
3287                 /* Too long */
3288                 bbr_reset_lt_bw_sampling(bbr, cts);
3289                 /* reason 3 is to reset sampling due too long of sampling */
3290                 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3291                 return;
3292         }
3293         del_time = d_time;
3294         bw = delivered;
3295         bw *= (uint64_t)USECS_IN_SECOND;
3296         bw /= del_time;
3297         bbr_lt_bw_samp_done(bbr, bw, cts, d_time);
3298 }
3299
3300 /*
3301  * Allocate a sendmap from our zone.
3302  */
3303 static struct bbr_sendmap *
3304 bbr_alloc(struct tcp_bbr *bbr)
3305 {
3306         struct bbr_sendmap *rsm;
3307
3308         BBR_STAT_INC(bbr_to_alloc);
3309         rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO));
3310         if (rsm) {
3311                 bbr->r_ctl.rc_num_maps_alloced++;
3312                 return (rsm);
3313         }
3314         if (bbr->r_ctl.rc_free_cnt) {
3315                 BBR_STAT_INC(bbr_to_alloc_emerg);
3316                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
3317                 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
3318                 bbr->r_ctl.rc_free_cnt--;
3319                 return (rsm);
3320         }
3321         BBR_STAT_INC(bbr_to_alloc_failed);
3322         return (NULL);
3323 }
3324
3325 static struct bbr_sendmap *
3326 bbr_alloc_full_limit(struct tcp_bbr *bbr)
3327 {
3328         if ((V_tcp_map_entries_limit > 0) &&
3329             (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3330                 BBR_STAT_INC(bbr_alloc_limited);
3331                 if (!bbr->alloc_limit_reported) {
3332                         bbr->alloc_limit_reported = 1;
3333                         BBR_STAT_INC(bbr_alloc_limited_conns);
3334                 }
3335                 return (NULL);
3336         }
3337         return (bbr_alloc(bbr));
3338 }
3339
3340
3341 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3342 static struct bbr_sendmap *
3343 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type)
3344 {
3345         struct bbr_sendmap *rsm;
3346
3347         if (limit_type) {
3348                 /* currently there is only one limit type */
3349                 if (V_tcp_map_split_limit > 0 &&
3350                     bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
3351                         BBR_STAT_INC(bbr_split_limited);
3352                         if (!bbr->alloc_limit_reported) {
3353                                 bbr->alloc_limit_reported = 1;
3354                                 BBR_STAT_INC(bbr_alloc_limited_conns);
3355                         }
3356                         return (NULL);
3357                 }
3358         }
3359
3360         /* allocate and mark in the limit type, if set */
3361         rsm = bbr_alloc(bbr);
3362         if (rsm != NULL && limit_type) {
3363                 rsm->r_limit_type = limit_type;
3364                 bbr->r_ctl.rc_num_split_allocs++;
3365         }
3366         return (rsm);
3367 }
3368
3369 static void
3370 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
3371 {
3372         if (rsm->r_limit_type) {
3373                 /* currently there is only one limit type */
3374                 bbr->r_ctl.rc_num_split_allocs--;
3375         }
3376         if (rsm->r_is_smallmap)
3377                 bbr->r_ctl.rc_num_small_maps_alloced--;
3378         if (bbr->r_ctl.rc_tlp_send == rsm)
3379                 bbr->r_ctl.rc_tlp_send = NULL;
3380         if (bbr->r_ctl.rc_resend == rsm) {
3381                 bbr->r_ctl.rc_resend = NULL;
3382         }
3383         if (bbr->r_ctl.rc_next == rsm)
3384                 bbr->r_ctl.rc_next = NULL;
3385         if (bbr->r_ctl.rc_sacklast == rsm)
3386                 bbr->r_ctl.rc_sacklast = NULL;
3387         if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
3388                 memset(rsm, 0, sizeof(struct bbr_sendmap));
3389                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
3390                 rsm->r_limit_type = 0;
3391                 bbr->r_ctl.rc_free_cnt++;
3392                 return;
3393         }
3394         bbr->r_ctl.rc_num_maps_alloced--;
3395         uma_zfree(bbr_zone, rsm);
3396 }
3397
3398 /*
3399  * Returns the BDP.
3400  */
3401 static uint64_t
3402 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) {
3403         /*
3404          * Calculate the bytes in flight needed given the bw (in bytes per
3405          * second) and the specifyed rtt in useconds. We need to put out the
3406          * returned value per RTT to match that rate. Gain will normaly
3407          * raise it up from there.
3408          *
3409          * This should not overflow as long as the bandwidth is below 1
3410          * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller
3411          * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30).
3412          */
3413         uint64_t usec_per_sec;
3414
3415         usec_per_sec = USECS_IN_SECOND;
3416         return ((rtt * bw) / usec_per_sec);
3417 }
3418
3419 /*
3420  * Return the initial cwnd.
3421  */
3422 static uint32_t
3423 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp)
3424 {
3425         uint32_t i_cwnd;
3426
3427         if (bbr->rc_init_win) {
3428                 i_cwnd = bbr->rc_init_win * tp->t_maxseg;
3429         } else if (V_tcp_initcwnd_segments)
3430                 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg),
3431                     max(2 * tp->t_maxseg, 14600));
3432         else if (V_tcp_do_rfc3390)
3433                 i_cwnd = min(4 * tp->t_maxseg,
3434                     max(2 * tp->t_maxseg, 4380));
3435         else {
3436                 /* Per RFC5681 Section 3.1 */
3437                 if (tp->t_maxseg > 2190)
3438                         i_cwnd = 2 * tp->t_maxseg;
3439                 else if (tp->t_maxseg > 1095)
3440                         i_cwnd = 3 * tp->t_maxseg;
3441                 else
3442                         i_cwnd = 4 * tp->t_maxseg;
3443         }
3444         return (i_cwnd);
3445 }
3446
3447 /*
3448  * Given a specified gain, return the target
3449  * cwnd based on that gain.
3450  */
3451 static uint32_t
3452 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw)
3453 {
3454         uint64_t bdp, rtt;
3455         uint32_t cwnd;
3456
3457         if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) ||
3458             (bbr_get_full_bw(bbr) == 0)) {
3459                 /* No measurements yet */
3460                 return (bbr_initial_cwnd(bbr, bbr->rc_tp));
3461         }
3462         /*
3463          * Get bytes per RTT needed (rttProp is normally in
3464          * bbr_cwndtarget_rtt_touse)
3465          */
3466         rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse);
3467         /* Get the bdp from the two values */
3468         bdp = bbr_get_bw_delay_prod(rtt, bw);
3469         /* Now apply the gain */
3470         cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT));
3471
3472         return (cwnd);
3473 }
3474
3475 static uint32_t
3476 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain)
3477 {
3478         uint32_t cwnd, mss;
3479
3480         mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
3481         /* Get the base cwnd with gain rounded to a mss */
3482         cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss);
3483         /*
3484          * Add in N (2 default since we do not have a
3485          * fq layer to trap packets in) quanta's per the I-D
3486          * section 4.2.3.2 quanta adjust.
3487          */
3488         cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs);
3489         if (bbr->rc_use_google) {
3490                 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3491                    (bbr_state_val(bbr) == BBR_SUB_GAIN)) {
3492                         /*
3493                          * The linux implementation adds
3494                          * an extra 2 x mss in gain cycle which
3495                          * is documented no-where except in the code.
3496                          * so we add more for Neal undocumented feature
3497                          */
3498                         cwnd += 2 * mss;
3499                 }
3500                 if ((cwnd / mss) & 0x1) {
3501                         /* Round up for odd num mss */
3502                         cwnd += mss;
3503                 }
3504         }
3505         /* Are we below the min cwnd? */
3506         if (cwnd < get_min_cwnd(bbr))
3507                 return (get_min_cwnd(bbr));
3508         return (cwnd);
3509 }
3510
3511 static uint16_t
3512 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain)
3513 {
3514         if (gain < 1)
3515                 gain = 1;
3516         return (gain);
3517 }
3518
3519 static uint32_t
3520 bbr_get_header_oh(struct tcp_bbr *bbr)
3521 {
3522         int seg_oh;
3523
3524         seg_oh = 0;
3525         if (bbr->r_ctl.rc_inc_tcp_oh) {
3526                 /* Do we include TCP overhead? */
3527                 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr));
3528         }
3529         if (bbr->r_ctl.rc_inc_ip_oh) {
3530                 /* Do we include IP overhead? */
3531 #ifdef INET6
3532                 if (bbr->r_is_v6)
3533                         seg_oh += sizeof(struct ip6_hdr);
3534                 else
3535 #endif
3536 #ifdef INET
3537                         seg_oh += sizeof(struct ip);
3538 #endif
3539         }
3540         if (bbr->r_ctl.rc_inc_enet_oh) {
3541                 /* Do we include the ethernet overhead?  */
3542                 seg_oh += sizeof(struct ether_header);
3543         }
3544         return(seg_oh);
3545 }
3546
3547
3548 static uint32_t
3549 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw)
3550 {
3551         uint64_t divor, res, tim;
3552
3553         if (useconds_time == 0)
3554                 return (0);
3555         gain = bbr_gain_adjust(bbr, gain);
3556         divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT;
3557         tim = useconds_time;
3558         res = (tim * bw * gain) / divor;
3559         if (res == 0)
3560                 res = 1;
3561         return ((uint32_t)res);
3562 }
3563
3564 /*
3565  * Given a gain and a length return the delay in useconds that
3566  * should be used to evenly space out packets
3567  * on the connection (based on the gain factor).
3568  */
3569 static uint32_t
3570 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog)
3571 {
3572         uint64_t bw, lentim, res;
3573         uint32_t usecs, srtt, over = 0;
3574         uint32_t seg_oh, num_segs, maxseg;
3575
3576         if (len == 0)
3577                 return (0);
3578
3579         maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
3580         num_segs = (len + maxseg - 1) / maxseg;
3581         if (bbr->rc_use_google == 0) {
3582                 seg_oh = bbr_get_header_oh(bbr);
3583                 len += (num_segs * seg_oh);
3584         }
3585         gain = bbr_gain_adjust(bbr, gain);
3586         bw = bbr_get_bw(bbr);
3587         if (bbr->rc_use_google) {
3588                 uint64_t cbw;
3589
3590                 /*
3591                  * Reduce the b/w by the google discount
3592                  * factor 10 = 1%.
3593                  */
3594                 cbw = bw *  (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount);
3595                 cbw /= (uint64_t)1000;
3596                 /* We don't apply a discount if it results in 0 */
3597                 if (cbw > 0)
3598                         bw = cbw;
3599         }
3600         lentim = ((uint64_t)len *
3601                   (uint64_t)USECS_IN_SECOND *
3602                   (uint64_t)BBR_UNIT);
3603         res = lentim / ((uint64_t)gain * bw);
3604         if (res == 0)
3605                 res = 1;
3606         usecs = (uint32_t)res;
3607         srtt = bbr_get_rtt(bbr, BBR_SRTT);
3608         if (bbr_hptsi_max_mul && bbr_hptsi_max_div &&
3609             (bbr->rc_use_google == 0) &&
3610             (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) {
3611                 /*
3612                  * We cannot let the delay be more than 1/2 the srtt time.
3613                  * Otherwise we cannot pace out or send properly.
3614                  */
3615                 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div;
3616                 BBR_STAT_INC(bbr_hpts_min_time);
3617         }
3618         if (!nolog)
3619                 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1);
3620         return (usecs);
3621 }
3622
3623 static void
3624 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack,
3625                  uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses)
3626 {
3627         INP_WLOCK_ASSERT(tp->t_inpcb);
3628         uint64_t bw;
3629         uint32_t cwnd, target_cwnd, saved_bytes, maxseg;
3630         int32_t meth;
3631
3632 #ifdef STATS
3633         if ((tp->t_flags & TF_GPUTINPROG) &&
3634             SEQ_GEQ(th->th_ack, tp->gput_ack)) {
3635                 /*
3636                  * Strech acks and compressed acks will cause this to
3637                  * oscillate but we are doing it the same way as the main
3638                  * stack so it will be compariable (though possibly not
3639                  * ideal).
3640                  */
3641                 int32_t cgput;
3642                 int64_t gput, time_stamp;
3643
3644                 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8;
3645                 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000));
3646                 cgput = gput / time_stamp;
3647                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
3648                                          cgput);
3649                 if (tp->t_stats_gput_prev > 0)
3650                         stats_voi_update_abs_s32(tp->t_stats,
3651                                                  VOI_TCP_GPUT_ND,
3652                                                  ((gput - tp->t_stats_gput_prev) * 100) /
3653                                                  tp->t_stats_gput_prev);
3654                 tp->t_flags &= ~TF_GPUTINPROG;
3655                 tp->t_stats_gput_prev = cgput;
3656         }
3657 #endif
3658         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3659             ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
3660                 /* We don't change anything in probe-rtt */
3661                 return;
3662         }
3663         maxseg = tp->t_maxseg - bbr->rc_last_options;
3664         saved_bytes = bytes_this_ack;
3665         bytes_this_ack += sack_changed;
3666         if (bytes_this_ack > prev_acked) {
3667                 bytes_this_ack -= prev_acked;
3668                 /*
3669                  * A byte ack'd gives us a full mss
3670                  * to be like linux i.e. they count packets.
3671                  */
3672                 if ((bytes_this_ack < maxseg) && bbr->rc_use_google)
3673                         bytes_this_ack = maxseg;
3674         } else {
3675                 /* Unlikely */
3676                 bytes_this_ack = 0;
3677         }
3678         cwnd = tp->snd_cwnd;
3679         bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3680         if (bw)
3681                 target_cwnd = bbr_get_target_cwnd(bbr,
3682                                                   bw,
3683                                                   (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain);
3684         else
3685                 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp);
3686         if (IN_RECOVERY(tp->t_flags) &&
3687             (bbr->bbr_prev_in_rec == 0)) {
3688                 /*
3689                  * We are entering recovery and
3690                  * thus packet conservation.
3691                  */
3692                 bbr->pkt_conservation = 1;
3693                 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime;
3694                 cwnd = ctf_flight_size(tp,
3695                                        (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3696                         bytes_this_ack;
3697         }
3698         if (IN_RECOVERY(tp->t_flags)) {
3699                 uint32_t flight;
3700
3701                 bbr->bbr_prev_in_rec = 1;
3702                 if (cwnd > losses) {
3703                         cwnd -= losses;
3704                         if (cwnd < maxseg)
3705                                 cwnd = maxseg;
3706                 } else
3707                         cwnd = maxseg;
3708                 flight = ctf_flight_size(tp,
3709                                          (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3710                 bbr_log_type_cwndupd(bbr, flight, 0,
3711                                      losses, 10, 0, 0, line);
3712                 if (bbr->pkt_conservation) {
3713                         uint32_t time_in;
3714
3715                         if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start))
3716                                 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start;
3717                         else
3718                                 time_in = 0;
3719
3720                         if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
3721                                 /* Clear packet conservation after an rttProp */
3722                                 bbr->pkt_conservation = 0;
3723                         } else {
3724                                 if ((flight + bytes_this_ack) > cwnd)
3725                                         cwnd = flight + bytes_this_ack;
3726                                 if (cwnd < get_min_cwnd(bbr))
3727                                         cwnd = get_min_cwnd(bbr);
3728                                 tp->snd_cwnd = cwnd;
3729                                 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed,
3730                                                      prev_acked, 1, target_cwnd, th->th_ack, line);
3731                                 return;
3732                         }
3733                 }
3734         } else
3735                 bbr->bbr_prev_in_rec = 0;
3736         if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) {
3737                 bbr->r_ctl.restrict_growth--;
3738                 if (bytes_this_ack > maxseg)
3739                         bytes_this_ack = maxseg;
3740         }
3741         if (bbr->rc_filled_pipe) {
3742                 /*
3743                  * Here we have exited startup and filled the pipe. We will
3744                  * thus allow the cwnd to shrink to the target. We hit here
3745                  * mostly.
3746                  */
3747                 uint32_t s_cwnd;
3748
3749                 meth = 2;
3750                 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd);
3751                 if (s_cwnd > cwnd)
3752                         cwnd = s_cwnd;
3753                 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing)
3754                         cwnd = s_cwnd;
3755         } else {
3756                 /*
3757                  * Here we are still in startup, we increase cwnd by what
3758                  * has been acked.
3759                  */
3760                 if ((cwnd < target_cwnd) ||
3761                     (bbr->rc_past_init_win == 0)) {
3762                         meth = 3;
3763                         cwnd += bytes_this_ack;
3764                 } else {
3765                         /*
3766                          * Method 4 means we are at target so no gain in
3767                          * startup and past the initial window.
3768                          */
3769                         meth = 4;
3770                 }
3771         }
3772         tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr));
3773         bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line);
3774 }
3775
3776 static void
3777 tcp_bbr_partialack(struct tcpcb *tp)
3778 {
3779         struct tcp_bbr *bbr;
3780
3781         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3782         INP_WLOCK_ASSERT(tp->t_inpcb);
3783         if (ctf_flight_size(tp,
3784                 (bbr->r_ctl.rc_sacked  + bbr->r_ctl.rc_lost_bytes)) <=
3785             tp->snd_cwnd) {
3786                 bbr->r_wanted_output = 1;
3787         }
3788 }
3789
3790 static void
3791 bbr_post_recovery(struct tcpcb *tp)
3792 {
3793         struct tcp_bbr *bbr;
3794         uint32_t  flight;
3795
3796         INP_WLOCK_ASSERT(tp->t_inpcb);
3797         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3798         /*
3799          * Here we just exit recovery.
3800          */
3801         EXIT_RECOVERY(tp->t_flags);
3802         /* Lock in our b/w reduction for the specified number of pkt-epochs */
3803         bbr->r_recovery_bw = 0;
3804         tp->snd_recover = tp->snd_una;
3805         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3806         bbr->pkt_conservation = 0;
3807         if (bbr->rc_use_google == 0) {
3808                 /*
3809                  * For non-google mode lets
3810                  * go ahead and make sure we clear
3811                  * the recovery state so if we
3812                  * bounce back in to recovery we
3813                  * will do PC.
3814                  */
3815                 bbr->bbr_prev_in_rec = 0;
3816         }
3817         bbr_log_type_exit_rec(bbr);
3818         if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
3819                 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
3820                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__);
3821         } else {
3822                 /* For probe-rtt case lets fix up its saved_cwnd */
3823                 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) {
3824                         bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent;
3825                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__);
3826                 }
3827         }
3828         flight = ctf_flight_size(tp,
3829                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3830         if ((bbr->rc_use_google == 0) &&
3831             bbr_do_red) {
3832                 uint64_t val, lr2use;
3833                 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd;
3834                 uint32_t *cwnd_p;
3835
3836                 if (bbr_get_rtt(bbr, BBR_SRTT)) {
3837                         val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000);
3838                         val /= bbr_get_rtt(bbr, BBR_SRTT);
3839                         ratio = (uint32_t)val;
3840                 } else
3841                         ratio = 1000;
3842
3843                 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div,
3844                                      bbr->r_ctl.recovery_lr, 21,
3845                                      ratio,
3846                                      bbr->r_ctl.rc_red_cwnd_pe,
3847                                      __LINE__);
3848                 if ((ratio < bbr_do_red) || (bbr_do_red == 0))
3849                         goto done;
3850                 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3851                      bbr_prtt_slam_cwnd) ||
3852                     (bbr_sub_drain_slam_cwnd &&
3853                      (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3854                      bbr->rc_hit_state_1 &&
3855                      (bbr_state_val(bbr) == BBR_SUB_DRAIN)) ||
3856                     ((bbr->rc_bbr_state == BBR_STATE_DRAIN) &&
3857                      bbr_slam_cwnd_in_main_drain)) {
3858                         /*
3859                          * Here we must poke at the saved cwnd
3860                          * as well as the cwnd.
3861                          */
3862                         cwnd = bbr->r_ctl.rc_saved_cwnd;
3863                         cwnd_p = &bbr->r_ctl.rc_saved_cwnd;
3864                 } else {
3865                         cwnd = tp->snd_cwnd;
3866                         cwnd_p = &tp->snd_cwnd;
3867                 }
3868                 maxseg = tp->t_maxseg - bbr->rc_last_options;
3869                 /* Add the overall lr with the recovery lr */
3870                 if (bbr->r_ctl.rc_lost == 0)
3871                         lr2use = 0;
3872                 else if (bbr->r_ctl.rc_delivered == 0)
3873                         lr2use = 1000;
3874                 else {
3875                         lr2use = bbr->r_ctl.rc_lost * 1000;
3876                         lr2use /= bbr->r_ctl.rc_delivered;
3877                 }
3878                 lr2use += bbr->r_ctl.recovery_lr;
3879                 acks_inflight = (flight / (maxseg * 2));
3880                 if (bbr_red_scale) {
3881                         lr2use *= bbr_get_rtt(bbr, BBR_SRTT);
3882                         lr2use /= bbr_red_scale;
3883                         if ((bbr_red_growth_restrict) &&
3884                             ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1))
3885                             bbr->r_ctl.restrict_growth += acks_inflight;
3886                 }
3887                 if (lr2use) {
3888                         val = (uint64_t)cwnd * lr2use;
3889                         val /= 1000;
3890                         if (cwnd > val)
3891                                 newcwnd = roundup((cwnd - val), maxseg);
3892                         else
3893                                 newcwnd = maxseg;
3894                 } else {
3895                         val = (uint64_t)cwnd * (uint64_t)bbr_red_mul;
3896                         val /= (uint64_t)bbr_red_div;
3897                         newcwnd = roundup((uint32_t)val, maxseg);
3898                 }
3899                 /* with standard delayed acks how many acks can I expect? */
3900                 if (bbr_drop_limit == 0) {
3901                         /*
3902                          * Anticpate how much we will
3903                          * raise the cwnd based on the acks.
3904                          */
3905                         if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) {
3906                                 /* We do enforce the min (with the acks) */
3907                                 newcwnd = (get_min_cwnd(bbr) - acks_inflight);
3908                         }
3909                 } else {
3910                         /*
3911                          * A strict drop limit of N is is inplace
3912                          */
3913                         if (newcwnd < (bbr_drop_limit * maxseg)) {
3914                                 newcwnd = bbr_drop_limit * maxseg;
3915                         }
3916                 }
3917                 /* For the next N acks do we restrict the growth */
3918                 *cwnd_p = newcwnd;
3919                 if (tp->snd_cwnd > newcwnd)
3920                         tp->snd_cwnd = newcwnd;
3921                 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22,
3922                                      (uint32_t)lr2use,
3923                                      bbr_get_rtt(bbr, BBR_SRTT), __LINE__);
3924                 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch;
3925         }
3926 done:
3927         bbr->r_ctl.recovery_lr = 0;
3928         if (flight <= tp->snd_cwnd) {
3929                 bbr->r_wanted_output = 1;
3930         }
3931         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3932 }
3933
3934 static void
3935 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts)
3936 {
3937         bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3938         /* Limit the drop in b/w to 1/2 our current filter. */
3939         if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate)
3940                 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate;
3941         if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2))
3942                 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2;
3943         tcp_bbr_tso_size_check(bbr, cts);
3944 }
3945
3946 static void
3947 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm)
3948 {
3949         struct tcp_bbr *bbr;
3950
3951         INP_WLOCK_ASSERT(tp->t_inpcb);
3952         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3953         switch (type) {
3954         case CC_NDUPACK:
3955                 if (!IN_RECOVERY(tp->t_flags)) {
3956                         tp->snd_recover = tp->snd_max;
3957                         /* Start a new epoch */
3958                         bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
3959                         if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) {
3960                                 /*
3961                                  * Move forward the lt epoch
3962                                  * so it won't count the truncated
3963                                  * epoch.
3964                                  */
3965                                 bbr->r_ctl.rc_lt_epoch++;
3966                         }
3967                         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
3968                                 /*
3969                                  * Just like the policer detection code
3970                                  * if we are in startup we must push
3971                                  * forward the last startup epoch
3972                                  * to hide the truncated PE.
3973                                  */
3974                                 bbr->r_ctl.rc_bbr_last_startup_epoch++;
3975                         }
3976                         bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd;
3977                         ENTER_RECOVERY(tp->t_flags);
3978                         bbr->rc_tlp_rtx_out = 0;
3979                         bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate;
3980                         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3981                         if (bbr->rc_inp->inp_in_hpts &&
3982                             ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) {
3983                                 /*
3984                                  * When we enter recovery, we need to restart
3985                                  * any timers. This may mean we gain an agg
3986                                  * early, which will be made up for at the last
3987                                  * rxt out.
3988                                  */
3989                                 bbr->rc_timer_first = 1;
3990                                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
3991                         }
3992                         /*
3993                          * Calculate a new cwnd based on to the current
3994                          * delivery rate with no gain. We get the bdp
3995                          * without gaining it up like we normally would and
3996                          * we use the last cur_del_rate.
3997                          */
3998                         if ((bbr->rc_use_google == 0) &&
3999                             (bbr->r_ctl.bbr_rttprobe_gain_val ||
4000                              (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) {
4001                                 tp->snd_cwnd = ctf_flight_size(tp,
4002                                                    (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
4003                                         (tp->t_maxseg - bbr->rc_last_options);
4004                                 if (tp->snd_cwnd < get_min_cwnd(bbr)) {
4005                                         /* We always gate to min cwnd */
4006                                         tp->snd_cwnd = get_min_cwnd(bbr);
4007                                 }
4008                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__);
4009                         }
4010                         bbr_log_type_enter_rec(bbr, rsm->r_start);
4011                 }
4012                 break;
4013         case CC_RTO_ERR:
4014                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4015                 /* RTO was unnecessary, so reset everything. */
4016                 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime);
4017                 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
4018                         tp->snd_cwnd = tp->snd_cwnd_prev;
4019                         tp->snd_ssthresh = tp->snd_ssthresh_prev;
4020                         tp->snd_recover = tp->snd_recover_prev;
4021                         tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
4022                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__);
4023                 }
4024                 tp->t_badrxtwin = 0;
4025                 break;
4026         }
4027 }
4028
4029 /*
4030  * Indicate whether this ack should be delayed.  We can delay the ack if
4031  * following conditions are met:
4032  *      - There is no delayed ack timer in progress.
4033  *      - Our last ack wasn't a 0-sized window. We never want to delay
4034  *        the ack that opens up a 0-sized window.
4035  *      - LRO wasn't used for this segment. We make sure by checking that the
4036  *        segment size is not larger than the MSS.
4037  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4038  *        connection.
4039  *      - The data being acked is less than a full segment (a stretch ack
4040  *        of more than a segment we should ack.
4041  *      - nsegs is 1 (if its more than that we received more than 1 ack).
4042  */
4043 #define DELAY_ACK(tp, bbr, nsegs)                               \
4044         (((tp->t_flags & TF_RXWIN0SENT) == 0) &&                \
4045          ((tp->t_flags & TF_DELACK) == 0) &&                    \
4046          ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) &&  \
4047          (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4048
4049 /*
4050  * Return the lowest RSM in the map of
4051  * packets still in flight that is not acked.
4052  * This should normally find on the first one
4053  * since we remove packets from the send
4054  * map after they are marked ACKED.
4055  */
4056 static struct bbr_sendmap *
4057 bbr_find_lowest_rsm(struct tcp_bbr *bbr)
4058 {
4059         struct bbr_sendmap *rsm;
4060
4061         /*
4062          * Walk the time-order transmitted list looking for an rsm that is
4063          * not acked. This will be the one that was sent the longest time
4064          * ago that is still outstanding.
4065          */
4066         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) {
4067                 if (rsm->r_flags & BBR_ACKED) {
4068                         continue;
4069                 }
4070                 goto finish;
4071         }
4072 finish:
4073         return (rsm);
4074 }
4075
4076 static struct bbr_sendmap *
4077 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
4078 {
4079         struct bbr_sendmap *prsm;
4080
4081         /*
4082          * Walk the sequence order list backward until we hit and arrive at
4083          * the highest seq not acked. In theory when this is called it
4084          * should be the last segment (which it was not).
4085          */
4086         prsm = rsm;
4087         TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4088                 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) {
4089                         continue;
4090                 }
4091                 return (prsm);
4092         }
4093         return (NULL);
4094 }
4095
4096 /*
4097  * Returns to the caller the number of microseconds that
4098  * the packet can be outstanding before we think we
4099  * should have had an ack returned.
4100  */
4101 static uint32_t
4102 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm)
4103 {
4104         /*
4105          * lro is the flag we use to determine if we have seen reordering.
4106          * If it gets set we have seen reordering. The reorder logic either
4107          * works in one of two ways:
4108          *
4109          * If reorder-fade is configured, then we track the last time we saw
4110          * re-ordering occur. If we reach the point where enough time as
4111          * passed we no longer consider reordering has occuring.
4112          *
4113          * Or if reorder-face is 0, then once we see reordering we consider
4114          * the connection to alway be subject to reordering and just set lro
4115          * to 1.
4116          *
4117          * In the end if lro is non-zero we add the extra time for
4118          * reordering in.
4119          */
4120         int32_t lro;
4121         uint32_t thresh, t_rxtcur;
4122
4123         if (srtt == 0)
4124                 srtt = 1;
4125         if (bbr->r_ctl.rc_reorder_ts) {
4126                 if (bbr->r_ctl.rc_reorder_fade) {
4127                         if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) {
4128                                 lro = cts - bbr->r_ctl.rc_reorder_ts;
4129                                 if (lro == 0) {
4130                                         /*
4131                                          * No time as passed since the last
4132                                          * reorder, mark it as reordering.
4133                                          */
4134                                         lro = 1;
4135                                 }
4136                         } else {
4137                                 /* Negative time? */
4138                                 lro = 0;
4139                         }
4140                         if (lro > bbr->r_ctl.rc_reorder_fade) {
4141                                 /* Turn off reordering seen too */
4142                                 bbr->r_ctl.rc_reorder_ts = 0;
4143                                 lro = 0;
4144                         }
4145                 } else {
4146                         /* Reodering does not fade */
4147                         lro = 1;
4148                 }
4149         } else {
4150                 lro = 0;
4151         }
4152         thresh = srtt + bbr->r_ctl.rc_pkt_delay;
4153         if (lro) {
4154                 /* It must be set, if not you get 1/4 rtt */
4155                 if (bbr->r_ctl.rc_reorder_shift)
4156                         thresh += (srtt >> bbr->r_ctl.rc_reorder_shift);
4157                 else
4158                         thresh += (srtt >> 2);
4159         } else {
4160                 thresh += 1000;
4161         }
4162         /* We don't let the rack timeout be above a RTO */
4163         if ((bbr->rc_tp)->t_srtt == 0)
4164                 t_rxtcur = BBR_INITIAL_RTO;
4165         else
4166                 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
4167         if (thresh > t_rxtcur) {
4168                 thresh = t_rxtcur;
4169         }
4170         /* And we don't want it above the RTO max either */
4171         if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4172                 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4173         }
4174         bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK);
4175         return (thresh);
4176 }
4177
4178 /*
4179  * Return to the caller the amount of time in mico-seconds
4180  * that should be used for the TLP timer from the last
4181  * send time of this packet.
4182  */
4183 static uint32_t
4184 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
4185     struct bbr_sendmap *rsm, uint32_t srtt,
4186     uint32_t cts)
4187 {
4188         uint32_t thresh, len, maxseg, t_rxtcur;
4189         struct bbr_sendmap *prsm;
4190
4191         if (srtt == 0)
4192                 srtt = 1;
4193         if (bbr->rc_tlp_threshold)
4194                 thresh = srtt + (srtt / bbr->rc_tlp_threshold);
4195         else
4196                 thresh = (srtt * 2);
4197         maxseg = tp->t_maxseg - bbr->rc_last_options;
4198         /* Get the previous sent packet, if any  */
4199         len = rsm->r_end - rsm->r_start;
4200
4201         /* 2.1 behavior */
4202         prsm = TAILQ_PREV(rsm, bbr_head, r_tnext);
4203         if (prsm && (len <= maxseg)) {
4204                 /*
4205                  * Two packets outstanding, thresh should be (2*srtt) +
4206                  * possible inter-packet delay (if any).
4207                  */
4208                 uint32_t inter_gap = 0;
4209                 int idx, nidx;
4210
4211                 idx = rsm->r_rtr_cnt - 1;
4212                 nidx = prsm->r_rtr_cnt - 1;
4213                 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
4214                         /* Yes it was sent later (or at the same time) */
4215                         inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
4216                 }
4217                 thresh += inter_gap;
4218         } else if (len <= maxseg) {
4219                 /*
4220                  * Possibly compensate for delayed-ack.
4221                  */
4222                 uint32_t alt_thresh;
4223
4224                 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time;
4225                 if (alt_thresh > thresh)
4226                         thresh = alt_thresh;
4227         }
4228         /* Not above the current  RTO */
4229         if (tp->t_srtt == 0)
4230                 t_rxtcur = BBR_INITIAL_RTO;
4231         else
4232                 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur);
4233
4234         bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP);
4235         /* Not above an RTO */
4236         if (thresh > t_rxtcur) {
4237                 thresh = t_rxtcur;
4238         }
4239         /* Not above a RTO max */
4240         if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4241                 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4242         }
4243         /* And now apply the user TLP min */
4244         if (thresh < bbr_tlp_min) {
4245                 thresh = bbr_tlp_min;
4246         }
4247         return (thresh);
4248 }
4249
4250 /*
4251  * Return one of three RTTs to use (in microseconds).
4252  */
4253 static __inline uint32_t
4254 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type)
4255 {
4256         uint32_t f_rtt;
4257         uint32_t srtt;
4258
4259         f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
4260         if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) {
4261                 /* We have no rtt at all */
4262                 if (bbr->rc_tp->t_srtt == 0)
4263                         f_rtt = BBR_INITIAL_RTO;
4264                 else
4265                         f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4266                 /*
4267                  * Since we don't know how good the rtt is apply a
4268                  * delayed-ack min
4269                  */
4270                 if (f_rtt < bbr_delayed_ack_time) {
4271                         f_rtt = bbr_delayed_ack_time;
4272                 }
4273         }
4274         /* Take the filter version or last measured pkt-rtt */
4275         if (rtt_type == BBR_RTT_PROP) {
4276                 srtt = f_rtt;
4277         } else if (rtt_type == BBR_RTT_PKTRTT) {
4278                 if (bbr->r_ctl.rc_pkt_epoch_rtt) {
4279                         srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
4280                 } else {
4281                         /* No pkt rtt yet */
4282                         srtt = f_rtt;
4283                 }
4284         } else if (rtt_type == BBR_RTT_RACK) {
4285                 srtt = bbr->r_ctl.rc_last_rtt;
4286                 /* We need to add in any internal delay for our timer */
4287                 if (bbr->rc_ack_was_delayed)
4288                         srtt += bbr->r_ctl.rc_ack_hdwr_delay;
4289         } else if (rtt_type == BBR_SRTT) {
4290                 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4291         } else {
4292                 /* TSNH */
4293                 srtt = f_rtt;
4294 #ifdef BBR_INVARIANTS
4295                 panic("Unknown rtt request type %d", rtt_type);
4296 #endif
4297         }
4298         return (srtt);
4299 }
4300
4301 static int
4302 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts)
4303 {
4304         uint32_t thresh;
4305
4306
4307         thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK),
4308                                       cts, rsm);
4309         if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) {
4310                 /* It is lost (past time) */
4311                 return (1);
4312         }
4313         return (0);
4314 }
4315
4316 /*
4317  * Return a sendmap if we need to retransmit something.
4318  */
4319 static struct bbr_sendmap *
4320 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4321 {
4322         /*
4323          * Check to see that we don't need to fall into recovery. We will
4324          * need to do so if our oldest transmit is past the time we should
4325          * have had an ack.
4326          */
4327
4328         struct bbr_sendmap *rsm;
4329         int32_t idx;
4330
4331         if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) {
4332                 /* Nothing outstanding that we know of */
4333                 return (NULL);
4334         }
4335         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
4336         if (rsm == NULL) {
4337                 /* Nothing in the transmit map */
4338                 return (NULL);
4339         }
4340         if (tp->t_flags & TF_SENTFIN) {
4341                 /* Fin restricted, don't find anything once a fin is sent */
4342                 return (NULL);
4343         }
4344         if (rsm->r_flags & BBR_ACKED) {
4345                 /*
4346                  * Ok the first one is acked (this really should not happen
4347                  * since we remove the from the tmap once they are acked)
4348                  */
4349                 rsm = bbr_find_lowest_rsm(bbr);
4350                 if (rsm == NULL)
4351                         return (NULL);
4352         }
4353         idx = rsm->r_rtr_cnt - 1;
4354         if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) {
4355                 /* Send timestamp is the same or less? can't be ready */
4356                 return (NULL);
4357         }
4358         /* Get our RTT time */
4359         if (bbr_is_lost(bbr, rsm, cts) &&
4360             ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
4361              (rsm->r_flags & BBR_SACK_PASSED))) {
4362                 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4363                         rsm->r_flags |= BBR_MARKED_LOST;
4364                         bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4365                         bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4366                 }
4367                 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm);
4368 #ifdef BBR_INVARIANTS
4369                 if ((rsm->r_end - rsm->r_start) == 0)
4370                         panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm);
4371 #endif
4372                 return (rsm);
4373         }
4374         return (NULL);
4375 }
4376
4377 /*
4378  * RACK Timer, here we simply do logging and house keeping.
4379  * the normal bbr_output_wtime() function will call the
4380  * appropriate thing to check if we need to do a RACK retransmit.
4381  * We return 1, saying don't proceed with bbr_output_wtime only
4382  * when all timers have been stopped (destroyed PCB?).
4383  */
4384 static int
4385 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4386 {
4387         /*
4388          * This timer simply provides an internal trigger to send out data.
4389          * The check_recovery_mode call will see if there are needed
4390          * retransmissions, if so we will enter fast-recovery. The output
4391          * call may or may not do the same thing depending on sysctl
4392          * settings.
4393          */
4394         uint32_t lost;
4395
4396         if (bbr->rc_all_timers_stopped) {
4397                 return (1);
4398         }
4399         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4400                 /* Its not time yet */
4401                 return (0);
4402         }
4403         BBR_STAT_INC(bbr_to_tot);
4404         lost = bbr->r_ctl.rc_lost;
4405         if (bbr->r_state && (bbr->r_state != tp->t_state))
4406                 bbr_set_state(tp, bbr, 0);
4407         bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK);
4408         if (bbr->r_ctl.rc_resend == NULL) {
4409                 /* Lets do the check here */
4410                 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
4411         }
4412         if (bbr_policer_call_from_rack_to)
4413                 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4414         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
4415         return (0);
4416 }
4417
4418 static __inline void
4419 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start)
4420 {
4421         int idx;
4422
4423         nrsm->r_start = start;
4424         nrsm->r_end = rsm->r_end;
4425         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
4426         nrsm->r_flags = rsm->r_flags;
4427         /* We don't transfer forward the SYN flag */
4428         nrsm->r_flags &= ~BBR_HAS_SYN;
4429         /* We move forward the FIN flag, not that this should happen */
4430         rsm->r_flags &= ~BBR_HAS_FIN;
4431         nrsm->r_dupack = rsm->r_dupack;
4432         nrsm->r_rtr_bytes = 0;
4433         nrsm->r_is_gain = rsm->r_is_gain;
4434         nrsm->r_is_drain = rsm->r_is_drain;
4435         nrsm->r_delivered = rsm->r_delivered;
4436         nrsm->r_ts_valid = rsm->r_ts_valid;
4437         nrsm->r_del_ack_ts = rsm->r_del_ack_ts;
4438         nrsm->r_del_time = rsm->r_del_time;
4439         nrsm->r_app_limited = rsm->r_app_limited;
4440         nrsm->r_first_sent_time = rsm->r_first_sent_time;
4441         nrsm->r_flight_at_send = rsm->r_flight_at_send;
4442         /* We split a piece the lower section looses any just_ret flag. */
4443         nrsm->r_bbr_state = rsm->r_bbr_state;
4444         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
4445                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
4446         }
4447         rsm->r_end = nrsm->r_start;
4448         idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
4449         idx /= 8;
4450         /* Check if we got too small */
4451         if ((rsm->r_is_smallmap == 0) &&
4452             ((rsm->r_end - rsm->r_start) <= idx)) {
4453                 bbr->r_ctl.rc_num_small_maps_alloced++;
4454                 rsm->r_is_smallmap = 1;
4455         }
4456         /* Check the new one as well */
4457         if ((nrsm->r_end - nrsm->r_start) <= idx) {
4458                 bbr->r_ctl.rc_num_small_maps_alloced++;
4459                 nrsm->r_is_smallmap = 1;
4460         }
4461 }
4462
4463 static int
4464 bbr_sack_mergable(struct bbr_sendmap *at,
4465                   uint32_t start, uint32_t end)
4466 {
4467         /*
4468          * Given a sack block defined by
4469          * start and end, and a current postion
4470          * at. Return 1 if either side of at
4471          * would show that the block is mergable
4472          * to that side. A block to be mergable
4473          * must have overlap with the start/end
4474          * and be in the SACK'd state.
4475          */
4476         struct bbr_sendmap *l_rsm;
4477         struct bbr_sendmap *r_rsm;
4478
4479         /* first get the either side blocks */
4480         l_rsm = TAILQ_PREV(at, bbr_head, r_next);
4481         r_rsm = TAILQ_NEXT(at, r_next);
4482         if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) {
4483                 /* Potentially mergeable */
4484                 if ((l_rsm->r_end == start) ||
4485                     (SEQ_LT(start, l_rsm->r_end) &&
4486                      SEQ_GT(end, l_rsm->r_end))) {
4487                             /*
4488                              * map blk   |------|
4489                              * sack blk         |------|
4490                              * <or>
4491                              * map blk   |------|
4492                              * sack blk      |------|
4493                              */
4494                             return (1);
4495                     }
4496         }
4497         if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) {
4498                 /* Potentially mergeable */
4499                 if ((r_rsm->r_start == end) ||
4500                     (SEQ_LT(start, r_rsm->r_start) &&
4501                      SEQ_GT(end, r_rsm->r_start))) {
4502                         /*
4503                          * map blk          |---------|
4504                          * sack blk    |----|
4505                          * <or>
4506                          * map blk          |---------|
4507                          * sack blk    |-------|
4508                          */
4509                         return (1);
4510                 }
4511         }
4512         return (0);
4513 }
4514
4515 static struct bbr_sendmap *
4516 bbr_merge_rsm(struct tcp_bbr *bbr,
4517               struct bbr_sendmap *l_rsm,
4518               struct bbr_sendmap *r_rsm)
4519 {
4520         /*
4521          * We are merging two ack'd RSM's,
4522          * the l_rsm is on the left (lower seq
4523          * values) and the r_rsm is on the right
4524          * (higher seq value). The simplest way
4525          * to merge these is to move the right
4526          * one into the left. I don't think there
4527          * is any reason we need to try to find
4528          * the oldest (or last oldest retransmitted).
4529          */
4530         l_rsm->r_end = r_rsm->r_end;
4531         if (l_rsm->r_dupack < r_rsm->r_dupack)
4532                 l_rsm->r_dupack = r_rsm->r_dupack;
4533         if (r_rsm->r_rtr_bytes)
4534                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
4535         if (r_rsm->r_in_tmap) {
4536                 /* This really should not happen */
4537                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext);
4538         }
4539         if (r_rsm->r_app_limited)
4540                 l_rsm->r_app_limited = r_rsm->r_app_limited;
4541         /* Now the flags */
4542         if (r_rsm->r_flags & BBR_HAS_FIN)
4543                 l_rsm->r_flags |= BBR_HAS_FIN;
4544         if (r_rsm->r_flags & BBR_TLP)
4545                 l_rsm->r_flags |= BBR_TLP;
4546         if (r_rsm->r_flags & BBR_RWND_COLLAPSED)
4547                 l_rsm->r_flags |= BBR_RWND_COLLAPSED;
4548         if (r_rsm->r_flags & BBR_MARKED_LOST) {
4549                 /* This really should not happen */
4550                 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start;
4551         }
4552         TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next);
4553         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
4554                 /* Transfer the split limit to the map we free */
4555                 r_rsm->r_limit_type = l_rsm->r_limit_type;
4556                 l_rsm->r_limit_type = 0;
4557         }
4558         bbr_free(bbr, r_rsm);
4559         return(l_rsm);
4560 }
4561
4562 /*
4563  * TLP Timer, here we simply setup what segment we want to
4564  * have the TLP expire on, the normal bbr_output_wtime() will then
4565  * send it out.
4566  *
4567  * We return 1, saying don't proceed with bbr_output_wtime only
4568  * when all timers have been stopped (destroyed PCB?).
4569  */
4570 static int
4571 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4572 {
4573         /*
4574          * Tail Loss Probe.
4575          */
4576         struct bbr_sendmap *rsm = NULL;
4577         struct socket *so;
4578         uint32_t amm;
4579         uint32_t out, avail;
4580         uint32_t maxseg;
4581         int collapsed_win = 0;
4582
4583         if (bbr->rc_all_timers_stopped) {
4584                 return (1);
4585         }
4586         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4587                 /* Its not time yet */
4588                 return (0);
4589         }
4590         if (ctf_progress_timeout_check(tp, true)) {
4591                 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4592                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4593                 return (1);
4594         }
4595         /* Did we somehow get into persists? */
4596         if (bbr->rc_in_persist) {
4597                 return (0);
4598         }
4599         if (bbr->r_state && (bbr->r_state != tp->t_state))
4600                 bbr_set_state(tp, bbr, 0);
4601         BBR_STAT_INC(bbr_tlp_tot);
4602         maxseg = tp->t_maxseg - bbr->rc_last_options;
4603 #ifdef KERN_TLS
4604         if (bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) {
4605                 /*
4606                  * For hardware TLS we do *not* want to send
4607                  * new data.
4608                  */
4609                 goto need_retran;
4610         }
4611 #endif
4612         /*
4613          * A TLP timer has expired. We have been idle for 2 rtts. So we now
4614          * need to figure out how to force a full MSS segment out.
4615          */
4616         so = tp->t_inpcb->inp_socket;
4617         avail = sbavail(&so->so_snd);
4618         out = ctf_outstanding(tp);
4619         if (out > tp->snd_wnd) {
4620                 /* special case, we need a retransmission */
4621                 collapsed_win = 1;
4622                 goto need_retran;
4623         }
4624         if (avail > out) {
4625                 /* New data is available */
4626                 amm = avail - out;
4627                 if (amm > maxseg) {
4628                         amm = maxseg;
4629                 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) {
4630                         /* not enough to fill a MTU and no-delay is off */
4631                         goto need_retran;
4632                 }
4633                 /* Set the send-new override */
4634                 if ((out + amm) <= tp->snd_wnd) {
4635                         bbr->rc_tlp_new_data = 1;
4636                 } else {
4637                         goto need_retran;
4638                 }
4639                 bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4640                 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max;
4641                 bbr->r_ctl.rc_tlp_send = NULL;
4642                 /* cap any slots */
4643                 BBR_STAT_INC(bbr_tlp_newdata);
4644                 goto send;
4645         }
4646 need_retran:
4647         /*
4648          * Ok we need to arrange the last un-acked segment to be re-sent, or
4649          * optionally the first un-acked segment.
4650          */
4651         if (collapsed_win == 0) {
4652                 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
4653                 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) {
4654                         rsm = bbr_find_high_nonack(bbr, rsm);
4655                 }
4656                 if (rsm == NULL) {
4657                         goto restore;
4658                 }
4659         } else {
4660                 /*
4661                  * We must find the last segment
4662                  * that was acceptable by the client.
4663                  */
4664                 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4665                         if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) {
4666                                 /* Found one */
4667                                 break;
4668                         }
4669                 }
4670                 if (rsm == NULL) {
4671                         /* None? if so send the first */
4672                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4673                         if (rsm == NULL)
4674                                 goto restore;
4675                 }
4676         }
4677         if ((rsm->r_end - rsm->r_start) > maxseg) {
4678                 /*
4679                  * We need to split this the last segment in two.
4680                  */
4681                 struct bbr_sendmap *nrsm;
4682
4683                 nrsm = bbr_alloc_full_limit(bbr);
4684                 if (nrsm == NULL) {
4685                         /*
4686                          * We can't get memory to split, we can either just
4687                          * not split it. Or retransmit the whole piece, lets
4688                          * do the large send (BTLP :-) ).
4689                          */
4690                         goto go_for_it;
4691                 }
4692                 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg));
4693                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
4694                 if (rsm->r_in_tmap) {
4695                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
4696                         nrsm->r_in_tmap = 1;
4697                 }
4698                 rsm->r_flags &= (~BBR_HAS_FIN);
4699                 rsm = nrsm;
4700         }
4701 go_for_it:
4702         bbr->r_ctl.rc_tlp_send = rsm;
4703         bbr->rc_tlp_rtx_out = 1;
4704         if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) {
4705                 bbr->r_ctl.rc_tlp_seg_send_cnt++;
4706                 tp->t_rxtshift++;
4707         } else {
4708                 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
4709                 bbr->r_ctl.rc_tlp_seg_send_cnt = 1;
4710         }
4711 send:
4712         if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
4713                 /*
4714                  * Can't [re]/transmit a segment we have retranmitted the
4715                  * max times. We need the retransmit timer to take over.
4716                  */
4717 restore:
4718                 bbr->rc_tlp_new_data = 0;
4719                 bbr->r_ctl.rc_tlp_send = NULL;
4720                 if (rsm)
4721                         rsm->r_flags &= ~BBR_TLP;
4722                 BBR_STAT_INC(bbr_tlp_retran_fail);
4723                 return (0);
4724         } else if (rsm) {
4725                 rsm->r_flags |= BBR_TLP;
4726         }
4727         if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) &&
4728             (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) {
4729                 /*
4730                  * We have retransmitted to many times for TLP. Switch to
4731                  * the regular RTO timer
4732                  */
4733                 goto restore;
4734         }
4735         bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP);
4736         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
4737         return (0);
4738 }
4739
4740 /*
4741  * Delayed ack Timer, here we simply need to setup the
4742  * ACK_NOW flag and remove the DELACK flag. From there
4743  * the output routine will send the ack out.
4744  *
4745  * We only return 1, saying don't proceed, if all timers
4746  * are stopped (destroyed PCB?).
4747  */
4748 static int
4749 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4750 {
4751         if (bbr->rc_all_timers_stopped) {
4752                 return (1);
4753         }
4754         bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK);
4755         tp->t_flags &= ~TF_DELACK;
4756         tp->t_flags |= TF_ACKNOW;
4757         KMOD_TCPSTAT_INC(tcps_delack);
4758         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
4759         return (0);
4760 }
4761
4762 /*
4763  * Here we send a KEEP-ALIVE like probe to the
4764  * peer, we do not send data.
4765  *
4766  * We only return 1, saying don't proceed, if all timers
4767  * are stopped (destroyed PCB?).
4768  */
4769 static int
4770 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4771 {
4772         struct tcptemp *t_template;
4773         int32_t retval = 1;
4774
4775         if (bbr->rc_all_timers_stopped) {
4776                 return (1);
4777         }
4778         if (bbr->rc_in_persist == 0)
4779                 return (0);
4780         KASSERT(tp->t_inpcb != NULL,
4781             ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
4782         /*
4783          * Persistence timer into zero window. Force a byte to be output, if
4784          * possible.
4785          */
4786         bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST);
4787         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
4788         KMOD_TCPSTAT_INC(tcps_persisttimeo);
4789         /*
4790          * Have we exceeded the user specified progress time?
4791          */
4792         if (ctf_progress_timeout_check(tp, true)) {
4793                 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4794                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4795                 goto out;
4796         }
4797         /*
4798          * Hack: if the peer is dead/unreachable, we do not time out if the
4799          * window is closed.  After a full backoff, drop the connection if
4800          * the idle time (no responses to probes) reaches the maximum
4801          * backoff that we would use if retransmitting.
4802          */
4803         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
4804             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
4805             ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
4806                 KMOD_TCPSTAT_INC(tcps_persistdrop);
4807                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4808                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4809                 goto out;
4810         }
4811         if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) &&
4812             tp->snd_una == tp->snd_max) {
4813                 bbr_exit_persist(tp, bbr, cts, __LINE__);
4814                 retval = 0;
4815                 goto out;
4816         }
4817         /*
4818          * If the user has closed the socket then drop a persisting
4819          * connection after a much reduced timeout.
4820          */
4821         if (tp->t_state > TCPS_CLOSE_WAIT &&
4822             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
4823                 KMOD_TCPSTAT_INC(tcps_persistdrop);
4824                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4825                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4826                 goto out;
4827         }
4828         t_template = tcpip_maketemplate(bbr->rc_inp);
4829         if (t_template) {
4830                 tcp_respond(tp, t_template->tt_ipgen,
4831                             &t_template->tt_t, (struct mbuf *)NULL,
4832                             tp->rcv_nxt, tp->snd_una - 1, 0);
4833                 /* This sends an ack */
4834                 if (tp->t_flags & TF_DELACK)
4835                         tp->t_flags &= ~TF_DELACK;
4836                 free(t_template, M_TEMP);
4837         }
4838         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
4839                 tp->t_rxtshift++;
4840         bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0);
4841 out:
4842         return (retval);
4843 }
4844
4845 /*
4846  * If a keepalive goes off, we had no other timers
4847  * happening. We always return 1 here since this
4848  * routine either drops the connection or sends
4849  * out a segment with respond.
4850  */
4851 static int
4852 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4853 {
4854         struct tcptemp *t_template;
4855         struct inpcb *inp;
4856
4857         if (bbr->rc_all_timers_stopped) {
4858                 return (1);
4859         }
4860         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
4861         inp = tp->t_inpcb;
4862         bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP);
4863         /*
4864          * Keep-alive timer went off; send something or drop connection if
4865          * idle for too long.
4866          */
4867         KMOD_TCPSTAT_INC(tcps_keeptimeo);
4868         if (tp->t_state < TCPS_ESTABLISHED)
4869                 goto dropit;
4870         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
4871             tp->t_state <= TCPS_CLOSING) {
4872                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
4873                         goto dropit;
4874                 /*
4875                  * Send a packet designed to force a response if the peer is
4876                  * up and reachable: either an ACK if the connection is
4877                  * still alive, or an RST if the peer has closed the
4878                  * connection due to timeout or reboot. Using sequence
4879                  * number tp->snd_una-1 causes the transmitted zero-length
4880                  * segment to lie outside the receive window; by the
4881                  * protocol spec, this requires the correspondent TCP to
4882                  * respond.
4883                  */
4884                 KMOD_TCPSTAT_INC(tcps_keepprobe);
4885                 t_template = tcpip_maketemplate(inp);
4886                 if (t_template) {
4887                         tcp_respond(tp, t_template->tt_ipgen,
4888                             &t_template->tt_t, (struct mbuf *)NULL,
4889                             tp->rcv_nxt, tp->snd_una - 1, 0);
4890                         free(t_template, M_TEMP);
4891                 }
4892         }
4893         bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0);
4894         return (1);
4895 dropit:
4896         KMOD_TCPSTAT_INC(tcps_keepdrops);
4897         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
4898         tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4899         return (1);
4900 }
4901
4902 /*
4903  * Retransmit helper function, clear up all the ack
4904  * flags and take care of important book keeping.
4905  */
4906 static void
4907 bbr_remxt_tmr(struct tcpcb *tp)
4908 {
4909         /*
4910          * The retransmit timer went off, all sack'd blocks must be
4911          * un-acked.
4912          */
4913         struct bbr_sendmap *rsm, *trsm = NULL;
4914         struct tcp_bbr *bbr;
4915         uint32_t cts, lost;
4916
4917         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
4918         cts = tcp_get_usecs(&bbr->rc_tv);
4919         lost = bbr->r_ctl.rc_lost;
4920         if (bbr->r_state && (bbr->r_state != tp->t_state))
4921                 bbr_set_state(tp, bbr, 0);
4922
4923         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
4924                 if (rsm->r_flags & BBR_ACKED) {
4925                         uint32_t old_flags;
4926
4927                         rsm->r_dupack = 0;
4928                         if (rsm->r_in_tmap == 0) {
4929                                 /* We must re-add it back to the tlist */
4930                                 if (trsm == NULL) {
4931                                         TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
4932                                 } else {
4933                                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext);
4934                                 }
4935                                 rsm->r_in_tmap = 1;
4936                         }
4937                         old_flags = rsm->r_flags;
4938                         rsm->r_flags |= BBR_RXT_CLEARED;
4939                         rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS);
4940                         bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
4941                 } else {
4942                         if ((tp->t_state < TCPS_ESTABLISHED) &&
4943                             (rsm->r_start == tp->snd_una)) {
4944                                 /*
4945                                  * Special case for TCP FO. Where
4946                                  * we sent more data beyond the snd_max.
4947                                  * We don't mark that as lost and stop here.
4948                                  */
4949                                 break;
4950                         }
4951                         if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4952                                 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4953                                 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4954                         }
4955                         if (bbr_marks_rxt_sack_passed) {
4956                                 /*
4957                                  * With this option, we will rack out
4958                                  * in 1ms increments the rest of the packets.
4959                                  */
4960                                 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST;
4961                                 rsm->r_flags &= ~BBR_WAS_SACKPASS;
4962                         } else {
4963                                 /*
4964                                  * With this option we only mark them lost
4965                                  * and remove all sack'd markings. We will run
4966                                  * another RXT or a TLP. This will cause
4967                                  * us to eventually send more based on what
4968                                  * ack's come in.
4969                                  */
4970                                 rsm->r_flags |= BBR_MARKED_LOST;
4971                                 rsm->r_flags &= ~BBR_WAS_SACKPASS;
4972                                 rsm->r_flags &= ~BBR_SACK_PASSED;
4973                         }
4974                 }
4975                 trsm = rsm;
4976         }
4977         bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4978         /* Clear the count (we just un-acked them) */
4979         bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR);
4980         bbr->rc_tlp_new_data = 0;
4981         bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4982         /* zap the behindness on a rxt */
4983         bbr->r_ctl.rc_hptsi_agg_delay = 0;
4984         bbr->r_agg_early_set = 0;
4985         bbr->r_ctl.rc_agg_early = 0;
4986         bbr->rc_tlp_rtx_out = 0;
4987         bbr->r_ctl.rc_sacked = 0;
4988         bbr->r_ctl.rc_sacklast = NULL;
4989         bbr->r_timer_override = 1;
4990         bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4991 }
4992
4993 /*
4994  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
4995  * we will setup to retransmit the lowest seq number outstanding.
4996  */
4997 static int
4998 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4999 {
5000         int32_t rexmt;
5001         int32_t retval = 0;
5002         bool isipv6;
5003
5004         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
5005         if (bbr->rc_all_timers_stopped) {
5006                 return (1);
5007         }
5008         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
5009             (tp->snd_una == tp->snd_max)) {
5010                 /* Nothing outstanding .. nothing to do */
5011                 return (0);
5012         }
5013         /*
5014          * Retransmission timer went off.  Message has not been acked within
5015          * retransmit interval.  Back off to a longer retransmit interval
5016          * and retransmit one segment.
5017          */
5018         if (ctf_progress_timeout_check(tp, true)) {
5019                 retval = 1;
5020                 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
5021                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
5022                 goto out;
5023         }
5024         bbr_remxt_tmr(tp);
5025         if ((bbr->r_ctl.rc_resend == NULL) ||
5026             ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) {
5027                 /*
5028                  * If the rwnd collapsed on
5029                  * the one we are retransmitting
5030                  * it does not count against the
5031                  * rxt count.
5032                  */
5033                 tp->t_rxtshift++;
5034         }
5035         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
5036                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
5037                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
5038                 retval = 1;
5039                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
5040                 tcp_set_inp_to_drop(bbr->rc_inp,
5041                     (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
5042                 goto out;
5043         }
5044         if (tp->t_state == TCPS_SYN_SENT) {
5045                 /*
5046                  * If the SYN was retransmitted, indicate CWND to be limited
5047                  * to 1 segment in cc_conn_init().
5048                  */
5049                 tp->snd_cwnd = 1;
5050         } else if (tp->t_rxtshift == 1) {
5051                 /*
5052                  * first retransmit; record ssthresh and cwnd so they can be
5053                  * recovered if this turns out to be a "bad" retransmit. A
5054                  * retransmit is considered "bad" if an ACK for this segment
5055                  * is received within RTT/2 interval; the assumption here is
5056                  * that the ACK was already in flight.  See "On Estimating
5057                  * End-to-End Network Path Properties" by Allman and Paxson
5058                  * for more details.
5059                  */
5060                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5061                 if (!IN_RECOVERY(tp->t_flags)) {
5062                         tp->snd_cwnd_prev = tp->snd_cwnd;
5063                         tp->snd_ssthresh_prev = tp->snd_ssthresh;
5064                         tp->snd_recover_prev = tp->snd_recover;
5065                         tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
5066                         tp->t_flags |= TF_PREVVALID;
5067                 } else {
5068                         tp->t_flags &= ~TF_PREVVALID;
5069                 }
5070                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5071         } else {
5072                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5073                 tp->t_flags &= ~TF_PREVVALID;
5074         }
5075         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
5076         if ((tp->t_state == TCPS_SYN_SENT) ||
5077             (tp->t_state == TCPS_SYN_RECEIVED))
5078                 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift];
5079         else
5080                 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
5081         TCPT_RANGESET(tp->t_rxtcur, rexmt,
5082             MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms),
5083             MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
5084         /*
5085          * We enter the path for PLMTUD if connection is established or, if
5086          * connection is FIN_WAIT_1 status, reason for the last is that if
5087          * amount of data we send is very small, we could send it in couple
5088          * of packets and process straight to FIN. In that case we won't
5089          * catch ESTABLISHED state.
5090          */
5091 #ifdef INET6
5092         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
5093 #else
5094         isipv6 = false;
5095 #endif
5096         if (((V_tcp_pmtud_blackhole_detect == 1) ||
5097             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
5098             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
5099             ((tp->t_state == TCPS_ESTABLISHED) ||
5100             (tp->t_state == TCPS_FIN_WAIT_1))) {
5101
5102                 /*
5103                  * Idea here is that at each stage of mtu probe (usually,
5104                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
5105                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
5106                  * should take care of that.
5107                  */
5108                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
5109                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
5110                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
5111                     tp->t_rxtshift % 2 == 0)) {
5112                         /*
5113                          * Enter Path MTU Black-hole Detection mechanism: -
5114                          * Disable Path MTU Discovery (IP "DF" bit). -
5115                          * Reduce MTU to lower value than what we negotiated
5116                          * with peer.
5117                          */
5118                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
5119                                 /*
5120                                  * Record that we may have found a black
5121                                  * hole.
5122                                  */
5123                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
5124                                 /* Keep track of previous MSS. */
5125                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
5126                         }
5127                         /*
5128                          * Reduce the MSS to blackhole value or to the
5129                          * default in an attempt to retransmit.
5130                          */
5131 #ifdef INET6
5132                         isipv6 = bbr->r_is_v6;
5133                         if (isipv6 &&
5134                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
5135                                 /* Use the sysctl tuneable blackhole MSS. */
5136                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
5137                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5138                         } else if (isipv6) {
5139                                 /* Use the default MSS. */
5140                                 tp->t_maxseg = V_tcp_v6mssdflt;
5141                                 /*
5142                                  * Disable Path MTU Discovery when we switch
5143                                  * to minmss.
5144                                  */
5145                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5146                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5147                         }
5148 #endif
5149 #if defined(INET6) && defined(INET)
5150                         else
5151 #endif
5152 #ifdef INET
5153                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
5154                                 /* Use the sysctl tuneable blackhole MSS. */
5155                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
5156                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5157                         } else {
5158                                 /* Use the default MSS. */
5159                                 tp->t_maxseg = V_tcp_mssdflt;
5160                                 /*
5161                                  * Disable Path MTU Discovery when we switch
5162                                  * to minmss.
5163                                  */
5164                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5165                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5166                         }
5167 #endif
5168                 } else {
5169                         /*
5170                          * If further retransmissions are still unsuccessful
5171                          * with a lowered MTU, maybe this isn't a blackhole
5172                          * and we restore the previous MSS and blackhole
5173                          * detection flags. The limit '6' is determined by
5174                          * giving each probe stage (1448, 1188, 524) 2
5175                          * chances to recover.
5176                          */
5177                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
5178                             (tp->t_rxtshift >= 6)) {
5179                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
5180                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
5181                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
5182                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
5183                         }
5184                 }
5185         }
5186         /*
5187          * Disable RFC1323 and SACK if we haven't got any response to our
5188          * third SYN to work-around some broken terminal servers (most of
5189          * which have hopefully been retired) that have bad VJ header
5190          * compression code which trashes TCP segments containing
5191          * unknown-to-them TCP options.
5192          */
5193         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
5194             (tp->t_rxtshift == 3))
5195                 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT);
5196         /*
5197          * If we backed off this far, our srtt estimate is probably bogus.
5198          * Clobber it so we'll take the next rtt measurement as our srtt;
5199          * move the current srtt into rttvar to keep the current retransmit
5200          * times until then.
5201          */
5202         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
5203 #ifdef INET6
5204                 if (bbr->r_is_v6)
5205                         in6_losing(tp->t_inpcb);
5206                 else
5207 #endif
5208                         in_losing(tp->t_inpcb);
5209                 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
5210                 tp->t_srtt = 0;
5211         }
5212         sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
5213         tp->snd_recover = tp->snd_max;
5214         tp->t_flags |= TF_ACKNOW;
5215         tp->t_rtttime = 0;
5216 out:
5217         return (retval);
5218 }
5219
5220 static int
5221 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling)
5222 {
5223         int32_t ret = 0;
5224         int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
5225
5226         if (timers == 0) {
5227                 return (0);
5228         }
5229         if (tp->t_state == TCPS_LISTEN) {
5230                 /* no timers on listen sockets */
5231                 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
5232                         return (0);
5233                 return (1);
5234         }
5235         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
5236                 uint32_t left;
5237
5238                 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
5239                         ret = -1;
5240                         bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5241                         return (0);
5242                 }
5243                 if (hpts_calling == 0) {
5244                         ret = -2;
5245                         bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5246                         return (0);
5247                 }
5248                 /*
5249                  * Ok our timer went off early and we are not paced false
5250                  * alarm, go back to sleep.
5251                  */
5252                 left = bbr->r_ctl.rc_timer_exp - cts;
5253                 ret = -3;
5254                 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling);
5255                 tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left));
5256                 return (1);
5257         }
5258         bbr->rc_tmr_stopped = 0;
5259         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
5260         if (timers & PACE_TMR_DELACK) {
5261                 ret = bbr_timeout_delack(tp, bbr, cts);
5262         } else if (timers & PACE_TMR_PERSIT) {
5263                 ret = bbr_timeout_persist(tp, bbr, cts);
5264         } else if (timers & PACE_TMR_RACK) {
5265                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5266                 ret = bbr_timeout_rack(tp, bbr, cts);
5267         } else if (timers & PACE_TMR_TLP) {
5268                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5269                 ret = bbr_timeout_tlp(tp, bbr, cts);
5270         } else if (timers & PACE_TMR_RXT) {
5271                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5272                 ret = bbr_timeout_rxt(tp, bbr, cts);
5273         } else if (timers & PACE_TMR_KEEP) {
5274                 ret = bbr_timeout_keepalive(tp, bbr, cts);
5275         }
5276         bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling);
5277         return (ret);
5278 }
5279
5280 static void
5281 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts)
5282 {
5283         if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
5284                 uint8_t hpts_removed = 0;
5285
5286                 if (bbr->rc_inp->inp_in_hpts &&
5287                     (bbr->rc_timer_first == 1)) {
5288                         /*
5289                          * If we are canceling timer's when we have the
5290                          * timer ahead of the output being paced. We also
5291                          * must remove ourselves from the hpts.
5292                          */
5293                         hpts_removed = 1;
5294                         tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
5295                         if (bbr->r_ctl.rc_last_delay_val) {
5296                                 /* Update the last hptsi delay too */
5297                                 uint32_t time_since_send;
5298
5299                                 if (TSTMP_GT(cts, bbr->rc_pacer_started))
5300                                         time_since_send = cts - bbr->rc_pacer_started;
5301                                 else
5302                                         time_since_send = 0;
5303                                 if (bbr->r_ctl.rc_last_delay_val > time_since_send) {
5304                                         /* Cut down our slot time */
5305                                         bbr->r_ctl.rc_last_delay_val -= time_since_send;
5306                                 } else {
5307                                         bbr->r_ctl.rc_last_delay_val = 0;
5308                                 }
5309                                 bbr->rc_pacer_started = cts;
5310                         }
5311                 }
5312                 bbr->rc_timer_first = 0;
5313                 bbr_log_to_cancel(bbr, line, cts, hpts_removed);
5314                 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
5315                 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
5316         }
5317 }
5318
5319 static void
5320 bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type)
5321 {
5322         struct tcp_bbr *bbr;
5323
5324         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
5325         bbr->rc_all_timers_stopped = 1;
5326         return;
5327 }
5328
5329 /*
5330  * stop all timers always returning 0.
5331  */
5332 static int
5333 bbr_stopall(struct tcpcb *tp)
5334 {
5335         return (0);
5336 }
5337
5338 static void
5339 bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
5340 {
5341         return;
5342 }
5343
5344 /*
5345  * return true if a bbr timer (rack or tlp) is active.
5346  */
5347 static int
5348 bbr_timer_active(struct tcpcb *tp, uint32_t timer_type)
5349 {
5350         return (0);
5351 }
5352
5353 static uint32_t
5354 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts)
5355 {
5356         struct bbr_sendmap *rsm;
5357
5358         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
5359         if ((rsm == NULL) || (u_rsm == rsm))
5360                 return (cts);
5361         return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
5362 }
5363
5364 static void
5365 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr,
5366      struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time)
5367 {
5368         int32_t idx;
5369
5370         rsm->r_rtr_cnt++;
5371         rsm->r_dupack = 0;
5372         if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) {
5373                 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS;
5374                 rsm->r_flags |= BBR_OVERMAX;
5375         }
5376         if (rsm->r_flags & BBR_RWND_COLLAPSED) {
5377                 /* Take off the collapsed flag at rxt */
5378                 rsm->r_flags &= ~BBR_RWND_COLLAPSED;
5379         }
5380         if (rsm->r_flags & BBR_MARKED_LOST) {
5381                 /* We have retransmitted, its no longer lost */
5382                 rsm->r_flags &= ~BBR_MARKED_LOST;
5383                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
5384         }
5385         if (rsm->r_flags & BBR_RXT_CLEARED) {
5386                 /*
5387                  * We hit a RXT timer on it and
5388                  * we cleared the "acked" flag.
5389                  * We now have it going back into
5390                  * flight, we can remove the cleared
5391                  * flag and possibly do accounting on
5392                  * this piece.
5393                  */
5394                 rsm->r_flags &= ~BBR_RXT_CLEARED;
5395         }
5396         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) {
5397                 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
5398                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
5399         }
5400         idx = rsm->r_rtr_cnt - 1;
5401         rsm->r_tim_lastsent[idx] = cts;
5402         rsm->r_pacing_delay = pacing_time;
5403         rsm->r_delivered = bbr->r_ctl.rc_delivered;
5404         rsm->r_ts_valid = bbr->rc_ts_valid;
5405         if (bbr->rc_ts_valid)
5406                 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
5407         if (bbr->r_ctl.r_app_limited_until)
5408                 rsm->r_app_limited = 1;
5409         else
5410                 rsm->r_app_limited = 0;
5411         if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
5412                 rsm->r_bbr_state = bbr_state_val(bbr);
5413         else
5414                 rsm->r_bbr_state = 8;
5415         if (rsm->r_flags & BBR_ACKED) {
5416                 /* Problably MTU discovery messing with us */
5417                 uint32_t old_flags;
5418
5419                 old_flags = rsm->r_flags;
5420                 rsm->r_flags &= ~BBR_ACKED;
5421                 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
5422                 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
5423                 if (bbr->r_ctl.rc_sacked == 0)
5424                         bbr->r_ctl.rc_sacklast = NULL;
5425         }
5426         if (rsm->r_in_tmap) {
5427                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5428         }
5429         TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5430         rsm->r_in_tmap = 1;
5431         if (rsm->r_flags & BBR_SACK_PASSED) {
5432                 /* We have retransmitted due to the SACK pass */
5433                 rsm->r_flags &= ~BBR_SACK_PASSED;
5434                 rsm->r_flags |= BBR_WAS_SACKPASS;
5435         }
5436         rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
5437         rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
5438                                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
5439         bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
5440         if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
5441                 rsm->r_is_gain = 1;
5442                 rsm->r_is_drain = 0;
5443         } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
5444                 rsm->r_is_drain = 1;
5445                 rsm->r_is_gain = 0;
5446         } else {
5447                 rsm->r_is_drain = 0;
5448                 rsm->r_is_gain = 0;
5449         }
5450         rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */
5451 }
5452
5453 /*
5454  * Returns 0, or the sequence where we stopped
5455  * updating. We also update the lenp to be the amount
5456  * of data left.
5457  */
5458
5459 static uint32_t
5460 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr,
5461     struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time)
5462 {
5463         /*
5464          * We (re-)transmitted starting at rsm->r_start for some length
5465          * (possibly less than r_end.
5466          */
5467         struct bbr_sendmap *nrsm;
5468         uint32_t c_end;
5469         int32_t len;
5470
5471         len = *lenp;
5472         c_end = rsm->r_start + len;
5473         if (SEQ_GEQ(c_end, rsm->r_end)) {
5474                 /*
5475                  * We retransmitted the whole piece or more than the whole
5476                  * slopping into the next rsm.
5477                  */
5478                 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5479                 if (c_end == rsm->r_end) {
5480                         *lenp = 0;
5481                         return (0);
5482                 } else {
5483                         int32_t act_len;
5484
5485                         /* Hangs over the end return whats left */
5486                         act_len = rsm->r_end - rsm->r_start;
5487                         *lenp = (len - act_len);
5488                         return (rsm->r_end);
5489                 }
5490                 /* We don't get out of this block. */
5491         }
5492         /*
5493          * Here we retransmitted less than the whole thing which means we
5494          * have to split this into what was transmitted and what was not.
5495          */
5496         nrsm = bbr_alloc_full_limit(bbr);
5497         if (nrsm == NULL) {
5498                 *lenp = 0;
5499                 return (0);
5500         }
5501         /*
5502          * So here we are going to take the original rsm and make it what we
5503          * retransmitted. nrsm will be the tail portion we did not
5504          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
5505          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
5506          * 1, 6 and the new piece will be 6, 11.
5507          */
5508         bbr_clone_rsm(bbr, nrsm, rsm, c_end);
5509         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
5510         nrsm->r_dupack = 0;
5511         if (rsm->r_in_tmap) {
5512                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
5513                 nrsm->r_in_tmap = 1;
5514         }
5515         rsm->r_flags &= (~BBR_HAS_FIN);
5516         bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5517         *lenp = 0;
5518         return (0);
5519 }
5520
5521 static uint64_t
5522 bbr_get_hardware_rate(struct tcp_bbr *bbr)
5523 {
5524         uint64_t bw;
5525
5526         bw = bbr_get_bw(bbr);
5527         bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN];
5528         bw /= (uint64_t)BBR_UNIT;
5529         return(bw);
5530 }
5531
5532 static void
5533 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts,
5534                        uint64_t act_rate, uint64_t rate_wanted)
5535 {
5536         /*
5537          * We could not get a full gains worth
5538          * of rate.
5539          */
5540         if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) {
5541                 /* we can't even get the real rate */
5542                 uint64_t red;
5543
5544                 bbr->skip_gain = 1;
5545                 bbr->gain_is_limited = 0;
5546                 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate;
5547                 if (red)
5548                         filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts);
5549         } else {
5550                 /* We can use a lower gain */
5551                 bbr->skip_gain = 0;
5552                 bbr->gain_is_limited = 1;
5553         }
5554 }
5555
5556 static void
5557 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts)
5558 {
5559         const struct tcp_hwrate_limit_table *nrte;
5560         int error, rate = -1;
5561
5562         if (bbr->r_ctl.crte == NULL)
5563                 return;
5564         if ((bbr->rc_inp->inp_route.ro_nh == NULL) ||
5565             (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) {
5566                 /* Lost our routes? */
5567                 /* Clear the way for a re-attempt */
5568                 bbr->bbr_attempt_hdwr_pace = 0;
5569 lost_rate:
5570                 bbr->gain_is_limited = 0;
5571                 bbr->skip_gain = 0;
5572                 bbr->bbr_hdrw_pacing = 0;
5573                 counter_u64_add(bbr_flows_whdwr_pacing, -1);
5574                 counter_u64_add(bbr_flows_nohdwr_pacing, 1);
5575                 tcp_bbr_tso_size_check(bbr, cts);
5576                 return;
5577         }
5578         rate = bbr_get_hardware_rate(bbr);
5579         nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte,
5580                                    bbr->rc_tp,
5581                                    bbr->rc_inp->inp_route.ro_nh->nh_ifp,
5582                                    rate,
5583                                    (RS_PACING_GEQ|RS_PACING_SUB_OK),
5584                                    &error);
5585         if (nrte == NULL) {
5586                 goto lost_rate;
5587         }
5588         if (nrte != bbr->r_ctl.crte) {
5589                 bbr->r_ctl.crte = nrte;
5590                 if (error == 0)  {
5591                         BBR_STAT_INC(bbr_hdwr_rl_mod_ok);
5592                         if (bbr->r_ctl.crte->rate < rate) {
5593                                 /* We have a problem */
5594                                 bbr_setup_less_of_rate(bbr, cts,
5595                                                        bbr->r_ctl.crte->rate, rate);
5596                         } else {
5597                                 /* We are good */
5598                                 bbr->gain_is_limited = 0;
5599                                 bbr->skip_gain = 0;
5600                         }
5601                 } else {
5602                         /* A failure should release the tag */
5603                         BBR_STAT_INC(bbr_hdwr_rl_mod_fail);
5604                         bbr->gain_is_limited = 0;
5605                         bbr->skip_gain = 0;
5606                         bbr->bbr_hdrw_pacing = 0;
5607                 }
5608                 bbr_type_log_hdwr_pacing(bbr,
5609                                          bbr->r_ctl.crte->ptbl->rs_ifp,
5610                                          rate,
5611                                          ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate),
5612                                          __LINE__,
5613                                          cts,
5614                                          error);
5615         }
5616 }
5617
5618 static void
5619 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts)
5620 {
5621         /*
5622          * If we have hardware pacing support
5623          * we need to factor that in for our
5624          * TSO size.
5625          */
5626         const struct tcp_hwrate_limit_table *rlp;
5627         uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay;
5628
5629         if ((bbr->bbr_hdrw_pacing == 0) ||
5630             (IN_RECOVERY(bbr->rc_tp->t_flags)) ||
5631             (bbr->r_ctl.crte == NULL))
5632                 return;
5633         if (bbr->hw_pacing_set == 0) {
5634                 /* Not yet by the hdwr pacing count delay */
5635                 return;
5636         }
5637         if (bbr_hdwr_pace_adjust == 0) {
5638                 /* No adjustment */
5639                 return;
5640         }
5641         rlp = bbr->r_ctl.crte;
5642         if (bbr->rc_tp->t_maxseg > bbr->rc_last_options)
5643                 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5644         else
5645                 maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5646         /*
5647          * So lets first get the
5648          * time we will take between
5649          * TSO sized sends currently without
5650          * hardware help.
5651          */
5652         cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT,
5653                         bbr->r_ctl.rc_pace_max_segs, cts, 1);
5654         hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg;
5655         hdwr_delay *= rlp->time_between;
5656         if (cur_delay > hdwr_delay)
5657                 delta = cur_delay - hdwr_delay;
5658         else
5659                 delta = 0;
5660         bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay,
5661                              (bbr->r_ctl.rc_pace_max_segs / maxseg),
5662                              1);
5663         if (delta &&
5664             (delta < (max(rlp->time_between,
5665                           bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) {
5666                 /*
5667                  * Now lets divide by the pacing
5668                  * time between each segment the
5669                  * hardware sends rounding up and
5670                  * derive a bytes from that. We multiply
5671                  * that by bbr_hdwr_pace_adjust to get
5672                  * more bang for our buck.
5673                  *
5674                  * The goal is to have the software pacer
5675                  * waiting no more than an additional
5676                  * pacing delay if we can (without the
5677                  * compensation i.e. x bbr_hdwr_pace_adjust).
5678                  */
5679                 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between),
5680                              (bbr->r_ctl.rc_pace_max_segs/maxseg));
5681                 seg_sz *= bbr_hdwr_pace_adjust;
5682                 if (bbr_hdwr_pace_floor &&
5683                     (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5684                         /* Currently hardware paces
5685                          * out rs_min_seg segments at a time.
5686                          * We need to make sure we always send at least
5687                          * a full burst of bbr_hdwr_pace_floor down.
5688                          */
5689                         seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5690                 }
5691                 seg_sz *= maxseg;
5692         } else if (delta == 0) {
5693                 /*
5694                  * The highest pacing rate is
5695                  * above our b/w gained. This means
5696                  * we probably are going quite fast at
5697                  * the hardware highest rate. Lets just multiply
5698                  * the calculated TSO size by the
5699                  * multiplier factor (its probably
5700                  * 4 segments in the default config for
5701                  * mlx).
5702                  */
5703                 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust;
5704                 if (bbr_hdwr_pace_floor &&
5705                     (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5706                         /* Currently hardware paces
5707                          * out rs_min_seg segments at a time.
5708                          * We need to make sure we always send at least
5709                          * a full burst of bbr_hdwr_pace_floor down.
5710                          */
5711                         seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5712                 }
5713         } else {
5714                 /*
5715                  * The pacing time difference is so
5716                  * big that the hardware will
5717                  * pace out more rapidly then we
5718                  * really want and then we
5719                  * will have a long delay. Lets just keep
5720                  * the same TSO size so its as if
5721                  * we were not using hdwr pacing (we
5722                  * just gain a bit of spacing from the
5723                  * hardware if seg_sz > 1).
5724                  */
5725                 seg_sz = bbr->r_ctl.rc_pace_max_segs;
5726         }
5727         if (seg_sz > bbr->r_ctl.rc_pace_max_segs)
5728                 new_tso = seg_sz;
5729         else
5730                 new_tso = bbr->r_ctl.rc_pace_max_segs;
5731         if (new_tso >= (PACE_MAX_IP_BYTES-maxseg))
5732                 new_tso = PACE_MAX_IP_BYTES - maxseg;
5733
5734         if (new_tso != bbr->r_ctl.rc_pace_max_segs) {
5735                 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0);
5736                 bbr->r_ctl.rc_pace_max_segs = new_tso;
5737         }
5738 }
5739
5740 static void
5741 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts)
5742 {
5743         uint64_t bw;
5744         uint32_t old_tso = 0, new_tso;
5745         uint32_t maxseg, bytes;
5746         uint32_t tls_seg=0;
5747         /*
5748          * Google/linux uses the following algorithm to determine
5749          * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18):
5750          *
5751          *  bytes = bw_in_bytes_per_second / 1000
5752          *  bytes = min(bytes, 64k)
5753          *  tso_segs = bytes / MSS
5754          *  if (bw < 1.2Mbs)
5755          *      min_tso_segs = 1
5756          *  else
5757          *      min_tso_segs = 2
5758          * tso_segs = max(tso_segs, min_tso_segs)
5759          *
5760          * * Note apply a device specific limit (we apply this in the
5761          *   tcp_m_copym).
5762          * Note that before the initial measurement is made google bursts out
5763          * a full iwnd just like new-reno/cubic.
5764          *
5765          * We do not use this algorithm. Instead we
5766          * use a two phased approach:
5767          *
5768          *  if ( bw <= per-tcb-cross-over)
5769          *     goal_tso =  calculate how much with this bw we
5770          *                 can send in goal-time seconds.
5771          *     if (goal_tso > mss)
5772          *         seg = goal_tso / mss
5773          *         tso = seg * mss
5774          *     else
5775          *         tso = mss
5776          *     if (tso > per-tcb-max)
5777          *         tso = per-tcb-max
5778          *  else if ( bw > 512Mbps)
5779          *     tso = max-tso (64k/mss)
5780          *  else
5781          *     goal_tso = bw / per-tcb-divsor
5782          *     seg = (goal_tso + mss-1)/mss
5783          *     tso = seg * mss
5784          *
5785          * if (tso < per-tcb-floor)
5786          *    tso = per-tcb-floor
5787          * if (tso > per-tcb-utter_max)
5788          *    tso = per-tcb-utter_max
5789          *
5790          * Note the default per-tcb-divisor is 1000 (same as google).
5791          * the goal cross over is 30Mbps however. To recreate googles
5792          * algorithm you need to set:
5793          *
5794          * cross-over = 23,168,000 bps
5795          * goal-time = 18000
5796          * per-tcb-max = 2
5797          * per-tcb-divisor = 1000
5798          * per-tcb-floor = 1
5799          *
5800          * This will get you "google bbr" behavior with respect to tso size.
5801          *
5802          * Note we do set anything TSO size until we are past the initial
5803          * window. Before that we gnerally use either a single MSS
5804          * or we use the full IW size (so we burst a IW at a time)
5805          * Also note that Hardware-TLS is special and does alternate
5806          * things to minimize PCI Bus Bandwidth use.
5807          */
5808
5809         if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) {
5810                 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5811         } else {
5812                 maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5813         }
5814 #ifdef KERN_TLS
5815         if (bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) {
5816                 tls_seg =  ctf_get_opt_tls_size(bbr->rc_inp->inp_socket, bbr->rc_tp->snd_wnd);
5817                 bbr->r_ctl.rc_pace_min_segs = (tls_seg + bbr->rc_last_options);
5818         }
5819 #endif
5820         old_tso = bbr->r_ctl.rc_pace_max_segs;
5821         if (bbr->rc_past_init_win == 0) {
5822                 /*
5823                  * Not enough data has been acknowledged to make a
5824                  * judgement unless we are hardware TLS. Set up
5825                  * the initial TSO based on if we are sending a
5826                  * full IW at once or not.
5827                  */
5828                 if (bbr->rc_use_google)
5829                         bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2);
5830                 else if (bbr->bbr_init_win_cheat)
5831                         bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp);
5832                 else
5833                         bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5834                 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg)
5835                         bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg;
5836 #ifdef KERN_TLS
5837                 if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) && tls_seg) {
5838                         /*
5839                          * For hardware TLS we set our min to the tls_seg size.
5840                          */
5841                         bbr->r_ctl.rc_pace_max_segs = tls_seg;
5842                         bbr->r_ctl.rc_pace_min_segs = tls_seg + bbr->rc_last_options;
5843                 }
5844 #endif
5845                 if (bbr->r_ctl.rc_pace_max_segs == 0) {
5846                         bbr->r_ctl.rc_pace_max_segs = maxseg;
5847                 }
5848                 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0);
5849 #ifdef KERN_TLS
5850                 if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) == 0)
5851 #endif
5852                         bbr_adjust_for_hw_pacing(bbr, cts);
5853                 return;
5854         }
5855         /**
5856          * Now lets set the TSO goal based on our delivery rate in
5857          * bytes per second. Note we only do this if
5858          * we have acked at least the initial cwnd worth of data.
5859          */
5860         bw = bbr_get_bw(bbr);
5861         if (IN_RECOVERY(bbr->rc_tp->t_flags) &&
5862              (bbr->rc_use_google == 0)) {
5863                 /* We clamp to one MSS in recovery */
5864                 new_tso = maxseg;
5865         } else if (bbr->rc_use_google) {
5866                 int min_tso_segs;
5867
5868                 /* Google considers the gain too */
5869                 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) {
5870                         bw *= bbr->r_ctl.rc_bbr_hptsi_gain;
5871                         bw /= BBR_UNIT;
5872                 }
5873                 bytes = bw / 1024;
5874                 if (bytes > (64 * 1024))
5875                         bytes = 64 * 1024;
5876                 new_tso = bytes / maxseg;
5877                 if (bw < ONE_POINT_TWO_MEG)
5878                         min_tso_segs = 1;
5879                 else
5880                         min_tso_segs = 2;
5881                 if (new_tso < min_tso_segs)
5882                         new_tso = min_tso_segs;
5883                 new_tso *= maxseg;
5884         } else if (bbr->rc_no_pacing) {
5885                 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg;
5886         } else if (bw <= bbr->r_ctl.bbr_cross_over) {
5887                 /*
5888                  * Calculate the worse case b/w TSO if we are inserting no
5889                  * more than a delay_target number of TSO's.
5890                  */
5891                 uint32_t tso_len, min_tso;
5892
5893                 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw);
5894                 if (tso_len > maxseg) {
5895                         new_tso = tso_len / maxseg;
5896                         if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max)
5897                                 new_tso = bbr->r_ctl.bbr_hptsi_segments_max;
5898                         new_tso *= maxseg;
5899                 } else {
5900                         /*
5901                          * less than a full sized frame yikes.. long rtt or
5902                          * low bw?
5903                          */
5904                         min_tso = bbr_minseg(bbr);
5905                         if ((tso_len > min_tso) && (bbr_all_get_min == 0))
5906                                 new_tso = rounddown(tso_len, min_tso);
5907                         else
5908                                 new_tso = min_tso;
5909                 }
5910         } else if (bw > FIVETWELVE_MBPS) {
5911                 /*
5912                  * This guy is so fast b/w wise that we can TSO as large as
5913                  * possible of segments that the NIC will allow.
5914                  */
5915                 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5916         } else {
5917                 /*
5918                  * This formula is based on attempting to send a segment or
5919                  * more every bbr_hptsi_per_second. The default is 1000
5920                  * which means you are targeting what you can send every 1ms
5921                  * based on the peers bw.
5922                  *
5923                  * If the number drops to say 500, then you are looking more
5924                  * at 2ms and you will raise how much we send in a single
5925                  * TSO thus saving CPU (less bbr_output_wtime() calls). The
5926                  * trade off of course is you will send more at once and
5927                  * thus tend to clump up the sends into larger "bursts"
5928                  * building a queue.
5929                  */
5930                 bw /= bbr->r_ctl.bbr_hptsi_per_second;
5931                 new_tso = roundup(bw, (uint64_t)maxseg);
5932                 /*
5933                  * Gate the floor to match what our lower than 48Mbps
5934                  * algorithm does. The ceiling (bbr_hptsi_segments_max) thus
5935                  * becomes the floor for this calculation.
5936                  */
5937                 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg))
5938                         new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg);
5939         }
5940         if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor)))
5941                 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor;
5942         if (new_tso > PACE_MAX_IP_BYTES)
5943                 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5944         /* Enforce an utter maximum if we are not HW-TLS */
5945 #ifdef KERN_TLS
5946         if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) == 0)
5947 #endif
5948                 if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) {
5949                         new_tso = bbr->r_ctl.bbr_utter_max * maxseg;
5950                 }
5951 #ifdef KERN_TLS
5952         if (tls_seg) {
5953                 /*
5954                  * Lets move the output size
5955                  * up to 1 or more TLS record sizes.
5956                  */
5957                 uint32_t temp;
5958
5959                 temp = roundup(new_tso, tls_seg);
5960                 new_tso = temp;
5961                 /* Back down if needed to under a full frame */
5962                 while (new_tso > PACE_MAX_IP_BYTES)
5963                         new_tso -= tls_seg;
5964         }
5965 #endif
5966         if (old_tso != new_tso) {
5967                 /* Only log changes */
5968                 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0);
5969                 bbr->r_ctl.rc_pace_max_segs = new_tso;
5970         }
5971 #ifdef KERN_TLS
5972         if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) &&
5973              tls_seg) {
5974                 bbr->r_ctl.rc_pace_min_segs = tls_seg + bbr->rc_last_options;
5975         } else
5976 #endif
5977                 /* We have hardware pacing and not hardware TLS! */
5978                 bbr_adjust_for_hw_pacing(bbr, cts);
5979 }
5980
5981 static void
5982 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len,
5983     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t cts,
5984     struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc,
5985     struct sockbuf *sb)
5986 {
5987
5988         struct bbr_sendmap *rsm, *nrsm;
5989         register uint32_t snd_max, snd_una;
5990         uint32_t pacing_time;
5991         /*
5992          * Add to the RACK log of packets in flight or retransmitted. If
5993          * there is a TS option we will use the TS echoed, if not we will
5994          * grab a TS.
5995          *
5996          * Retransmissions will increment the count and move the ts to its
5997          * proper place. Note that if options do not include TS's then we
5998          * won't be able to effectively use the ACK for an RTT on a retran.
5999          *
6000          * Notes about r_start and r_end. Lets consider a send starting at
6001          * sequence 1 for 10 bytes. In such an example the r_start would be
6002          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
6003          * This means that r_end is actually the first sequence for the next
6004          * slot (11).
6005          *
6006          */
6007         INP_WLOCK_ASSERT(tp->t_inpcb);
6008         if (err) {
6009                 /*
6010                  * We don't log errors -- we could but snd_max does not
6011                  * advance in this case either.
6012                  */
6013                 return;
6014         }
6015         if (th_flags & TH_RST) {
6016                 /*
6017                  * We don't log resets and we return immediately from
6018                  * sending
6019                  */
6020                 *abandon = 1;
6021                 return;
6022         }
6023         snd_una = tp->snd_una;
6024         if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) {
6025                 /*
6026                  * The call to bbr_log_output is made before bumping
6027                  * snd_max. This means we can record one extra byte on a SYN
6028                  * or FIN if seq_out is adding more on and a FIN is present
6029                  * (and we are not resending).
6030                  */
6031                 if ((th_flags & TH_SYN) && (tp->iss == seq_out))
6032                         len++;
6033                 if (th_flags & TH_FIN)
6034                         len++;
6035         }
6036         if (SEQ_LEQ((seq_out + len), snd_una)) {
6037                 /* Are sending an old segment to induce an ack (keep-alive)? */
6038                 return;
6039         }
6040         if (SEQ_LT(seq_out, snd_una)) {
6041                 /* huh? should we panic? */
6042                 uint32_t end;
6043
6044                 end = seq_out + len;
6045                 seq_out = snd_una;
6046                 len = end - seq_out;
6047         }
6048         snd_max = tp->snd_max;
6049         if (len == 0) {
6050                 /* We don't log zero window probes */
6051                 return;
6052         }
6053         pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1);
6054         /* First question is it a retransmission? */
6055         if (seq_out == snd_max) {
6056 again:
6057                 rsm = bbr_alloc(bbr);
6058                 if (rsm == NULL) {
6059                         return;
6060                 }
6061                 rsm->r_flags = 0;
6062                 if (th_flags & TH_SYN)
6063                         rsm->r_flags |= BBR_HAS_SYN;
6064                 if (th_flags & TH_FIN)
6065                         rsm->r_flags |= BBR_HAS_FIN;
6066                 rsm->r_tim_lastsent[0] = cts;
6067                 rsm->r_rtr_cnt = 1;
6068                 rsm->r_rtr_bytes = 0;
6069                 rsm->r_start = seq_out;
6070                 rsm->r_end = rsm->r_start + len;
6071                 rsm->r_dupack = 0;
6072                 rsm->r_delivered = bbr->r_ctl.rc_delivered;
6073                 rsm->r_pacing_delay = pacing_time;
6074                 rsm->r_ts_valid = bbr->rc_ts_valid;
6075                 if (bbr->rc_ts_valid)
6076                         rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
6077                 rsm->r_del_time = bbr->r_ctl.rc_del_time;
6078                 if (bbr->r_ctl.r_app_limited_until)
6079                         rsm->r_app_limited = 1;
6080                 else
6081                         rsm->r_app_limited = 0;
6082                 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
6083                 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
6084                                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
6085                 /*
6086                  * Here we must also add in this rsm since snd_max
6087                  * is updated after we return from a new send.
6088                  */
6089                 rsm->r_flight_at_send += len;
6090                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
6091                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
6092                 rsm->r_in_tmap = 1;
6093                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
6094                         rsm->r_bbr_state = bbr_state_val(bbr);
6095                 else
6096                         rsm->r_bbr_state = 8;
6097                 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
6098                         rsm->r_is_gain = 1;
6099                         rsm->r_is_drain = 0;
6100                 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
6101                         rsm->r_is_drain = 1;
6102                         rsm->r_is_gain = 0;
6103                 } else {
6104                         rsm->r_is_drain = 0;
6105                         rsm->r_is_gain = 0;
6106                 }
6107                 return;
6108         }
6109         /*
6110          * If we reach here its a retransmission and we need to find it.
6111          */
6112 more:
6113         if (hintrsm && (hintrsm->r_start == seq_out)) {
6114                 rsm = hintrsm;
6115                 hintrsm = NULL;
6116         } else if (bbr->r_ctl.rc_next) {
6117                 /* We have a hint from a previous run */
6118                 rsm = bbr->r_ctl.rc_next;
6119         } else {
6120                 /* No hints sorry */
6121                 rsm = NULL;
6122         }
6123         if ((rsm) && (rsm->r_start == seq_out)) {
6124                 /*
6125                  * We used rc_next or hintrsm  to retransmit, hopefully the
6126                  * likely case.
6127                  */
6128                 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6129                 if (len == 0) {
6130                         return;
6131                 } else {
6132                         goto more;
6133                 }
6134         }
6135         /* Ok it was not the last pointer go through it the hard way. */
6136         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6137                 if (rsm->r_start == seq_out) {
6138                         seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6139                         bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
6140                         if (len == 0) {
6141                                 return;
6142                         } else {
6143                                 continue;
6144                         }
6145                 }
6146                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
6147                         /* Transmitted within this piece */
6148                         /*
6149                          * Ok we must split off the front and then let the
6150                          * update do the rest
6151                          */
6152                         nrsm = bbr_alloc_full_limit(bbr);
6153                         if (nrsm == NULL) {
6154                                 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
6155                                 return;
6156                         }
6157                         /*
6158                          * copy rsm to nrsm and then trim the front of rsm
6159                          * to not include this part.
6160                          */
6161                         bbr_clone_rsm(bbr, nrsm, rsm, seq_out);
6162                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
6163                         if (rsm->r_in_tmap) {
6164                                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6165                                 nrsm->r_in_tmap = 1;
6166                         }
6167                         rsm->r_flags &= (~BBR_HAS_FIN);
6168                         seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time);
6169                         if (len == 0) {
6170                                 return;
6171                         }
6172                 }
6173         }
6174         /*
6175          * Hmm not found in map did they retransmit both old and on into the
6176          * new?
6177          */
6178         if (seq_out == tp->snd_max) {
6179                 goto again;
6180         } else if (SEQ_LT(seq_out, tp->snd_max)) {
6181 #ifdef BBR_INVARIANTS
6182                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
6183                     seq_out, len, tp->snd_una, tp->snd_max);
6184                 printf("Starting Dump of all rack entries\n");
6185                 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6186                         printf("rsm:%p start:%u end:%u\n",
6187                             rsm, rsm->r_start, rsm->r_end);
6188                 }
6189                 printf("Dump complete\n");
6190                 panic("seq_out not found rack:%p tp:%p",
6191                     bbr, tp);
6192 #endif
6193         } else {
6194 #ifdef BBR_INVARIANTS
6195                 /*
6196                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
6197                  * flag)
6198                  */
6199                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
6200                     seq_out, len, tp->snd_max, tp);
6201 #endif
6202         }
6203 }
6204
6205 static void
6206 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt)
6207 {
6208         /*
6209          * Collapse timeout back the cum-ack moved.
6210          */
6211         tp->t_rxtshift = 0;
6212         tp->t_softerror = 0;
6213 }
6214
6215
6216 static void
6217 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin)
6218 {
6219         bbr->rtt_valid = 1;
6220         bbr->r_ctl.cur_rtt = rtt_usecs;
6221         bbr->r_ctl.ts_in = tsin;
6222         if (rsm_send_time)
6223                 bbr->r_ctl.cur_rtt_send_time = rsm_send_time;
6224 }
6225
6226 static void
6227 bbr_make_timestamp_determination(struct tcp_bbr *bbr)
6228 {
6229         /**
6230          * We have in our bbr control:
6231          * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp).
6232          * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts).
6233          * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts)
6234          * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time)
6235          *
6236          * Now we can calculate the time between the sends by doing:
6237          *
6238          * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts
6239          *
6240          * And the peer's time between receiving them by doing:
6241          *
6242          * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp
6243          *
6244          * We want to figure out if the timestamp values are in msec, 10msec or usec.
6245          * We also may find that we can't use the timestamps if say we see
6246          * that the peer_delta indicates that though we may have taken 10ms to
6247          * pace out the data, it only saw 1ms between the two packets. This would
6248          * indicate that somewhere on the path is a batching entity that is giving
6249          * out time-slices of the actual b/w. This would mean we could not use
6250          * reliably the peers timestamps.
6251          *
6252          * We expect delta > peer_delta initially. Until we figure out the
6253          * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio.
6254          * If we place 1000 there then its a ms vs our usec. If we place 10000 there
6255          * then its 10ms vs our usec. If the peer is running a usec clock we would
6256          * put a 1 there. If the value is faster then ours, we will disable the
6257          * use of timestamps (though we could revist this later if we find it to be not
6258          * just an isolated one or two flows)).
6259          *
6260          * To detect the batching middle boxes we will come up with our compensation and
6261          * if with it in place, we find the peer is drastically off (by some margin) in
6262          * the smaller direction, then we will assume the worst case and disable use of timestamps.
6263          *
6264          */
6265         uint64_t delta, peer_delta, delta_up;
6266
6267         delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts;
6268         if (delta < bbr_min_usec_delta) {
6269                 /*
6270                  * Have not seen a min amount of time
6271                  * between our send times so we can
6272                  * make a determination of the timestamp
6273                  * yet.
6274                  */
6275                 return;
6276         }
6277         peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp;
6278         if (peer_delta < bbr_min_peer_delta) {
6279                 /*
6280                  * We may have enough in the form of
6281                  * our delta but the peers number
6282                  * has not changed that much. It could
6283                  * be its clock ratio is such that
6284                  * we need more data (10ms tick) or
6285                  * there may be other compression scenarios
6286                  * going on. In any event we need the
6287                  * spread to be larger.
6288                  */
6289                 return;
6290         }
6291         /* Ok lets first see which way our delta is going */
6292         if (peer_delta > delta) {
6293                 /* Very unlikely, the peer without
6294                  * compensation shows that it saw
6295                  * the two sends arrive further apart
6296                  * then we saw then in micro-seconds.
6297                  */
6298                 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) {
6299                         /* well it looks like the peer is a micro-second clock. */
6300                         bbr->rc_ts_clock_set = 1;
6301                         bbr->r_ctl.bbr_peer_tsratio = 1;
6302                 } else {
6303                         bbr->rc_ts_cant_be_used = 1;
6304                         bbr->rc_ts_clock_set = 1;
6305                 }
6306                 return;
6307         }
6308         /* Ok we know that the peer_delta is smaller than our send distance */
6309         bbr->rc_ts_clock_set = 1;
6310         /* First question is it within the percentage that they are using usec time? */
6311         delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent;
6312         if ((peer_delta + delta_up) >= delta) {
6313                 /* Its a usec clock */
6314                 bbr->r_ctl.bbr_peer_tsratio = 1;
6315                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6316                 return;
6317         }
6318         /* Ok if not usec, what about 10usec (though unlikely)? */
6319         delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent;
6320         if (((peer_delta * 10) + delta_up) >= delta) {
6321                 bbr->r_ctl.bbr_peer_tsratio = 10;
6322                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6323                 return;
6324         }
6325         /* And what about 100usec (though again unlikely)? */
6326         delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent;
6327         if (((peer_delta * 100) + delta_up) >= delta) {
6328                 bbr->r_ctl.bbr_peer_tsratio = 100;
6329                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6330                 return;
6331         }
6332         /* And how about 1 msec (the most likely one)? */
6333         delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent;
6334         if (((peer_delta * 1000) + delta_up) >= delta) {
6335                 bbr->r_ctl.bbr_peer_tsratio = 1000;
6336                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6337                 return;
6338         }
6339         /* Ok if not msec could it be 10 msec? */
6340         delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent;
6341         if (((peer_delta * 10000) + delta_up) >= delta) {
6342                 bbr->r_ctl.bbr_peer_tsratio = 10000;
6343                 return;
6344         }
6345         /* If we fall down here the clock tick so slowly we can't use it */
6346         bbr->rc_ts_cant_be_used = 1;
6347         bbr->r_ctl.bbr_peer_tsratio = 0;
6348         bbr_log_tstmp_validation(bbr, peer_delta, delta);
6349 }
6350
6351 /*
6352  * Collect new round-trip time estimate
6353  * and update averages and current timeout.
6354  */
6355 static void
6356 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts)
6357 {
6358         int32_t delta;
6359         uint32_t rtt, tsin;
6360         int32_t rtt_ticks;
6361
6362
6363         if (bbr->rtt_valid == 0)
6364                 /* No valid sample */
6365                 return;
6366
6367         rtt = bbr->r_ctl.cur_rtt;
6368         tsin = bbr->r_ctl.ts_in;
6369         if (bbr->rc_prtt_set_ts) {
6370                 /*
6371                  * We are to force feed the rttProp filter due
6372                  * to an entry into PROBE_RTT. This assures
6373                  * that the times are sync'd between when we
6374                  * go into PROBE_RTT and the filter expiration.
6375                  *
6376                  * Google does not use a true filter, so they do
6377                  * this implicitly since they only keep one value
6378                  * and when they enter probe-rtt they update the
6379                  * value to the newest rtt.
6380                  */
6381                 uint32_t rtt_prop;
6382
6383                 bbr->rc_prtt_set_ts = 0;
6384                 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
6385                 if (rtt > rtt_prop)
6386                         filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts);
6387                 else
6388                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6389         }
6390         if (bbr->rc_ack_was_delayed)
6391                 rtt += bbr->r_ctl.rc_ack_hdwr_delay;
6392
6393         if (rtt < bbr->r_ctl.rc_lowest_rtt)
6394                 bbr->r_ctl.rc_lowest_rtt = rtt;
6395         bbr_log_rtt_sample(bbr, rtt, tsin);
6396         if (bbr->r_init_rtt) {
6397                 /*
6398                  * The initial rtt is not-trusted, nuke it and lets get
6399                  * our first valid measurement in.
6400                  */
6401                 bbr->r_init_rtt = 0;
6402                 tp->t_srtt = 0;
6403         }
6404         if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) {
6405                 /*
6406                  * So we have not yet figured out
6407                  * what the peers TSTMP value is
6408                  * in (most likely ms). We need a
6409                  * series of cum-ack's to determine
6410                  * this reliably.
6411                  */
6412                 if (bbr->rc_ack_is_cumack) {
6413                         if (bbr->rc_ts_data_set) {
6414                                 /* Lets attempt to determine the timestamp granularity. */
6415                                 bbr_make_timestamp_determination(bbr);
6416                         } else {
6417                                 bbr->rc_ts_data_set = 1;
6418                                 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts;
6419                                 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time;
6420                         }
6421                 } else {
6422                         /*
6423                          * We have to have consecutive acks
6424                          * reset any "filled" state to none.
6425                          */
6426                         bbr->rc_ts_data_set = 0;
6427                 }
6428         }
6429         /* Round it up */
6430         rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1)));
6431         if (rtt_ticks == 0)
6432                 rtt_ticks = 1;
6433         if (tp->t_srtt != 0) {
6434                 /*
6435                  * srtt is stored as fixed point with 5 bits after the
6436                  * binary point (i.e., scaled by 8).  The following magic is
6437                  * equivalent to the smoothing algorithm in rfc793 with an
6438                  * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
6439                  * Adjust rtt to origin 0.
6440                  */
6441
6442                 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT)
6443                     - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
6444
6445                 tp->t_srtt += delta;
6446                 if (tp->t_srtt <= 0)
6447                         tp->t_srtt = 1;
6448
6449                 /*
6450                  * We accumulate a smoothed rtt variance (actually, a
6451                  * smoothed mean difference), then set the retransmit timer
6452                  * to smoothed rtt + 4 times the smoothed variance. rttvar
6453                  * is stored as fixed point with 4 bits after the binary
6454                  * point (scaled by 16).  The following is equivalent to
6455                  * rfc793 smoothing with an alpha of .75 (rttvar =
6456                  * rttvar*3/4 + |delta| / 4).  This replaces rfc793's
6457                  * wired-in beta.
6458                  */
6459                 if (delta < 0)
6460                         delta = -delta;
6461                 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
6462                 tp->t_rttvar += delta;
6463                 if (tp->t_rttvar <= 0)
6464                         tp->t_rttvar = 1;
6465                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
6466                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6467         } else {
6468                 /*
6469                  * No rtt measurement yet - use the unsmoothed rtt. Set the
6470                  * variance to half the rtt (so our first retransmit happens
6471                  * at 3*rtt).
6472                  */
6473                 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT;
6474                 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1);
6475                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6476         }
6477         KMOD_TCPSTAT_INC(tcps_rttupdated);
6478         tp->t_rttupdated++;
6479 #ifdef STATS
6480         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks));
6481 #endif
6482         /*
6483          * the retransmit should happen at rtt + 4 * rttvar. Because of the
6484          * way we do the smoothing, srtt and rttvar will each average +1/2
6485          * tick of bias.  When we compute the retransmit timer, we want 1/2
6486          * tick of rounding and 1 extra tick because of +-1/2 tick
6487          * uncertainty in the firing of the timer.  The bias will give us
6488          * exactly the 1.5 tick we need.  But, because the bias is
6489          * statistical, we have to test that we don't drop below the minimum
6490          * feasible timer (which is 2 ticks).
6491          */
6492         TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
6493             max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2),
6494             MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
6495
6496         /*
6497          * We received an ack for a packet that wasn't retransmitted; it is
6498          * probably safe to discard any error indications we've received
6499          * recently.  This isn't quite right, but close enough for now (a
6500          * route might have failed after we sent a segment, and the return
6501          * path might not be symmetrical).
6502          */
6503         tp->t_softerror = 0;
6504         rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
6505         if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt)
6506                 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt;
6507 }
6508
6509 static void
6510 bbr_earlier_retran(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm,
6511                    uint32_t t, uint32_t cts, int ack_type)
6512 {
6513         /*
6514          * For this RSM, we acknowledged the data from a previous
6515          * transmission, not the last one we made. This means we did a false
6516          * retransmit.
6517          */
6518         if (rsm->r_flags & BBR_HAS_FIN) {
6519                 /*
6520                  * The sending of the FIN often is multiple sent when we
6521                  * have everything outstanding ack'd. We ignore this case
6522                  * since its over now.
6523                  */
6524                 return;
6525         }
6526         if (rsm->r_flags & BBR_TLP) {
6527                 /*
6528                  * We expect TLP's to have this occur often
6529                  */
6530                 bbr->rc_tlp_rtx_out = 0;
6531                 return;
6532         }
6533         if (ack_type != BBR_CUM_ACKED) {
6534                 /*
6535                  * If it was not a cum-ack we
6536                  * don't really know for sure since
6537                  * the timestamp could be from some
6538                  * other transmission.
6539                  */
6540                 return;
6541         }
6542
6543         if (rsm->r_flags & BBR_WAS_SACKPASS) {
6544                 /*
6545                  * We retransmitted based on a sack and the earlier
6546                  * retransmission ack'd it - re-ordering is occuring.
6547                  */
6548                 BBR_STAT_INC(bbr_reorder_seen);
6549                 bbr->r_ctl.rc_reorder_ts = cts;
6550         }
6551         /* Back down the loss count */
6552         if (rsm->r_flags & BBR_MARKED_LOST) {
6553                 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
6554                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
6555                 rsm->r_flags &= ~BBR_MARKED_LOST;
6556                 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
6557                         /* LT sampling also needs adjustment */
6558                         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
6559         }
6560         /***** RRS HERE ************************/
6561         /* Do we need to do this???            */
6562         /* bbr_reset_lt_bw_sampling(bbr, cts); */
6563         /***** RRS HERE ************************/
6564         BBR_STAT_INC(bbr_badfr);
6565         BBR_STAT_ADD(bbr_badfr_bytes, (rsm->r_end - rsm->r_start));
6566 }
6567
6568
6569 static void
6570 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line)
6571 {
6572         bbr->r_ctl.rc_rtt_shrinks = cts;
6573         if (bbr_can_force_probertt &&
6574             (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
6575             ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
6576                 /*
6577                  * We should enter probe-rtt its been too long
6578                  * since we have been there.
6579                  */
6580                 bbr_enter_probe_rtt(bbr, cts, __LINE__);
6581         } else
6582                 bbr_check_probe_rtt_limits(bbr, cts);
6583 }
6584
6585 static void
6586 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts)
6587 {
6588         uint64_t orig_bw;
6589
6590         if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) {
6591                 /* We never apply a zero measurment */
6592                 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0,
6593                                     0, 0, 0, 0, 0, 0);
6594                 return;
6595         }
6596         if (bbr->r_ctl.r_measurement_count < 0xffffffff)
6597                 bbr->r_ctl.r_measurement_count++;
6598         orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
6599         apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch);
6600         bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw,
6601                             (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate),
6602                             0, 0, 0, 0, 0, 0);
6603         if (orig_bw &&
6604             (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) {
6605                 if (bbr->bbr_hdrw_pacing) {
6606                         /*
6607                          * Apply a new rate to the hardware
6608                          * possibly.
6609                          */
6610                         bbr_update_hardware_pacing_rate(bbr, cts);
6611                 }
6612                 bbr_set_state_target(bbr, __LINE__);
6613                 tcp_bbr_tso_size_check(bbr, cts);
6614                 if (bbr->r_recovery_bw)  {
6615                         bbr_setup_red_bw(bbr, cts);
6616                         bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW);
6617                 }
6618         } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate))
6619                 tcp_bbr_tso_size_check(bbr, cts);
6620 }
6621
6622 static void
6623 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6624 {
6625         if (bbr->rc_in_persist == 0) {
6626                 /* We log only when not in persist */
6627                 /* Translate to a Bytes Per Second */
6628                 uint64_t tim, bw, ts_diff, ts_bw;
6629                 uint32_t upper, lower, delivered;
6630
6631                 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6632                         tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6633                 else
6634                         tim = 1;
6635                 /*
6636                  * Now that we have processed the tim (skipping the sample
6637                  * or possibly updating the time, go ahead and
6638                  * calculate the cdr.
6639                  */
6640                 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6641                 bw = (uint64_t)delivered;
6642                 bw *= (uint64_t)USECS_IN_SECOND;
6643                 bw /= tim;
6644                 if (bw == 0) {
6645                         /* We must have a calculatable amount */
6646                         return;
6647                 }
6648                 upper = (bw >> 32) & 0x00000000ffffffff;
6649                 lower = bw & 0x00000000ffffffff;
6650                 /*
6651                  * If we are using this b/w shove it in now so we
6652                  * can see in the trace viewer if it gets over-ridden.
6653                  */
6654                 if (rsm->r_ts_valid &&
6655                     bbr->rc_ts_valid &&
6656                     bbr->rc_ts_clock_set &&
6657                     (bbr->rc_ts_cant_be_used == 0) &&
6658                     bbr->rc_use_ts_limit) {
6659                         ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1);
6660                         ts_diff *= bbr->r_ctl.bbr_peer_tsratio;
6661                         if ((delivered == 0) ||
6662                             (rtt < 1000)) {
6663                                 /* Can't use the ts */
6664                                 bbr_log_type_bbrupd(bbr, 61, cts,
6665                                                     ts_diff,
6666                                                     bbr->r_ctl.last_inbound_ts,
6667                                                     rsm->r_del_ack_ts, 0,
6668                                                     0, 0, 0, delivered);
6669                         } else {
6670                                 ts_bw = (uint64_t)delivered;
6671                                 ts_bw *= (uint64_t)USECS_IN_SECOND;
6672                                 ts_bw /= ts_diff;
6673                                 bbr_log_type_bbrupd(bbr, 62, cts,
6674                                                     (ts_bw >> 32),
6675                                                     (ts_bw & 0xffffffff), 0, 0,
6676                                                     0, 0, ts_diff, delivered);
6677                                 if ((bbr->ts_can_raise) &&
6678                                     (ts_bw > bw)) {
6679                                         bbr_log_type_bbrupd(bbr, 8, cts,
6680                                                             delivered,
6681                                                             ts_diff,
6682                                                             (bw >> 32),
6683                                                             (bw & 0x00000000ffffffff),
6684                                                             0, 0, 0, 0);
6685                                         bw = ts_bw;
6686                                 } else if (ts_bw && (ts_bw < bw)) {
6687                                         bbr_log_type_bbrupd(bbr, 7, cts,
6688                                                             delivered,
6689                                                             ts_diff,
6690                                                             (bw >> 32),
6691                                                             (bw & 0x00000000ffffffff),
6692                                                             0, 0, 0, 0);
6693                                         bw = ts_bw;
6694                                 }
6695                         }
6696                 }
6697                 if (rsm->r_first_sent_time &&
6698                     TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6699                         uint64_t sbw, sti;
6700                         /*
6701                          * We use what was in flight at the time of our
6702                          * send  and the size of this send to figure
6703                          * out what we have been sending at (amount).
6704                          * For the time we take from the time of
6705                          * the send of the first send outstanding
6706                          * until this send plus this sends pacing
6707                          * time. This gives us a good calculation
6708                          * as to the rate we have been sending at.
6709                          */
6710
6711                         sbw = (uint64_t)(rsm->r_flight_at_send);
6712                         sbw *= (uint64_t)USECS_IN_SECOND;
6713                         sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6714                         sti += rsm->r_pacing_delay;
6715                         sbw /= sti;
6716                         if (sbw < bw) {
6717                                 bbr_log_type_bbrupd(bbr, 6, cts,
6718                                                     delivered,
6719                                                     (uint32_t)sti,
6720                                                     (bw >> 32),
6721                                                     (uint32_t)bw,
6722                                                     rsm->r_first_sent_time, 0, (sbw >> 32),
6723                                                     (uint32_t)sbw);
6724                                 bw = sbw;
6725                         }
6726                 }
6727                 /* Use the google algorithm for b/w measurements */
6728                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6729                 if ((rsm->r_app_limited == 0) ||
6730                     (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) {
6731                         tcp_bbr_commit_bw(bbr, cts);
6732                         bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6733                                             0, 0, 0, 0,  bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6734                 }
6735         }
6736 }
6737
6738 static void
6739 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6740 {
6741         if (bbr->rc_in_persist == 0) {
6742                 /* We log only when not in persist */
6743                 /* Translate to a Bytes Per Second */
6744                 uint64_t tim, bw;
6745                 uint32_t upper, lower, delivered;
6746                 int no_apply = 0;
6747
6748                 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6749                         tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6750                 else
6751                         tim = 1;
6752                 /*
6753                  * Now that we have processed the tim (skipping the sample
6754                  * or possibly updating the time, go ahead and
6755                  * calculate the cdr.
6756                  */
6757                 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6758                 bw = (uint64_t)delivered;
6759                 bw *= (uint64_t)USECS_IN_SECOND;
6760                 bw /= tim;
6761                 if (tim < bbr->r_ctl.rc_lowest_rtt) {
6762                         bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6763                                             tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6764
6765                         no_apply = 1;
6766                 }
6767                 upper = (bw >> 32) & 0x00000000ffffffff;
6768                 lower = bw & 0x00000000ffffffff;
6769                 /*
6770                  * If we are using this b/w shove it in now so we
6771                  * can see in the trace viewer if it gets over-ridden.
6772                  */
6773                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6774                 /* Gate by the sending rate */
6775                 if (rsm->r_first_sent_time &&
6776                     TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6777                         uint64_t sbw, sti;
6778                         /*
6779                          * We use what was in flight at the time of our
6780                          * send  and the size of this send to figure
6781                          * out what we have been sending at (amount).
6782                          * For the time we take from the time of
6783                          * the send of the first send outstanding
6784                          * until this send plus this sends pacing
6785                          * time. This gives us a good calculation
6786                          * as to the rate we have been sending at.
6787                          */
6788
6789                         sbw = (uint64_t)(rsm->r_flight_at_send);
6790                         sbw *= (uint64_t)USECS_IN_SECOND;
6791                         sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6792                         sti += rsm->r_pacing_delay;
6793                         sbw /= sti;
6794                         if (sbw < bw) {
6795                                 bbr_log_type_bbrupd(bbr, 6, cts,
6796                                                     delivered,
6797                                                     (uint32_t)sti,
6798                                                     (bw >> 32),
6799                                                     (uint32_t)bw,
6800                                                     rsm->r_first_sent_time, 0, (sbw >> 32),
6801                                                     (uint32_t)sbw);
6802                                 bw = sbw;
6803                         }
6804                         if ((sti > tim) &&
6805                             (sti < bbr->r_ctl.rc_lowest_rtt)) {
6806                                 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6807                                                     (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6808                                 no_apply = 1;
6809                         } else
6810                                 no_apply = 0;
6811                 }
6812                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6813                 if ((no_apply == 0) &&
6814                     ((rsm->r_app_limited == 0) ||
6815                      (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) {
6816                         tcp_bbr_commit_bw(bbr, cts);
6817                         bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6818                                             0, 0, 0, 0, bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6819                 }
6820         }
6821 }
6822
6823
6824 static void
6825 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin,
6826     uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to)
6827 {
6828         uint64_t old_rttprop;
6829
6830         /* Update our delivery time and amount */
6831         bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start);
6832         bbr->r_ctl.rc_del_time = cts;
6833         if (rtt == 0) {
6834                 /*
6835                  * 0 means its a retransmit, for now we don't use these for
6836                  * the rest of BBR.
6837                  */
6838                 return;
6839         }
6840         if ((bbr->rc_use_google == 0) &&
6841             (match != BBR_RTT_BY_EXACTMATCH) &&
6842             (match != BBR_RTT_BY_TIMESTAMP)){
6843                 /*
6844                  * We get a lot of rtt updates, lets not pay attention to
6845                  * any that are not an exact match. That way we don't have
6846                  * to worry about timestamps and the whole nonsense of
6847                  * unsure if its a retransmission etc (if we ever had the
6848                  * timestamp fixed to always have the last thing sent this
6849                  * would not be a issue).
6850                  */
6851                 return;
6852         }
6853         if ((bbr_no_retran && bbr->rc_use_google) &&
6854             (match != BBR_RTT_BY_EXACTMATCH) &&
6855             (match != BBR_RTT_BY_TIMESTAMP)){
6856                 /*
6857                  * We only do measurements in google mode
6858                  * with bbr_no_retran on for sure things.
6859                  */
6860                 return;
6861         }
6862         /* Only update srtt if we know by exact match */
6863         tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin);
6864         if (ack_type == BBR_CUM_ACKED)
6865                 bbr->rc_ack_is_cumack = 1;
6866         else
6867                 bbr->rc_ack_is_cumack = 0;
6868         old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP);
6869         /*
6870          * Note the following code differs to the original
6871          * BBR spec. It calls for <= not <. However after a
6872          * long discussion in email with Neal, he acknowledged
6873          * that it should be < than so that we will have flows
6874          * going into probe-rtt (we were seeing cases where that
6875          * did not happen and caused ugly things to occur). We
6876          * have added this agreed upon fix to our code base.
6877          */
6878         if (rtt < old_rttprop) {
6879                 /* Update when we last saw a rtt drop */
6880                 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0);
6881                 bbr_set_reduced_rtt(bbr, cts, __LINE__);
6882         }
6883         bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts,
6884             match, rsm->r_start, rsm->r_flags);
6885         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6886         if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) {
6887                 /*
6888                  * The RTT-prop moved, reset the target (may be a
6889                  * nop for some states).
6890                  */
6891                 bbr_set_state_target(bbr, __LINE__);
6892                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)
6893                         bbr_log_rtt_shrinks(bbr, cts, 0, 0,
6894                                             __LINE__, BBR_RTTS_NEW_TARGET, 0);
6895                 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP))
6896                         /* It went up */
6897                         bbr_check_probe_rtt_limits(bbr, cts);
6898         }
6899         if ((bbr->rc_use_google == 0) &&
6900             (match == BBR_RTT_BY_TIMESTAMP)) {
6901                 /*
6902                  * We don't do b/w update with
6903                  * these since they are not really
6904                  * reliable.
6905                  */
6906                 return;
6907         }
6908         if (bbr->r_ctl.r_app_limited_until &&
6909             (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) {
6910                 /* We are no longer app-limited */
6911                 bbr->r_ctl.r_app_limited_until = 0;
6912         }
6913         if (bbr->rc_use_google) {
6914                 bbr_google_measurement(bbr, rsm, rtt, cts);
6915         } else {
6916                 bbr_nf_measurement(bbr, rsm, rtt, cts);
6917         }
6918 }
6919
6920 /*
6921  * Convert a timestamp that the main stack
6922  * uses (milliseconds) into one that bbr uses
6923  * (microseconds). Return that converted timestamp.
6924  */
6925 static uint32_t
6926 bbr_ts_convert(uint32_t cts) {
6927         uint32_t sec, msec;
6928
6929         sec = cts / MS_IN_USEC;
6930         msec = cts - (MS_IN_USEC * sec);
6931         return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC));
6932 }
6933
6934 /*
6935  * Return 0 if we did not update the RTT time, return
6936  * 1 if we did.
6937  */
6938 static int
6939 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr,
6940     struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack)
6941 {
6942         int32_t i;
6943         uint32_t t, uts = 0;
6944
6945         if ((rsm->r_flags & BBR_ACKED) ||
6946             (rsm->r_flags & BBR_WAS_RENEGED) ||
6947             (rsm->r_flags & BBR_RXT_CLEARED)) {
6948                 /* Already done */
6949                 return (0);
6950         }
6951         if (rsm->r_rtr_cnt == 1) {
6952                 /*
6953                  * Only one transmit. Hopefully the normal case.
6954                  */
6955                 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0]))
6956                         t = cts - rsm->r_tim_lastsent[0];
6957                 else
6958                         t = 1;
6959                 if ((int)t <= 0)
6960                         t = 1;
6961                 bbr->r_ctl.rc_last_rtt = t;
6962                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6963                                     BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to);
6964                 return (1);
6965         }
6966         /* Convert to usecs */
6967         if ((bbr_can_use_ts_for_rtt == 1) &&
6968             (bbr->rc_use_google == 1) &&
6969             (ack_type == BBR_CUM_ACKED) &&
6970             (to->to_flags & TOF_TS) &&
6971             (to->to_tsecr != 0)) {
6972
6973                 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr;
6974                 if (t < 1)
6975                         t = 1;
6976                 t *= MS_IN_USEC;
6977                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6978                                     BBR_RTT_BY_TIMESTAMP,
6979                                     rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)],
6980                                     ack_type, to);
6981                 return (1);
6982         }
6983         uts = bbr_ts_convert(to->to_tsecr);
6984         if ((to->to_flags & TOF_TS) &&
6985             (to->to_tsecr != 0) &&
6986             (ack_type == BBR_CUM_ACKED) &&
6987             ((rsm->r_flags & BBR_OVERMAX) == 0)) {
6988                 /*
6989                  * Now which timestamp does it match? In this block the ACK
6990                  * may be coming from a previous transmission.
6991                  */
6992                 uint32_t fudge;
6993
6994                 fudge = BBR_TIMER_FUDGE;
6995                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
6996                         if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) &&
6997                             (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) {
6998                                 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6999                                         t = cts - rsm->r_tim_lastsent[i];
7000                                 else
7001                                         t = 1;
7002                                 if ((int)t <= 0)
7003                                         t = 1;
7004                                 bbr->r_ctl.rc_last_rtt = t;
7005                                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING,
7006                                                     rsm->r_tim_lastsent[i], ack_type, to);
7007                                 if ((i + 1) < rsm->r_rtr_cnt) {
7008                                         /* Likely */
7009                                         bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type);
7010                                 } else if (rsm->r_flags & BBR_TLP) {
7011                                         bbr->rc_tlp_rtx_out = 0;
7012                                 }
7013                                 return (1);
7014                         }
7015                 }
7016                 /* Fall through if we can't find a matching timestamp */
7017         }
7018         /*
7019          * Ok its a SACK block that we retransmitted. or a windows
7020          * machine without timestamps. We can tell nothing from the
7021          * time-stamp since its not there or the time the peer last
7022          * recieved a segment that moved forward its cum-ack point.
7023          *
7024          * Lets look at the last retransmit and see what we can tell
7025          * (with BBR for space we only keep 2 note we have to keep
7026          * at least 2 so the map can not be condensed more).
7027          */
7028         i = rsm->r_rtr_cnt - 1;
7029         if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
7030                 t = cts - rsm->r_tim_lastsent[i];
7031         else
7032                 goto not_sure;
7033         if (t < bbr->r_ctl.rc_lowest_rtt) {
7034                 /*
7035                  * We retransmitted and the ack came back in less
7036                  * than the smallest rtt we have observed in the
7037                  * windowed rtt. We most likey did an improper
7038                  * retransmit as outlined in 4.2 Step 3 point 2 in
7039                  * the rack-draft.
7040                  *
7041                  * Use the prior transmission to update all the
7042                  * information as long as there is only one prior
7043                  * transmission.
7044                  */
7045                 if ((rsm->r_flags & BBR_OVERMAX) == 0) {
7046 #ifdef BBR_INVARIANTS
7047                         if (rsm->r_rtr_cnt == 1)
7048                                 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags);
7049 #endif
7050                         i = rsm->r_rtr_cnt - 2;
7051                         if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
7052                                 t = cts - rsm->r_tim_lastsent[i];
7053                         else
7054                                 t = 1;
7055                         bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET,
7056                                             rsm->r_tim_lastsent[i], ack_type, to);
7057                         bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type);
7058                 } else {
7059                         /*
7060                          * Too many prior transmissions, just
7061                          * updated BBR delivered
7062                          */
7063 not_sure:
7064                         bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
7065                                             BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
7066                 }
7067         } else {
7068                 /*
7069                  * We retransmitted it and the retransmit did the
7070                  * job.
7071                  */
7072                 if (rsm->r_flags & BBR_TLP)
7073                         bbr->rc_tlp_rtx_out = 0;
7074                 if ((rsm->r_flags & BBR_OVERMAX) == 0)
7075                         bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts,
7076                                             BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to);
7077                 else
7078                         bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
7079                                             BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
7080                 return (1);
7081         }
7082         return (0);
7083 }
7084
7085 /*
7086  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7087  */
7088 static void
7089 bbr_log_sack_passed(struct tcpcb *tp,
7090     struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
7091 {
7092         struct bbr_sendmap *nrsm;
7093
7094         nrsm = rsm;
7095         TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap,
7096             bbr_head, r_tnext) {
7097                 if (nrsm == rsm) {
7098                         /* Skip orginal segment he is acked */
7099                         continue;
7100                 }
7101                 if (nrsm->r_flags & BBR_ACKED) {
7102                         /* Skip ack'd segments */
7103                         continue;
7104                 }
7105                 if (nrsm->r_flags & BBR_SACK_PASSED) {
7106                         /*
7107                          * We found one that is already marked
7108                          * passed, we have been here before and
7109                          * so all others below this are marked.
7110                          */
7111                         break;
7112                 }
7113                 BBR_STAT_INC(bbr_sack_passed);
7114                 nrsm->r_flags |= BBR_SACK_PASSED;
7115                 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) &&
7116                     bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) {
7117                         bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start;
7118                         bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start;
7119                         nrsm->r_flags |= BBR_MARKED_LOST;
7120                 }
7121                 nrsm->r_flags &= ~BBR_WAS_SACKPASS;
7122         }
7123 }
7124
7125 /*
7126  * Returns the number of bytes that were
7127  * newly ack'd by sack blocks.
7128  */
7129 static uint32_t
7130 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack,
7131     struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts)
7132 {
7133         int32_t times = 0;
7134         uint32_t start, end, maxseg, changed = 0;
7135         struct bbr_sendmap *rsm, *nrsm;
7136         int32_t used_ref = 1;
7137         uint8_t went_back = 0, went_fwd = 0;
7138
7139         maxseg = tp->t_maxseg - bbr->rc_last_options;
7140         start = sack->start;
7141         end = sack->end;
7142         rsm = *prsm;
7143         if (rsm == NULL)
7144                 used_ref = 0;
7145
7146         /* Do we locate the block behind where we last were? */
7147         if (rsm && SEQ_LT(start, rsm->r_start)) {
7148                 went_back = 1;
7149                 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
7150                         if (SEQ_GEQ(start, rsm->r_start) &&
7151                             SEQ_LT(start, rsm->r_end)) {
7152                                 goto do_rest_ofb;
7153                         }
7154                 }
7155         }
7156 start_at_beginning:
7157         went_fwd = 1;
7158         /*
7159          * Ok lets locate the block where this guy is fwd from rsm (if its
7160          * set)
7161          */
7162         TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) {
7163                 if (SEQ_GEQ(start, rsm->r_start) &&
7164                     SEQ_LT(start, rsm->r_end)) {
7165                         break;
7166                 }
7167         }
7168 do_rest_ofb:
7169         if (rsm == NULL) {
7170                 /*
7171                  * This happens when we get duplicate sack blocks with the
7172                  * same end. For example SACK 4: 100 SACK 3: 100 The sort
7173                  * will not change there location so we would just start at
7174                  * the end of the first one and get lost.
7175                  */
7176                 if (tp->t_flags & TF_SENTFIN) {
7177                         /*
7178                          * Check to see if we have not logged the FIN that
7179                          * went out.
7180                          */
7181                         nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7182                         if (nrsm && (nrsm->r_end + 1) == tp->snd_max) {
7183                                 /*
7184                                  * Ok we did not get the FIN logged.
7185                                  */
7186                                 nrsm->r_end++;
7187                                 rsm = nrsm;
7188                                 goto do_rest_ofb;
7189                         }
7190                 }
7191                 if (times == 1) {
7192 #ifdef BBR_INVARIANTS
7193                         panic("tp:%p bbr:%p sack:%p to:%p prsm:%p",
7194                             tp, bbr, sack, to, prsm);
7195 #else
7196                         goto out;
7197 #endif
7198                 }
7199                 times++;
7200                 BBR_STAT_INC(bbr_sack_proc_restart);
7201                 rsm = NULL;
7202                 goto start_at_beginning;
7203         }
7204         /* Ok we have an ACK for some piece of rsm */
7205         if (rsm->r_start != start) {
7206                 /*
7207                  * Need to split this in two pieces the before and after.
7208                  */
7209                 if (bbr_sack_mergable(rsm, start, end))
7210                         nrsm = bbr_alloc_full_limit(bbr);
7211                 else
7212                         nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7213                 if (nrsm == NULL) {
7214                         /* We could not allocate ignore the sack */
7215                         struct sackblk blk;
7216
7217                         blk.start = start;
7218                         blk.end = end;
7219                         sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7220                         goto out;
7221                 }
7222                 bbr_clone_rsm(bbr, nrsm, rsm, start);
7223                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7224                 if (rsm->r_in_tmap) {
7225                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7226                         nrsm->r_in_tmap = 1;
7227                 }
7228                 rsm->r_flags &= (~BBR_HAS_FIN);
7229                 rsm = nrsm;
7230         }
7231         if (SEQ_GEQ(end, rsm->r_end)) {
7232                 /*
7233                  * The end of this block is either beyond this guy or right
7234                  * at this guy.
7235                  */
7236                 if ((rsm->r_flags & BBR_ACKED) == 0) {
7237                         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7238                         changed += (rsm->r_end - rsm->r_start);
7239                         bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7240                         bbr_log_sack_passed(tp, bbr, rsm);
7241                         if (rsm->r_flags & BBR_MARKED_LOST) {
7242                                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7243                         }
7244                         /* Is Reordering occuring? */
7245                         if (rsm->r_flags & BBR_SACK_PASSED) {
7246                                 BBR_STAT_INC(bbr_reorder_seen);
7247                                 bbr->r_ctl.rc_reorder_ts = cts;
7248                                 if (rsm->r_flags & BBR_MARKED_LOST) {
7249                                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7250                                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7251                                                 /* LT sampling also needs adjustment */
7252                                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7253                                 }
7254                         }
7255                         rsm->r_flags |= BBR_ACKED;
7256                         rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7257                         if (rsm->r_in_tmap) {
7258                                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7259                                 rsm->r_in_tmap = 0;
7260                         }
7261                 }
7262                 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7263                 if (end == rsm->r_end) {
7264                         /* This block only - done */
7265                         goto out;
7266                 }
7267                 /* There is more not coverend by this rsm move on */
7268                 start = rsm->r_end;
7269                 nrsm = TAILQ_NEXT(rsm, r_next);
7270                 rsm = nrsm;
7271                 times = 0;
7272                 goto do_rest_ofb;
7273         }
7274         if (rsm->r_flags & BBR_ACKED) {
7275                 /* Been here done that */
7276                 goto out;
7277         }
7278         /* Ok we need to split off this one at the tail */
7279         if (bbr_sack_mergable(rsm, start, end))
7280                 nrsm = bbr_alloc_full_limit(bbr);
7281         else
7282                 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7283         if (nrsm == NULL) {
7284                 /* failed XXXrrs what can we do but loose the sack info? */
7285                 struct sackblk blk;
7286
7287                 blk.start = start;
7288                 blk.end = end;
7289                 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7290                 goto out;
7291         }
7292         /* Clone it */
7293         bbr_clone_rsm(bbr, nrsm, rsm, end);
7294         /* The sack block does not cover this guy fully */
7295         rsm->r_flags &= (~BBR_HAS_FIN);
7296         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7297         if (rsm->r_in_tmap) {
7298                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7299                 nrsm->r_in_tmap = 1;
7300         }
7301         nrsm->r_dupack = 0;
7302         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7303         bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7304         changed += (rsm->r_end - rsm->r_start);
7305         bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7306         bbr_log_sack_passed(tp, bbr, rsm);
7307         /* Is Reordering occuring? */
7308         if (rsm->r_flags & BBR_MARKED_LOST) {
7309                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7310         }
7311         if (rsm->r_flags & BBR_SACK_PASSED) {
7312                 BBR_STAT_INC(bbr_reorder_seen);
7313                 bbr->r_ctl.rc_reorder_ts = cts;
7314                 if (rsm->r_flags & BBR_MARKED_LOST) {
7315                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7316                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7317                                 /* LT sampling also needs adjustment */
7318                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7319                 }
7320         }
7321         rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7322         rsm->r_flags |= BBR_ACKED;
7323         if (rsm->r_in_tmap) {
7324                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7325                 rsm->r_in_tmap = 0;
7326         }
7327 out:
7328         if (rsm && (rsm->r_flags & BBR_ACKED)) {
7329                 /*
7330                  * Now can we merge this newly acked
7331                  * block with either the previous or
7332                  * next block?
7333                  */
7334                 nrsm = TAILQ_NEXT(rsm, r_next);
7335                 if (nrsm &&
7336                     (nrsm->r_flags & BBR_ACKED)) {
7337                         /* yep this and next can be merged */
7338                         rsm = bbr_merge_rsm(bbr, rsm, nrsm);
7339                 }
7340                 /* Now what about the previous? */
7341                 nrsm = TAILQ_PREV(rsm, bbr_head, r_next);
7342                 if (nrsm &&
7343                     (nrsm->r_flags & BBR_ACKED)) {
7344                         /* yep the previous and this can be merged */
7345                         rsm = bbr_merge_rsm(bbr, nrsm, rsm);
7346                 }
7347         }
7348         if (used_ref == 0) {
7349                 BBR_STAT_INC(bbr_sack_proc_all);
7350         } else {
7351                 BBR_STAT_INC(bbr_sack_proc_short);
7352         }
7353         if (went_fwd && went_back) {
7354                 BBR_STAT_INC(bbr_sack_search_both);
7355         } else if (went_fwd) {
7356                 BBR_STAT_INC(bbr_sack_search_fwd);
7357         } else if (went_back) {
7358                 BBR_STAT_INC(bbr_sack_search_back);
7359         }
7360         /* Save off where the next seq is */
7361         if (rsm)
7362                 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next);
7363         else
7364                 bbr->r_ctl.rc_sacklast = NULL;
7365         *prsm = rsm;
7366         return (changed);
7367 }
7368
7369
7370 static void inline
7371 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack)
7372 {
7373         struct bbr_sendmap *tmap;
7374
7375         BBR_STAT_INC(bbr_reneges_seen);
7376         tmap = NULL;
7377         while (rsm && (rsm->r_flags & BBR_ACKED)) {
7378                 /* Its no longer sacked, mark it so */
7379                 uint32_t oflags;
7380                 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7381 #ifdef BBR_INVARIANTS
7382                 if (rsm->r_in_tmap) {
7383                         panic("bbr:%p rsm:%p flags:0x%x in tmap?",
7384                             bbr, rsm, rsm->r_flags);
7385                 }
7386 #endif
7387                 oflags = rsm->r_flags;
7388                 if (rsm->r_flags & BBR_MARKED_LOST) {
7389                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7390                         bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7391                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7392                                 /* LT sampling also needs adjustment */
7393                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7394                 }
7395                 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST);
7396                 rsm->r_flags |= BBR_WAS_RENEGED;
7397                 rsm->r_flags |= BBR_RXT_CLEARED;
7398                 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__);
7399                 /* Rebuild it into our tmap */
7400                 if (tmap == NULL) {
7401                         TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7402                         tmap = rsm;
7403                 } else {
7404                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext);
7405                         tmap = rsm;
7406                 }
7407                 tmap->r_in_tmap = 1;
7408                 /*
7409                  * XXXrrs Delivered? Should we do anything here?
7410                  *
7411                  * Of course we don't on a rxt timeout so maybe its ok that
7412                  * we don't?
7413                  *
7414                  * For now lets not.
7415                  */
7416                 rsm = TAILQ_NEXT(rsm, r_next);
7417         }
7418         /*
7419          * Now lets possibly clear the sack filter so we start recognizing
7420          * sacks that cover this area.
7421          */
7422         sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack);
7423 }
7424
7425 static void
7426 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to)
7427 {
7428         struct tcp_bbr *bbr;
7429         struct bbr_sendmap *rsm;
7430         uint32_t cts;
7431
7432         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7433         cts = bbr->r_ctl.rc_rcvtime;
7434         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7435         if (rsm && (rsm->r_flags & BBR_HAS_SYN)) {
7436                 if ((rsm->r_end - rsm->r_start) <= 1) {
7437                         /* Log out the SYN completely */
7438                         bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7439                         rsm->r_rtr_bytes = 0;
7440                         TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7441                         if (rsm->r_in_tmap) {
7442                                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7443                                 rsm->r_in_tmap = 0;
7444                         }
7445                         if (bbr->r_ctl.rc_next == rsm) {
7446                                 /* scoot along the marker */
7447                                 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7448                         }
7449                         if (to != NULL)
7450                                 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0);
7451                         bbr_free(bbr, rsm);
7452                 } else {
7453                         /* There is more (Fast open)? strip out SYN. */
7454                         rsm->r_flags &= ~BBR_HAS_SYN;
7455                         rsm->r_start++;
7456                 }
7457         }
7458 }
7459
7460 /*
7461  * Returns the number of bytes that were
7462  * acknowledged by SACK blocks.
7463  */
7464
7465 static uint32_t
7466 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th,
7467     uint32_t *prev_acked)
7468 {
7469         uint32_t changed, last_seq, entered_recovery = 0;
7470         struct tcp_bbr *bbr;
7471         struct bbr_sendmap *rsm;
7472         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
7473         register uint32_t th_ack;
7474         int32_t i, j, k, new_sb, num_sack_blks = 0;
7475         uint32_t cts, acked, ack_point, sack_changed = 0;
7476         uint32_t p_maxseg, maxseg, p_acked = 0;
7477
7478         INP_WLOCK_ASSERT(tp->t_inpcb);
7479         if (th->th_flags & TH_RST) {
7480                 /* We don't log resets */
7481                 return (0);
7482         }
7483         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7484         cts = bbr->r_ctl.rc_rcvtime;
7485
7486         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7487         changed = 0;
7488         maxseg = tp->t_maxseg - bbr->rc_last_options;
7489         p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg);
7490         th_ack = th->th_ack;
7491         if (SEQ_GT(th_ack, tp->snd_una)) {
7492                 acked = th_ack - tp->snd_una;
7493                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__);
7494                 bbr->rc_tp->t_acktime = ticks;
7495         } else
7496                 acked = 0;
7497         if (SEQ_LEQ(th_ack, tp->snd_una)) {
7498                 /* Only sent here for sack processing */
7499                 goto proc_sack;
7500         }
7501         if (rsm && SEQ_GT(th_ack, rsm->r_start)) {
7502                 changed = th_ack - rsm->r_start;
7503         } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) {
7504                 /*
7505                  * For the SYN incoming case we will not have called
7506                  * tcp_output for the sending of the SYN, so there will be
7507                  * no map. All other cases should probably be a panic.
7508                  */
7509                 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) {
7510                         /*
7511                          * We have a timestamp that can be used to generate
7512                          * an initial RTT.
7513                          */
7514                         uint32_t ts, now, rtt;
7515
7516                         ts = bbr_ts_convert(to->to_tsecr);
7517                         now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv));
7518                         rtt = now - ts;
7519                         if (rtt < 1)
7520                                 rtt = 1;
7521                         bbr_log_type_bbrrttprop(bbr, rtt,
7522                                                 tp->iss, 0, cts,
7523                                                 BBR_RTT_BY_TIMESTAMP, tp->iss, 0);
7524                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
7525                         changed = 1;
7526                         bbr->r_wanted_output = 1;
7527                         goto out;
7528                 }
7529                 goto proc_sack;
7530         } else if (rsm == NULL) {
7531                 goto out;
7532         }
7533         if (changed) {
7534                 /*
7535                  * The ACK point is advancing to th_ack, we must drop off
7536                  * the packets in the rack log and calculate any eligble
7537                  * RTT's.
7538                  */
7539                 bbr->r_wanted_output = 1;
7540 more:
7541                 if (rsm == NULL) {
7542
7543                         if (tp->t_flags & TF_SENTFIN) {
7544                                 /* if we send a FIN we will not hav a map */
7545                                 goto proc_sack;
7546                         }
7547 #ifdef BBR_INVARIANTS
7548                         panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n",
7549                             tp,
7550                             th, tp->t_state, bbr,
7551                             tp->snd_una, tp->snd_max, changed);
7552 #endif
7553                         goto proc_sack;
7554                 }
7555         }
7556         if (SEQ_LT(th_ack, rsm->r_start)) {
7557                 /* Huh map is missing this */
7558 #ifdef BBR_INVARIANTS
7559                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n",
7560                     rsm->r_start,
7561                     th_ack, tp->t_state,
7562                     bbr->r_state, bbr);
7563                 panic("th-ack is bad bbr:%p tp:%p", bbr, tp);
7564 #endif
7565                 goto proc_sack;
7566         } else if (th_ack == rsm->r_start) {
7567                 /* None here to ack */
7568                 goto proc_sack;
7569         }
7570         /*
7571          * Clear the dup ack counter, it will
7572          * either be freed or if there is some
7573          * remaining we need to start it at zero.
7574          */
7575         rsm->r_dupack = 0;
7576         /* Now do we consume the whole thing? */
7577         if (SEQ_GEQ(th_ack, rsm->r_end)) {
7578                 /* Its all consumed. */
7579                 uint32_t left;
7580
7581                 if (rsm->r_flags & BBR_ACKED) {
7582                         /*
7583                          * It was acked on the scoreboard -- remove it from
7584                          * total
7585                          */
7586                         p_acked += (rsm->r_end - rsm->r_start);
7587                         bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7588                         if (bbr->r_ctl.rc_sacked == 0)
7589                                 bbr->r_ctl.rc_sacklast = NULL;
7590                 } else {
7591                         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack);
7592                         if (rsm->r_flags & BBR_MARKED_LOST) {
7593                                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7594                         }
7595                         if (rsm->r_flags & BBR_SACK_PASSED) {
7596                                 /*
7597                                  * There are acked segments ACKED on the
7598                                  * scoreboard further up. We are seeing
7599                                  * reordering.
7600                                  */
7601                                 BBR_STAT_INC(bbr_reorder_seen);
7602                                 bbr->r_ctl.rc_reorder_ts = cts;
7603                                 if (rsm->r_flags & BBR_MARKED_LOST) {
7604                                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7605                                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7606                                                 /* LT sampling also needs adjustment */
7607                                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7608                                 }
7609                         }
7610                         rsm->r_flags &= ~BBR_MARKED_LOST;
7611                 }
7612                 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7613                 rsm->r_rtr_bytes = 0;
7614                 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7615                 if (rsm->r_in_tmap) {
7616                         TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7617                         rsm->r_in_tmap = 0;
7618                 }
7619                 if (bbr->r_ctl.rc_next == rsm) {
7620                         /* scoot along the marker */
7621                         bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7622                 }
7623                 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7624                 /* Adjust the packet counts */
7625                 left = th_ack - rsm->r_end;
7626                 /* Free back to zone */
7627                 bbr_free(bbr, rsm);
7628                 if (left) {
7629                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7630                         goto more;
7631                 }
7632                 goto proc_sack;
7633         }
7634         if (rsm->r_flags & BBR_ACKED) {
7635                 /*
7636                  * It was acked on the scoreboard -- remove it from total
7637                  * for the part being cum-acked.
7638                  */
7639                 p_acked += (rsm->r_end - rsm->r_start);
7640                 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
7641                 if (bbr->r_ctl.rc_sacked == 0)
7642                         bbr->r_ctl.rc_sacklast = NULL;
7643         } else {
7644                 /*
7645                  * It was acked up to th_ack point for the first time
7646                  */
7647                 struct bbr_sendmap lrsm;
7648
7649                 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap));
7650                 lrsm.r_end = th_ack;
7651                 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack);
7652         }
7653         if ((rsm->r_flags & BBR_MARKED_LOST) &&
7654             ((rsm->r_flags & BBR_ACKED) == 0)) {
7655                 /*
7656                  * It was marked lost and partly ack'd now
7657                  * for the first time. We lower the rc_lost_bytes
7658                  * and still leave it MARKED.
7659                  */
7660                 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start;
7661         }
7662         bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7663         bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7664         rsm->r_rtr_bytes = 0;
7665         /* adjust packet count */
7666         rsm->r_start = th_ack;
7667 proc_sack:
7668         /* Check for reneging */
7669         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7670         if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) {
7671                 /*
7672                  * The peer has moved snd_una up to the edge of this send,
7673                  * i.e. one that it had previously acked. The only way that
7674                  * can be true if the peer threw away data (space issues)
7675                  * that it had previously sacked (else it would have given
7676                  * us snd_una up to (rsm->r_end). We need to undo the acked
7677                  * markings here.
7678                  *
7679                  * Note we have to look to make sure th_ack is our
7680                  * rsm->r_start in case we get an old ack where th_ack is
7681                  * behind snd_una.
7682                  */
7683                 bbr_peer_reneges(bbr, rsm, th->th_ack);
7684         }
7685         if ((to->to_flags & TOF_SACK) == 0) {
7686                 /* We are done nothing left to log */
7687                 goto out;
7688         }
7689         rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7690         if (rsm) {
7691                 last_seq = rsm->r_end;
7692         } else {
7693                 last_seq = tp->snd_max;
7694         }
7695         /* Sack block processing */
7696         if (SEQ_GT(th_ack, tp->snd_una))
7697                 ack_point = th_ack;
7698         else
7699                 ack_point = tp->snd_una;
7700         for (i = 0; i < to->to_nsacks; i++) {
7701                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
7702                     &sack, sizeof(sack));
7703                 sack.start = ntohl(sack.start);
7704                 sack.end = ntohl(sack.end);
7705                 if (SEQ_GT(sack.end, sack.start) &&
7706                     SEQ_GT(sack.start, ack_point) &&
7707                     SEQ_LT(sack.start, tp->snd_max) &&
7708                     SEQ_GT(sack.end, ack_point) &&
7709                     SEQ_LEQ(sack.end, tp->snd_max)) {
7710                         if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) &&
7711                             (SEQ_LT(sack.end, last_seq)) &&
7712                             ((sack.end - sack.start) < (p_maxseg / 8))) {
7713                                 /*
7714                                  * Not the last piece and its smaller than
7715                                  * 1/8th of a p_maxseg. We ignore this.
7716                                  */
7717                                 BBR_STAT_INC(bbr_runt_sacks);
7718                                 continue;
7719                         }
7720                         sack_blocks[num_sack_blks] = sack;
7721                         num_sack_blks++;
7722 #ifdef NETFLIX_STATS
7723                 } else if (SEQ_LEQ(sack.start, th_ack) &&
7724                     SEQ_LEQ(sack.end, th_ack)) {
7725                         /*
7726                          * Its a D-SACK block.
7727                          */
7728                         tcp_record_dsack(sack.start, sack.end);
7729 #endif
7730                 }
7731         }
7732         if (num_sack_blks == 0)
7733                 goto out;
7734         /*
7735          * Sort the SACK blocks so we can update the rack scoreboard with
7736          * just one pass.
7737          */
7738         new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks,
7739                                   num_sack_blks, th->th_ack);
7740         ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks);
7741         BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks);
7742         BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb));
7743         num_sack_blks = new_sb;
7744         if (num_sack_blks < 2) {
7745                 goto do_sack_work;
7746         }
7747         /* Sort the sacks */
7748         for (i = 0; i < num_sack_blks; i++) {
7749                 for (j = i + 1; j < num_sack_blks; j++) {
7750                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
7751                                 sack = sack_blocks[i];
7752                                 sack_blocks[i] = sack_blocks[j];
7753                                 sack_blocks[j] = sack;
7754                         }
7755                 }
7756         }
7757         /*
7758          * Now are any of the sack block ends the same (yes some
7759          * implememtations send these)?
7760          */
7761 again:
7762         if (num_sack_blks > 1) {
7763                 for (i = 0; i < num_sack_blks; i++) {
7764                         for (j = i + 1; j < num_sack_blks; j++) {
7765                                 if (sack_blocks[i].end == sack_blocks[j].end) {
7766                                         /*
7767                                          * Ok these two have the same end we
7768                                          * want the smallest end and then
7769                                          * throw away the larger and start
7770                                          * again.
7771                                          */
7772                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
7773                                                 /*
7774                                                  * The second block covers
7775                                                  * more area use that
7776                                                  */
7777                                                 sack_blocks[i].start = sack_blocks[j].start;
7778                                         }
7779                                         /*
7780                                          * Now collapse out the dup-sack and
7781                                          * lower the count
7782                                          */
7783                                         for (k = (j + 1); k < num_sack_blks; k++) {
7784                                                 sack_blocks[j].start = sack_blocks[k].start;
7785                                                 sack_blocks[j].end = sack_blocks[k].end;
7786                                                 j++;
7787                                         }
7788                                         num_sack_blks--;
7789                                         goto again;
7790                                 }
7791                         }
7792                 }
7793         }
7794 do_sack_work:
7795         rsm = bbr->r_ctl.rc_sacklast;
7796         for (i = 0; i < num_sack_blks; i++) {
7797                 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts);
7798                 if (acked) {
7799                         bbr->r_wanted_output = 1;
7800                         changed += acked;
7801                         sack_changed += acked;
7802                 }
7803         }
7804 out:
7805         *prev_acked = p_acked;
7806         if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) {
7807                 /*
7808                  * Ok we have a high probability that we need to go in to
7809                  * recovery since we have data sack'd
7810                  */
7811                 struct bbr_sendmap *rsm;
7812
7813                 rsm = bbr_check_recovery_mode(tp, bbr, cts);
7814                 if (rsm) {
7815                         /* Enter recovery */
7816                         entered_recovery = 1;
7817                         bbr->r_wanted_output = 1;
7818                         /*
7819                          * When we enter recovery we need to assure we send
7820                          * one packet.
7821                          */
7822                         if (bbr->r_ctl.rc_resend == NULL) {
7823                                 bbr->r_ctl.rc_resend = rsm;
7824                         }
7825                 }
7826         }
7827         if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) {
7828                 /*
7829                  * See if we need to rack-retransmit anything if so set it
7830                  * up as the thing to resend assuming something else is not
7831                  * already in that position.
7832                  */
7833                 if (bbr->r_ctl.rc_resend == NULL) {
7834                         bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
7835                 }
7836         }
7837         /*
7838          * We return the amount that changed via sack, this is used by the
7839          * ack-received code to augment what was changed between th_ack <->
7840          * snd_una.
7841          */
7842         return (sack_changed);
7843 }
7844
7845 static void
7846 bbr_strike_dupack(struct tcp_bbr *bbr)
7847 {
7848         struct bbr_sendmap *rsm;
7849
7850         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
7851         if (rsm && (rsm->r_dupack < 0xff)) {
7852                 rsm->r_dupack++;
7853                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD)
7854                         bbr->r_wanted_output = 1;
7855         }
7856 }
7857
7858 /*
7859  * Return value of 1, we do not need to call bbr_process_data().
7860  * return value of 0, bbr_process_data can be called.
7861  * For ret_val if its 0 the TCB is locked and valid, if its non-zero
7862  * its unlocked and probably unsafe to touch the TCB.
7863  */
7864 static int
7865 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
7866     struct tcpcb *tp, struct tcpopt *to,
7867     uint32_t tiwin, int32_t tlen,
7868     int32_t * ofia, int32_t thflags, int32_t * ret_val)
7869 {
7870         int32_t ourfinisacked = 0;
7871         int32_t acked_amount;
7872         uint16_t nsegs;
7873         int32_t acked;
7874         uint32_t lost, sack_changed = 0;
7875         struct mbuf *mfree;
7876         struct tcp_bbr *bbr;
7877         uint32_t prev_acked = 0;
7878
7879         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7880         lost = bbr->r_ctl.rc_lost;
7881         nsegs = max(1, m->m_pkthdr.lro_nsegs);
7882         if (SEQ_GT(th->th_ack, tp->snd_max)) {
7883                 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
7884                 bbr->r_wanted_output = 1;
7885                 return (1);
7886         }
7887         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
7888                 /* Process the ack */
7889                 if (bbr->rc_in_persist)
7890                         tp->t_rxtshift = 0;
7891                 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd))
7892                         bbr_strike_dupack(bbr);
7893                 sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
7894         }
7895         bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost));
7896         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
7897                 /*
7898                  * Old ack, behind the last one rcv'd or a duplicate ack
7899                  * with SACK info.
7900                  */
7901                 if (th->th_ack == tp->snd_una) {
7902                         bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0);
7903                         if (bbr->r_state == TCPS_SYN_SENT) {
7904                                 /*
7905                                  * Special case on where we sent SYN. When
7906                                  * the SYN-ACK is processed in syn_sent
7907                                  * state it bumps the snd_una. This causes
7908                                  * us to hit here even though we did ack 1
7909                                  * byte.
7910                                  *
7911                                  * Go through the nothing left case so we
7912                                  * send data.
7913                                  */
7914                                 goto nothing_left;
7915                         }
7916                 }
7917                 return (0);
7918         }
7919         /*
7920          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
7921          * something we sent.
7922          */
7923         if (tp->t_flags & TF_NEEDSYN) {
7924                 /*
7925                  * T/TCP: Connection was half-synchronized, and our SYN has
7926                  * been ACK'd (so connection is now fully synchronized).  Go
7927                  * to non-starred state, increment snd_una for ACK of SYN,
7928                  * and check if we can do window scaling.
7929                  */
7930                 tp->t_flags &= ~TF_NEEDSYN;
7931                 tp->snd_una++;
7932                 /* Do window scaling? */
7933                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
7934                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
7935                         tp->rcv_scale = tp->request_r_scale;
7936                         /* Send window already scaled. */
7937                 }
7938         }
7939         INP_WLOCK_ASSERT(tp->t_inpcb);
7940
7941         acked = BYTES_THIS_ACK(tp, th);
7942         KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
7943         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
7944
7945         /*
7946          * If we just performed our first retransmit, and the ACK arrives
7947          * within our recovery window, then it was a mistake to do the
7948          * retransmit in the first place.  Recover our original cwnd and
7949          * ssthresh, and proceed to transmit where we left off.
7950          */
7951         if (tp->t_flags & TF_PREVVALID) {
7952                 tp->t_flags &= ~TF_PREVVALID;
7953                 if (tp->t_rxtshift == 1 &&
7954                     (int)(ticks - tp->t_badrxtwin) < 0)
7955                         bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
7956         }
7957         SOCKBUF_LOCK(&so->so_snd);
7958         acked_amount = min(acked, (int)sbavail(&so->so_snd));
7959         tp->snd_wnd -= acked_amount;
7960         mfree = sbcut_locked(&so->so_snd, acked_amount);
7961         /* NB: sowwakeup_locked() does an implicit unlock. */
7962         sowwakeup_locked(so);
7963         m_freem(mfree);
7964         if (SEQ_GT(th->th_ack, tp->snd_una)) {
7965                 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
7966         }
7967         tp->snd_una = th->th_ack;
7968         bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost));
7969         if (IN_RECOVERY(tp->t_flags)) {
7970                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
7971                     (SEQ_LT(th->th_ack, tp->snd_max))) {
7972                         tcp_bbr_partialack(tp);
7973                 } else {
7974                         bbr_post_recovery(tp);
7975                 }
7976         }
7977         if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
7978                 tp->snd_recover = tp->snd_una;
7979         }
7980         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
7981                 tp->snd_nxt = tp->snd_max;
7982         }
7983         if (tp->snd_una == tp->snd_max) {
7984                 /* Nothing left outstanding */
7985 nothing_left:
7986                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
7987                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
7988                         bbr->rc_tp->t_acktime = 0;
7989                 if ((sbused(&so->so_snd) == 0) &&
7990                     (tp->t_flags & TF_SENTFIN)) {
7991                         ourfinisacked = 1;
7992                 }
7993                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
7994                 if (bbr->rc_in_persist == 0) {
7995                         bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
7996                 }
7997                 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
7998                 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
7999                 /*
8000                  * We invalidate the last ack here since we
8001                  * don't want to transfer forward the time
8002                  * for our sum's calculations.
8003                  */
8004                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
8005                     (sbavail(&so->so_snd) == 0) &&
8006                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
8007                         /*
8008                          * The socket was gone and the peer sent data, time
8009                          * to reset him.
8010                          */
8011                         *ret_val = 1;
8012                         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
8013                         /* tcp_close will kill the inp pre-log the Reset */
8014                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
8015                         tp = tcp_close(tp);
8016                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
8017                         BBR_STAT_INC(bbr_dropped_af_data);
8018                         return (1);
8019                 }
8020                 /* Set need output so persist might get set */
8021                 bbr->r_wanted_output = 1;
8022         }
8023         if (ofia)
8024                 *ofia = ourfinisacked;
8025         return (0);
8026 }
8027
8028 static void
8029 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
8030 {
8031         if (bbr->rc_in_persist == 0) {
8032                 bbr_timer_cancel(bbr, __LINE__, cts);
8033                 bbr->r_ctl.rc_last_delay_val = 0;
8034                 tp->t_rxtshift = 0;
8035                 bbr->rc_in_persist = 1;
8036                 bbr->r_ctl.rc_went_idle_time = cts;
8037                 /* We should be capped when rw went to 0 but just in case */
8038                 bbr_log_type_pesist(bbr, cts, 0, line, 1);
8039                 /* Time freezes for the state, so do the accounting now */
8040                 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
8041                         uint32_t time_in;
8042
8043                         time_in = cts - bbr->r_ctl.rc_bbr_state_time;
8044                         if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
8045                                 int32_t idx;
8046
8047                                 idx = bbr_state_val(bbr);
8048                                 counter_u64_add(bbr_state_time[(idx + 5)], time_in);
8049                         } else {
8050                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
8051                         }
8052                 }
8053                 bbr->r_ctl.rc_bbr_state_time = cts;
8054         }
8055 }
8056
8057 static void
8058 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time)
8059 {
8060         /*
8061          * Note that if idle time does not exceed our
8062          * threshold, we do nothing continuing the state
8063          * transitions we were last walking through.
8064          */
8065         if (idle_time >= bbr_idle_restart_threshold) {
8066                 if (bbr->rc_use_idle_restart) {
8067                         bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT;
8068                         /*
8069                          * Set our target using BBR_UNIT, so
8070                          * we increase at a dramatic rate but
8071                          * we stop when we get the pipe
8072                          * full again for our current b/w estimate.
8073                          */
8074                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
8075                         bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
8076                         bbr_set_state_target(bbr, __LINE__);
8077                         /* Now setup our gains to ramp up */
8078                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
8079                         bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
8080                         bbr_log_type_statechange(bbr, cts, __LINE__);
8081                 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
8082                         bbr_substate_change(bbr, cts, __LINE__, 1);
8083                 }
8084         }
8085 }
8086
8087 static void
8088 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
8089 {
8090         uint32_t idle_time;
8091
8092         if (bbr->rc_in_persist == 0)
8093                 return;
8094         idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time);
8095         bbr->rc_in_persist = 0;
8096         bbr->rc_hit_state_1 = 0;
8097         bbr->r_ctl.rc_del_time = cts;
8098         /*
8099          * We invalidate the last ack here since we
8100          * don't want to transfer forward the time
8101          * for our sum's calculations.
8102          */
8103         if (bbr->rc_inp->inp_in_hpts) {
8104                 tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
8105                 bbr->rc_timer_first = 0;
8106                 bbr->r_ctl.rc_hpts_flags = 0;
8107                 bbr->r_ctl.rc_last_delay_val = 0;
8108                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
8109                 bbr->r_agg_early_set = 0;
8110                 bbr->r_ctl.rc_agg_early = 0;
8111         }
8112         bbr_log_type_pesist(bbr, cts, idle_time, line, 0);
8113         if (idle_time >= bbr_rtt_probe_time) {
8114                 /*
8115                  * This qualifies as a RTT_PROBE session since we drop the
8116                  * data outstanding to nothing and waited more than
8117                  * bbr_rtt_probe_time.
8118                  */
8119                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0);
8120                 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts;
8121         }
8122         tp->t_rxtshift = 0;
8123         /*
8124          * If in probeBW and we have persisted more than an RTT lets do
8125          * special handling.
8126          */
8127         /* Force a time based epoch */
8128         bbr_set_epoch(bbr, cts, __LINE__);
8129         /*
8130          * Setup the lost so we don't count anything against the guy
8131          * we have been stuck with during persists.
8132          */
8133         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
8134         /* Time un-freezes for the state */
8135         bbr->r_ctl.rc_bbr_state_time = cts;
8136         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) ||
8137             (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) {
8138                 /*
8139                  * If we are going back to probe-bw
8140                  * or probe_rtt, we may need to possibly
8141                  * do a fast restart.
8142                  */
8143                 bbr_restart_after_idle(bbr, cts, idle_time);
8144         }
8145 }
8146
8147 static void
8148 bbr_collapsed_window(struct tcp_bbr *bbr)
8149 {
8150         /*
8151          * Now we must walk the
8152          * send map and divide the
8153          * ones left stranded. These
8154          * guys can't cause us to abort
8155          * the connection and are really
8156          * "unsent". However if a buggy
8157          * client actually did keep some
8158          * of the data i.e. collapsed the win
8159          * and refused to ack and then opened
8160          * the win and acked that data. We would
8161          * get into an ack war, the simplier
8162          * method then of just pretending we
8163          * did not send those segments something
8164          * won't work.
8165          */
8166         struct bbr_sendmap *rsm, *nrsm;
8167         tcp_seq max_seq;
8168         uint32_t maxseg;
8169         int can_split = 0;
8170         int fnd = 0;
8171
8172         maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
8173         max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd;
8174         bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0);
8175         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
8176                 /* Find the first seq past or at maxseq */
8177                 if (rsm->r_flags & BBR_RWND_COLLAPSED)
8178                         rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8179                 if (SEQ_GEQ(max_seq, rsm->r_start) &&
8180                     SEQ_GEQ(rsm->r_end, max_seq)) {
8181                         fnd = 1;
8182                         break;
8183                 }
8184         }
8185         bbr->rc_has_collapsed = 0;
8186         if (!fnd) {
8187                 /* Nothing to do strange */
8188                 return;
8189         }
8190         /*
8191          * Now can we split?
8192          *
8193          * We don't want to split if splitting
8194          * would generate too many small segments
8195          * less we let an attacker fragment our
8196          * send_map and leave us out of memory.
8197          */
8198         if ((max_seq != rsm->r_start) &&
8199             (max_seq != rsm->r_end)){
8200                 /* can we split? */
8201                 int res1, res2;
8202
8203                 res1 = max_seq - rsm->r_start;
8204                 res2 = rsm->r_end - max_seq;
8205                 if ((res1 >= (maxseg/8)) &&
8206                     (res2 >= (maxseg/8))) {
8207                         /* No small pieces here */
8208                         can_split = 1;
8209                 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) {
8210                         /* We are under the limit */
8211                         can_split = 1;
8212                 }
8213         }
8214         /* Ok do we need to split this rsm? */
8215         if (max_seq == rsm->r_start) {
8216                 /* It's this guy no split required */
8217                 nrsm = rsm;
8218         } else if (max_seq == rsm->r_end) {
8219                 /* It's the next one no split required. */
8220                 nrsm = TAILQ_NEXT(rsm, r_next);
8221                 if (nrsm == NULL) {
8222                         /* Huh? */
8223                         return;
8224                 }
8225         } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) {
8226                 /* yep we need to split it */
8227                 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
8228                 if (nrsm == NULL) {
8229                         /* failed XXXrrs what can we do mark the whole? */
8230                         nrsm = rsm;
8231                         goto no_split;
8232                 }
8233                 /* Clone it */
8234                 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0);
8235                 bbr_clone_rsm(bbr, nrsm, rsm, max_seq);
8236                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
8237                 if (rsm->r_in_tmap) {
8238                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8239                         nrsm->r_in_tmap = 1;
8240                 }
8241         } else {
8242                 /*
8243                  * Split not allowed just start here just
8244                  * use this guy.
8245                  */
8246                 nrsm = rsm;
8247         }
8248 no_split:
8249         BBR_STAT_INC(bbr_collapsed_win);
8250         /* reuse fnd as a count */
8251         fnd = 0;
8252         TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) {
8253                 nrsm->r_flags |= BBR_RWND_COLLAPSED;
8254                 fnd++;
8255                 bbr->rc_has_collapsed = 1;
8256         }
8257         bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd);
8258 }
8259
8260 static void
8261 bbr_un_collapse_window(struct tcp_bbr *bbr)
8262 {
8263         struct bbr_sendmap *rsm;
8264         int cleared = 0;
8265
8266         TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
8267                 if (rsm->r_flags & BBR_RWND_COLLAPSED) {
8268                         /* Clear the flag */
8269                         rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8270                         cleared++;
8271                 } else
8272                         break;
8273         }
8274         bbr_log_type_rwnd_collapse(bbr,
8275                                    (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared);
8276         bbr->rc_has_collapsed = 0;
8277 }
8278
8279 /*
8280  * Return value of 1, the TCB is unlocked and most
8281  * likely gone, return value of 0, the TCB is still
8282  * locked.
8283  */
8284 static int
8285 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
8286     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
8287     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8288 {
8289         /*
8290          * Update window information. Don't look at window if no ACK: TAC's
8291          * send garbage on first SYN.
8292          */
8293         uint16_t nsegs;
8294         int32_t tfo_syn;
8295         struct tcp_bbr *bbr;
8296
8297         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8298         INP_WLOCK_ASSERT(tp->t_inpcb);
8299         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8300         if ((thflags & TH_ACK) &&
8301             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
8302             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
8303             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
8304                 /* keep track of pure window updates */
8305                 if (tlen == 0 &&
8306                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
8307                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
8308                 tp->snd_wnd = tiwin;
8309                 tp->snd_wl1 = th->th_seq;
8310                 tp->snd_wl2 = th->th_ack;
8311                 if (tp->snd_wnd > tp->max_sndwnd)
8312                         tp->max_sndwnd = tp->snd_wnd;
8313                 bbr->r_wanted_output = 1;
8314         } else if (thflags & TH_ACK) {
8315                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
8316                         tp->snd_wnd = tiwin;
8317                         tp->snd_wl1 = th->th_seq;
8318                         tp->snd_wl2 = th->th_ack;
8319                 }
8320         }
8321         if (tp->snd_wnd < ctf_outstanding(tp))
8322                 /* The peer collapsed its window on us */
8323                 bbr_collapsed_window(bbr);
8324         else if (bbr->rc_has_collapsed)
8325                 bbr_un_collapse_window(bbr);
8326         /* Was persist timer active and now we have window space? */
8327         if ((bbr->rc_in_persist != 0) &&
8328             (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8329                                 bbr_minseg(bbr)))) {
8330                 /*
8331                  * Make the rate persist at end of persist mode if idle long
8332                  * enough
8333                  */
8334                 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8335
8336                 /* Make sure we output to start the timer */
8337                 bbr->r_wanted_output = 1;
8338         }
8339         /* Do we need to enter persist? */
8340         if ((bbr->rc_in_persist == 0) &&
8341             (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8342             TCPS_HAVEESTABLISHED(tp->t_state) &&
8343             (tp->snd_max == tp->snd_una) &&
8344             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8345             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8346                 /* No send window.. we must enter persist */
8347                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8348         }
8349         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
8350                 m_freem(m);
8351                 return (0);
8352         }
8353         /*
8354          * We don't support urgent data but
8355          * drag along the up just to make sure
8356          * if there is a stack switch no one
8357          * is surprised.
8358          */
8359         tp->rcv_up = tp->rcv_nxt;
8360         INP_WLOCK_ASSERT(tp->t_inpcb);
8361
8362         /*
8363          * Process the segment text, merging it into the TCP sequencing
8364          * queue, and arranging for acknowledgment of receipt if necessary.
8365          * This process logically involves adjusting tp->rcv_wnd as data is
8366          * presented to the user (this happens in tcp_usrreq.c, case
8367          * PRU_RCVD).  If a FIN has already been received on this connection
8368          * then we just ignore the text.
8369          */
8370         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
8371                    IS_FASTOPEN(tp->t_flags));
8372         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
8373             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8374                 tcp_seq save_start = th->th_seq;
8375                 tcp_seq save_rnxt  = tp->rcv_nxt;
8376                 int     save_tlen  = tlen;
8377
8378                 m_adj(m, drop_hdrlen);  /* delayed header drop */
8379                 /*
8380                  * Insert segment which includes th into TCP reassembly
8381                  * queue with control block tp.  Set thflags to whether
8382                  * reassembly now includes a segment with FIN.  This handles
8383                  * the common case inline (segment is the next to be
8384                  * received on an established connection, and the queue is
8385                  * empty), avoiding linkage into and removal from the queue
8386                  * and repetition of various conversions. Set DELACK for
8387                  * segments received in order, but ack immediately when
8388                  * segments are out of order (so fast retransmit can work).
8389                  */
8390                 if (th->th_seq == tp->rcv_nxt &&
8391                     SEGQ_EMPTY(tp) &&
8392                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
8393                     tfo_syn)) {
8394 #ifdef NETFLIX_SB_LIMITS
8395                         u_int mcnt, appended;
8396
8397                         if (so->so_rcv.sb_shlim) {
8398                                 mcnt = m_memcnt(m);
8399                                 appended = 0;
8400                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8401                                     CFO_NOSLEEP, NULL) == false) {
8402                                         counter_u64_add(tcp_sb_shlim_fails, 1);
8403                                         m_freem(m);
8404                                         return (0);
8405                                 }
8406                         }
8407
8408 #endif
8409                         if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) {
8410                                 bbr->bbr_segs_rcvd += max(1, nsegs);
8411                                 tp->t_flags |= TF_DELACK;
8412                                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8413                         } else {
8414                                 bbr->r_wanted_output = 1;
8415                                 tp->t_flags |= TF_ACKNOW;
8416                         }
8417                         tp->rcv_nxt += tlen;
8418                         if (tlen &&
8419                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8420                             (tp->t_fbyte_in == 0)) {
8421                                 tp->t_fbyte_in = ticks;
8422                                 if (tp->t_fbyte_in == 0)
8423                                         tp->t_fbyte_in = 1;
8424                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
8425                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8426                         }
8427                         thflags = th->th_flags & TH_FIN;
8428                         KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8429                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8430                         SOCKBUF_LOCK(&so->so_rcv);
8431                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
8432                                 m_freem(m);
8433                         else
8434 #ifdef NETFLIX_SB_LIMITS
8435                                 appended =
8436 #endif
8437                                         sbappendstream_locked(&so->so_rcv, m, 0);
8438                         /* NB: sorwakeup_locked() does an implicit unlock. */
8439                         sorwakeup_locked(so);
8440 #ifdef NETFLIX_SB_LIMITS
8441                         if (so->so_rcv.sb_shlim && appended != mcnt)
8442                                 counter_fo_release(so->so_rcv.sb_shlim,
8443                                     mcnt - appended);
8444 #endif
8445                 } else {
8446                         /*
8447                          * XXX: Due to the header drop above "th" is
8448                          * theoretically invalid by now.  Fortunately
8449                          * m_adj() doesn't actually frees any mbufs when
8450                          * trimming from the head.
8451                          */
8452                         tcp_seq temp = save_start;
8453                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
8454                         tp->t_flags |= TF_ACKNOW;
8455                 }
8456                 if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0)) {
8457                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
8458                                 /*
8459                                  * DSACK actually handled in the fastpath
8460                                  * above.
8461                                  */
8462                                 tcp_update_sack_list(tp, save_start,
8463                                     save_start + save_tlen);
8464                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
8465                                 if ((tp->rcv_numsacks >= 1) &&
8466                                     (tp->sackblks[0].end == save_start)) {
8467                                         /*
8468                                          * Partial overlap, recorded at todrop
8469                                          * above.
8470                                          */
8471                                         tcp_update_sack_list(tp,
8472                                             tp->sackblks[0].start,
8473                                             tp->sackblks[0].end);
8474                                 } else {
8475                                         tcp_update_dsack_list(tp, save_start,
8476                                             save_start + save_tlen);
8477                                 }
8478                         } else if (tlen >= save_tlen) {
8479                                 /* Update of sackblks. */
8480                                 tcp_update_dsack_list(tp, save_start,
8481                                     save_start + save_tlen);
8482                         } else if (tlen > 0) {
8483                                 tcp_update_dsack_list(tp, save_start,
8484                                     save_start + tlen);
8485                         }
8486                 }
8487         } else {
8488                 m_freem(m);
8489                 thflags &= ~TH_FIN;
8490         }
8491
8492         /*
8493          * If FIN is received ACK the FIN and let the user know that the
8494          * connection is closing.
8495          */
8496         if (thflags & TH_FIN) {
8497                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8498                         socantrcvmore(so);
8499                         /*
8500                          * If connection is half-synchronized (ie NEEDSYN
8501                          * flag on) then delay ACK, so it may be piggybacked
8502                          * when SYN is sent. Otherwise, since we received a
8503                          * FIN then no more input can be expected, send ACK
8504                          * now.
8505                          */
8506                         if (tp->t_flags & TF_NEEDSYN) {
8507                                 tp->t_flags |= TF_DELACK;
8508                                 bbr_timer_cancel(bbr,
8509                                     __LINE__, bbr->r_ctl.rc_rcvtime);
8510                         } else {
8511                                 tp->t_flags |= TF_ACKNOW;
8512                         }
8513                         tp->rcv_nxt++;
8514                 }
8515                 switch (tp->t_state) {
8516
8517                         /*
8518                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
8519                          * CLOSE_WAIT state.
8520                          */
8521                 case TCPS_SYN_RECEIVED:
8522                         tp->t_starttime = ticks;
8523                         /* FALLTHROUGH */
8524                 case TCPS_ESTABLISHED:
8525                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
8526                         break;
8527
8528                         /*
8529                          * If still in FIN_WAIT_1 STATE FIN has not been
8530                          * acked so enter the CLOSING state.
8531                          */
8532                 case TCPS_FIN_WAIT_1:
8533                         tcp_state_change(tp, TCPS_CLOSING);
8534                         break;
8535
8536                         /*
8537                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
8538                          * starting the time-wait timer, turning off the
8539                          * other standard timers.
8540                          */
8541                 case TCPS_FIN_WAIT_2:
8542                         bbr->rc_timer_first = 1;
8543                         bbr_timer_cancel(bbr,
8544                             __LINE__, bbr->r_ctl.rc_rcvtime);
8545                         INP_WLOCK_ASSERT(tp->t_inpcb);
8546                         tcp_twstart(tp);
8547                         return (1);
8548                 }
8549         }
8550         /*
8551          * Return any desired output.
8552          */
8553         if ((tp->t_flags & TF_ACKNOW) ||
8554             (sbavail(&so->so_snd) > ctf_outstanding(tp))) {
8555                 bbr->r_wanted_output = 1;
8556         }
8557         INP_WLOCK_ASSERT(tp->t_inpcb);
8558         return (0);
8559 }
8560
8561 /*
8562  * Here nothing is really faster, its just that we
8563  * have broken out the fast-data path also just like
8564  * the fast-ack. Return 1 if we processed the packet
8565  * return 0 if you need to take the "slow-path".
8566  */
8567 static int
8568 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
8569     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8570     uint32_t tiwin, int32_t nxt_pkt)
8571 {
8572         uint16_t nsegs;
8573         int32_t newsize = 0;    /* automatic sockbuf scaling */
8574         struct tcp_bbr *bbr;
8575 #ifdef NETFLIX_SB_LIMITS
8576         u_int mcnt, appended;
8577 #endif
8578 #ifdef TCPDEBUG
8579         /*
8580          * The size of tcp_saveipgen must be the size of the max ip header,
8581          * now IPv6.
8582          */
8583         u_char tcp_saveipgen[IP6_HDR_LEN];
8584         struct tcphdr tcp_savetcp;
8585         short ostate = 0;
8586
8587 #endif
8588         /* On the hpts and we would have called output */
8589         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8590
8591         /*
8592          * If last ACK falls within this segment's sequence numbers, record
8593          * the timestamp. NOTE that the test is modified according to the
8594          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8595          */
8596         if (bbr->r_ctl.rc_resend != NULL) {
8597                 return (0);
8598         }
8599         if (tiwin && tiwin != tp->snd_wnd) {
8600                 return (0);
8601         }
8602         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
8603                 return (0);
8604         }
8605         if (__predict_false((to->to_flags & TOF_TS) &&
8606             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
8607                 return (0);
8608         }
8609         if (__predict_false((th->th_ack != tp->snd_una))) {
8610                 return (0);
8611         }
8612         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
8613                 return (0);
8614         }
8615         if ((to->to_flags & TOF_TS) != 0 &&
8616             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8617                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
8618                 tp->ts_recent = to->to_tsval;
8619         }
8620         /*
8621          * This is a pure, in-sequence data packet with nothing on the
8622          * reassembly queue and we have enough buffer space to take it.
8623          */
8624         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8625
8626 #ifdef NETFLIX_SB_LIMITS
8627         if (so->so_rcv.sb_shlim) {
8628                 mcnt = m_memcnt(m);
8629                 appended = 0;
8630                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8631                     CFO_NOSLEEP, NULL) == false) {
8632                         counter_u64_add(tcp_sb_shlim_fails, 1);
8633                         m_freem(m);
8634                         return (1);
8635                 }
8636         }
8637 #endif
8638         /* Clean receiver SACK report if present */
8639         if (tp->rcv_numsacks)
8640                 tcp_clean_sackreport(tp);
8641         KMOD_TCPSTAT_INC(tcps_preddat);
8642         tp->rcv_nxt += tlen;
8643         if (tlen &&
8644             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8645             (tp->t_fbyte_in == 0)) {
8646                 tp->t_fbyte_in = ticks;
8647                 if (tp->t_fbyte_in == 0)
8648                         tp->t_fbyte_in = 1;
8649                 if (tp->t_fbyte_out && tp->t_fbyte_in)
8650                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8651         }
8652         /*
8653          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
8654          */
8655         tp->snd_wl1 = th->th_seq;
8656         /*
8657          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
8658          */
8659         tp->rcv_up = tp->rcv_nxt;
8660         KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8661         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8662 #ifdef TCPDEBUG
8663         if (so->so_options & SO_DEBUG)
8664                 tcp_trace(TA_INPUT, ostate, tp,
8665                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
8666 #endif
8667         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
8668
8669         /* Add data to socket buffer. */
8670         SOCKBUF_LOCK(&so->so_rcv);
8671         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8672                 m_freem(m);
8673         } else {
8674                 /*
8675                  * Set new socket buffer size. Give up when limit is
8676                  * reached.
8677                  */
8678                 if (newsize)
8679                         if (!sbreserve_locked(&so->so_rcv,
8680                             newsize, so, NULL))
8681                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
8682                 m_adj(m, drop_hdrlen);  /* delayed header drop */
8683
8684 #ifdef NETFLIX_SB_LIMITS
8685                 appended =
8686 #endif
8687                         sbappendstream_locked(&so->so_rcv, m, 0);
8688                 ctf_calc_rwin(so, tp);
8689         }
8690         /* NB: sorwakeup_locked() does an implicit unlock. */
8691         sorwakeup_locked(so);
8692 #ifdef NETFLIX_SB_LIMITS
8693         if (so->so_rcv.sb_shlim && mcnt != appended)
8694                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
8695 #endif
8696         if (DELAY_ACK(tp, bbr, nsegs)) {
8697                 bbr->bbr_segs_rcvd += max(1, nsegs);
8698                 tp->t_flags |= TF_DELACK;
8699                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8700         } else {
8701                 bbr->r_wanted_output = 1;
8702                 tp->t_flags |= TF_ACKNOW;
8703         }
8704         return (1);
8705 }
8706
8707 /*
8708  * This subfunction is used to try to highly optimize the
8709  * fast path. We again allow window updates that are
8710  * in sequence to remain in the fast-path. We also add
8711  * in the __predict's to attempt to help the compiler.
8712  * Note that if we return a 0, then we can *not* process
8713  * it and the caller should push the packet into the
8714  * slow-path. If we return 1, then all is well and
8715  * the packet is fully processed.
8716  */
8717 static int
8718 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8719     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8720     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
8721 {
8722         int32_t acked;
8723         uint16_t nsegs;
8724         uint32_t sack_changed;
8725 #ifdef TCPDEBUG
8726         /*
8727          * The size of tcp_saveipgen must be the size of the max ip header,
8728          * now IPv6.
8729          */
8730         u_char tcp_saveipgen[IP6_HDR_LEN];
8731         struct tcphdr tcp_savetcp;
8732         short ostate = 0;
8733
8734 #endif
8735         uint32_t prev_acked = 0;
8736         struct tcp_bbr *bbr;
8737
8738         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8739                 /* Old ack, behind (or duplicate to) the last one rcv'd */
8740                 return (0);
8741         }
8742         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
8743                 /* Above what we have sent? */
8744                 return (0);
8745         }
8746         if (__predict_false(tiwin == 0)) {
8747                 /* zero window */
8748                 return (0);
8749         }
8750         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
8751                 /* We need a SYN or a FIN, unlikely.. */
8752                 return (0);
8753         }
8754         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
8755                 /* Timestamp is behind .. old ack with seq wrap? */
8756                 return (0);
8757         }
8758         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
8759                 /* Still recovering */
8760                 return (0);
8761         }
8762         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8763         if (__predict_false(bbr->r_ctl.rc_resend != NULL)) {
8764                 /* We are retransmitting */
8765                 return (0);
8766         }
8767         if (__predict_false(bbr->rc_in_persist != 0)) {
8768                 /* In persist mode */
8769                 return (0);
8770         }
8771         if (bbr->r_ctl.rc_sacked) {
8772                 /* We have sack holes on our scoreboard */
8773                 return (0);
8774         }
8775         /* Ok if we reach here, we can process a fast-ack */
8776         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8777         sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
8778         /*
8779          * We never detect loss in fast ack [we can't
8780          * have a sack and can't be in recovery so
8781          * we always pass 0 (nothing detected)].
8782          */
8783         bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0);
8784         /* Did the window get updated? */
8785         if (tiwin != tp->snd_wnd) {
8786                 tp->snd_wnd = tiwin;
8787                 tp->snd_wl1 = th->th_seq;
8788                 if (tp->snd_wnd > tp->max_sndwnd)
8789                         tp->max_sndwnd = tp->snd_wnd;
8790         }
8791         /* Do we need to exit persists? */
8792         if ((bbr->rc_in_persist != 0) &&
8793             (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8794                                bbr_minseg(bbr)))) {
8795                 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8796                 bbr->r_wanted_output = 1;
8797         }
8798         /* Do we need to enter persists? */
8799         if ((bbr->rc_in_persist == 0) &&
8800             (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8801             TCPS_HAVEESTABLISHED(tp->t_state) &&
8802             (tp->snd_max == tp->snd_una) &&
8803             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8804             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8805                 /* No send window.. we must enter persist */
8806                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8807         }
8808         /*
8809          * If last ACK falls within this segment's sequence numbers, record
8810          * the timestamp. NOTE that the test is modified according to the
8811          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8812          */
8813         if ((to->to_flags & TOF_TS) != 0 &&
8814             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8815                 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime;
8816                 tp->ts_recent = to->to_tsval;
8817         }
8818         /*
8819          * This is a pure ack for outstanding data.
8820          */
8821         KMOD_TCPSTAT_INC(tcps_predack);
8822
8823         /*
8824          * "bad retransmit" recovery.
8825          */
8826         if (tp->t_flags & TF_PREVVALID) {
8827                 tp->t_flags &= ~TF_PREVVALID;
8828                 if (tp->t_rxtshift == 1 &&
8829                     (int)(ticks - tp->t_badrxtwin) < 0)
8830                         bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
8831         }
8832         /*
8833          * Recalculate the transmit timer / rtt.
8834          *
8835          * Some boxes send broken timestamp replies during the SYN+ACK
8836          * phase, ignore timestamps of 0 or we could calculate a huge RTT
8837          * and blow up the retransmit timer.
8838          */
8839         acked = BYTES_THIS_ACK(tp, th);
8840
8841 #ifdef TCP_HHOOK
8842         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
8843         hhook_run_tcp_est_in(tp, th, to);
8844 #endif
8845
8846         KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
8847         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
8848         sbdrop(&so->so_snd, acked);
8849
8850         if (SEQ_GT(th->th_ack, tp->snd_una))
8851                 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
8852         tp->snd_una = th->th_ack;
8853         if (tp->snd_wnd < ctf_outstanding(tp))
8854                 /* The peer collapsed its window on us */
8855                 bbr_collapsed_window(bbr);
8856         else if (bbr->rc_has_collapsed)
8857                 bbr_un_collapse_window(bbr);
8858
8859         if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
8860                 tp->snd_recover = tp->snd_una;
8861         }
8862         bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0);
8863         /*
8864          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
8865          */
8866         tp->snd_wl2 = th->th_ack;
8867         m_freem(m);
8868         /*
8869          * If all outstanding data are acked, stop retransmit timer,
8870          * otherwise restart timer using current (possibly backed-off)
8871          * value. If process is waiting for space, wakeup/selwakeup/signal.
8872          * If data are ready to send, let tcp_output decide between more
8873          * output or persist.
8874          */
8875 #ifdef TCPDEBUG
8876         if (so->so_options & SO_DEBUG)
8877                 tcp_trace(TA_INPUT, ostate, tp,
8878                     (void *)tcp_saveipgen,
8879                     &tcp_savetcp, 0);
8880 #endif
8881         /* Wake up the socket if we have room to write more */
8882         sowwakeup(so);
8883         if (tp->snd_una == tp->snd_max) {
8884                 /* Nothing left outstanding */
8885                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
8886                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
8887                         bbr->rc_tp->t_acktime = 0;
8888                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8889                 if (bbr->rc_in_persist == 0) {
8890                         bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
8891                 }
8892                 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
8893                 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
8894                 /*
8895                  * We invalidate the last ack here since we
8896                  * don't want to transfer forward the time
8897                  * for our sum's calculations.
8898                  */
8899                 bbr->r_wanted_output = 1;
8900         }
8901         if (sbavail(&so->so_snd)) {
8902                 bbr->r_wanted_output = 1;
8903         }
8904         return (1);
8905 }
8906
8907 /*
8908  * Return value of 1, the TCB is unlocked and most
8909  * likely gone, return value of 0, the TCB is still
8910  * locked.
8911  */
8912 static int
8913 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
8914     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8915     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
8916 {
8917         int32_t todrop;
8918         int32_t ourfinisacked = 0;
8919         struct tcp_bbr *bbr;
8920         int32_t ret_val = 0;
8921
8922         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8923         ctf_calc_rwin(so, tp);
8924         /*
8925          * If the state is SYN_SENT: if seg contains an ACK, but not for our
8926          * SYN, drop the input. if seg contains a RST, then drop the
8927          * connection. if seg does not contain SYN, then drop it. Otherwise
8928          * this is an acceptable SYN segment initialize tp->rcv_nxt and
8929          * tp->irs if seg contains ack then advance tp->snd_una. BRR does
8930          * not support ECN so we will not say we are capable. if SYN has
8931          * been acked change to ESTABLISHED else SYN_RCVD state arrange for
8932          * segment to be acked (eventually) continue processing rest of
8933          * data/controls, beginning with URG
8934          */
8935         if ((thflags & TH_ACK) &&
8936             (SEQ_LEQ(th->th_ack, tp->iss) ||
8937             SEQ_GT(th->th_ack, tp->snd_max))) {
8938                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8939                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
8940                 return (1);
8941         }
8942         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
8943                 TCP_PROBE5(connect__refused, NULL, tp,
8944                     mtod(m, const char *), tp, th);
8945                 tp = tcp_drop(tp, ECONNREFUSED);
8946                 ctf_do_drop(m, tp);
8947                 return (1);
8948         }
8949         if (thflags & TH_RST) {
8950                 ctf_do_drop(m, tp);
8951                 return (1);
8952         }
8953         if (!(thflags & TH_SYN)) {
8954                 ctf_do_drop(m, tp);
8955                 return (1);
8956         }
8957         tp->irs = th->th_seq;
8958         tcp_rcvseqinit(tp);
8959         if (thflags & TH_ACK) {
8960                 int tfo_partial = 0;
8961
8962                 KMOD_TCPSTAT_INC(tcps_connects);
8963                 soisconnected(so);
8964 #ifdef MAC
8965                 mac_socketpeer_set_from_mbuf(m, so);
8966 #endif
8967                 /* Do window scaling on this connection? */
8968                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
8969                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
8970                         tp->rcv_scale = tp->request_r_scale;
8971                 }
8972                 tp->rcv_adv += min(tp->rcv_wnd,
8973                     TCP_MAXWIN << tp->rcv_scale);
8974                 /*
8975                  * If not all the data that was sent in the TFO SYN
8976                  * has been acked, resend the remainder right away.
8977                  */
8978                 if (IS_FASTOPEN(tp->t_flags) &&
8979                     (tp->snd_una != tp->snd_max)) {
8980                         tp->snd_nxt = th->th_ack;
8981                         tfo_partial = 1;
8982                 }
8983                 /*
8984                  * If there's data, delay ACK; if there's also a FIN ACKNOW
8985                  * will be turned on later.
8986                  */
8987                 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) {
8988                         bbr->bbr_segs_rcvd += 1;
8989                         tp->t_flags |= TF_DELACK;
8990                         bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8991                 } else {
8992                         bbr->r_wanted_output = 1;
8993                         tp->t_flags |= TF_ACKNOW;
8994                 }
8995                 if (SEQ_GT(th->th_ack, tp->iss)) {
8996                         /*
8997                          * The SYN is acked
8998                          * handle it specially.
8999                          */
9000                         bbr_log_syn(tp, to);
9001                 }
9002                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
9003                         /*
9004                          * We advance snd_una for the
9005                          * fast open case. If th_ack is
9006                          * acknowledging data beyond
9007                          * snd_una we can't just call
9008                          * ack-processing since the
9009                          * data stream in our send-map
9010                          * will start at snd_una + 1 (one
9011                          * beyond the SYN). If its just
9012                          * equal we don't need to do that
9013                          * and there is no send_map.
9014                          */
9015                         tp->snd_una++;
9016                 }
9017                 /*
9018                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
9019                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
9020                  */
9021                 tp->t_starttime = ticks;
9022                 if (tp->t_flags & TF_NEEDFIN) {
9023                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
9024                         tp->t_flags &= ~TF_NEEDFIN;
9025                         thflags &= ~TH_SYN;
9026                 } else {
9027                         tcp_state_change(tp, TCPS_ESTABLISHED);
9028                         TCP_PROBE5(connect__established, NULL, tp,
9029                             mtod(m, const char *), tp, th);
9030                         cc_conn_init(tp);
9031                 }
9032         } else {
9033                 /*
9034                  * Received initial SYN in SYN-SENT[*] state => simultaneous
9035                  * open.  If segment contains CC option and there is a
9036                  * cached CC, apply TAO test. If it succeeds, connection is *
9037                  * half-synchronized. Otherwise, do 3-way handshake:
9038                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
9039                  * there was no CC option, clear cached CC value.
9040                  */
9041                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
9042                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
9043         }
9044         INP_WLOCK_ASSERT(tp->t_inpcb);
9045         /*
9046          * Advance th->th_seq to correspond to first data byte. If data,
9047          * trim to stay within window, dropping FIN if necessary.
9048          */
9049         th->th_seq++;
9050         if (tlen > tp->rcv_wnd) {
9051                 todrop = tlen - tp->rcv_wnd;
9052                 m_adj(m, -todrop);
9053                 tlen = tp->rcv_wnd;
9054                 thflags &= ~TH_FIN;
9055                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
9056                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
9057         }
9058         tp->snd_wl1 = th->th_seq - 1;
9059         tp->rcv_up = th->th_seq;
9060         /*
9061          * Client side of transaction: already sent SYN and data. If the
9062          * remote host used T/TCP to validate the SYN, our data will be
9063          * ACK'd; if so, enter normal data segment processing in the middle
9064          * of step 5, ack processing. Otherwise, goto step 6.
9065          */
9066         if (thflags & TH_ACK) {
9067                 if ((to->to_flags & TOF_TS) != 0) {
9068                         uint32_t t, rtt;
9069
9070                         t = tcp_tv_to_mssectick(&bbr->rc_tv);
9071                         if (TSTMP_GEQ(t, to->to_tsecr)) {
9072                                 rtt = t - to->to_tsecr;
9073                                 if (rtt == 0) {
9074                                         rtt = 1;
9075                                 }
9076                                 rtt *= MS_IN_USEC;
9077                                 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9078                                 apply_filter_min_small(&bbr->r_ctl.rc_rttprop,
9079                                                        rtt, bbr->r_ctl.rc_rcvtime);
9080                         }
9081                 }
9082                 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
9083                         return (ret_val);
9084                 /* We may have changed to FIN_WAIT_1 above */
9085                 if (tp->t_state == TCPS_FIN_WAIT_1) {
9086                         /*
9087                          * In FIN_WAIT_1 STATE in addition to the processing
9088                          * for the ESTABLISHED state if our FIN is now
9089                          * acknowledged then enter FIN_WAIT_2.
9090                          */
9091                         if (ourfinisacked) {
9092                                 /*
9093                                  * If we can't receive any more data, then
9094                                  * closing user can proceed. Starting the
9095                                  * timer is contrary to the specification,
9096                                  * but if we don't get a FIN we'll hang
9097                                  * forever.
9098                                  *
9099                                  * XXXjl: we should release the tp also, and
9100                                  * use a compressed state.
9101                                  */
9102                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9103                                         soisdisconnected(so);
9104                                         tcp_timer_activate(tp, TT_2MSL,
9105                                             (tcp_fast_finwait2_recycle ?
9106                                             tcp_finwait2_timeout :
9107                                             TP_MAXIDLE(tp)));
9108                                 }
9109                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
9110                         }
9111                 }
9112         }
9113         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9114             tiwin, thflags, nxt_pkt));
9115 }
9116
9117 /*
9118  * Return value of 1, the TCB is unlocked and most
9119  * likely gone, return value of 0, the TCB is still
9120  * locked.
9121  */
9122 static int
9123 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
9124                 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9125                 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9126 {
9127         int32_t ourfinisacked = 0;
9128         int32_t ret_val;
9129         struct tcp_bbr *bbr;
9130
9131         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9132         ctf_calc_rwin(so, tp);
9133         if ((thflags & TH_ACK) &&
9134             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
9135              SEQ_GT(th->th_ack, tp->snd_max))) {
9136                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9137                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9138                 return (1);
9139         }
9140         if (IS_FASTOPEN(tp->t_flags)) {
9141                 /*
9142                  * When a TFO connection is in SYN_RECEIVED, the only valid
9143                  * packets are the initial SYN, a retransmit/copy of the
9144                  * initial SYN (possibly with a subset of the original
9145                  * data), a valid ACK, a FIN, or a RST.
9146                  */
9147                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
9148                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9149                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9150                         return (1);
9151                 } else if (thflags & TH_SYN) {
9152                         /* non-initial SYN is ignored */
9153                         if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
9154                             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
9155                             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
9156                                 ctf_do_drop(m, NULL);
9157                                 return (0);
9158                         }
9159                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
9160                         ctf_do_drop(m, NULL);
9161                         return (0);
9162                 }
9163         }
9164         if ((thflags & TH_RST) ||
9165             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9166                 return (ctf_process_rst(m, th, so, tp));
9167         /*
9168          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9169          * it's less than ts_recent, drop it.
9170          */
9171         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9172             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9173                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9174                         return (ret_val);
9175         }
9176         /*
9177          * In the SYN-RECEIVED state, validate that the packet belongs to
9178          * this connection before trimming the data to fit the receive
9179          * window.  Check the sequence number versus IRS since we know the
9180          * sequence numbers haven't wrapped.  This is a partial fix for the
9181          * "LAND" DoS attack.
9182          */
9183         if (SEQ_LT(th->th_seq, tp->irs)) {
9184                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9185                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9186                 return (1);
9187         }
9188         INP_WLOCK_ASSERT(tp->t_inpcb);
9189         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9190                 return (ret_val);
9191         }
9192         /*
9193          * If last ACK falls within this segment's sequence numbers, record
9194          * its timestamp. NOTE: 1) That the test incorporates suggestions
9195          * from the latest proposal of the tcplw@cray.com list (Braden
9196          * 1993/04/26). 2) That updating only on newer timestamps interferes
9197          * with our earlier PAWS tests, so this check should be solely
9198          * predicated on the sequence space of this segment. 3) That we
9199          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9200          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9201          * SEG.Len, This modified check allows us to overcome RFC1323's
9202          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9203          * p.869. In such cases, we can still calculate the RTT correctly
9204          * when RCV.NXT == Last.ACK.Sent.
9205          */
9206         if ((to->to_flags & TOF_TS) != 0 &&
9207             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9208             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9209                     ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9210                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9211                 tp->ts_recent = to->to_tsval;
9212         }
9213         tp->snd_wnd = tiwin;
9214         /*
9215          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9216          * is on (half-synchronized state), then queue data for later
9217          * processing; else drop segment and return.
9218          */
9219         if ((thflags & TH_ACK) == 0) {
9220                 if (IS_FASTOPEN(tp->t_flags)) {
9221                         cc_conn_init(tp);
9222                 }
9223                 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9224                                          tiwin, thflags, nxt_pkt));
9225         }
9226         KMOD_TCPSTAT_INC(tcps_connects);
9227         soisconnected(so);
9228         /* Do window scaling? */
9229         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9230             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9231                 tp->rcv_scale = tp->request_r_scale;
9232         }
9233         /*
9234          * ok for the first time in lets see if we can use the ts to figure
9235          * out what the initial RTT was.
9236          */
9237         if ((to->to_flags & TOF_TS) != 0) {
9238                 uint32_t t, rtt;
9239
9240                 t = tcp_tv_to_mssectick(&bbr->rc_tv);
9241                 if (TSTMP_GEQ(t, to->to_tsecr)) {
9242                         rtt = t - to->to_tsecr;
9243                         if (rtt == 0) {
9244                                 rtt = 1;
9245                         }
9246                         rtt *= MS_IN_USEC;
9247                         tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9248                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime);
9249                 }
9250         }
9251         /* Drop off any SYN in the send map (probably not there)  */
9252         if (thflags & TH_ACK)
9253                 bbr_log_syn(tp, to);
9254         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
9255
9256                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
9257                 tp->t_tfo_pending = NULL;
9258         }
9259         /*
9260          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
9261          * FIN-WAIT-1
9262          */
9263         tp->t_starttime = ticks;
9264         if (tp->t_flags & TF_NEEDFIN) {
9265                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
9266                 tp->t_flags &= ~TF_NEEDFIN;
9267         } else {
9268                 tcp_state_change(tp, TCPS_ESTABLISHED);
9269                 TCP_PROBE5(accept__established, NULL, tp,
9270                            mtod(m, const char *), tp, th);
9271                 /*
9272                  * TFO connections call cc_conn_init() during SYN
9273                  * processing.  Calling it again here for such connections
9274                  * is not harmless as it would undo the snd_cwnd reduction
9275                  * that occurs when a TFO SYN|ACK is retransmitted.
9276                  */
9277                 if (!IS_FASTOPEN(tp->t_flags))
9278                         cc_conn_init(tp);
9279         }
9280         /*
9281          * Account for the ACK of our SYN prior to
9282          * regular ACK processing below, except for
9283          * simultaneous SYN, which is handled later.
9284          */
9285         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
9286                 tp->snd_una++;
9287         /*
9288          * If segment contains data or ACK, will call tcp_reass() later; if
9289          * not, do so now to pass queued data to user.
9290          */
9291         if (tlen == 0 && (thflags & TH_FIN) == 0)
9292                 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
9293                         (struct mbuf *)0);
9294         tp->snd_wl1 = th->th_seq - 1;
9295         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9296                 return (ret_val);
9297         }
9298         if (tp->t_state == TCPS_FIN_WAIT_1) {
9299                 /* We could have went to FIN_WAIT_1 (or EST) above */
9300                 /*
9301                  * In FIN_WAIT_1 STATE in addition to the processing for the
9302                  * ESTABLISHED state if our FIN is now acknowledged then
9303                  * enter FIN_WAIT_2.
9304                  */
9305                 if (ourfinisacked) {
9306                         /*
9307                          * If we can't receive any more data, then closing
9308                          * user can proceed. Starting the timer is contrary
9309                          * to the specification, but if we don't get a FIN
9310                          * we'll hang forever.
9311                          *
9312                          * XXXjl: we should release the tp also, and use a
9313                          * compressed state.
9314                          */
9315                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9316                                 soisdisconnected(so);
9317                                 tcp_timer_activate(tp, TT_2MSL,
9318                                                    (tcp_fast_finwait2_recycle ?
9319                                                     tcp_finwait2_timeout :
9320                                                     TP_MAXIDLE(tp)));
9321                         }
9322                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
9323                 }
9324         }
9325         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9326                                  tiwin, thflags, nxt_pkt));
9327 }
9328
9329 /*
9330  * Return value of 1, the TCB is unlocked and most
9331  * likely gone, return value of 0, the TCB is still
9332  * locked.
9333  */
9334 static int
9335 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
9336     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9337     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9338 {
9339         struct tcp_bbr *bbr;
9340         int32_t ret_val;
9341
9342         /*
9343          * Header prediction: check for the two common cases of a
9344          * uni-directional data xfer.  If the packet has no control flags,
9345          * is in-sequence, the window didn't change and we're not
9346          * retransmitting, it's a candidate.  If the length is zero and the
9347          * ack moved forward, we're the sender side of the xfer.  Just free
9348          * the data acked & wake any higher level process that was blocked
9349          * waiting for space.  If the length is non-zero and the ack didn't
9350          * move, we're the receiver side.  If we're getting packets in-order
9351          * (the reassembly queue is empty), add the data toc The socket
9352          * buffer and note that we need a delayed ack. Make sure that the
9353          * hidden state-flags are also off. Since we check for
9354          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
9355          */
9356         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9357         if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) {
9358                 /*
9359                  * If we have delived under 4 segments increase the initial
9360                  * window if raised by the peer. We use this to determine
9361                  * dynamic and static rwnd's at the end of a connection.
9362                  */
9363                 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd);
9364         }
9365         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
9366             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) &&
9367             __predict_true(SEGQ_EMPTY(tp)) &&
9368             __predict_true(th->th_seq == tp->rcv_nxt)) {
9369                 if (tlen == 0) {
9370                         if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
9371                             tiwin, nxt_pkt, iptos)) {
9372                                 return (0);
9373                         }
9374                 } else {
9375                         if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
9376                             tiwin, nxt_pkt)) {
9377                                 return (0);
9378                         }
9379                 }
9380         }
9381         ctf_calc_rwin(so, tp);
9382
9383         if ((thflags & TH_RST) ||
9384             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9385                 return (ctf_process_rst(m, th, so, tp));
9386         /*
9387          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9388          * synchronized state.
9389          */
9390         if (thflags & TH_SYN) {
9391                 ctf_challenge_ack(m, th, tp, &ret_val);
9392                 return (ret_val);
9393         }
9394         /*
9395          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9396          * it's less than ts_recent, drop it.
9397          */
9398         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9399             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9400                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9401                         return (ret_val);
9402         }
9403         INP_WLOCK_ASSERT(tp->t_inpcb);
9404         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9405                 return (ret_val);
9406         }
9407         /*
9408          * If last ACK falls within this segment's sequence numbers, record
9409          * its timestamp. NOTE: 1) That the test incorporates suggestions
9410          * from the latest proposal of the tcplw@cray.com list (Braden
9411          * 1993/04/26). 2) That updating only on newer timestamps interferes
9412          * with our earlier PAWS tests, so this check should be solely
9413          * predicated on the sequence space of this segment. 3) That we
9414          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9415          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9416          * SEG.Len, This modified check allows us to overcome RFC1323's
9417          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9418          * p.869. In such cases, we can still calculate the RTT correctly
9419          * when RCV.NXT == Last.ACK.Sent.
9420          */
9421         if ((to->to_flags & TOF_TS) != 0 &&
9422             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9423             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9424             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9425                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9426                 tp->ts_recent = to->to_tsval;
9427         }
9428         /*
9429          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9430          * is on (half-synchronized state), then queue data for later
9431          * processing; else drop segment and return.
9432          */
9433         if ((thflags & TH_ACK) == 0) {
9434                 if (tp->t_flags & TF_NEEDSYN) {
9435                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9436                             tiwin, thflags, nxt_pkt));
9437                 } else if (tp->t_flags & TF_ACKNOW) {
9438                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9439                         bbr->r_wanted_output = 1;
9440                         return (ret_val);
9441                 } else {
9442                         ctf_do_drop(m, NULL);
9443                         return (0);
9444                 }
9445         }
9446         /*
9447          * Ack processing.
9448          */
9449         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9450                 return (ret_val);
9451         }
9452         if (sbavail(&so->so_snd)) {
9453                 if (ctf_progress_timeout_check(tp, true)) {
9454                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9455                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9456                         return (1);
9457                 }
9458         }
9459         /* State changes only happen in bbr_process_data() */
9460         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9461             tiwin, thflags, nxt_pkt));
9462 }
9463
9464 /*
9465  * Return value of 1, the TCB is unlocked and most
9466  * likely gone, return value of 0, the TCB is still
9467  * locked.
9468  */
9469 static int
9470 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
9471     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9472     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9473 {
9474         struct tcp_bbr *bbr;
9475         int32_t ret_val;
9476
9477         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9478         ctf_calc_rwin(so, tp);
9479         if ((thflags & TH_RST) ||
9480             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9481                 return (ctf_process_rst(m, th, so, tp));
9482         /*
9483          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9484          * synchronized state.
9485          */
9486         if (thflags & TH_SYN) {
9487                 ctf_challenge_ack(m, th, tp, &ret_val);
9488                 return (ret_val);
9489         }
9490         /*
9491          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9492          * it's less than ts_recent, drop it.
9493          */
9494         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9495             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9496                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9497                         return (ret_val);
9498         }
9499         INP_WLOCK_ASSERT(tp->t_inpcb);
9500         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9501                 return (ret_val);
9502         }
9503         /*
9504          * If last ACK falls within this segment's sequence numbers, record
9505          * its timestamp. NOTE: 1) That the test incorporates suggestions
9506          * from the latest proposal of the tcplw@cray.com list (Braden
9507          * 1993/04/26). 2) That updating only on newer timestamps interferes
9508          * with our earlier PAWS tests, so this check should be solely
9509          * predicated on the sequence space of this segment. 3) That we
9510          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9511          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9512          * SEG.Len, This modified check allows us to overcome RFC1323's
9513          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9514          * p.869. In such cases, we can still calculate the RTT correctly
9515          * when RCV.NXT == Last.ACK.Sent.
9516          */
9517         if ((to->to_flags & TOF_TS) != 0 &&
9518             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9519             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9520             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9521                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9522                 tp->ts_recent = to->to_tsval;
9523         }
9524         /*
9525          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9526          * is on (half-synchronized state), then queue data for later
9527          * processing; else drop segment and return.
9528          */
9529         if ((thflags & TH_ACK) == 0) {
9530                 if (tp->t_flags & TF_NEEDSYN) {
9531                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9532                             tiwin, thflags, nxt_pkt));
9533                 } else if (tp->t_flags & TF_ACKNOW) {
9534                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9535                         bbr->r_wanted_output = 1;
9536                         return (ret_val);
9537                 } else {
9538                         ctf_do_drop(m, NULL);
9539                         return (0);
9540                 }
9541         }
9542         /*
9543          * Ack processing.
9544          */
9545         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9546                 return (ret_val);
9547         }
9548         if (sbavail(&so->so_snd)) {
9549                 if (ctf_progress_timeout_check(tp, true)) {
9550                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9551                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9552                         return (1);
9553                 }
9554         }
9555         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9556             tiwin, thflags, nxt_pkt));
9557 }
9558
9559 static int
9560 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr,
9561     struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so)
9562 {
9563
9564         if (bbr->rc_allow_data_af_clo == 0) {
9565 close_now:
9566                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9567                 /* tcp_close will kill the inp pre-log the Reset */
9568                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9569                 tp = tcp_close(tp);
9570                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
9571                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
9572                 return (1);
9573         }
9574         if (sbavail(&so->so_snd) == 0)
9575                 goto close_now;
9576         /* Ok we allow data that is ignored and a followup reset */
9577         tp->rcv_nxt = th->th_seq + *tlen;
9578         tp->t_flags2 |= TF2_DROP_AF_DATA;
9579         bbr->r_wanted_output = 1;
9580         *tlen = 0;
9581         return (0);
9582 }
9583
9584 /*
9585  * Return value of 1, the TCB is unlocked and most
9586  * likely gone, return value of 0, the TCB is still
9587  * locked.
9588  */
9589 static int
9590 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
9591     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9592     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9593 {
9594         int32_t ourfinisacked = 0;
9595         int32_t ret_val;
9596         struct tcp_bbr *bbr;
9597
9598         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9599         ctf_calc_rwin(so, tp);
9600         if ((thflags & TH_RST) ||
9601             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9602                 return (ctf_process_rst(m, th, so, tp));
9603         /*
9604          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9605          * synchronized state.
9606          */
9607         if (thflags & TH_SYN) {
9608                 ctf_challenge_ack(m, th, tp, &ret_val);
9609                 return (ret_val);
9610         }
9611         /*
9612          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9613          * it's less than ts_recent, drop it.
9614          */
9615         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9616             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9617                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9618                         return (ret_val);
9619         }
9620         INP_WLOCK_ASSERT(tp->t_inpcb);
9621         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9622                 return (ret_val);
9623         }
9624         /*
9625          * If new data are received on a connection after the user processes
9626          * are gone, then RST the other end.
9627          */
9628         if ((so->so_state & SS_NOFDREF) && tlen) {
9629                 /*
9630                  * We call a new function now so we might continue and setup
9631                  * to reset at all data being ack'd.
9632                  */
9633                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9634                         return (1);
9635         }
9636         /*
9637          * If last ACK falls within this segment's sequence numbers, record
9638          * its timestamp. NOTE: 1) That the test incorporates suggestions
9639          * from the latest proposal of the tcplw@cray.com list (Braden
9640          * 1993/04/26). 2) That updating only on newer timestamps interferes
9641          * with our earlier PAWS tests, so this check should be solely
9642          * predicated on the sequence space of this segment. 3) That we
9643          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9644          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9645          * SEG.Len, This modified check allows us to overcome RFC1323's
9646          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9647          * p.869. In such cases, we can still calculate the RTT correctly
9648          * when RCV.NXT == Last.ACK.Sent.
9649          */
9650         if ((to->to_flags & TOF_TS) != 0 &&
9651             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9652             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9653             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9654                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9655                 tp->ts_recent = to->to_tsval;
9656         }
9657         /*
9658          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9659          * is on (half-synchronized state), then queue data for later
9660          * processing; else drop segment and return.
9661          */
9662         if ((thflags & TH_ACK) == 0) {
9663                 if (tp->t_flags & TF_NEEDSYN) {
9664                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9665                             tiwin, thflags, nxt_pkt));
9666                 } else if (tp->t_flags & TF_ACKNOW) {
9667                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9668                         bbr->r_wanted_output = 1;
9669                         return (ret_val);
9670                 } else {
9671                         ctf_do_drop(m, NULL);
9672                         return (0);
9673                 }
9674         }
9675         /*
9676          * Ack processing.
9677          */
9678         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9679                 return (ret_val);
9680         }
9681         if (ourfinisacked) {
9682                 /*
9683                  * If we can't receive any more data, then closing user can
9684                  * proceed. Starting the timer is contrary to the
9685                  * specification, but if we don't get a FIN we'll hang
9686                  * forever.
9687                  *
9688                  * XXXjl: we should release the tp also, and use a
9689                  * compressed state.
9690                  */
9691                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9692                         soisdisconnected(so);
9693                         tcp_timer_activate(tp, TT_2MSL,
9694                             (tcp_fast_finwait2_recycle ?
9695                             tcp_finwait2_timeout :
9696                             TP_MAXIDLE(tp)));
9697                 }
9698                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
9699         }
9700         if (sbavail(&so->so_snd)) {
9701                 if (ctf_progress_timeout_check(tp, true)) {
9702                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9703                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9704                         return (1);
9705                 }
9706         }
9707         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9708             tiwin, thflags, nxt_pkt));
9709 }
9710
9711 /*
9712  * Return value of 1, the TCB is unlocked and most
9713  * likely gone, return value of 0, the TCB is still
9714  * locked.
9715  */
9716 static int
9717 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
9718     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9719     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9720 {
9721         int32_t ourfinisacked = 0;
9722         int32_t ret_val;
9723         struct tcp_bbr *bbr;
9724
9725         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9726         ctf_calc_rwin(so, tp);
9727         if ((thflags & TH_RST) ||
9728             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9729                 return (ctf_process_rst(m, th, so, tp));
9730         /*
9731          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9732          * synchronized state.
9733          */
9734         if (thflags & TH_SYN) {
9735                 ctf_challenge_ack(m, th, tp, &ret_val);
9736                 return (ret_val);
9737         }
9738         /*
9739          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9740          * it's less than ts_recent, drop it.
9741          */
9742         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9743             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9744                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9745                         return (ret_val);
9746         }
9747         INP_WLOCK_ASSERT(tp->t_inpcb);
9748         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9749                 return (ret_val);
9750         }
9751         /*
9752          * If new data are received on a connection after the user processes
9753          * are gone, then RST the other end.
9754          */
9755         if ((so->so_state & SS_NOFDREF) && tlen) {
9756                 /*
9757                  * We call a new function now so we might continue and setup
9758                  * to reset at all data being ack'd.
9759                  */
9760                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9761                         return (1);
9762         }
9763         /*
9764          * If last ACK falls within this segment's sequence numbers, record
9765          * its timestamp. NOTE: 1) That the test incorporates suggestions
9766          * from the latest proposal of the tcplw@cray.com list (Braden
9767          * 1993/04/26). 2) That updating only on newer timestamps interferes
9768          * with our earlier PAWS tests, so this check should be solely
9769          * predicated on the sequence space of this segment. 3) That we
9770          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9771          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9772          * SEG.Len, This modified check allows us to overcome RFC1323's
9773          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9774          * p.869. In such cases, we can still calculate the RTT correctly
9775          * when RCV.NXT == Last.ACK.Sent.
9776          */
9777         if ((to->to_flags & TOF_TS) != 0 &&
9778             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9779             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9780             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9781                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9782                 tp->ts_recent = to->to_tsval;
9783         }
9784         /*
9785          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9786          * is on (half-synchronized state), then queue data for later
9787          * processing; else drop segment and return.
9788          */
9789         if ((thflags & TH_ACK) == 0) {
9790                 if (tp->t_flags & TF_NEEDSYN) {
9791                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9792                             tiwin, thflags, nxt_pkt));
9793                 } else if (tp->t_flags & TF_ACKNOW) {
9794                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9795                         bbr->r_wanted_output = 1;
9796                         return (ret_val);
9797                 } else {
9798                         ctf_do_drop(m, NULL);
9799                         return (0);
9800                 }
9801         }
9802         /*
9803          * Ack processing.
9804          */
9805         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9806                 return (ret_val);
9807         }
9808         if (ourfinisacked) {
9809                 tcp_twstart(tp);
9810                 m_freem(m);
9811                 return (1);
9812         }
9813         if (sbavail(&so->so_snd)) {
9814                 if (ctf_progress_timeout_check(tp, true)) {
9815                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9816                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9817                         return (1);
9818                 }
9819         }
9820         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9821             tiwin, thflags, nxt_pkt));
9822 }
9823
9824 /*
9825  * Return value of 1, the TCB is unlocked and most
9826  * likely gone, return value of 0, the TCB is still
9827  * locked.
9828  */
9829 static int
9830 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9831     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9832     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9833 {
9834         int32_t ourfinisacked = 0;
9835         int32_t ret_val;
9836         struct tcp_bbr *bbr;
9837
9838         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9839         ctf_calc_rwin(so, tp);
9840         if ((thflags & TH_RST) ||
9841             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9842                 return (ctf_process_rst(m, th, so, tp));
9843         /*
9844          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9845          * synchronized state.
9846          */
9847         if (thflags & TH_SYN) {
9848                 ctf_challenge_ack(m, th, tp, &ret_val);
9849                 return (ret_val);
9850         }
9851         /*
9852          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9853          * it's less than ts_recent, drop it.
9854          */
9855         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9856             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9857                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9858                         return (ret_val);
9859         }
9860         INP_WLOCK_ASSERT(tp->t_inpcb);
9861         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9862                 return (ret_val);
9863         }
9864         /*
9865          * If new data are received on a connection after the user processes
9866          * are gone, then RST the other end.
9867          */
9868         if ((so->so_state & SS_NOFDREF) && tlen) {
9869                 /*
9870                  * We call a new function now so we might continue and setup
9871                  * to reset at all data being ack'd.
9872                  */
9873                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9874                         return (1);
9875         }
9876         /*
9877          * If last ACK falls within this segment's sequence numbers, record
9878          * its timestamp. NOTE: 1) That the test incorporates suggestions
9879          * from the latest proposal of the tcplw@cray.com list (Braden
9880          * 1993/04/26). 2) That updating only on newer timestamps interferes
9881          * with our earlier PAWS tests, so this check should be solely
9882          * predicated on the sequence space of this segment. 3) That we
9883          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9884          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9885          * SEG.Len, This modified check allows us to overcome RFC1323's
9886          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9887          * p.869. In such cases, we can still calculate the RTT correctly
9888          * when RCV.NXT == Last.ACK.Sent.
9889          */
9890         if ((to->to_flags & TOF_TS) != 0 &&
9891             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9892             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9893             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9894                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9895                 tp->ts_recent = to->to_tsval;
9896         }
9897         /*
9898          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9899          * is on (half-synchronized state), then queue data for later
9900          * processing; else drop segment and return.
9901          */
9902         if ((thflags & TH_ACK) == 0) {
9903                 if (tp->t_flags & TF_NEEDSYN) {
9904                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9905                             tiwin, thflags, nxt_pkt));
9906                 } else if (tp->t_flags & TF_ACKNOW) {
9907                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9908                         bbr->r_wanted_output = 1;
9909                         return (ret_val);
9910                 } else {
9911                         ctf_do_drop(m, NULL);
9912                         return (0);
9913                 }
9914         }
9915         /*
9916          * case TCPS_LAST_ACK: Ack processing.
9917          */
9918         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9919                 return (ret_val);
9920         }
9921         if (ourfinisacked) {
9922                 tp = tcp_close(tp);
9923                 ctf_do_drop(m, tp);
9924                 return (1);
9925         }
9926         if (sbavail(&so->so_snd)) {
9927                 if (ctf_progress_timeout_check(tp, true)) {
9928                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9929                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9930                         return (1);
9931                 }
9932         }
9933         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9934             tiwin, thflags, nxt_pkt));
9935 }
9936
9937
9938 /*
9939  * Return value of 1, the TCB is unlocked and most
9940  * likely gone, return value of 0, the TCB is still
9941  * locked.
9942  */
9943 static int
9944 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
9945     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9946     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9947 {
9948         int32_t ourfinisacked = 0;
9949         int32_t ret_val;
9950         struct tcp_bbr *bbr;
9951
9952         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9953         ctf_calc_rwin(so, tp);
9954         /* Reset receive buffer auto scaling when not in bulk receive mode. */
9955         if ((thflags & TH_RST) ||
9956             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9957                 return (ctf_process_rst(m, th, so, tp));
9958
9959         /*
9960          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9961          * synchronized state.
9962          */
9963         if (thflags & TH_SYN) {
9964                 ctf_challenge_ack(m, th, tp, &ret_val);
9965                 return (ret_val);
9966         }
9967         INP_WLOCK_ASSERT(tp->t_inpcb);
9968         /*
9969          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9970          * it's less than ts_recent, drop it.
9971          */
9972         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9973             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9974                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9975                         return (ret_val);
9976         }
9977         INP_WLOCK_ASSERT(tp->t_inpcb);
9978         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9979                 return (ret_val);
9980         }
9981         /*
9982          * If new data are received on a connection after the user processes
9983          * are gone, then we may RST the other end depending on the outcome
9984          * of bbr_check_data_after_close.
9985          */
9986         if ((so->so_state & SS_NOFDREF) &&
9987             tlen) {
9988                 /*
9989                  * We call a new function now so we might continue and setup
9990                  * to reset at all data being ack'd.
9991                  */
9992                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9993                         return (1);
9994         }
9995         INP_WLOCK_ASSERT(tp->t_inpcb);
9996         /*
9997          * If last ACK falls within this segment's sequence numbers, record
9998          * its timestamp. NOTE: 1) That the test incorporates suggestions
9999          * from the latest proposal of the tcplw@cray.com list (Braden
10000          * 1993/04/26). 2) That updating only on newer timestamps interferes
10001          * with our earlier PAWS tests, so this check should be solely
10002          * predicated on the sequence space of this segment. 3) That we
10003          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
10004          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
10005          * SEG.Len, This modified check allows us to overcome RFC1323's
10006          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
10007          * p.869. In such cases, we can still calculate the RTT correctly
10008          * when RCV.NXT == Last.ACK.Sent.
10009          */
10010         INP_WLOCK_ASSERT(tp->t_inpcb);
10011         if ((to->to_flags & TOF_TS) != 0 &&
10012             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
10013             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
10014             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
10015                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
10016                 tp->ts_recent = to->to_tsval;
10017         }
10018         /*
10019          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
10020          * is on (half-synchronized state), then queue data for later
10021          * processing; else drop segment and return.
10022          */
10023         if ((thflags & TH_ACK) == 0) {
10024                 if (tp->t_flags & TF_NEEDSYN) {
10025                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
10026                             tiwin, thflags, nxt_pkt));
10027                 } else if (tp->t_flags & TF_ACKNOW) {
10028                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
10029                         bbr->r_wanted_output = 1;
10030                         return (ret_val);
10031                 } else {
10032                         ctf_do_drop(m, NULL);
10033                         return (0);
10034                 }
10035         }
10036         /*
10037          * Ack processing.
10038          */
10039         INP_WLOCK_ASSERT(tp->t_inpcb);
10040         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
10041                 return (ret_val);
10042         }
10043         if (sbavail(&so->so_snd)) {
10044                 if (ctf_progress_timeout_check(tp, true)) {
10045                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
10046                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10047                         return (1);
10048                 }
10049         }
10050         INP_WLOCK_ASSERT(tp->t_inpcb);
10051         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
10052             tiwin, thflags, nxt_pkt));
10053 }
10054
10055 static void
10056 bbr_stop_all_timers(struct tcpcb *tp)
10057 {
10058         struct tcp_bbr *bbr;
10059
10060         /*
10061          * Assure no timers are running.
10062          */
10063         if (tcp_timer_active(tp, TT_PERSIST)) {
10064                 /* We enter in persists, set the flag appropriately */
10065                 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10066                 bbr->rc_in_persist = 1;
10067         }
10068         tcp_timer_suspend(tp, TT_PERSIST);
10069         tcp_timer_suspend(tp, TT_REXMT);
10070         tcp_timer_suspend(tp, TT_KEEP);
10071         tcp_timer_suspend(tp, TT_DELACK);
10072 }
10073
10074 static void
10075 bbr_google_mode_on(struct tcp_bbr *bbr)
10076 {
10077         bbr->rc_use_google = 1;
10078         bbr->rc_no_pacing = 0;
10079         bbr->r_ctl.bbr_google_discount = bbr_google_discount;
10080         bbr->r_use_policer = bbr_policer_detection_enabled;
10081         bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10082         bbr->bbr_use_rack_cheat = 0;
10083         bbr->r_ctl.rc_incr_tmrs = 0;
10084         bbr->r_ctl.rc_inc_tcp_oh = 0;
10085         bbr->r_ctl.rc_inc_ip_oh = 0;
10086         bbr->r_ctl.rc_inc_enet_oh = 0;
10087         reset_time(&bbr->r_ctl.rc_delrate,
10088                    BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10089         reset_time_small(&bbr->r_ctl.rc_rttprop,
10090                          (11 * USECS_IN_SECOND));
10091         tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
10092 }
10093
10094 static void
10095 bbr_google_mode_off(struct tcp_bbr *bbr)
10096 {
10097         bbr->rc_use_google = 0;
10098         bbr->r_ctl.bbr_google_discount = 0;
10099         bbr->no_pacing_until = bbr_no_pacing_until;
10100         bbr->r_use_policer = 0;
10101         if (bbr->no_pacing_until)
10102                 bbr->rc_no_pacing = 1;
10103         else
10104                 bbr->rc_no_pacing = 0;
10105         if (bbr_use_rack_resend_cheat)
10106                 bbr->bbr_use_rack_cheat = 1;
10107         else
10108                 bbr->bbr_use_rack_cheat = 0;
10109         if (bbr_incr_timers)
10110                 bbr->r_ctl.rc_incr_tmrs = 1;
10111         else
10112                 bbr->r_ctl.rc_incr_tmrs = 0;
10113         if (bbr_include_tcp_oh)
10114                 bbr->r_ctl.rc_inc_tcp_oh = 1;
10115         else
10116                 bbr->r_ctl.rc_inc_tcp_oh = 0;
10117         if (bbr_include_ip_oh)
10118                 bbr->r_ctl.rc_inc_ip_oh = 1;
10119         else
10120                 bbr->r_ctl.rc_inc_ip_oh = 0;
10121         if (bbr_include_enet_oh)
10122                 bbr->r_ctl.rc_inc_enet_oh = 1;
10123         else
10124                 bbr->r_ctl.rc_inc_enet_oh = 0;
10125         bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10126         reset_time(&bbr->r_ctl.rc_delrate,
10127                    bbr_num_pktepo_for_del_limit);
10128         reset_time_small(&bbr->r_ctl.rc_rttprop,
10129                          (bbr_filter_len_sec * USECS_IN_SECOND));
10130         tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
10131 }
10132 /*
10133  * Return 0 on success, non-zero on failure
10134  * which indicates the error (usually no memory).
10135  */
10136 static int
10137 bbr_init(struct tcpcb *tp)
10138 {
10139         struct tcp_bbr *bbr = NULL;
10140         struct inpcb *inp;
10141         uint32_t cts;
10142
10143         tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO));
10144         if (tp->t_fb_ptr == NULL) {
10145                 /*
10146                  * We need to allocate memory but cant. The INP and INP_INFO
10147                  * locks and they are recusive (happens during setup. So a
10148                  * scheme to drop the locks fails :(
10149                  *
10150                  */
10151                 return (ENOMEM);
10152         }
10153         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10154         bbr->rtt_valid = 0;
10155         inp = tp->t_inpcb;
10156         inp->inp_flags2 |= INP_CANNOT_DO_ECN;
10157         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
10158         TAILQ_INIT(&bbr->r_ctl.rc_map);
10159         TAILQ_INIT(&bbr->r_ctl.rc_free);
10160         TAILQ_INIT(&bbr->r_ctl.rc_tmap);
10161         bbr->rc_tp = tp;
10162         if (tp->t_inpcb) {
10163                 bbr->rc_inp = tp->t_inpcb;
10164         }
10165         cts = tcp_get_usecs(&bbr->rc_tv);
10166         tp->t_acktime = 0;
10167         bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close;
10168         bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade;
10169         bbr->rc_tlp_threshold = bbr_tlp_thresh;
10170         bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh;
10171         bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay;
10172         bbr->r_ctl.rc_min_to = bbr_min_to;
10173         bbr->rc_bbr_state = BBR_STATE_STARTUP;
10174         bbr->r_ctl.bbr_lost_at_state = 0;
10175         bbr->r_ctl.rc_lost_at_startup = 0;
10176         bbr->rc_all_timers_stopped = 0;
10177         bbr->r_ctl.rc_bbr_lastbtlbw = 0;
10178         bbr->r_ctl.rc_pkt_epoch_del = 0;
10179         bbr->r_ctl.rc_pkt_epoch = 0;
10180         bbr->r_ctl.rc_lowest_rtt = 0xffffffff;
10181         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain;
10182         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
10183         bbr->r_ctl.rc_went_idle_time = cts;
10184         bbr->rc_pacer_started = cts;
10185         bbr->r_ctl.rc_pkt_epoch_time = cts;
10186         bbr->r_ctl.rc_rcvtime = cts;
10187         bbr->r_ctl.rc_bbr_state_time = cts;
10188         bbr->r_ctl.rc_del_time = cts;
10189         bbr->r_ctl.rc_tlp_rxt_last_time = cts;
10190         bbr->r_ctl.last_in_probertt = cts;
10191         bbr->skip_gain = 0;
10192         bbr->gain_is_limited = 0;
10193         bbr->no_pacing_until = bbr_no_pacing_until;
10194         if (bbr->no_pacing_until)
10195                 bbr->rc_no_pacing = 1;
10196         if (bbr_use_google_algo) {
10197                 bbr->rc_no_pacing = 0;
10198                 bbr->rc_use_google = 1;
10199                 bbr->r_ctl.bbr_google_discount = bbr_google_discount;
10200                 bbr->r_use_policer = bbr_policer_detection_enabled;
10201         } else {
10202                 bbr->rc_use_google = 0;
10203                 bbr->r_ctl.bbr_google_discount = 0;
10204                 bbr->r_use_policer = 0;
10205         }
10206         if (bbr_ts_limiting)
10207                 bbr->rc_use_ts_limit = 1;
10208         else
10209                 bbr->rc_use_ts_limit = 0;
10210         if (bbr_ts_can_raise)
10211                 bbr->ts_can_raise = 1;
10212         else
10213                 bbr->ts_can_raise = 0;
10214         if (V_tcp_delack_enabled == 1)
10215                 tp->t_delayed_ack = 2;
10216         else if (V_tcp_delack_enabled == 0)
10217                 tp->t_delayed_ack = 0;
10218         else if (V_tcp_delack_enabled < 100)
10219                 tp->t_delayed_ack = V_tcp_delack_enabled;
10220         else
10221                 tp->t_delayed_ack = 2;
10222         if (bbr->rc_use_google == 0)
10223                 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10224         else
10225                 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10226         bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms;
10227         bbr->rc_max_rto_sec = bbr_rto_max_sec;
10228         bbr->rc_init_win = bbr_def_init_win;
10229         if (tp->t_flags & TF_REQ_TSTMP)
10230                 bbr->rc_last_options = TCP_TS_OVERHEAD;
10231         bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options;
10232         bbr->r_ctl.rc_high_rwnd = tp->snd_wnd;
10233         bbr->r_init_rtt = 1;
10234
10235         counter_u64_add(bbr_flows_nohdwr_pacing, 1);
10236         if (bbr_allow_hdwr_pacing)
10237                 bbr->bbr_hdw_pace_ena = 1;
10238         else
10239                 bbr->bbr_hdw_pace_ena = 0;
10240         if (bbr_sends_full_iwnd)
10241                 bbr->bbr_init_win_cheat = 1;
10242         else
10243                 bbr->bbr_init_win_cheat = 0;
10244         bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max;
10245         bbr->r_ctl.rc_drain_pg = bbr_drain_gain;
10246         bbr->r_ctl.rc_startup_pg = bbr_high_gain;
10247         bbr->rc_loss_exit = bbr_exit_startup_at_loss;
10248         bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain;
10249         bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second;
10250         bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar;
10251         bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max;
10252         bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor;
10253         bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min;
10254         bbr->r_ctl.bbr_cross_over = bbr_cross_over;
10255         bbr->r_ctl.rc_rtt_shrinks = cts;
10256         if (bbr->rc_use_google) {
10257                 setup_time_filter(&bbr->r_ctl.rc_delrate,
10258                                   FILTER_TYPE_MAX,
10259                                   BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10260                 setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10261                                         FILTER_TYPE_MIN, (11 * USECS_IN_SECOND));
10262         } else {
10263                 setup_time_filter(&bbr->r_ctl.rc_delrate,
10264                                   FILTER_TYPE_MAX,
10265                                   bbr_num_pktepo_for_del_limit);
10266                 setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10267                                         FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND));
10268         }
10269         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0);
10270         if (bbr_uses_idle_restart)
10271                 bbr->rc_use_idle_restart = 1;
10272         else
10273                 bbr->rc_use_idle_restart = 0;
10274         bbr->r_ctl.rc_bbr_cur_del_rate = 0;
10275         bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps;
10276         if (bbr_resends_use_tso)
10277                 bbr->rc_resends_use_tso = 1;
10278 #ifdef NETFLIX_PEAKRATE
10279         tp->t_peakrate_thr = tp->t_maxpeakrate;
10280 #endif
10281         if (tp->snd_una != tp->snd_max) {
10282                 /* Create a send map for the current outstanding data */
10283                 struct bbr_sendmap *rsm;
10284
10285                 rsm = bbr_alloc(bbr);
10286                 if (rsm == NULL) {
10287                         uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10288                         tp->t_fb_ptr = NULL;
10289                         return (ENOMEM);
10290                 }
10291                 rsm->r_flags = BBR_OVERMAX;
10292                 rsm->r_tim_lastsent[0] = cts;
10293                 rsm->r_rtr_cnt = 1;
10294                 rsm->r_rtr_bytes = 0;
10295                 rsm->r_start = tp->snd_una;
10296                 rsm->r_end = tp->snd_max;
10297                 rsm->r_dupack = 0;
10298                 rsm->r_delivered = bbr->r_ctl.rc_delivered;
10299                 rsm->r_ts_valid = 0;
10300                 rsm->r_del_ack_ts = tp->ts_recent;
10301                 rsm->r_del_time = cts;
10302                 if (bbr->r_ctl.r_app_limited_until)
10303                         rsm->r_app_limited = 1;
10304                 else
10305                         rsm->r_app_limited = 0;
10306                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
10307                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
10308                 rsm->r_in_tmap = 1;
10309                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
10310                         rsm->r_bbr_state = bbr_state_val(bbr);
10311                 else
10312                         rsm->r_bbr_state = 8;
10313         }
10314         if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0))
10315                 bbr->bbr_use_rack_cheat = 1;
10316         if (bbr_incr_timers && (bbr->rc_use_google == 0))
10317                 bbr->r_ctl.rc_incr_tmrs = 1;
10318         if (bbr_include_tcp_oh && (bbr->rc_use_google == 0))
10319                 bbr->r_ctl.rc_inc_tcp_oh = 1;
10320         if (bbr_include_ip_oh && (bbr->rc_use_google == 0))
10321                 bbr->r_ctl.rc_inc_ip_oh = 1;
10322         if (bbr_include_enet_oh && (bbr->rc_use_google == 0))
10323                 bbr->r_ctl.rc_inc_enet_oh = 1;
10324
10325         bbr_log_type_statechange(bbr, cts, __LINE__);
10326         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
10327             (tp->t_srtt)) {
10328                 uint32_t rtt;
10329
10330                 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
10331                 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
10332         }
10333         /* announce the settings and state */
10334         bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT);
10335         tcp_bbr_tso_size_check(bbr, cts);
10336         /*
10337          * Now call the generic function to start a timer. This will place
10338          * the TCB on the hptsi wheel if a timer is needed with appropriate
10339          * flags.
10340          */
10341         bbr_stop_all_timers(tp);
10342         bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0);
10343         return (0);
10344 }
10345
10346 /*
10347  * Return 0 if we can accept the connection. Return
10348  * non-zero if we can't handle the connection. A EAGAIN
10349  * means you need to wait until the connection is up.
10350  * a EADDRNOTAVAIL means we can never handle the connection
10351  * (no SACK).
10352  */
10353 static int
10354 bbr_handoff_ok(struct tcpcb *tp)
10355 {
10356         if ((tp->t_state == TCPS_CLOSED) ||
10357             (tp->t_state == TCPS_LISTEN)) {
10358                 /* Sure no problem though it may not stick */
10359                 return (0);
10360         }
10361         if ((tp->t_state == TCPS_SYN_SENT) ||
10362             (tp->t_state == TCPS_SYN_RECEIVED)) {
10363                 /*
10364                  * We really don't know you have to get to ESTAB or beyond
10365                  * to tell.
10366                  */
10367                 return (EAGAIN);
10368         }
10369         if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) {
10370                 return (0);
10371         }
10372         /*
10373          * If we reach here we don't do SACK on this connection so we can
10374          * never do rack.
10375          */
10376         return (EINVAL);
10377 }
10378
10379 static void
10380 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged)
10381 {
10382         if (tp->t_fb_ptr) {
10383                 uint32_t calc;
10384                 struct tcp_bbr *bbr;
10385                 struct bbr_sendmap *rsm;
10386
10387                 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10388                 if (bbr->r_ctl.crte)
10389                         tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
10390                 bbr_log_flowend(bbr);
10391                 bbr->rc_tp = NULL;
10392                 if (tp->t_inpcb) {
10393                         /* Backout any flags2 we applied */
10394                         tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN;
10395                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
10396                         tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
10397                 }
10398                 if (bbr->bbr_hdrw_pacing)
10399                         counter_u64_add(bbr_flows_whdwr_pacing, -1);
10400                 else
10401                         counter_u64_add(bbr_flows_nohdwr_pacing, -1);
10402                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10403                 while (rsm) {
10404                         TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
10405                         uma_zfree(bbr_zone, rsm);
10406                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10407                 }
10408                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10409                 while (rsm) {
10410                         TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
10411                         uma_zfree(bbr_zone, rsm);
10412                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10413                 }
10414                 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd;
10415                 if (calc > (bbr->r_ctl.rc_init_rwnd / 10))
10416                         BBR_STAT_INC(bbr_dynamic_rwnd);
10417                 else
10418                         BBR_STAT_INC(bbr_static_rwnd);
10419                 bbr->r_ctl.rc_free_cnt = 0;
10420                 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10421                 tp->t_fb_ptr = NULL;
10422         }
10423         /* Make sure snd_nxt is correctly set */
10424         tp->snd_nxt = tp->snd_max;
10425 }
10426
10427 static void
10428 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win)
10429 {
10430         switch (tp->t_state) {
10431         case TCPS_SYN_SENT:
10432                 bbr->r_state = TCPS_SYN_SENT;
10433                 bbr->r_substate = bbr_do_syn_sent;
10434                 break;
10435         case TCPS_SYN_RECEIVED:
10436                 bbr->r_state = TCPS_SYN_RECEIVED;
10437                 bbr->r_substate = bbr_do_syn_recv;
10438                 break;
10439         case TCPS_ESTABLISHED:
10440                 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd);
10441                 bbr->r_state = TCPS_ESTABLISHED;
10442                 bbr->r_substate = bbr_do_established;
10443                 break;
10444         case TCPS_CLOSE_WAIT:
10445                 bbr->r_state = TCPS_CLOSE_WAIT;
10446                 bbr->r_substate = bbr_do_close_wait;
10447                 break;
10448         case TCPS_FIN_WAIT_1:
10449                 bbr->r_state = TCPS_FIN_WAIT_1;
10450                 bbr->r_substate = bbr_do_fin_wait_1;
10451                 break;
10452         case TCPS_CLOSING:
10453                 bbr->r_state = TCPS_CLOSING;
10454                 bbr->r_substate = bbr_do_closing;
10455                 break;
10456         case TCPS_LAST_ACK:
10457                 bbr->r_state = TCPS_LAST_ACK;
10458                 bbr->r_substate = bbr_do_lastack;
10459                 break;
10460         case TCPS_FIN_WAIT_2:
10461                 bbr->r_state = TCPS_FIN_WAIT_2;
10462                 bbr->r_substate = bbr_do_fin_wait_2;
10463                 break;
10464         case TCPS_LISTEN:
10465         case TCPS_CLOSED:
10466         case TCPS_TIME_WAIT:
10467         default:
10468                 break;
10469         };
10470 }
10471
10472 static void
10473 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog)
10474 {
10475         /*
10476          * Now what state are we going into now? Is there adjustments
10477          * needed?
10478          */
10479         int32_t old_state, old_gain;
10480
10481
10482         old_state = bbr_state_val(bbr);
10483         old_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
10484         if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) {
10485                 /* Save the lowest srtt we saw in our end of the sub-state */
10486                 bbr->rc_hit_state_1 = 0;
10487                 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff)
10488                         bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state;
10489         }
10490         bbr->rc_bbr_substate++;
10491         if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) {
10492                 /* Cycle back to first state-> gain */
10493                 bbr->rc_bbr_substate = 0;
10494         }
10495         if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10496                 /*
10497                  * We enter the gain(5/4) cycle (possibly less if
10498                  * shallow buffer detection is enabled)
10499                  */
10500                 if (bbr->skip_gain) {
10501                         /*
10502                          * Hardware pacing has set our rate to
10503                          * the max and limited our b/w just
10504                          * do level i.e. no gain.
10505                          */
10506                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1];
10507                 } else if (bbr->gain_is_limited &&
10508                            bbr->bbr_hdrw_pacing &&
10509                            bbr->r_ctl.crte) {
10510                         /*
10511                          * We can't gain above the hardware pacing
10512                          * rate which is less than our rate + the gain
10513                          * calculate the gain needed to reach the hardware
10514                          * pacing rate..
10515                          */
10516                         uint64_t bw, rate, gain_calc;
10517
10518                         bw = bbr_get_bw(bbr);
10519                         rate = bbr->r_ctl.crte->rate;
10520                         if ((rate > bw) &&
10521                             (((bw *  (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) {
10522                                 gain_calc = (rate * BBR_UNIT) / bw;
10523                                 if (gain_calc < BBR_UNIT)
10524                                         gain_calc = BBR_UNIT;
10525                                 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc;
10526                         } else {
10527                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10528                         }
10529                 } else
10530                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10531                 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) {
10532                         bbr->r_ctl.rc_bbr_state_atflight = cts;
10533                 } else
10534                         bbr->r_ctl.rc_bbr_state_atflight = 0;
10535         } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10536                 bbr->rc_hit_state_1 = 1;
10537                 bbr->r_ctl.rc_exta_time_gd = 0;
10538                 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10539                                                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10540                 if (bbr_state_drain_2_tar) {
10541                         bbr->r_ctl.rc_bbr_state_atflight = 0;
10542                 } else
10543                         bbr->r_ctl.rc_bbr_state_atflight = cts;
10544                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN];
10545         } else {
10546                 /* All other cycles hit here 2-7 */
10547                 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) {
10548                         if (bbr_sub_drain_slam_cwnd &&
10549                             (bbr->rc_use_google == 0) &&
10550                             (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10551                                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10552                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10553                         }
10554                         if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP))
10555                                 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) -
10556                                                                bbr_get_rtt(bbr, BBR_RTT_PROP));
10557                         else
10558                                 bbr->r_ctl.rc_exta_time_gd = 0;
10559                         if (bbr->r_ctl.rc_exta_time_gd) {
10560                                 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd;
10561                                 /* Now chop up the time for each state (div by 7) */
10562                                 bbr->r_ctl.rc_level_state_extra /= 7;
10563                                 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) {
10564                                         /* Add a randomization */
10565                                         bbr_randomize_extra_state_time(bbr);
10566                                 }
10567                         }
10568                 }
10569                 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10570                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)];
10571         }
10572         if (bbr->rc_use_google) {
10573                 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10574         }
10575         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10576         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10577         if (dolog)
10578                 bbr_log_type_statechange(bbr, cts, line);
10579
10580         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10581                 uint32_t time_in;
10582
10583                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10584                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
10585                         counter_u64_add(bbr_state_time[(old_state + 5)], time_in);
10586                 } else {
10587                         counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10588                 }
10589         }
10590         bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
10591         bbr_set_state_target(bbr, __LINE__);
10592         if (bbr_sub_drain_slam_cwnd &&
10593             (bbr->rc_use_google == 0) &&
10594             (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10595                 /* Slam down the cwnd */
10596                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10597                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10598                 if (bbr_sub_drain_app_limit) {
10599                         /* Go app limited if we are on a long drain */
10600                         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered +
10601                                                           ctf_flight_size(bbr->rc_tp,
10602                                                               (bbr->r_ctl.rc_sacked +
10603                                                                bbr->r_ctl.rc_lost_bytes)));
10604                 }
10605                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10606         }
10607         if (bbr->rc_lt_use_bw) {
10608                 /* In policed mode we clamp pacing_gain to BBR_UNIT */
10609                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10610         }
10611         /* Google changes TSO size every cycle */
10612         if (bbr->rc_use_google)
10613                 tcp_bbr_tso_size_check(bbr, cts);
10614         bbr->r_ctl.gain_epoch = cts;
10615         bbr->r_ctl.rc_bbr_state_time = cts;
10616         bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch;
10617 }
10618
10619 static void
10620 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10621 {
10622         if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) &&
10623             (google_allow_early_out == 1) &&
10624             (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) {
10625                 /* We have reached out target flight size possibly early */
10626                 goto change_state;
10627         }
10628         if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10629                 return;
10630         }
10631         if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) {
10632                 /*
10633                  * Must be a rttProp movement forward before
10634                  * we can change states.
10635                  */
10636                 return;
10637         }
10638         if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10639                 /*
10640                  * The needed time has passed but for
10641                  * the gain cycle extra rules apply:
10642                  * 1) If we have seen loss, we exit
10643                  * 2) If we have not reached the target
10644                  *    we stay in GAIN (gain-to-target).
10645                  */
10646                 if (google_consider_lost && losses)
10647                         goto change_state;
10648                 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) {
10649                         return;
10650                 }
10651         }
10652 change_state:
10653         /* For gain we must reach our target, all others last 1 rttProp */
10654         bbr_substate_change(bbr, cts, __LINE__, 1);
10655 }
10656
10657 static void
10658 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10659 {
10660         uint32_t flight, bbr_cur_cycle_time;
10661
10662         if (bbr->rc_use_google) {
10663                 bbr_set_probebw_google_gains(bbr, cts, losses);
10664                 return;
10665         }
10666         if (cts == 0) {
10667                 /*
10668                  * Never alow cts to be 0 we
10669                  * do this so we can judge if
10670                  * we have set a timestamp.
10671                  */
10672                 cts = 1;
10673         }
10674         if (bbr_state_is_pkt_epoch)
10675                 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
10676         else
10677                 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP);
10678
10679         if (bbr->r_ctl.rc_bbr_state_atflight == 0) {
10680                 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10681                         flight = ctf_flight_size(bbr->rc_tp,
10682                                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10683                         if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) {
10684                                 /* Keep it slam down */
10685                                 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) {
10686                                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10687                                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10688                                 }
10689                                 if (bbr_sub_drain_app_limit) {
10690                                         /* Go app limited if we are on a long drain */
10691                                         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight);
10692                                 }
10693                         }
10694                         if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) &&
10695                             (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) ||
10696                              (flight >= bbr->r_ctl.flightsize_at_drain))) {
10697                                 /*
10698                                  * Still here after the same time as
10699                                  * the gain. We need to drain harder
10700                                  * for the next srtt. Reduce by a set amount
10701                                  * the gain drop is capped at DRAIN states
10702                                  * value (88).
10703                                  */
10704                                 bbr->r_ctl.flightsize_at_drain = flight;
10705                                 if (bbr_drain_drop_mul &&
10706                                     bbr_drain_drop_div &&
10707                                     (bbr_drain_drop_mul < bbr_drain_drop_div)) {
10708                                         /* Use your specific drop value (def 4/5 = 20%) */
10709                                         bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul;
10710                                         bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div;
10711                                 } else {
10712                                         /* You get drop of 20% */
10713                                         bbr->r_ctl.rc_bbr_hptsi_gain *= 4;
10714                                         bbr->r_ctl.rc_bbr_hptsi_gain /= 5;
10715                                 }
10716                                 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) {
10717                                         /* Reduce our gain again to the bottom  */
10718                                         bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
10719                                 }
10720                                 bbr_log_exit_gain(bbr, cts, 4);
10721                                 /*
10722                                  * Extend out so we wait another
10723                                  * epoch before dropping again.
10724                                  */
10725                                 bbr->r_ctl.gain_epoch = cts;
10726                         }
10727                         if (flight <= bbr->r_ctl.rc_target_at_state) {
10728                                 if (bbr_sub_drain_slam_cwnd &&
10729                                     (bbr->rc_use_google == 0) &&
10730                                     (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10731                                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10732                                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10733                                 }
10734                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10735                                 bbr_log_exit_gain(bbr, cts, 3);
10736                         }
10737                 } else {
10738                         /* Its a gain  */
10739                         if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) {
10740                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10741                                 goto change_state;
10742                         }
10743                         if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) ||
10744                             ((ctf_outstanding(bbr->rc_tp) +  bbr->rc_tp->t_maxseg - 1) >=
10745                              bbr->rc_tp->snd_wnd)) {
10746                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10747                                 bbr_log_exit_gain(bbr, cts, 2);
10748                         }
10749                 }
10750                 /**
10751                  * We fall through and return always one of two things has
10752                  * occured.
10753                  * 1) We are still not at target
10754                  *    <or>
10755                  * 2) We reached the target and set rc_bbr_state_atflight
10756                  *    which means we no longer hit this block
10757                  *    next time we are called.
10758                  */
10759                 return;
10760         }
10761 change_state:
10762         if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time))
10763                 return;
10764         if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) {
10765                 /* Less than a full time-period has passed */
10766                 return;
10767         }
10768         if (bbr->r_ctl.rc_level_state_extra &&
10769             (bbr_state_val(bbr) > BBR_SUB_DRAIN) &&
10770             ((cts - bbr->r_ctl.rc_bbr_state_time) <
10771              (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10772                 /* Less than a full time-period + extra has passed */
10773                 return;
10774         }
10775         if (bbr_gain_gets_extra_too &&
10776             bbr->r_ctl.rc_level_state_extra &&
10777             (bbr_state_val(bbr) == BBR_SUB_GAIN) &&
10778             ((cts - bbr->r_ctl.rc_bbr_state_time) <
10779              (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10780                 /* Less than a full time-period + extra has passed */
10781                 return;
10782         }
10783         bbr_substate_change(bbr, cts, __LINE__, 1);
10784 }
10785
10786 static uint32_t
10787 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain)
10788 {
10789         uint32_t mss, tar;
10790
10791         if (bbr->rc_use_google) {
10792                 /* Google just uses the cwnd target */
10793                 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain);
10794         } else {
10795                 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options),
10796                           bbr->r_ctl.rc_pace_max_segs);
10797                 /* Get the base cwnd with gain rounded to a mss */
10798                 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr),
10799                                                       gain), mss);
10800                 /* Make sure it is within our min */
10801                 if (tar < get_min_cwnd(bbr))
10802                         return (get_min_cwnd(bbr));
10803         }
10804         return (tar);
10805 }
10806
10807 static void
10808 bbr_set_state_target(struct tcp_bbr *bbr, int line)
10809 {
10810         uint32_t tar, meth;
10811
10812         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
10813             ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
10814                 /* Special case using old probe-rtt method */
10815                 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10816                 meth = 1;
10817         } else {
10818                 /* Non-probe-rtt case and reduced probe-rtt  */
10819                 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
10820                     (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) {
10821                         /* For gain cycle we use the hptsi gain */
10822                         tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10823                         meth = 2;
10824                 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) {
10825                         /*
10826                          * If configured, or for google all other states
10827                          * get BBR_UNIT.
10828                          */
10829                         tar = bbr_get_a_state_target(bbr, BBR_UNIT);
10830                         meth = 3;
10831                 } else {
10832                         /*
10833                          * Or we set a target based on the pacing gain
10834                          * for non-google mode and default (non-configured).
10835                          * Note we don't set a target goal below drain (192).
10836                          */
10837                         if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN])  {
10838                                 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]);
10839                                 meth = 4;
10840                         } else {
10841                                 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10842                                 meth = 5;
10843                         }
10844                 }
10845         }
10846         bbr_log_set_of_state_target(bbr, tar, line, meth);
10847         bbr->r_ctl.rc_target_at_state = tar;
10848 }
10849
10850 static void
10851 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
10852 {
10853         /* Change to probe_rtt */
10854         uint32_t time_in;
10855
10856         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10857         bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10858                                              (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10859         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain
10860                                           + bbr->r_ctl.rc_delivered);
10861         /* Setup so we force feed the filter */
10862         if (bbr->rc_use_google || bbr_probertt_sets_rtt)
10863                 bbr->rc_prtt_set_ts = 1;
10864         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10865                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10866                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10867         }
10868         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0);
10869         bbr->r_ctl.rc_rtt_shrinks = cts;
10870         bbr->r_ctl.last_in_probertt = cts;
10871         bbr->r_ctl.rc_probertt_srttchktim = cts;
10872         bbr->r_ctl.rc_bbr_state_time = cts;
10873         bbr->rc_bbr_state = BBR_STATE_PROBE_RTT;
10874         /* We need to force the filter to update */
10875
10876         if ((bbr_sub_drain_slam_cwnd) &&
10877             bbr->rc_hit_state_1 &&
10878             (bbr->rc_use_google == 0) &&
10879             (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10880                 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd)
10881                         bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10882         } else
10883                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10884         /* Update the lost */
10885         bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10886         if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){
10887                 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN)  */
10888                 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10889                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10890                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10891                 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10892                 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6);
10893                 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd;
10894         } else {
10895                 /*
10896                  * We bring it down slowly by using a hptsi gain that is
10897                  * probably 75%. This will slowly float down our outstanding
10898                  * without tampering with the cwnd.
10899                  */
10900                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
10901                 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10902                 bbr_set_state_target(bbr, __LINE__);
10903                 if (bbr_prtt_slam_cwnd &&
10904                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
10905                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10906                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10907                 }
10908         }
10909         if (ctf_flight_size(bbr->rc_tp,
10910                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
10911             bbr->r_ctl.rc_target_at_state) {
10912                 /* We are at target */
10913                 bbr->r_ctl.rc_bbr_enters_probertt = cts;
10914         } else {
10915                 /* We need to come down to reach target before our time begins */
10916                 bbr->r_ctl.rc_bbr_enters_probertt = 0;
10917         }
10918         bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch;
10919         BBR_STAT_INC(bbr_enter_probertt);
10920         bbr_log_exit_gain(bbr, cts, 0);
10921         bbr_log_type_statechange(bbr, cts, line);
10922 }
10923
10924 static void
10925 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts)
10926 {
10927         /*
10928          * Sanity check on probe-rtt intervals.
10929          * In crazy situations where we are competing
10930          * against new-reno flows with huge buffers
10931          * our rtt-prop interval could come to dominate
10932          * things if we can't get through a full set
10933          * of cycles, we need to adjust it.
10934          */
10935         if (bbr_can_adjust_probertt &&
10936             (bbr->rc_use_google == 0)) {
10937                 uint16_t val = 0;
10938                 uint32_t cur_rttp, fval, newval, baseval;
10939
10940                 /* Are we to small and go into probe-rtt to often? */
10941                 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1));
10942                 cur_rttp = roundup(baseval, USECS_IN_SECOND);
10943                 fval = bbr_filter_len_sec * USECS_IN_SECOND;
10944                 if (bbr_is_ratio == 0) {
10945                         if (fval > bbr_rtt_probe_limit)
10946                                 newval = cur_rttp + (fval - bbr_rtt_probe_limit);
10947                         else
10948                                 newval = cur_rttp;
10949                 } else {
10950                         int mul;
10951
10952                         mul = fval / bbr_rtt_probe_limit;
10953                         newval = cur_rttp * mul;
10954                 }
10955                 if (cur_rttp >  bbr->r_ctl.rc_probertt_int) {
10956                         bbr->r_ctl.rc_probertt_int = cur_rttp;
10957                         reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10958                         val = 1;
10959                 } else {
10960                         /*
10961                          * No adjustments were made
10962                          * do we need to shrink it?
10963                          */
10964                         if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) {
10965                                 if (cur_rttp <= bbr_rtt_probe_limit) {
10966                                         /*
10967                                          * Things have calmed down lets
10968                                          * shrink all the way to default
10969                                          */
10970                                         bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10971                                         reset_time_small(&bbr->r_ctl.rc_rttprop,
10972                                                          (bbr_filter_len_sec * USECS_IN_SECOND));
10973                                         cur_rttp = bbr_rtt_probe_limit;
10974                                         newval = (bbr_filter_len_sec * USECS_IN_SECOND);
10975                                         val = 2;
10976                                 } else {
10977                                         /*
10978                                          * Well does some adjustment make sense?
10979                                          */
10980                                         if (cur_rttp < bbr->r_ctl.rc_probertt_int) {
10981                                                 /* We can reduce interval time some */
10982                                                 bbr->r_ctl.rc_probertt_int = cur_rttp;
10983                                                 reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10984                                                 val = 3;
10985                                         }
10986                                 }
10987                         }
10988                 }
10989                 if (val)
10990                         bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val);
10991         }
10992 }
10993
10994 static void
10995 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
10996 {
10997         /* Exit probe-rtt */
10998
10999         if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) {
11000                 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
11001                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11002         }
11003         bbr_log_exit_gain(bbr, cts, 1);
11004         bbr->rc_hit_state_1 = 0;
11005         bbr->r_ctl.rc_rtt_shrinks = cts;
11006         bbr->r_ctl.last_in_probertt = cts;
11007         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0);
11008         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11009         bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp,
11010                                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
11011                                           bbr->r_ctl.rc_delivered);
11012         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11013                 uint32_t time_in;
11014
11015                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11016                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11017         }
11018         if (bbr->rc_filled_pipe) {
11019                 /* Switch to probe_bw */
11020                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11021                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11022                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
11023                 bbr_substate_change(bbr, cts, __LINE__, 0);
11024                 bbr_log_type_statechange(bbr, cts, __LINE__);
11025         } else {
11026                 /* Back to startup */
11027                 bbr->rc_bbr_state = BBR_STATE_STARTUP;
11028                 bbr->r_ctl.rc_bbr_state_time = cts;
11029                 /*
11030                  * We don't want to give a complete free 3
11031                  * measurements until we exit, so we use
11032                  * the number of pe's we were in probe-rtt
11033                  * to add to the startup_epoch. That way
11034                  * we will still retain the old state.
11035                  */
11036                 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt);
11037                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11038                 /* Make sure to use the lower pg when shifting back in */
11039                 if (bbr->r_ctl.rc_lost &&
11040                     bbr_use_lower_gain_in_startup &&
11041                     (bbr->rc_use_google == 0))
11042                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
11043                 else
11044                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
11045                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
11046                 /* Probably not needed but set it anyway */
11047                 bbr_set_state_target(bbr, __LINE__);
11048                 bbr_log_type_statechange(bbr, cts, __LINE__);
11049                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11050                     bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0);
11051         }
11052         bbr_check_probe_rtt_limits(bbr, cts);
11053 }
11054
11055 static int32_t inline
11056 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts)
11057 {
11058         if ((bbr->rc_past_init_win == 1) &&
11059             (bbr->rc_in_persist == 0) &&
11060             (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) {
11061                 return (1);
11062         }
11063         if (bbr_can_force_probertt &&
11064             (bbr->rc_in_persist == 0) &&
11065             (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
11066             ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
11067                 return (1);
11068         }
11069         return (0);
11070 }
11071
11072
11073 static int32_t
11074 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t  pkt_epoch)
11075 {
11076         uint64_t btlbw, gain;
11077         if (pkt_epoch == 0) {
11078                 /*
11079                  * Need to be on a pkt-epoch to continue.
11080                  */
11081                 return (0);
11082         }
11083         btlbw = bbr_get_full_bw(bbr);
11084         gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11085                  (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11086         if (btlbw >= gain) {
11087                 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11088                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11089                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11090                 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
11091         }
11092         if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)
11093                 return (1);
11094         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11095                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
11096         return(0);
11097 }
11098
11099 static int32_t inline
11100 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch)
11101 {
11102         /* Have we gained 25% in the last 3 packet based epoch's? */
11103         uint64_t btlbw, gain;
11104         int do_exit;
11105         int delta, rtt_gain;
11106
11107         if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11108             (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11109                 /*
11110                  * This qualifies as a RTT_PROBE session since we drop the
11111                  * data outstanding to nothing and waited more than
11112                  * bbr_rtt_probe_time.
11113                  */
11114                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11115                 bbr_set_reduced_rtt(bbr, cts, __LINE__);
11116         }
11117         if (bbr_should_enter_probe_rtt(bbr, cts)) {
11118                 bbr_enter_probe_rtt(bbr, cts, __LINE__);
11119                 return (0);
11120         }
11121         if (bbr->rc_use_google)
11122                 return (bbr_google_startup(bbr, cts,  pkt_epoch));
11123
11124         if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11125             (bbr_use_lower_gain_in_startup)) {
11126                 /* Drop to a lower gain 1.5 x since we saw loss */
11127                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
11128         }
11129         if (pkt_epoch == 0) {
11130                 /*
11131                  * Need to be on a pkt-epoch to continue.
11132                  */
11133                 return (0);
11134         }
11135         if (bbr_rtt_gain_thresh) {
11136                 /*
11137                  * Do we allow a flow to stay
11138                  * in startup with no loss and no
11139                  * gain in rtt over a set threshold?
11140                  */
11141                 if (bbr->r_ctl.rc_pkt_epoch_rtt &&
11142                     bbr->r_ctl.startup_last_srtt &&
11143                     (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) {
11144                         delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt;
11145                         rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt;
11146                 } else
11147                         rtt_gain = 0;
11148                 if ((bbr->r_ctl.startup_last_srtt == 0)  ||
11149                     (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt))
11150                         /* First time or new lower value */
11151                         bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
11152
11153                 if ((bbr->r_ctl.rc_lost == 0) &&
11154                     (rtt_gain < bbr_rtt_gain_thresh)) {
11155                         /*
11156                          * No loss, and we are under
11157                          * our gain threhold for
11158                          * increasing RTT.
11159                          */
11160                         if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11161                                 bbr->r_ctl.rc_bbr_last_startup_epoch++;
11162                         bbr_log_startup_event(bbr, cts, rtt_gain,
11163                                               delta, bbr->r_ctl.startup_last_srtt, 10);
11164                         return (0);
11165                 }
11166         }
11167         if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) &&
11168             (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) &&
11169             (!IN_RECOVERY(bbr->rc_tp->t_flags))) {
11170                 /*
11171                  * We only assess if we have a new measurment when
11172                  * we have no loss and are not in recovery.
11173                  * Drag up by one our last_startup epoch so we will hold
11174                  * the number of non-gain we have already accumulated.
11175                  */
11176                 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11177                         bbr->r_ctl.rc_bbr_last_startup_epoch++;
11178                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11179                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9);
11180                 return (0);
11181         }
11182         /* Case where we reduced the lost (bad retransmit) */
11183         if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost)
11184                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11185         bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count;
11186         btlbw = bbr_get_full_bw(bbr);
11187         if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower)
11188                 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11189                          (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11190         else
11191                 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11192                          (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11193         do_exit = 0;
11194         if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw)
11195                 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
11196         if (btlbw >= gain) {
11197                 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11198                 /* Update the lost so we won't exit in next set of tests */
11199                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11200                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11201                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11202         }
11203         if ((bbr->rc_loss_exit &&
11204              (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11205              (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) &&
11206             ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) {
11207                 /*
11208                  * If we had no gain,  we had loss and that loss was above
11209                  * our threshould, the rwnd is not constrained, and we have
11210                  * had at least 3 packet epochs exit. Note that this is
11211                  * switched off by sysctl. Google does not do this by the
11212                  * way.
11213                  */
11214                 if ((ctf_flight_size(bbr->rc_tp,
11215                          (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
11216                      (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) {
11217                         do_exit = 1;
11218                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11219                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4);
11220                 } else {
11221                         /* Just record an updated loss value */
11222                         bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11223                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11224                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5);
11225                 }
11226         } else
11227                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11228         if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) ||
11229             do_exit) {
11230                 /* Return 1 to exit the startup state. */
11231                 return (1);
11232         }
11233         /* Stay in startup */
11234         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11235                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
11236         return (0);
11237 }
11238
11239 static void
11240 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses)
11241 {
11242         /*
11243          * A tick occured in the rtt epoch do we need to do anything?
11244          */
11245 #ifdef BBR_INVARIANTS
11246         if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
11247             (bbr->rc_bbr_state != BBR_STATE_DRAIN) &&
11248             (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) &&
11249             (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
11250             (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) {
11251                 /* Debug code? */
11252                 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state);
11253         }
11254 #endif
11255         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
11256                 /* Do we exit the startup state? */
11257                 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) {
11258                         uint32_t time_in;
11259
11260                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11261                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6);
11262                         bbr->rc_filled_pipe = 1;
11263                         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11264                         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11265
11266                                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11267                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11268                         } else
11269                                 time_in = 0;
11270                         if (bbr->rc_no_pacing)
11271                                 bbr->rc_no_pacing = 0;
11272                         bbr->r_ctl.rc_bbr_state_time = cts;
11273                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg;
11274                         bbr->rc_bbr_state = BBR_STATE_DRAIN;
11275                         bbr_set_state_target(bbr, __LINE__);
11276                         if ((bbr->rc_use_google == 0) &&
11277                             bbr_slam_cwnd_in_main_drain) {
11278                                 /* Here we don't have to worry about probe-rtt */
11279                                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
11280                                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11281                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11282                         }
11283                         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
11284                         bbr_log_type_statechange(bbr, cts, __LINE__);
11285                         if (ctf_flight_size(bbr->rc_tp,
11286                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
11287                             bbr->r_ctl.rc_target_at_state) {
11288                                 /*
11289                                  * Switch to probe_bw if we are already
11290                                  * there
11291                                  */
11292                                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11293                                 bbr_substate_change(bbr, cts, __LINE__, 0);
11294                                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11295                                 bbr_log_type_statechange(bbr, cts, __LINE__);
11296                         }
11297                 }
11298         } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) {
11299                 uint32_t inflight;
11300                 struct tcpcb *tp;
11301
11302                 tp = bbr->rc_tp;
11303                 inflight = ctf_flight_size(tp,
11304                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11305                 if (inflight >= bbr->r_ctl.rc_target_at_state) {
11306                         /* We have reached a flight of the cwnd target */
11307                         bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11308                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11309                         bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
11310                         bbr_set_state_target(bbr, __LINE__);
11311                         /*
11312                          * Rig it so we don't do anything crazy and
11313                          * start fresh with a new randomization.
11314                          */
11315                         bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
11316                         bbr->rc_bbr_substate = BBR_SUB_LEVEL6;
11317                         bbr_substate_change(bbr, cts, __LINE__, 1);
11318                 }
11319         } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) {
11320                 /* Has in-flight reached the bdp (or less)? */
11321                 uint32_t inflight;
11322                 struct tcpcb *tp;
11323
11324                 tp = bbr->rc_tp;
11325                 inflight = ctf_flight_size(tp,
11326                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11327                 if ((bbr->rc_use_google == 0) &&
11328                     bbr_slam_cwnd_in_main_drain &&
11329                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11330                         /*
11331                          * Here we don't have to worry about probe-rtt
11332                          * re-slam it, but keep it slammed down.
11333                          */
11334                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11335                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11336                 }
11337                 if (inflight <= bbr->r_ctl.rc_target_at_state) {
11338                         /* We have drained */
11339                         bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11340                         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11341                         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11342                                 uint32_t time_in;
11343
11344                                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11345                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11346                         }
11347                         if ((bbr->rc_use_google == 0) &&
11348                             bbr_slam_cwnd_in_main_drain &&
11349                             (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
11350                                 /* Restore the cwnd */
11351                                 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
11352                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11353                         }
11354                         /* Setup probe-rtt has being done now RRS-HERE */
11355                         bbr->r_ctl.rc_rtt_shrinks = cts;
11356                         bbr->r_ctl.last_in_probertt = cts;
11357                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0);
11358                         /* Randomly pick a sub-state */
11359                         bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11360                         bbr_substate_change(bbr, cts, __LINE__, 0);
11361                         bbr_log_type_statechange(bbr, cts, __LINE__);
11362                 }
11363         } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) {
11364                 uint32_t flight;
11365
11366                 flight = ctf_flight_size(bbr->rc_tp,
11367                              (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11368                 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered);
11369                 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) &&
11370                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11371                         /*
11372                          * We must keep cwnd at the desired MSS.
11373                          */
11374                         bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
11375                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11376                 } else if ((bbr_prtt_slam_cwnd) &&
11377                            (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11378                         /* Re-slam it */
11379                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11380                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11381                 }
11382                 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) {
11383                         /* Has outstanding reached our target? */
11384                         if (flight <= bbr->r_ctl.rc_target_at_state) {
11385                                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0);
11386                                 bbr->r_ctl.rc_bbr_enters_probertt = cts;
11387                                 /* If time is exactly 0, be 1usec off */
11388                                 if (bbr->r_ctl.rc_bbr_enters_probertt == 0)
11389                                         bbr->r_ctl.rc_bbr_enters_probertt = 1;
11390                                 if (bbr->rc_use_google == 0) {
11391                                         /*
11392                                          * Restore any lowering that as occured to
11393                                          * reach here
11394                                          */
11395                                         if (bbr->r_ctl.bbr_rttprobe_gain_val)
11396                                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
11397                                         else
11398                                                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11399                                 }
11400                         }
11401                         if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) &&
11402                             (bbr->rc_use_google == 0) &&
11403                             bbr->r_ctl.bbr_rttprobe_gain_val &&
11404                             (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) ||
11405                              (flight >= bbr->r_ctl.flightsize_at_drain))) {
11406                                 /*
11407                                  * We have doddled with our current hptsi
11408                                  * gain an srtt and have still not made it
11409                                  * to target, or we have increased our flight.
11410                                  * Lets reduce the gain by xx%
11411                                  * flooring the reduce at DRAIN (based on
11412                                  * mul/div)
11413                                  */
11414                                 int red;
11415
11416                                 bbr->r_ctl.flightsize_at_drain = flight;
11417                                 bbr->r_ctl.rc_probertt_srttchktim = cts;
11418                                 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1);
11419                                 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) {
11420                                         /* Reduce our gain again */
11421                                         bbr->r_ctl.rc_bbr_hptsi_gain -= red;
11422                                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0);
11423                                 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) {
11424                                         /* one more chance before we give up */
11425                                         bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
11426                                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0);
11427                                 } else {
11428                                         /* At the very bottom */
11429                                         bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1);
11430                                 }
11431                         }
11432                 }
11433                 if (bbr->r_ctl.rc_bbr_enters_probertt &&
11434                     (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) &&
11435                     ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) {
11436                         /* Time to exit probe RTT normally */
11437                         bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts);
11438                 }
11439         } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
11440                 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11441                     (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11442                         /*
11443                          * This qualifies as a RTT_PROBE session since we
11444                          * drop the data outstanding to nothing and waited
11445                          * more than bbr_rtt_probe_time.
11446                          */
11447                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11448                         bbr_set_reduced_rtt(bbr, cts, __LINE__);
11449                 }
11450                 if (bbr_should_enter_probe_rtt(bbr, cts)) {
11451                         bbr_enter_probe_rtt(bbr, cts, __LINE__);
11452                 } else {
11453                         bbr_set_probebw_gains(bbr, cts, losses);
11454                 }
11455         }
11456 }
11457
11458 static void
11459 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses)
11460 {
11461         int32_t epoch = 0;
11462
11463         if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
11464                 bbr_set_epoch(bbr, cts, line);
11465                 /* At each epoch doe lt bw sampling */
11466                 epoch = 1;
11467         }
11468         bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses);
11469 }
11470
11471 static int
11472 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
11473     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
11474     int32_t nxt_pkt, struct timeval *tv)
11475 {
11476         int32_t thflags, retval;
11477         uint32_t cts, lcts;
11478         uint32_t tiwin;
11479         struct tcpopt to;
11480         struct tcp_bbr *bbr;
11481         struct bbr_sendmap *rsm;
11482         struct timeval ltv;
11483         int32_t did_out = 0;
11484         int32_t in_recovery;
11485         uint16_t nsegs;
11486         int32_t prev_state;
11487         uint32_t lost;
11488
11489         nsegs = max(1, m->m_pkthdr.lro_nsegs);
11490         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11491         /* add in our stats */
11492         kern_prefetch(bbr, &prev_state);
11493         prev_state = 0;
11494         thflags = th->th_flags;
11495         /*
11496          * If this is either a state-changing packet or current state isn't
11497          * established, we require a write lock on tcbinfo.  Otherwise, we
11498          * allow the tcbinfo to be in either alocked or unlocked, as the
11499          * caller may have unnecessarily acquired a write lock due to a
11500          * race.
11501          */
11502         INP_WLOCK_ASSERT(tp->t_inpcb);
11503         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
11504             __func__));
11505         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
11506             __func__));
11507
11508         tp->t_rcvtime = ticks;
11509         /*
11510          * Unscale the window into a 32-bit value. For the SYN_SENT state
11511          * the scale is zero.
11512          */
11513         tiwin = th->th_win << tp->snd_scale;
11514 #ifdef STATS
11515         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
11516 #endif
11517         /*
11518          * Parse options on any incoming segment.
11519          */
11520         tcp_dooptions(&to, (u_char *)(th + 1),
11521             (th->th_off << 2) - sizeof(struct tcphdr),
11522             (thflags & TH_SYN) ? TO_SYN : 0);
11523
11524         if (m->m_flags & M_TSTMP) {
11525                 /* Prefer the hardware timestamp if present */
11526                 struct timespec ts;
11527
11528                 mbuf_tstmp2timespec(m, &ts);
11529                 bbr->rc_tv.tv_sec = ts.tv_sec;
11530                 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11531                 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11532         } else if (m->m_flags & M_TSTMP_LRO) {
11533                 /* Next the arrival timestamp */
11534                 struct timespec ts;
11535
11536                 mbuf_tstmp2timespec(m, &ts);
11537                 bbr->rc_tv.tv_sec = ts.tv_sec;
11538                 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11539                 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11540         } else {
11541                 /*
11542                  * Ok just get the current time.
11543                  */
11544                 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv);
11545         }
11546         /*
11547          * If echoed timestamp is later than the current time, fall back to
11548          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
11549          * were used when this connection was established.
11550          */
11551         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
11552                 to.to_tsecr -= tp->ts_offset;
11553                 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv)))
11554                         to.to_tsecr = 0;
11555         }
11556         /*
11557          * If its the first time in we need to take care of options and
11558          * verify we can do SACK for rack!
11559          */
11560         if (bbr->r_state == 0) {
11561                 /*
11562                  * Process options only when we get SYN/ACK back. The SYN
11563                  * case for incoming connections is handled in tcp_syncache.
11564                  * According to RFC1323 the window field in a SYN (i.e., a
11565                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
11566                  * this is traditional behavior, may need to be cleaned up.
11567                  */
11568                 if (bbr->rc_inp == NULL) {
11569                         bbr->rc_inp = tp->t_inpcb;
11570                 }
11571                 /*
11572                  * We need to init rc_inp here since its not init'd when
11573                  * bbr_init is called
11574                  */
11575                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
11576                         if ((to.to_flags & TOF_SCALE) &&
11577                             (tp->t_flags & TF_REQ_SCALE)) {
11578                                 tp->t_flags |= TF_RCVD_SCALE;
11579                                 tp->snd_scale = to.to_wscale;
11580                         } else
11581                                 tp->t_flags &= ~TF_REQ_SCALE;
11582                         /*
11583                          * Initial send window.  It will be updated with the
11584                          * next incoming segment to the scaled value.
11585                          */
11586                         tp->snd_wnd = th->th_win;
11587                         if ((to.to_flags & TOF_TS) &&
11588                             (tp->t_flags & TF_REQ_TSTMP)) {
11589                                 tp->t_flags |= TF_RCVD_TSTMP;
11590                                 tp->ts_recent = to.to_tsval;
11591                                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
11592                         } else
11593                             tp->t_flags &= ~TF_REQ_TSTMP;
11594                         if (to.to_flags & TOF_MSS)
11595                                 tcp_mss(tp, to.to_mss);
11596                         if ((tp->t_flags & TF_SACK_PERMIT) &&
11597                             (to.to_flags & TOF_SACKPERM) == 0)
11598                                 tp->t_flags &= ~TF_SACK_PERMIT;
11599                         if (IS_FASTOPEN(tp->t_flags)) {
11600                                 if (to.to_flags & TOF_FASTOPEN) {
11601                                         uint16_t mss;
11602
11603                                         if (to.to_flags & TOF_MSS)
11604                                                 mss = to.to_mss;
11605                                         else
11606                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
11607                                                         mss = TCP6_MSS;
11608                                                 else
11609                                                         mss = TCP_MSS;
11610                                         tcp_fastopen_update_cache(tp, mss,
11611                                             to.to_tfo_len, to.to_tfo_cookie);
11612                                 } else
11613                                         tcp_fastopen_disable_path(tp);
11614                         }
11615                 }
11616                 /*
11617                  * At this point we are at the initial call. Here we decide
11618                  * if we are doing RACK or not. We do this by seeing if
11619                  * TF_SACK_PERMIT is set, if not rack is *not* possible and
11620                  * we switch to the default code.
11621                  */
11622                 if ((tp->t_flags & TF_SACK_PERMIT) == 0) {
11623                         /* Bail */
11624                         tcp_switch_back_to_default(tp);
11625                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
11626                             tlen, iptos);
11627                         return (1);
11628                 }
11629                 /* Set the flag */
11630                 bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
11631                 tcp_set_hpts(tp->t_inpcb);
11632                 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack);
11633         }
11634         if (thflags & TH_ACK) {
11635                 /* Track ack types */
11636                 if (to.to_flags & TOF_SACK)
11637                         BBR_STAT_INC(bbr_acks_with_sacks);
11638                 else
11639                         BBR_STAT_INC(bbr_plain_acks);
11640         }
11641         /*
11642          * This is the one exception case where we set the rack state
11643          * always. All other times (timers etc) we must have a rack-state
11644          * set (so we assure we have done the checks above for SACK).
11645          */
11646         if (thflags & TH_FIN)
11647                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
11648         if (bbr->r_state != tp->t_state)
11649                 bbr_set_state(tp, bbr, tiwin);
11650
11651         if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL)
11652                 kern_prefetch(rsm, &prev_state);
11653         prev_state = bbr->r_state;
11654         bbr->rc_ack_was_delayed = 0;
11655         lost = bbr->r_ctl.rc_lost;
11656         bbr->rc_is_pkt_epoch_now = 0;
11657         if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) {
11658                 /* Get the real time into lcts and figure the real delay */
11659                 lcts = tcp_get_usecs(&ltv);
11660                 if (TSTMP_GT(lcts, cts)) {
11661                         bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts;
11662                         bbr->rc_ack_was_delayed = 1;
11663                         if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay,
11664                                      bbr->r_ctl.highest_hdwr_delay))
11665                                 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay;
11666                 } else {
11667                         bbr->r_ctl.rc_ack_hdwr_delay = 0;
11668                         bbr->rc_ack_was_delayed = 0;
11669                 }
11670         } else {
11671                 bbr->r_ctl.rc_ack_hdwr_delay = 0;
11672                 bbr->rc_ack_was_delayed = 0;
11673         }
11674         bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m);
11675         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
11676                 retval = 0;
11677                 m_freem(m);
11678                 goto done_with_input;
11679         }
11680         /*
11681          * If a segment with the ACK-bit set arrives in the SYN-SENT state
11682          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
11683          */
11684         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
11685             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
11686                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11687                 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11688                 return (1);
11689         }
11690         in_recovery = IN_RECOVERY(tp->t_flags);
11691         if (tiwin > bbr->r_ctl.rc_high_rwnd)
11692                 bbr->r_ctl.rc_high_rwnd = tiwin;
11693 #ifdef BBR_INVARIANTS
11694         if ((tp->t_inpcb->inp_flags & INP_DROPPED) ||
11695             (tp->t_inpcb->inp_flags2 & INP_FREED)) {
11696                 panic("tp:%p bbr:%p given a dropped inp:%p",
11697                     tp, bbr, tp->t_inpcb);
11698         }
11699 #endif
11700         bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp,
11701                                             (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11702         bbr->rtt_valid = 0;
11703         if (to.to_flags & TOF_TS) {
11704                 bbr->rc_ts_valid = 1;
11705                 bbr->r_ctl.last_inbound_ts = to.to_tsval;
11706         } else {
11707                 bbr->rc_ts_valid = 0;
11708                 bbr->r_ctl.last_inbound_ts = 0;
11709         }
11710         retval = (*bbr->r_substate) (m, th, so,
11711             tp, &to, drop_hdrlen,
11712             tlen, tiwin, thflags, nxt_pkt, iptos);
11713 #ifdef BBR_INVARIANTS
11714         if ((retval == 0) &&
11715             (tp->t_inpcb == NULL)) {
11716                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
11717                     retval, tp, prev_state);
11718         }
11719 #endif
11720         if (nxt_pkt == 0)
11721                 BBR_STAT_INC(bbr_rlock_left_ret0);
11722         else
11723                 BBR_STAT_INC(bbr_rlock_left_ret1);
11724         if (retval == 0) {
11725                 /*
11726                  * If retval is 1 the tcb is unlocked and most likely the tp
11727                  * is gone.
11728                  */
11729                 INP_WLOCK_ASSERT(tp->t_inpcb);
11730                 tcp_bbr_xmit_timer_commit(bbr, tp, cts);
11731                 if (bbr->rc_is_pkt_epoch_now)
11732                         bbr_set_pktepoch(bbr, cts, __LINE__);
11733                 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost));
11734                 if (nxt_pkt == 0) {
11735                         if (bbr->r_wanted_output != 0) {
11736                                 bbr->rc_output_starts_timer = 0;
11737                                 did_out = 1;
11738                                 (void)tp->t_fb->tfb_tcp_output(tp);
11739                         } else
11740                                 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0);
11741                 }
11742                 if ((nxt_pkt == 0) &&
11743                     ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
11744                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
11745                      (tp->t_flags & TF_DELACK) ||
11746                      ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
11747                       (tp->t_state <= TCPS_CLOSING)))) {
11748                         /*
11749                          * We could not send (probably in the hpts but
11750                          * stopped the timer)?
11751                          */
11752                         if ((tp->snd_max == tp->snd_una) &&
11753                             ((tp->t_flags & TF_DELACK) == 0) &&
11754                             (bbr->rc_inp->inp_in_hpts) &&
11755                             (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11756                                 /*
11757                                  * keep alive not needed if we are hptsi
11758                                  * output yet
11759                                  */
11760                                 ;
11761                         } else {
11762                                 if (bbr->rc_inp->inp_in_hpts) {
11763                                         tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
11764                                         if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11765                                             (TSTMP_GT(lcts, bbr->rc_pacer_started))) {
11766                                                 uint32_t del;
11767
11768                                                 del = lcts - bbr->rc_pacer_started;
11769                                                 if (bbr->r_ctl.rc_last_delay_val > del) {
11770                                                         BBR_STAT_INC(bbr_force_timer_start);
11771                                                         bbr->r_ctl.rc_last_delay_val -= del;
11772                                                         bbr->rc_pacer_started = lcts;
11773                                                 } else {
11774                                                         /* We are late */
11775                                                         bbr->r_ctl.rc_last_delay_val = 0;
11776                                                         BBR_STAT_INC(bbr_force_output);
11777                                                         (void)tp->t_fb->tfb_tcp_output(tp);
11778                                                 }
11779                                         }
11780                                 }
11781                                 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val,
11782                                     0);
11783                         }
11784                 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) {
11785                         /* Do we have the correct timer running? */
11786                         bbr_timer_audit(tp, bbr, lcts, &so->so_snd);
11787                 }
11788                 /* Do we have a new state */
11789                 if (bbr->r_state != tp->t_state)
11790                         bbr_set_state(tp, bbr, tiwin);
11791 done_with_input:
11792                 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out);
11793                 if (did_out)
11794                         bbr->r_wanted_output = 0;
11795 #ifdef BBR_INVARIANTS
11796                 if (tp->t_inpcb == NULL) {
11797                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
11798                             did_out,
11799                             retval, tp, prev_state);
11800                 }
11801 #endif
11802         }
11803         return (retval);
11804 }
11805
11806 static void
11807 bbr_log_type_hrdwtso(struct tcpcb *tp, struct tcp_bbr *bbr, int len, int mod, int what_we_can_send)
11808 {
11809         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
11810                 union tcp_log_stackspecific log;
11811                 struct timeval tv;
11812                 uint32_t cts;
11813
11814                 cts = tcp_get_usecs(&tv);
11815                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
11816                 log.u_bbr.flex1 = bbr->r_ctl.rc_pace_min_segs;
11817                 log.u_bbr.flex2 = what_we_can_send;
11818                 log.u_bbr.flex3 = bbr->r_ctl.rc_pace_max_segs;
11819                 log.u_bbr.flex4 = len;
11820                 log.u_bbr.flex5 = 0;
11821                 log.u_bbr.flex7 = mod;
11822                 log.u_bbr.flex8 = 1;
11823                 TCP_LOG_EVENTP(tp, NULL,
11824                     &tp->t_inpcb->inp_socket->so_rcv,
11825                     &tp->t_inpcb->inp_socket->so_snd,
11826                     TCP_HDWR_TLS, 0,
11827                     0, &log, false, &tv);
11828         }
11829 }
11830
11831 static void
11832 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
11833     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
11834 {
11835         struct timeval tv;
11836         int retval;
11837
11838         /* First lets see if we have old packets */
11839         if (tp->t_in_pkt) {
11840                 if (ctf_do_queued_segments(so, tp, 1)) {
11841                         m_freem(m);
11842                         return;
11843                 }
11844         }
11845         if (m->m_flags & M_TSTMP_LRO) {
11846                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
11847                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
11848         } else {
11849                 /* Should not be should we kassert instead? */
11850                 tcp_get_usecs(&tv);
11851         }
11852         retval = bbr_do_segment_nounlock(m, th, so, tp,
11853                                          drop_hdrlen, tlen, iptos, 0, &tv);
11854         if (retval == 0)
11855                 INP_WUNLOCK(tp->t_inpcb);
11856 }
11857
11858 /*
11859  * Return how much data can be sent without violating the
11860  * cwnd or rwnd.
11861  */
11862
11863 static inline uint32_t
11864 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin,
11865     uint32_t avail, int32_t sb_offset, uint32_t cts)
11866 {
11867         uint32_t len;
11868
11869         if (ctf_outstanding(tp) >= tp->snd_wnd) {
11870                 /* We never want to go over our peers rcv-window */
11871                 len = 0;
11872         } else {
11873                 uint32_t flight;
11874
11875                 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11876                 if (flight >= sendwin) {
11877                         /*
11878                          * We have in flight what we are allowed by cwnd (if
11879                          * it was rwnd blocking it would have hit above out
11880                          * >= tp->snd_wnd).
11881                          */
11882                         return (0);
11883                 }
11884                 len = sendwin - flight;
11885                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
11886                         /* We would send too much (beyond the rwnd) */
11887                         len = tp->snd_wnd - ctf_outstanding(tp);
11888                 }
11889                 if ((len + sb_offset) > avail) {
11890                         /*
11891                          * We don't have that much in the SB, how much is
11892                          * there?
11893                          */
11894                         len = avail - sb_offset;
11895                 }
11896         }
11897         return (len);
11898 }
11899
11900 static inline void
11901 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11902 {
11903 #ifdef NETFLIX_STATS
11904         KMOD_TCPSTAT_INC(tcps_sndpack_error);
11905         KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len);
11906 #endif
11907 }
11908
11909 static inline void
11910 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11911 {
11912         if (error) {
11913                 bbr_do_error_accounting(tp, bbr, rsm, len, error);
11914                 return;
11915         }
11916         if (rsm) {
11917                 if (rsm->r_flags & BBR_TLP) {
11918                         /*
11919                          * TLP should not count in retran count, but in its
11920                          * own bin
11921                          */
11922 #ifdef NETFLIX_STATS
11923                         tp->t_sndtlppack++;
11924                         tp->t_sndtlpbyte += len;
11925                         KMOD_TCPSTAT_INC(tcps_tlpresends);
11926                         KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len);
11927 #endif
11928                 } else {
11929                         /* Retransmit */
11930                         tp->t_sndrexmitpack++;
11931                         KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
11932                         KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
11933 #ifdef STATS
11934                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
11935                             len);
11936 #endif
11937                 }
11938                 /*
11939                  * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is
11940                  * sub-state
11941                  */
11942                 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len);
11943                 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) {
11944                         /* Non probe_bw log in 1, 2, or 4. */
11945                         counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len);
11946                 } else {
11947                         /*
11948                          * Log our probe state 3, and log also 5-13 to show
11949                          * us the recovery sub-state for the send. This
11950                          * means that 3 == (5+6+7+8+9+10+11+12+13)
11951                          */
11952                         counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len);
11953                         counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len);
11954                 }
11955                 /* Place in both 16's the totals of retransmitted */
11956                 counter_u64_add(bbr_state_lost[16], len);
11957                 counter_u64_add(bbr_state_resend[16], len);
11958                 /* Place in 17's the total sent */
11959                 counter_u64_add(bbr_state_resend[17], len);
11960                 counter_u64_add(bbr_state_lost[17], len);
11961
11962         } else {
11963                 /* New sends */
11964                 KMOD_TCPSTAT_INC(tcps_sndpack);
11965                 KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
11966                 /* Place in 17's the total sent */
11967                 counter_u64_add(bbr_state_resend[17], len);
11968                 counter_u64_add(bbr_state_lost[17], len);
11969 #ifdef STATS
11970                 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
11971                     len);
11972 #endif
11973         }
11974 }
11975
11976 static void
11977 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level)
11978 {
11979         if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) {
11980                 /*
11981                  * Limit the cwnd to not be above N x the target plus whats
11982                  * is outstanding. The target is based on the current b/w
11983                  * estimate.
11984                  */
11985                 uint32_t target;
11986
11987                 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT);
11988                 target += ctf_outstanding(tp);
11989                 target *= bbr_target_cwnd_mult_limit;
11990                 if (tp->snd_cwnd > target)
11991                         tp->snd_cwnd = target;
11992                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__);
11993         }
11994 }
11995
11996 static int
11997 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg)
11998 {
11999         /*
12000          * "adv" is the amount we could increase the window, taking into
12001          * account that we are limited by TCP_MAXWIN << tp->rcv_scale.
12002          */
12003         int32_t adv;
12004         int32_t oldwin;
12005
12006         adv = recwin;
12007         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
12008                 oldwin = (tp->rcv_adv - tp->rcv_nxt);
12009                 if (adv > oldwin)
12010                         adv -= oldwin;
12011                 else {
12012                         /* We can't increase the window */
12013                         adv = 0;
12014                 }
12015         } else
12016                 oldwin = 0;
12017
12018         /*
12019          * If the new window size ends up being the same as or less
12020          * than the old size when it is scaled, then don't force
12021          * a window update.
12022          */
12023         if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
12024                 return (0);
12025
12026         if (adv >= (2 * maxseg) &&
12027             (adv >= (so->so_rcv.sb_hiwat / 4) ||
12028             recwin <= (so->so_rcv.sb_hiwat / 8) ||
12029             so->so_rcv.sb_hiwat <= 8 * maxseg)) {
12030                 return (1);
12031         }
12032         if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat)
12033                 return (1);
12034         return (0);
12035 }
12036
12037 /*
12038  * Return 0 on success and a errno on failure to send.
12039  * Note that a 0 return may not mean we sent anything
12040  * if the TCB was on the hpts. A non-zero return
12041  * does indicate the error we got from ip[6]_output.
12042  */
12043 static int
12044 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv)
12045 {
12046         struct socket *so;
12047         int32_t len;
12048         uint32_t cts;
12049         uint32_t recwin, sendwin;
12050         int32_t sb_offset;
12051         int32_t flags, abandon, error = 0;
12052         struct tcp_log_buffer *lgb = NULL;
12053         struct mbuf *m;
12054         struct mbuf *mb;
12055         uint32_t if_hw_tsomaxsegcount = 0;
12056         uint32_t if_hw_tsomaxsegsize = 0;
12057         uint32_t if_hw_tsomax = 0;
12058         struct ip *ip = NULL;
12059 #ifdef TCPDEBUG
12060         struct ipovly *ipov = NULL;
12061 #endif
12062         struct tcp_bbr *bbr;
12063         struct tcphdr *th;
12064 #ifdef NETFLIX_TCPOUDP
12065         struct udphdr *udp = NULL;
12066 #endif
12067         u_char opt[TCP_MAXOLEN];
12068         unsigned ipoptlen, optlen, hdrlen;
12069 #ifdef NETFLIX_TCPOUDP
12070         unsigned ulen;
12071 #endif
12072         uint32_t bbr_seq;
12073         uint32_t delay_calc=0;
12074         uint8_t doing_tlp = 0;
12075         uint8_t local_options;
12076 #ifdef BBR_INVARIANTS
12077         uint8_t doing_retran_from = 0;
12078         uint8_t picked_up_retran = 0;
12079 #endif
12080         uint8_t wanted_cookie = 0;
12081         uint8_t more_to_rxt=0;
12082         int32_t prefetch_so_done = 0;
12083         int32_t prefetch_rsm = 0;
12084         uint32_t what_we_can = 0;
12085         uint32_t tot_len = 0;
12086         uint32_t rtr_cnt = 0;
12087         uint32_t maxseg, pace_max_segs, p_maxseg;
12088         int32_t csum_flags;
12089         int32_t hw_tls;
12090 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12091         unsigned ipsec_optlen = 0;
12092
12093 #endif
12094         volatile int32_t sack_rxmit;
12095         struct bbr_sendmap *rsm = NULL;
12096         int32_t tso, mtu;
12097         int force_tso = 0;
12098         struct tcpopt to;
12099         int32_t slot = 0;
12100         struct inpcb *inp;
12101         struct sockbuf *sb;
12102         uint32_t hpts_calling;
12103 #ifdef INET6
12104         struct ip6_hdr *ip6 = NULL;
12105         int32_t isipv6;
12106 #endif
12107         uint8_t app_limited = BBR_JR_SENT_DATA;
12108         uint8_t filled_all = 0;
12109         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
12110         /* We take a cache hit here */
12111         memcpy(&bbr->rc_tv, tv, sizeof(struct timeval));
12112         cts = tcp_tv_to_usectick(&bbr->rc_tv);
12113         inp = bbr->rc_inp;
12114         so = inp->inp_socket;
12115         sb = &so->so_snd;
12116 #ifdef KERN_TLS
12117         if (sb->sb_flags & SB_TLS_IFNET)
12118                 hw_tls = 1;
12119         else
12120 #endif
12121                 hw_tls = 0;
12122         kern_prefetch(sb, &maxseg);
12123         maxseg = tp->t_maxseg - bbr->rc_last_options;
12124         if (bbr_minseg(bbr) < maxseg) {
12125                 tcp_bbr_tso_size_check(bbr, cts);
12126         }
12127         /* Remove any flags that indicate we are pacing on the inp  */
12128         pace_max_segs = bbr->r_ctl.rc_pace_max_segs;
12129         p_maxseg = min(maxseg, pace_max_segs);
12130         INP_WLOCK_ASSERT(inp);
12131 #ifdef TCP_OFFLOAD
12132         if (tp->t_flags & TF_TOE)
12133                 return (tcp_offload_output(tp));
12134 #endif
12135
12136 #ifdef INET6
12137         if (bbr->r_state) {
12138                 /* Use the cache line loaded if possible */
12139                 isipv6 = bbr->r_is_v6;
12140         } else {
12141                 isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
12142         }
12143 #endif
12144         if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
12145             inp->inp_in_hpts) {
12146                 /*
12147                  * We are on the hpts for some timer but not hptsi output.
12148                  * Possibly remove from the hpts so we can send/recv etc.
12149                  */
12150                 if ((tp->t_flags & TF_ACKNOW) == 0) {
12151                         /*
12152                          * No immediate demand right now to send an ack, but
12153                          * the user may have read, making room for new data
12154                          * (a window update). If so we may want to cancel
12155                          * whatever timer is running (KEEP/DEL-ACK?) and
12156                          * continue to send out a window update. Or we may
12157                          * have gotten more data into the socket buffer to
12158                          * send.
12159                          */
12160                         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12161                                       (long)TCP_MAXWIN << tp->rcv_scale);
12162                         if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) &&
12163                             ((tcp_outflags[tp->t_state] & TH_RST) == 0) &&
12164                             ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <=
12165                             (tp->snd_max - tp->snd_una))) {
12166                                 /*
12167                                  * Nothing new to send and no window update
12168                                  * is needed to send. Lets just return and
12169                                  * let the timer-run off.
12170                                  */
12171                                 return (0);
12172                         }
12173                 }
12174                 tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12175                 bbr_timer_cancel(bbr, __LINE__, cts);
12176         }
12177         if (bbr->r_ctl.rc_last_delay_val) {
12178                 /* Calculate a rough delay for early escape to sending  */
12179                 if (SEQ_GT(cts, bbr->rc_pacer_started))
12180                         delay_calc = cts - bbr->rc_pacer_started;
12181                 if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12182                         delay_calc -= bbr->r_ctl.rc_last_delay_val;
12183                 else
12184                         delay_calc = 0;
12185         }
12186         /* Mark that we have called bbr_output(). */
12187         if ((bbr->r_timer_override) ||
12188             (tp->t_state < TCPS_ESTABLISHED)) {
12189                 /* Timeouts or early states are exempt */
12190                 if (inp->inp_in_hpts)
12191                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12192         } else if (inp->inp_in_hpts) {
12193                 if ((bbr->r_ctl.rc_last_delay_val) &&
12194                     (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
12195                     delay_calc) {
12196                         /*
12197                          * We were being paced for output and the delay has
12198                          * already exceeded when we were supposed to be
12199                          * called, lets go ahead and pull out of the hpts
12200                          * and call output.
12201                          */
12202                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1);
12203                         bbr->r_ctl.rc_last_delay_val = 0;
12204                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12205                 } else if (tp->t_state == TCPS_CLOSED) {
12206                         bbr->r_ctl.rc_last_delay_val = 0;
12207                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12208                 } else {
12209                         /*
12210                          * On the hpts, you shall not pass! even if ACKNOW
12211                          * is on, we will when the hpts fires, unless of
12212                          * course we are overdue.
12213                          */
12214                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1);
12215                         return (0);
12216                 }
12217         }
12218         bbr->rc_cwnd_limited = 0;
12219         if (bbr->r_ctl.rc_last_delay_val) {
12220                 /* recalculate the real delay and deal with over/under  */
12221                 if (SEQ_GT(cts, bbr->rc_pacer_started))
12222                         delay_calc = cts - bbr->rc_pacer_started;
12223                 else
12224                         delay_calc = 0;
12225                 if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12226                         /* Setup the delay which will be added in */
12227                         delay_calc -= bbr->r_ctl.rc_last_delay_val;
12228                 else {
12229                         /*
12230                          * We are early setup to adjust
12231                          * our slot time.
12232                          */
12233                         uint64_t merged_val;
12234
12235                         bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc);
12236                         bbr->r_agg_early_set = 1;
12237                         if (bbr->r_ctl.rc_hptsi_agg_delay) {
12238                                 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) {
12239                                         /* Nope our previous late cancels out the early */
12240                                         bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early;
12241                                         bbr->r_agg_early_set = 0;
12242                                         bbr->r_ctl.rc_agg_early = 0;
12243                                 } else {
12244                                         bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay;
12245                                         bbr->r_ctl.rc_hptsi_agg_delay = 0;
12246                                 }
12247                         }
12248                         merged_val = bbr->rc_pacer_started;
12249                         merged_val <<= 32;
12250                         merged_val |= bbr->r_ctl.rc_last_delay_val;
12251                         bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls,
12252                                                  bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val,
12253                                                  bbr->r_agg_early_set, 3);
12254                         bbr->r_ctl.rc_last_delay_val = 0;
12255                         BBR_STAT_INC(bbr_early);
12256                         delay_calc = 0;
12257                 }
12258         } else {
12259                 /* We were not delayed due to hptsi */
12260                 if (bbr->r_agg_early_set)
12261                         bbr->r_ctl.rc_agg_early = 0;
12262                 bbr->r_agg_early_set = 0;
12263                 delay_calc = 0;
12264         }
12265         if (delay_calc) {
12266                 /*
12267                  * We had a hptsi delay which means we are falling behind on
12268                  * sending at the expected rate. Calculate an extra amount
12269                  * of data we can send, if any, to put us back on track.
12270                  */
12271                 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay)
12272                         bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff;
12273                 else
12274                         bbr->r_ctl.rc_hptsi_agg_delay += delay_calc;
12275         }
12276         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12277         if ((tp->snd_una == tp->snd_max) &&
12278             (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
12279             (sbavail(sb))) {
12280                 /*
12281                  * Ok we have been idle with nothing outstanding
12282                  * we possibly need to start fresh with either a new
12283                  * suite of states or a fast-ramp up.
12284                  */
12285                 bbr_restart_after_idle(bbr,
12286                                        cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time));
12287         }
12288         /*
12289          * Now was there a hptsi delay where we are behind? We only count
12290          * being behind if: a) We are not in recovery. b) There was a delay.
12291          * <and> c) We had room to send something.
12292          *
12293          */
12294         hpts_calling = inp->inp_hpts_calls;
12295         inp->inp_hpts_calls = 0;
12296         if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
12297                 if (bbr_process_timers(tp, bbr, cts, hpts_calling)) {
12298                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1);
12299                         return (0);
12300                 }
12301         }
12302         bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12303         if (hpts_calling &&
12304             (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12305                 bbr->r_ctl.rc_last_delay_val = 0;
12306         }
12307         bbr->r_timer_override = 0;
12308         bbr->r_wanted_output = 0;
12309         /*
12310          * For TFO connections in SYN_RECEIVED, only allow the initial
12311          * SYN|ACK and those sent by the retransmit timer.
12312          */
12313         if (IS_FASTOPEN(tp->t_flags) &&
12314             ((tp->t_state == TCPS_SYN_RECEIVED) ||
12315              (tp->t_state == TCPS_SYN_SENT)) &&
12316             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
12317             (tp->t_rxtshift == 0)) {    /* not a retransmit */
12318                 len = 0;
12319                 goto just_return_nolock;
12320         }
12321         /*
12322          * Before sending anything check for a state update. For hpts
12323          * calling without input this is important. If its input calling
12324          * then this was already done.
12325          */
12326         if (bbr->rc_use_google == 0)
12327                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12328 again:
12329         /*
12330          * If we've recently taken a timeout, snd_max will be greater than
12331          * snd_max. BBR in general does not pay much attention to snd_nxt
12332          * for historic reasons the persist timer still uses it. This means
12333          * we have to look at it. All retransmissions that are not persits
12334          * use the rsm that needs to be sent so snd_nxt is ignored. At the
12335          * end of this routine we pull snd_nxt always up to snd_max.
12336          */
12337         doing_tlp = 0;
12338 #ifdef BBR_INVARIANTS
12339         doing_retran_from = picked_up_retran = 0;
12340 #endif
12341         error = 0;
12342         tso = 0;
12343         slot = 0;
12344         mtu = 0;
12345         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12346         sb_offset = tp->snd_max - tp->snd_una;
12347         flags = tcp_outflags[tp->t_state];
12348         sack_rxmit = 0;
12349         len = 0;
12350         rsm = NULL;
12351         if (flags & TH_RST) {
12352                 SOCKBUF_LOCK(sb);
12353                 goto send;
12354         }
12355 recheck_resend:
12356         while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
12357                 /* We need to always have one in reserve */
12358                 rsm = bbr_alloc(bbr);
12359                 if (rsm == NULL) {
12360                         error = ENOMEM;
12361                         /* Lie to get on the hpts */
12362                         tot_len = tp->t_maxseg;
12363                         if (hpts_calling)
12364                                 /* Retry in a ms */
12365                                 slot = 1001;
12366                         goto just_return_nolock;
12367                 }
12368                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
12369                 bbr->r_ctl.rc_free_cnt++;
12370                 rsm = NULL;
12371         }
12372         /* What do we send, a resend? */
12373         if (bbr->r_ctl.rc_resend == NULL) {
12374                 /* Check for rack timeout */
12375                 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
12376                 if (bbr->r_ctl.rc_resend) {
12377 #ifdef BBR_INVARIANTS
12378                         picked_up_retran = 1;
12379 #endif
12380                         bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend);
12381                 }
12382         }
12383         if (bbr->r_ctl.rc_resend) {
12384                 rsm = bbr->r_ctl.rc_resend;
12385 #ifdef BBR_INVARIANTS
12386                 doing_retran_from = 1;
12387 #endif
12388                 /* Remove any TLP flags its a RACK or T-O */
12389                 rsm->r_flags &= ~BBR_TLP;
12390                 bbr->r_ctl.rc_resend = NULL;
12391                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
12392 #ifdef BBR_INVARIANTS
12393                         panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n",
12394                             tp, bbr, rsm, rsm->r_start, tp->snd_una);
12395                         goto recheck_resend;
12396 #else
12397                         /* TSNH */
12398                         rsm = NULL;
12399                         goto recheck_resend;
12400 #endif
12401                 }
12402                 rtr_cnt++;
12403                 if (rsm->r_flags & BBR_HAS_SYN) {
12404                         /* Only retransmit a SYN by itself */
12405                         len = 0;
12406                         if ((flags & TH_SYN) == 0) {
12407                                 /* Huh something is wrong */
12408                                 rsm->r_start++;
12409                                 if (rsm->r_start == rsm->r_end) {
12410                                         /* Clean it up, somehow we missed the ack? */
12411                                         bbr_log_syn(tp, NULL);
12412                                 } else {
12413                                         /* TFO with data? */
12414                                         rsm->r_flags &= ~BBR_HAS_SYN;
12415                                         len = rsm->r_end - rsm->r_start;
12416                                 }
12417                         } else {
12418                                 /* Retransmitting SYN */
12419                                 rsm = NULL;
12420                                 SOCKBUF_LOCK(sb);
12421                                 goto send;
12422                         }
12423                 } else
12424                         len = rsm->r_end - rsm->r_start;
12425                 if ((bbr->rc_resends_use_tso == 0) &&
12426 #ifdef KERN_TLS
12427                     ((sb->sb_flags & SB_TLS_IFNET) == 0) &&
12428 #endif
12429                     (len > maxseg)) {
12430                         len = maxseg;
12431                         more_to_rxt = 1;
12432                 }
12433                 sb_offset = rsm->r_start - tp->snd_una;
12434                 if (len > 0) {
12435                         sack_rxmit = 1;
12436                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
12437                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
12438                             min(len, maxseg));
12439                 } else {
12440                         /* I dont think this can happen */
12441                         rsm = NULL;
12442                         goto recheck_resend;
12443                 }
12444                 BBR_STAT_INC(bbr_resends_set);
12445         } else if (bbr->r_ctl.rc_tlp_send) {
12446                 /*
12447                  * Tail loss probe
12448                  */
12449                 doing_tlp = 1;
12450                 rsm = bbr->r_ctl.rc_tlp_send;
12451                 bbr->r_ctl.rc_tlp_send = NULL;
12452                 sack_rxmit = 1;
12453                 len = rsm->r_end - rsm->r_start;
12454                 rtr_cnt++;
12455                 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12456                         len = maxseg;
12457
12458                 if (SEQ_GT(tp->snd_una, rsm->r_start)) {
12459 #ifdef BBR_INVARIANTS
12460                         panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u",
12461                             tp, bbr, tp->snd_una, rsm, rsm->r_start);
12462 #else
12463                         /* TSNH */
12464                         rsm = NULL;
12465                         goto recheck_resend;
12466 #endif
12467                 }
12468                 sb_offset = rsm->r_start - tp->snd_una;
12469                 BBR_STAT_INC(bbr_tlp_set);
12470         }
12471         /*
12472          * Enforce a connection sendmap count limit if set
12473          * as long as we are not retransmiting.
12474          */
12475         if ((rsm == NULL) &&
12476             (V_tcp_map_entries_limit > 0) &&
12477             (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
12478                 BBR_STAT_INC(bbr_alloc_limited);
12479                 if (!bbr->alloc_limit_reported) {
12480                         bbr->alloc_limit_reported = 1;
12481                         BBR_STAT_INC(bbr_alloc_limited_conns);
12482                 }
12483                 goto just_return_nolock;
12484         }
12485 #ifdef BBR_INVARIANTS
12486         if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) {
12487                 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u",
12488                     tp, bbr, rsm, sb_offset, len);
12489         }
12490 #endif
12491         /*
12492          * Get standard flags, and add SYN or FIN if requested by 'hidden'
12493          * state flags.
12494          */
12495         if (tp->t_flags & TF_NEEDFIN && (rsm == NULL))
12496                 flags |= TH_FIN;
12497         if (tp->t_flags & TF_NEEDSYN)
12498                 flags |= TH_SYN;
12499
12500         if (rsm && (rsm->r_flags & BBR_HAS_FIN)) {
12501                 /* we are retransmitting the fin */
12502                 len--;
12503                 if (len) {
12504                         /*
12505                          * When retransmitting data do *not* include the
12506                          * FIN. This could happen from a TLP probe if we
12507                          * allowed data with a FIN.
12508                          */
12509                         flags &= ~TH_FIN;
12510                 }
12511         } else if (rsm) {
12512                 if (flags & TH_FIN)
12513                         flags &= ~TH_FIN;
12514         }
12515         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
12516                 void *end_rsm;
12517
12518                 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
12519                 if (end_rsm)
12520                         kern_prefetch(end_rsm, &prefetch_rsm);
12521                 prefetch_rsm = 1;
12522         }
12523         SOCKBUF_LOCK(sb);
12524         /*
12525          * If snd_nxt == snd_max and we have transmitted a FIN, the
12526          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
12527          * negative length.  This can also occur when TCP opens up its
12528          * congestion window while receiving additional duplicate acks after
12529          * fast-retransmit because TCP will reset snd_nxt to snd_max after
12530          * the fast-retransmit.
12531          *
12532          * In the normal retransmit-FIN-only case, however, snd_nxt will be
12533          * set to snd_una, the sb_offset will be 0, and the length may wind
12534          * up 0.
12535          *
12536          * If sack_rxmit is true we are retransmitting from the scoreboard
12537          * in which case len is already set.
12538          */
12539         if (sack_rxmit == 0) {
12540                 uint32_t avail;
12541
12542                 avail = sbavail(sb);
12543                 if (SEQ_GT(tp->snd_max, tp->snd_una))
12544                         sb_offset = tp->snd_max - tp->snd_una;
12545                 else
12546                         sb_offset = 0;
12547                 if (bbr->rc_tlp_new_data) {
12548                         /* TLP is forcing out new data */
12549                         uint32_t tlplen;
12550
12551                         doing_tlp = 1;
12552                         tlplen = maxseg;
12553
12554                         if (tlplen > (uint32_t)(avail - sb_offset)) {
12555                                 tlplen = (uint32_t)(avail - sb_offset);
12556                         }
12557                         if (tlplen > tp->snd_wnd) {
12558                                 len = tp->snd_wnd;
12559                         } else {
12560                                 len = tlplen;
12561                         }
12562                         bbr->rc_tlp_new_data = 0;
12563                 } else {
12564                         what_we_can = len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts);
12565                         if ((len < p_maxseg) &&
12566                             (bbr->rc_in_persist == 0) &&
12567                             (ctf_outstanding(tp) >= (2 * p_maxseg)) &&
12568                             ((avail - sb_offset) >= p_maxseg)) {
12569                                 /*
12570                                  * We are not completing whats in the socket
12571                                  * buffer (i.e. there is at least a segment
12572                                  * waiting to send) and we have 2 or more
12573                                  * segments outstanding. There is no sense
12574                                  * of sending a little piece. Lets defer and
12575                                  * and wait until we can send a whole
12576                                  * segment.
12577                                  */
12578                                 len = 0;
12579                         }
12580                         if (bbr->rc_in_persist) {
12581                                 /*
12582                                  * We are in persists, figure out if
12583                                  * a retransmit is available (maybe the previous
12584                                  * persists we sent) or if we have to send new
12585                                  * data.
12586                                  */
12587                                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
12588                                 if (rsm) {
12589                                         len = rsm->r_end - rsm->r_start;
12590                                         if (rsm->r_flags & BBR_HAS_FIN)
12591                                                 len--;
12592                                         if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12593                                                 len = maxseg;
12594                                         if (len > 1)
12595                                                 BBR_STAT_INC(bbr_persist_reneg);
12596                                         /*
12597                                          * XXXrrs we could force the len to
12598                                          * 1 byte here to cause the chunk to
12599                                          * split apart.. but that would then
12600                                          * mean we always retransmit it as
12601                                          * one byte even after the window
12602                                          * opens.
12603                                          */
12604                                         sack_rxmit = 1;
12605                                         sb_offset = rsm->r_start - tp->snd_una;
12606                                 } else {
12607                                         /*
12608                                          * First time through in persists or peer
12609                                          * acked our one byte. Though we do have
12610                                          * to have something in the sb.
12611                                          */
12612                                         len = 1;
12613                                         sb_offset = 0;
12614                                         if (avail == 0)
12615                                             len = 0;
12616                                 }
12617                         }
12618                 }
12619         }
12620         if (prefetch_so_done == 0) {
12621                 kern_prefetch(so, &prefetch_so_done);
12622                 prefetch_so_done = 1;
12623         }
12624         /*
12625          * Lop off SYN bit if it has already been sent.  However, if this is
12626          * SYN-SENT state and if segment contains data and if we don't know
12627          * that foreign host supports TAO, suppress sending segment.
12628          */
12629         if ((flags & TH_SYN) && (rsm == NULL) &&
12630             SEQ_GT(tp->snd_max, tp->snd_una)) {
12631                 if (tp->t_state != TCPS_SYN_RECEIVED)
12632                         flags &= ~TH_SYN;
12633                 /*
12634                  * When sending additional segments following a TFO SYN|ACK,
12635                  * do not include the SYN bit.
12636                  */
12637                 if (IS_FASTOPEN(tp->t_flags) &&
12638                     (tp->t_state == TCPS_SYN_RECEIVED))
12639                         flags &= ~TH_SYN;
12640                 sb_offset--, len++;
12641                 if (sbavail(sb) == 0)
12642                         len = 0;
12643         } else if ((flags & TH_SYN) && rsm) {
12644                 /*
12645                  * Subtract one from the len for the SYN being
12646                  * retransmitted.
12647                  */
12648                 len--;
12649         }
12650         /*
12651          * Be careful not to send data and/or FIN on SYN segments. This
12652          * measure is needed to prevent interoperability problems with not
12653          * fully conformant TCP implementations.
12654          */
12655         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
12656                 len = 0;
12657                 flags &= ~TH_FIN;
12658         }
12659         /*
12660          * On TFO sockets, ensure no data is sent in the following cases:
12661          *
12662          *  - When retransmitting SYN|ACK on a passively-created socket
12663          *  - When retransmitting SYN on an actively created socket
12664          *  - When sending a zero-length cookie (cookie request) on an
12665          *    actively created socket
12666          *  - When the socket is in the CLOSED state (RST is being sent)
12667          */
12668         if (IS_FASTOPEN(tp->t_flags) &&
12669             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
12670              ((tp->t_state == TCPS_SYN_SENT) &&
12671               (tp->t_tfo_client_cookie_len == 0)) ||
12672              (flags & TH_RST))) {
12673                 len = 0;
12674                 sack_rxmit = 0;
12675                 rsm = NULL;
12676         }
12677         /* Without fast-open there should never be data sent on a SYN */
12678         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags)))
12679                 len = 0;
12680         if (len <= 0) {
12681                 /*
12682                  * If FIN has been sent but not acked, but we haven't been
12683                  * called to retransmit, len will be < 0.  Otherwise, window
12684                  * shrank after we sent into it.  If window shrank to 0,
12685                  * cancel pending retransmit, pull snd_nxt back to (closed)
12686                  * window, and set the persist timer if it isn't already
12687                  * going.  If the window didn't close completely, just wait
12688                  * for an ACK.
12689                  *
12690                  * We also do a general check here to ensure that we will
12691                  * set the persist timer when we have data to send, but a
12692                  * 0-byte window. This makes sure the persist timer is set
12693                  * even if the packet hits one of the "goto send" lines
12694                  * below.
12695                  */
12696                 len = 0;
12697                 if ((tp->snd_wnd == 0) &&
12698                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12699                     (tp->snd_una == tp->snd_max) &&
12700                     (sb_offset < (int)sbavail(sb))) {
12701                         /*
12702                          * Not enough room in the rwnd to send
12703                          * a paced segment out.
12704                          */
12705                         bbr_enter_persist(tp, bbr, cts, __LINE__);
12706                 }
12707         } else if ((rsm == NULL) &&
12708                    (doing_tlp == 0) &&
12709                    (len < bbr->r_ctl.rc_pace_max_segs)) {
12710                 /*
12711                  * We are not sending a full segment for
12712                  * some reason. Should we not send anything (think
12713                  * sws or persists)?
12714                  */
12715                 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12716                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12717                     (len < (int)(sbavail(sb) - sb_offset))) {
12718                         /*
12719                          * Here the rwnd is less than
12720                          * the pacing size, this is not a retransmit,
12721                          * we are established and
12722                          * the send is not the last in the socket buffer
12723                          * lets not send, and possibly enter persists.
12724                          */
12725                         len = 0;
12726                         if (tp->snd_max == tp->snd_una)
12727                                 bbr_enter_persist(tp, bbr, cts, __LINE__);
12728                 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) &&
12729                            (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12730                                                  bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12731                            (len < (int)(sbavail(sb) - sb_offset)) &&
12732                            (len < bbr_minseg(bbr))) {
12733                         /*
12734                          * Here we are not retransmitting, and
12735                          * the cwnd is not so small that we could
12736                          * not send at least a min size (rxt timer
12737                          * not having gone off), We have 2 segments or
12738                          * more already in flight, its not the tail end
12739                          * of the socket buffer  and the cwnd is blocking
12740                          * us from sending out minimum pacing segment size.
12741                          * Lets not send anything.
12742                          */
12743                         bbr->rc_cwnd_limited = 1;
12744                         len = 0;
12745                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
12746                             min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12747                            (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12748                                                  bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12749                            (len < (int)(sbavail(sb) - sb_offset)) &&
12750                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
12751                         /*
12752                          * Here we have a send window but we have
12753                          * filled it up and we can't send another pacing segment.
12754                          * We also have in flight more than 2 segments
12755                          * and we are not completing the sb i.e. we allow
12756                          * the last bytes of the sb to go out even if
12757                          * its not a full pacing segment.
12758                          */
12759                         len = 0;
12760                 }
12761         }
12762         /* len will be >= 0 after this point. */
12763         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
12764         tcp_sndbuf_autoscale(tp, so, sendwin);
12765         /*
12766          *
12767          */
12768         if (bbr->rc_in_persist &&
12769             len &&
12770             (rsm == NULL) &&
12771             (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) {
12772                 /*
12773                  * We are in persist, not doing a retransmit and don't have enough space
12774                  * yet to send a full TSO. So is it at the end of the sb
12775                  * if so we need to send else nuke to 0 and don't send.
12776                  */
12777                 int sbleft;
12778                 if (sbavail(sb) > sb_offset)
12779                         sbleft = sbavail(sb) - sb_offset;
12780                 else
12781                         sbleft = 0;
12782                 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) {
12783                         /* not at end of sb lets not send */
12784                         len = 0;
12785                 }
12786         }
12787         /*
12788          * Decide if we can use TCP Segmentation Offloading (if supported by
12789          * hardware).
12790          *
12791          * TSO may only be used if we are in a pure bulk sending state.  The
12792          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
12793          * options prevent using TSO.  With TSO the TCP header is the same
12794          * (except for the sequence number) for all generated packets.  This
12795          * makes it impossible to transmit any options which vary per
12796          * generated segment or packet.
12797          *
12798          * IPv4 handling has a clear separation of ip options and ip header
12799          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen()
12800          * does the right thing below to provide length of just ip options
12801          * and thus checking for ipoptlen is enough to decide if ip options
12802          * are present.
12803          */
12804 #ifdef INET6
12805         if (isipv6)
12806                 ipoptlen = ip6_optlen(inp);
12807         else
12808 #endif
12809         if (inp->inp_options)
12810                 ipoptlen = inp->inp_options->m_len -
12811                     offsetof(struct ipoption, ipopt_list);
12812         else
12813                 ipoptlen = 0;
12814 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12815         /*
12816          * Pre-calculate here as we save another lookup into the darknesses
12817          * of IPsec that way and can actually decide if TSO is ok.
12818          */
12819 #ifdef INET6
12820         if (isipv6 && IPSEC_ENABLED(ipv6))
12821                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
12822 #ifdef INET
12823         else
12824 #endif
12825 #endif                          /* INET6 */
12826 #ifdef INET
12827         if (IPSEC_ENABLED(ipv4))
12828                 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
12829 #endif                          /* INET */
12830 #endif                          /* IPSEC */
12831 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12832         ipoptlen += ipsec_optlen;
12833 #endif
12834         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
12835             (len > maxseg) &&
12836             (tp->t_port == 0) &&
12837             ((tp->t_flags & TF_SIGNATURE) == 0) &&
12838             tp->rcv_numsacks == 0 &&
12839             ipoptlen == 0)
12840                 tso = 1;
12841
12842         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12843             (long)TCP_MAXWIN << tp->rcv_scale);
12844         /*
12845          * Sender silly window avoidance.   We transmit under the following
12846          * conditions when len is non-zero:
12847          *
12848          * - We have a full segment (or more with TSO) - This is the last
12849          * buffer in a write()/send() and we are either idle or running
12850          * NODELAY - we've timed out (e.g. persist timer) - we have more
12851          * then 1/2 the maximum send window's worth of data (receiver may be
12852          * limited the window size) - we need to retransmit
12853          */
12854         if (rsm)
12855                 goto send;
12856         if (len) {
12857                 if (sack_rxmit)
12858                         goto send;
12859                 if (len >= p_maxseg)
12860                         goto send;
12861                 /*
12862                  * NOTE! on localhost connections an 'ack' from the remote
12863                  * end may occur synchronously with the output and cause us
12864                  * to flush a buffer queued with moretocome.  XXX
12865                  *
12866                  */
12867                 if (((tp->t_flags & TF_MORETOCOME) == 0) &&     /* normal case */
12868                     ((tp->t_flags & TF_NODELAY) ||
12869                     ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) &&
12870                     (tp->t_flags & TF_NOPUSH) == 0) {
12871                         goto send;
12872                 }
12873                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
12874                         goto send;
12875                 }
12876                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
12877                         goto send;
12878                 }
12879         }
12880         /*
12881          * Sending of standalone window updates.
12882          *
12883          * Window updates are important when we close our window due to a
12884          * full socket buffer and are opening it again after the application
12885          * reads data from it.  Once the window has opened again and the
12886          * remote end starts to send again the ACK clock takes over and
12887          * provides the most current window information.
12888          *
12889          * We must avoid the silly window syndrome whereas every read from
12890          * the receive buffer, no matter how small, causes a window update
12891          * to be sent.  We also should avoid sending a flurry of window
12892          * updates when the socket buffer had queued a lot of data and the
12893          * application is doing small reads.
12894          *
12895          * Prevent a flurry of pointless window updates by only sending an
12896          * update when we can increase the advertized window by more than
12897          * 1/4th of the socket buffer capacity.  When the buffer is getting
12898          * full or is very small be more aggressive and send an update
12899          * whenever we can increase by two mss sized segments. In all other
12900          * situations the ACK's to new incoming data will carry further
12901          * window increases.
12902          *
12903          * Don't send an independent window update if a delayed ACK is
12904          * pending (it will get piggy-backed on it) or the remote side
12905          * already has done a half-close and won't send more data.  Skip
12906          * this if the connection is in T/TCP half-open state.
12907          */
12908         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
12909             !(tp->t_flags & TF_DELACK) &&
12910             !TCPS_HAVERCVDFIN(tp->t_state)) {
12911                 /* Check to see if we should do a window update */
12912                 if (bbr_window_update_needed(tp, so, recwin, maxseg))
12913                         goto send;
12914         }
12915         /*
12916          * Send if we owe the peer an ACK, RST, SYN.  ACKNOW
12917          * is also a catch-all for the retransmit timer timeout case.
12918          */
12919         if (tp->t_flags & TF_ACKNOW) {
12920                 goto send;
12921         }
12922         if (flags & TH_RST) {
12923                 /* Always send a RST if one is due */
12924                 goto send;
12925         }
12926         if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) {
12927                 goto send;
12928         }
12929         /*
12930          * If our state indicates that FIN should be sent and we have not
12931          * yet done so, then we need to send.
12932          */
12933         if (flags & TH_FIN &&
12934             ((tp->t_flags & TF_SENTFIN) == 0)) {
12935                 goto send;
12936         }
12937         /*
12938          * No reason to send a segment, just return.
12939          */
12940 just_return:
12941         SOCKBUF_UNLOCK(sb);
12942 just_return_nolock:
12943         if (tot_len)
12944                 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
12945         if (bbr->rc_no_pacing)
12946                 slot = 0;
12947         if (tot_len == 0) {
12948                 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >=
12949                     tp->snd_wnd) {
12950                         BBR_STAT_INC(bbr_rwnd_limited);
12951                         app_limited = BBR_JR_RWND_LIMITED;
12952                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12953                         if ((bbr->rc_in_persist == 0) &&
12954                             TCPS_HAVEESTABLISHED(tp->t_state) &&
12955                             (tp->snd_max == tp->snd_una) &&
12956                             sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
12957                                 /* No send window.. we must enter persist */
12958                                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
12959                         }
12960                 } else if (ctf_outstanding(tp) >= sbavail(sb)) {
12961                         BBR_STAT_INC(bbr_app_limited);
12962                         app_limited = BBR_JR_APP_LIMITED;
12963                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12964                 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12965                                                  bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) {
12966                         BBR_STAT_INC(bbr_cwnd_limited);
12967                         app_limited = BBR_JR_CWND_LIMITED;
12968                         bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12969                                                                         bbr->r_ctl.rc_lost_bytes)));
12970                         bbr->rc_cwnd_limited = 1;
12971                 } else {
12972                         BBR_STAT_INC(bbr_app_limited);
12973                         app_limited = BBR_JR_APP_LIMITED;
12974                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12975                 }
12976                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
12977                 bbr->r_agg_early_set = 0;
12978                 bbr->r_ctl.rc_agg_early = 0;
12979                 bbr->r_ctl.rc_last_delay_val = 0;
12980         } else if (bbr->rc_use_google == 0)
12981                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12982         /* Are we app limited? */
12983         if ((app_limited == BBR_JR_APP_LIMITED) ||
12984             (app_limited == BBR_JR_RWND_LIMITED)) {
12985                 /**
12986                  * We are application limited.
12987                  */
12988                 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12989                                                                        bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered);
12990         }
12991         if (tot_len == 0)
12992                 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1);
12993         /* Dont update the time if we did not send */
12994         bbr->r_ctl.rc_last_delay_val = 0;
12995         bbr->rc_output_starts_timer = 1;
12996         bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len);
12997         bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len);
12998         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12999                 /* Make sure snd_nxt is drug up */
13000                 tp->snd_nxt = tp->snd_max;
13001         }
13002         return (error);
13003
13004 send:
13005         if (doing_tlp == 0) {
13006                 /*
13007                  * Data not a TLP, and its not the rxt firing. If it is the
13008                  * rxt firing, we want to leave the tlp_in_progress flag on
13009                  * so we don't send another TLP. It has to be a rack timer
13010                  * or normal send (response to acked data) to clear the tlp
13011                  * in progress flag.
13012                  */
13013                 bbr->rc_tlp_in_progress = 0;
13014                 bbr->rc_tlp_rtx_out = 0;
13015         } else {
13016                 /*
13017                  * Its a TLP.
13018                  */
13019                 bbr->rc_tlp_in_progress = 1;
13020         }
13021         bbr_timer_cancel(bbr, __LINE__, cts);
13022         if (rsm == NULL) {
13023                 if (sbused(sb) > 0) {
13024                         /*
13025                          * This is sub-optimal. We only send a stand alone
13026                          * FIN on its own segment.
13027                          */
13028                         if (flags & TH_FIN) {
13029                                 flags &= ~TH_FIN;
13030                                 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) {
13031                                         /* Lets not send this */
13032                                         slot = 0;
13033                                         goto just_return;
13034                                 }
13035                         }
13036                 }
13037         } else {
13038                 /*
13039                  * We do *not* send a FIN on a retransmit if it has data.
13040                  * The if clause here where len > 1 should never come true.
13041                  */
13042                 if ((len > 0) &&
13043                     (((rsm->r_flags & BBR_HAS_FIN) == 0) &&
13044                     (flags & TH_FIN))) {
13045                         flags &= ~TH_FIN;
13046                         len--;
13047                 }
13048         }
13049         SOCKBUF_LOCK_ASSERT(sb);
13050         if (len > 0) {
13051                 if ((tp->snd_una == tp->snd_max) &&
13052                     (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
13053                         /*
13054                          * This qualifies as a RTT_PROBE session since we
13055                          * drop the data outstanding to nothing and waited
13056                          * more than bbr_rtt_probe_time.
13057                          */
13058                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
13059                         bbr_set_reduced_rtt(bbr, cts, __LINE__);
13060                 }
13061                 if (len >= maxseg)
13062                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
13063                 else
13064                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
13065         }
13066         /*
13067          * Before ESTABLISHED, force sending of initial options unless TCP
13068          * set not to do any options. NOTE: we assume that the IP/TCP header
13069          * plus TCP options always fit in a single mbuf, leaving room for a
13070          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
13071          * + optlen <= MCLBYTES
13072          */
13073         optlen = 0;
13074 #ifdef INET6
13075         if (isipv6)
13076                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
13077         else
13078 #endif
13079                 hdrlen = sizeof(struct tcpiphdr);
13080
13081         /*
13082          * Compute options for segment. We only have to care about SYN and
13083          * established connection segments.  Options for SYN-ACK segments
13084          * are handled in TCP syncache.
13085          */
13086         to.to_flags = 0;
13087         local_options = 0;
13088         if ((tp->t_flags & TF_NOOPT) == 0) {
13089                 /* Maximum segment size. */
13090                 if (flags & TH_SYN) {
13091                         to.to_mss = tcp_mssopt(&inp->inp_inc);
13092 #ifdef NETFLIX_TCPOUDP
13093                         if (tp->t_port)
13094                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
13095 #endif
13096                         to.to_flags |= TOF_MSS;
13097                         /*
13098                          * On SYN or SYN|ACK transmits on TFO connections,
13099                          * only include the TFO option if it is not a
13100                          * retransmit, as the presence of the TFO option may
13101                          * have caused the original SYN or SYN|ACK to have
13102                          * been dropped by a middlebox.
13103                          */
13104                         if (IS_FASTOPEN(tp->t_flags) &&
13105                             (tp->t_rxtshift == 0)) {
13106                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
13107                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
13108                                         to.to_tfo_cookie =
13109                                             (u_int8_t *)&tp->t_tfo_cookie.server;
13110                                         to.to_flags |= TOF_FASTOPEN;
13111                                         wanted_cookie = 1;
13112                                 } else if (tp->t_state == TCPS_SYN_SENT) {
13113                                         to.to_tfo_len =
13114                                             tp->t_tfo_client_cookie_len;
13115                                         to.to_tfo_cookie =
13116                                             tp->t_tfo_cookie.client;
13117                                         to.to_flags |= TOF_FASTOPEN;
13118                                         wanted_cookie = 1;
13119                                 }
13120                         }
13121                 }
13122                 /* Window scaling. */
13123                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
13124                         to.to_wscale = tp->request_r_scale;
13125                         to.to_flags |= TOF_SCALE;
13126                 }
13127                 /* Timestamps. */
13128                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
13129                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
13130                         to.to_tsval =   tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset;
13131                         to.to_tsecr = tp->ts_recent;
13132                         to.to_flags |= TOF_TS;
13133                         local_options += TCPOLEN_TIMESTAMP + 2;
13134                 }
13135                 /* Set receive buffer autosizing timestamp. */
13136                 if (tp->rfbuf_ts == 0 &&
13137                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
13138                         tp->rfbuf_ts =  tcp_tv_to_mssectick(&bbr->rc_tv);
13139                 /* Selective ACK's. */
13140                 if (flags & TH_SYN)
13141                         to.to_flags |= TOF_SACKPERM;
13142                 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13143                     tp->rcv_numsacks > 0) {
13144                         to.to_flags |= TOF_SACK;
13145                         to.to_nsacks = tp->rcv_numsacks;
13146                         to.to_sacks = (u_char *)tp->sackblks;
13147                 }
13148 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13149                 /* TCP-MD5 (RFC2385). */
13150                 if (tp->t_flags & TF_SIGNATURE)
13151                         to.to_flags |= TOF_SIGNATURE;
13152 #endif                          /* TCP_SIGNATURE */
13153
13154                 /* Processing the options. */
13155                 hdrlen += (optlen = tcp_addoptions(&to, opt));
13156                 /*
13157                  * If we wanted a TFO option to be added, but it was unable
13158                  * to fit, ensure no data is sent.
13159                  */
13160                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
13161                     !(to.to_flags & TOF_FASTOPEN))
13162                         len = 0;
13163         }
13164 #ifdef NETFLIX_TCPOUDP
13165         if (tp->t_port) {
13166                 if (V_tcp_udp_tunneling_port == 0) {
13167                         /* The port was removed?? */
13168                         SOCKBUF_UNLOCK(&so->so_snd);
13169                         return (EHOSTUNREACH);
13170                 }
13171                 hdrlen += sizeof(struct udphdr);
13172         }
13173 #endif
13174 #ifdef INET6
13175         if (isipv6)
13176                 ipoptlen = ip6_optlen(tp->t_inpcb);
13177         else
13178 #endif
13179         if (tp->t_inpcb->inp_options)
13180                 ipoptlen = tp->t_inpcb->inp_options->m_len -
13181                     offsetof(struct ipoption, ipopt_list);
13182         else
13183                 ipoptlen = 0;
13184         ipoptlen = 0;
13185 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
13186         ipoptlen += ipsec_optlen;
13187 #endif
13188         if (bbr->rc_last_options != local_options) {
13189                 /*
13190                  * Cache the options length this generally does not change
13191                  * on a connection. We use this to calculate TSO.
13192                  */
13193                 bbr->rc_last_options = local_options;
13194         }
13195         maxseg = tp->t_maxseg - (ipoptlen + optlen);
13196         p_maxseg = min(maxseg, pace_max_segs);
13197         /*
13198          * Adjust data length if insertion of options will bump the packet
13199          * length beyond the t_maxseg length. Clear the FIN bit because we
13200          * cut off the tail of the segment.
13201          */
13202 #ifdef KERN_TLS
13203         /* force TSO for so TLS offload can get mss */
13204         if (sb->sb_flags & SB_TLS_IFNET) {
13205                 force_tso = 1;
13206         }
13207 #endif
13208
13209         if (len > maxseg) {
13210                 if (len != 0 && (flags & TH_FIN)) {
13211                         flags &= ~TH_FIN;
13212                 }
13213                 if (tso) {
13214                         uint32_t moff;
13215                         int32_t max_len;
13216
13217                         /* extract TSO information */
13218                         if_hw_tsomax = tp->t_tsomax;
13219                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
13220                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
13221                         KASSERT(ipoptlen == 0,
13222                             ("%s: TSO can't do IP options", __func__));
13223
13224                         /*
13225                          * Check if we should limit by maximum payload
13226                          * length:
13227                          */
13228                         if (if_hw_tsomax != 0) {
13229                                 /* compute maximum TSO length */
13230                                 max_len = (if_hw_tsomax - hdrlen -
13231                                     max_linkhdr);
13232                                 if (max_len <= 0) {
13233                                         len = 0;
13234                                 } else if (len > max_len) {
13235                                         len = max_len;
13236                                 }
13237                         }
13238                         /*
13239                          * Prevent the last segment from being fractional
13240                          * unless the send sockbuf can be emptied:
13241                          */
13242                         if (((sb_offset + len) < sbavail(sb)) &&
13243                             (hw_tls == 0)) {
13244                                 moff = len % (uint32_t)maxseg;
13245                                 if (moff != 0) {
13246                                         len -= moff;
13247                                 }
13248                         }
13249                         /*
13250                          * In case there are too many small fragments don't
13251                          * use TSO:
13252                          */
13253                         if (len <= maxseg) {
13254                                 len = maxseg;
13255                                 tso = 0;
13256                         }
13257                 } else {
13258                         /* Not doing TSO */
13259                         if (optlen + ipoptlen >= tp->t_maxseg) {
13260                                 /*
13261                                  * Since we don't have enough space to put
13262                                  * the IP header chain and the TCP header in
13263                                  * one packet as required by RFC 7112, don't
13264                                  * send it. Also ensure that at least one
13265                                  * byte of the payload can be put into the
13266                                  * TCP segment.
13267                                  */
13268                                 SOCKBUF_UNLOCK(&so->so_snd);
13269                                 error = EMSGSIZE;
13270                                 sack_rxmit = 0;
13271                                 goto out;
13272                         }
13273                         len = maxseg;
13274                 }
13275         } else {
13276                 /* Not doing TSO */
13277                 if_hw_tsomaxsegcount = 0;
13278                 tso = 0;
13279         }
13280         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
13281             ("%s: len > IP_MAXPACKET", __func__));
13282 #ifdef DIAGNOSTIC
13283 #ifdef INET6
13284         if (max_linkhdr + hdrlen > MCLBYTES)
13285 #else
13286         if (max_linkhdr + hdrlen > MHLEN)
13287 #endif
13288                 panic("tcphdr too big");
13289 #endif
13290         /*
13291          * This KASSERT is here to catch edge cases at a well defined place.
13292          * Before, those had triggered (random) panic conditions further
13293          * down.
13294          */
13295 #ifdef BBR_INVARIANTS
13296         if (sack_rxmit) {
13297                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
13298                         panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u",
13299                             rsm, tp, bbr, rsm->r_start, tp->snd_una);
13300                 }
13301         }
13302 #endif
13303         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
13304         if ((len == 0) &&
13305             (flags & TH_FIN) &&
13306             (sbused(sb))) {
13307                 /*
13308                  * We have outstanding data, don't send a fin by itself!.
13309                  */
13310                 slot = 0;
13311                 goto just_return;
13312         }
13313         /*
13314          * Grab a header mbuf, attaching a copy of data to be transmitted,
13315          * and initialize the header from the template for sends on this
13316          * connection.
13317          */
13318         if (len) {
13319                 uint32_t moff;
13320                 uint32_t orig_len;
13321
13322                 /*
13323                  * We place a limit on sending with hptsi.
13324                  */
13325                 if ((rsm == NULL) && len > pace_max_segs)
13326                         len = pace_max_segs;
13327                 if (len <= maxseg)
13328                         tso = 0;
13329 #ifdef INET6
13330                 if (MHLEN < hdrlen + max_linkhdr)
13331                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
13332                 else
13333 #endif
13334                         m = m_gethdr(M_NOWAIT, MT_DATA);
13335
13336                 if (m == NULL) {
13337                         BBR_STAT_INC(bbr_failed_mbuf_aloc);
13338                         bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13339                         SOCKBUF_UNLOCK(sb);
13340                         error = ENOBUFS;
13341                         sack_rxmit = 0;
13342                         goto out;
13343                 }
13344                 m->m_data += max_linkhdr;
13345                 m->m_len = hdrlen;
13346                 /*
13347                  * Start the m_copy functions from the closest mbuf to the
13348                  * sb_offset in the socket buffer chain.
13349                  */
13350                 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) {
13351 #ifdef BBR_INVARIANTS
13352                         if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0)))
13353                                 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u",
13354                                     tp, bbr, len, sb_offset, sbavail(sb), rsm,
13355                                     doing_retran_from,
13356                                     picked_up_retran,
13357                                     doing_tlp);
13358
13359 #endif
13360                         /*
13361                          * In this messed up situation we have two choices,
13362                          * a) pretend the send worked, and just start timers
13363                          * and what not (not good since that may lead us
13364                          * back here a lot). <or> b) Send the lowest segment
13365                          * in the map. <or> c) Drop the connection. Lets do
13366                          * <b> which if it continues to happen will lead to
13367                          * <c> via timeouts.
13368                          */
13369                         BBR_STAT_INC(bbr_offset_recovery);
13370                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
13371                         sb_offset = 0;
13372                         if (rsm == NULL) {
13373                                 sack_rxmit = 0;
13374                                 len = sbavail(sb);
13375                         } else {
13376                                 sack_rxmit = 1;
13377                                 if (rsm->r_start != tp->snd_una) {
13378                                         /*
13379                                          * Things are really messed up, <c>
13380                                          * is the only thing to do.
13381                                          */
13382                                         BBR_STAT_INC(bbr_offset_drop);
13383                                         tcp_set_inp_to_drop(inp, EFAULT);
13384                                         SOCKBUF_UNLOCK(sb);
13385                                         (void)m_free(m);
13386                                         return (0);
13387                                 }
13388                                 len = rsm->r_end - rsm->r_start;
13389                         }
13390                         if (len > sbavail(sb))
13391                                 len = sbavail(sb);
13392                         if (len > maxseg)
13393                                 len = maxseg;
13394                 }
13395                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
13396                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
13397                         m_copydata(mb, moff, (int)len,
13398                             mtod(m, caddr_t)+hdrlen);
13399                         if (rsm == NULL)
13400                                 sbsndptr_adv(sb, mb, len);
13401                         m->m_len += len;
13402                 } else {
13403                         struct sockbuf *msb;
13404
13405                         if (rsm)
13406                                 msb = NULL;
13407                         else
13408                                 msb = sb;
13409 #ifdef BBR_INVARIANTS
13410                         if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) {
13411                                 if (rsm) {
13412                                         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 ",
13413                                             tp, bbr, len, moff,
13414                                             sbavail(sb), rsm,
13415                                             tp->snd_una, rsm->r_flags, rsm->r_start,
13416                                             doing_retran_from,
13417                                             picked_up_retran,
13418                                             doing_tlp, sack_rxmit);
13419                                 } else {
13420                                         panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u",
13421                                             tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una);
13422                                 }
13423                         }
13424 #endif
13425                         orig_len = len;
13426                         m->m_next = tcp_m_copym(
13427                                 mb, moff, &len,
13428                                 if_hw_tsomaxsegcount,
13429                                 if_hw_tsomaxsegsize, msb,
13430                                 ((rsm == NULL) ? hw_tls : 0)
13431 #ifdef NETFLIX_COPY_ARGS
13432                                 , &filled_all
13433 #endif
13434                                 );
13435                         if (len <= maxseg && !force_tso) {
13436                                 /*
13437                                  * Must have ran out of mbufs for the copy
13438                                  * shorten it to no longer need tso. Lets
13439                                  * not put on sendalot since we are low on
13440                                  * mbufs.
13441                                  */
13442                                 tso = 0;
13443                         }
13444                         if (m->m_next == NULL) {
13445                                 SOCKBUF_UNLOCK(sb);
13446                                 (void)m_free(m);
13447                                 error = ENOBUFS;
13448                                 sack_rxmit = 0;
13449                                 goto out;
13450                         }
13451                 }
13452 #ifdef BBR_INVARIANTS
13453                 if (tso && len < maxseg) {
13454                         panic("tp:%p tso on, but len:%d < maxseg:%d",
13455                             tp, len, maxseg);
13456                 }
13457                 if (tso && if_hw_tsomaxsegcount) {
13458                         int32_t seg_cnt = 0;
13459                         struct mbuf *foo;
13460
13461                         foo = m;
13462                         while (foo) {
13463                                 seg_cnt++;
13464                                 foo = foo->m_next;
13465                         }
13466                         if (seg_cnt > if_hw_tsomaxsegcount) {
13467                                 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount);
13468                         }
13469                 }
13470 #endif
13471                 /*
13472                  * If we're sending everything we've got, set PUSH. (This
13473                  * will keep happy those implementations which only give
13474                  * data to the user when a buffer fills or a PUSH comes in.)
13475                  */
13476                 if (sb_offset + len == sbused(sb) &&
13477                     sbused(sb) &&
13478                     !(flags & TH_SYN)) {
13479                         flags |= TH_PUSH;
13480                 }
13481                 SOCKBUF_UNLOCK(sb);
13482         } else {
13483                 SOCKBUF_UNLOCK(sb);
13484                 if (tp->t_flags & TF_ACKNOW)
13485                         KMOD_TCPSTAT_INC(tcps_sndacks);
13486                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
13487                         KMOD_TCPSTAT_INC(tcps_sndctrl);
13488                 else
13489                         KMOD_TCPSTAT_INC(tcps_sndwinup);
13490
13491                 m = m_gethdr(M_NOWAIT, MT_DATA);
13492                 if (m == NULL) {
13493                         BBR_STAT_INC(bbr_failed_mbuf_aloc);
13494                         bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13495                         error = ENOBUFS;
13496                         /* Fudge the send time since we could not send */
13497                         sack_rxmit = 0;
13498                         goto out;
13499                 }
13500 #ifdef INET6
13501                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
13502                     MHLEN >= hdrlen) {
13503                         M_ALIGN(m, hdrlen);
13504                 } else
13505 #endif
13506                         m->m_data += max_linkhdr;
13507                 m->m_len = hdrlen;
13508         }
13509         SOCKBUF_UNLOCK_ASSERT(sb);
13510         m->m_pkthdr.rcvif = (struct ifnet *)0;
13511 #ifdef MAC
13512         mac_inpcb_create_mbuf(inp, m);
13513 #endif
13514 #ifdef INET6
13515         if (isipv6) {
13516                 ip6 = mtod(m, struct ip6_hdr *);
13517 #ifdef NETFLIX_TCPOUDP
13518                 if (tp->t_port) {
13519                         udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr));
13520                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13521                         udp->uh_dport = tp->t_port;
13522                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
13523                         udp->uh_ulen = htons(ulen);
13524                         th = (struct tcphdr *)(udp + 1);
13525                 } else {
13526 #endif
13527                         th = (struct tcphdr *)(ip6 + 1);
13528
13529 #ifdef NETFLIX_TCPOUDP
13530                 }
13531 #endif
13532                 tcpip_fillheaders(inp,
13533 #ifdef NETFLIX_TCPOUDP
13534                                   tp->t_port,
13535 #endif
13536                                   ip6, th);
13537         } else
13538 #endif                          /* INET6 */
13539         {
13540                 ip = mtod(m, struct ip *);
13541 #ifdef TCPDEBUG
13542                 ipov = (struct ipovly *)ip;
13543 #endif
13544 #ifdef NETFLIX_TCPOUDP
13545                 if (tp->t_port) {
13546                         udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip));
13547                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13548                         udp->uh_dport = tp->t_port;
13549                         ulen = hdrlen + len - sizeof(struct ip);
13550                         udp->uh_ulen = htons(ulen);
13551                         th = (struct tcphdr *)(udp + 1);
13552                 } else
13553 #endif
13554                         th = (struct tcphdr *)(ip + 1);
13555                 tcpip_fillheaders(inp,
13556 #ifdef NETFLIX_TCPOUDP
13557                                   tp->t_port,
13558 #endif
13559                                   ip, th);
13560         }
13561         /*
13562          * If we are doing retransmissions, then snd_nxt will not reflect
13563          * the first unsent octet.  For ACK only packets, we do not want the
13564          * sequence number of the retransmitted packet, we want the sequence
13565          * number of the next unsent octet.  So, if there is no data (and no
13566          * SYN or FIN), use snd_max instead of snd_nxt when filling in
13567          * ti_seq.  But if we are in persist state, snd_max might reflect
13568          * one byte beyond the right edge of the window, so use snd_nxt in
13569          * that case, since we know we aren't doing a retransmission.
13570          * (retransmit and persist are mutually exclusive...)
13571          */
13572         if (sack_rxmit == 0) {
13573                 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) {
13574                         /* New data (including new persists) */
13575                         th->th_seq = htonl(tp->snd_max);
13576                         bbr_seq = tp->snd_max;
13577                 } else if (flags & TH_SYN) {
13578                         /* Syn's always send from iss */
13579                         th->th_seq = htonl(tp->iss);
13580                         bbr_seq = tp->iss;
13581                 } else if (flags & TH_FIN) {
13582                         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) {
13583                                 /*
13584                                  * If we sent the fin already its 1 minus
13585                                  * snd_max
13586                                  */
13587                                 th->th_seq = (htonl(tp->snd_max - 1));
13588                                 bbr_seq = (tp->snd_max - 1);
13589                         } else {
13590                                 /* First time FIN use snd_max */
13591                                 th->th_seq = htonl(tp->snd_max);
13592                                 bbr_seq = tp->snd_max;
13593                         }
13594                 } else if (flags & TH_RST) {
13595                         /*
13596                          * For a Reset send the last cum ack in sequence
13597                          * (this like any other choice may still generate a
13598                          * challenge ack, if a ack-update packet is in
13599                          * flight).
13600                          */
13601                         th->th_seq = htonl(tp->snd_una);
13602                         bbr_seq = tp->snd_una;
13603                 } else {
13604                         /*
13605                          * len == 0 and not persist we use snd_max, sending
13606                          * an ack unless we have sent the fin then its 1
13607                          * minus.
13608                          */
13609                         /*
13610                          * XXXRRS Question if we are in persists and we have
13611                          * nothing outstanding to send and we have not sent
13612                          * a FIN, we will send an ACK. In such a case it
13613                          * might be better to send (tp->snd_una - 1) which
13614                          * would force the peer to ack.
13615                          */
13616                         if (tp->t_flags & TF_SENTFIN) {
13617                                 th->th_seq = htonl(tp->snd_max - 1);
13618                                 bbr_seq = (tp->snd_max - 1);
13619                         } else {
13620                                 th->th_seq = htonl(tp->snd_max);
13621                                 bbr_seq = tp->snd_max;
13622                         }
13623                 }
13624         } else {
13625                 /* All retransmits use the rsm to guide the send */
13626                 th->th_seq = htonl(rsm->r_start);
13627                 bbr_seq = rsm->r_start;
13628         }
13629         th->th_ack = htonl(tp->rcv_nxt);
13630         if (optlen) {
13631                 bcopy(opt, th + 1, optlen);
13632                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
13633         }
13634         th->th_flags = flags;
13635         /*
13636          * Calculate receive window.  Don't shrink window, but avoid silly
13637          * window syndrome.
13638          */
13639         if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) &&
13640                                   recwin < maxseg)))
13641                 recwin = 0;
13642         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
13643             recwin < (tp->rcv_adv - tp->rcv_nxt))
13644                 recwin = (tp->rcv_adv - tp->rcv_nxt);
13645         if (recwin > TCP_MAXWIN << tp->rcv_scale)
13646                 recwin = TCP_MAXWIN << tp->rcv_scale;
13647
13648         /*
13649          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
13650          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
13651          * handled in syncache.
13652          */
13653         if (flags & TH_SYN)
13654                 th->th_win = htons((u_short)
13655                     (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
13656         else {
13657                 /* Avoid shrinking window with window scaling. */
13658                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
13659                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
13660         }
13661         /*
13662          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
13663          * window.  This may cause the remote transmitter to stall.  This
13664          * flag tells soreceive() to disable delayed acknowledgements when
13665          * draining the buffer.  This can occur if the receiver is
13666          * attempting to read more data than can be buffered prior to
13667          * transmitting on the connection.
13668          */
13669         if (th->th_win == 0) {
13670                 tp->t_sndzerowin++;
13671                 tp->t_flags |= TF_RXWIN0SENT;
13672         } else
13673                 tp->t_flags &= ~TF_RXWIN0SENT;
13674         /*
13675          * We don't support urgent data, but drag along
13676          * the pointer in case of a stack switch.
13677          */
13678         tp->snd_up = tp->snd_una;
13679
13680 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13681         if (to.to_flags & TOF_SIGNATURE) {
13682                 /*
13683                  * Calculate MD5 signature and put it into the place
13684                  * determined before. NOTE: since TCP options buffer doesn't
13685                  * point into mbuf's data, calculate offset and use it.
13686                  */
13687                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
13688                     (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
13689                         /*
13690                          * Do not send segment if the calculation of MD5
13691                          * digest has failed.
13692                          */
13693                         goto out;
13694                 }
13695         }
13696 #endif
13697
13698         /*
13699          * Put TCP length in extended header, and then checksum extended
13700          * header and data.
13701          */
13702         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
13703 #ifdef INET6
13704         if (isipv6) {
13705                 /*
13706                  * ip6_plen is not need to be filled now, and will be filled
13707                  * in ip6_output.
13708                  */
13709 #ifdef NETFLIX_TCPOUDP
13710                 if (tp->t_port) {
13711                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
13712                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13713                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
13714                         th->th_sum = htons(0);
13715                         UDPSTAT_INC(udps_opackets);
13716                 } else {
13717 #endif
13718                         csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
13719                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13720                         th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
13721                             optlen + len, IPPROTO_TCP, 0);
13722 #ifdef NETFLIX_TCPOUDP
13723                 }
13724 #endif
13725         }
13726 #endif
13727 #if defined(INET6) && defined(INET)
13728         else
13729 #endif
13730 #ifdef INET
13731         {
13732 #ifdef NETFLIX_TCPOUDP
13733                 if (tp->t_port) {
13734                         m->m_pkthdr.csum_flags = CSUM_UDP;
13735                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13736                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
13737                             ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
13738                         th->th_sum = htons(0);
13739                         UDPSTAT_INC(udps_opackets);
13740                 } else {
13741 #endif
13742                         csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP;
13743                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13744                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
13745                             ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
13746                             IPPROTO_TCP + len + optlen));
13747 #ifdef NETFLIX_TCPOUDP
13748                 }
13749 #endif
13750                 /* IP version must be set here for ipv4/ipv6 checking later */
13751                 KASSERT(ip->ip_v == IPVERSION,
13752                     ("%s: IP version incorrect: %d", __func__, ip->ip_v));
13753         }
13754 #endif
13755
13756         /*
13757          * Enable TSO and specify the size of the segments. The TCP pseudo
13758          * header checksum is always provided. XXX: Fixme: This is currently
13759          * not the case for IPv6.
13760          */
13761         if (tso || force_tso) {
13762                 KASSERT(force_tso || len > maxseg,
13763                     ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg));
13764                 m->m_pkthdr.csum_flags |= CSUM_TSO;
13765                 csum_flags |= CSUM_TSO;
13766                 m->m_pkthdr.tso_segsz = maxseg;
13767         }
13768         KASSERT(len + hdrlen == m_length(m, NULL),
13769             ("%s: mbuf chain different than expected: %d + %u != %u",
13770             __func__, len, hdrlen, m_length(m, NULL)));
13771
13772 #ifdef TCP_HHOOK
13773         /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */
13774         hhook_run_tcp_est_out(tp, th, &to, len, tso);
13775 #endif
13776 #ifdef TCPDEBUG
13777         /*
13778          * Trace.
13779          */
13780         if (so->so_options & SO_DEBUG) {
13781                 u_short save = 0;
13782
13783 #ifdef INET6
13784                 if (!isipv6)
13785 #endif
13786                 {
13787                         save = ipov->ih_len;
13788                         ipov->ih_len = htons(m->m_pkthdr.len    /* - hdrlen +
13789                               * (th->th_off << 2) */ );
13790                 }
13791                 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
13792 #ifdef INET6
13793                 if (!isipv6)
13794 #endif
13795                         ipov->ih_len = save;
13796         }
13797 #endif                          /* TCPDEBUG */
13798
13799         /* Log to the black box */
13800         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13801                 union tcp_log_stackspecific log;
13802
13803                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
13804                 /* Record info on type of transmission */
13805                 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay;
13806                 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3);
13807                 log.u_bbr.flex3 = maxseg;
13808                 log.u_bbr.flex4 = delay_calc;
13809                 /* Encode filled_all into the upper flex5 bit */
13810                 log.u_bbr.flex5 = bbr->rc_past_init_win;
13811                 log.u_bbr.flex5 <<= 1;
13812                 log.u_bbr.flex5 |= bbr->rc_no_pacing;
13813                 log.u_bbr.flex5 <<= 29;
13814                 if (filled_all)
13815                         log.u_bbr.flex5 |= 0x80000000;
13816                 log.u_bbr.flex5 |= tp->t_maxseg;
13817                 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs;
13818                 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr);
13819                 /* lets poke in the low and the high here for debugging */
13820                 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
13821                 if (rsm || sack_rxmit) {
13822                         if (doing_tlp)
13823                                 log.u_bbr.flex8 = 2;
13824                         else
13825                                 log.u_bbr.flex8 = 1;
13826                 } else {
13827                         log.u_bbr.flex8 = 0;
13828                 }
13829                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
13830                     len, &log, false, NULL, NULL, 0, tv);
13831         } else {
13832                 lgb = NULL;
13833         }
13834         /*
13835          * Fill in IP length and desired time to live and send to IP level.
13836          * There should be a better way to handle ttl and tos; we could keep
13837          * them in the template, but need a way to checksum without them.
13838          */
13839         /*
13840          * m->m_pkthdr.len should have been set before cksum calcuration,
13841          * because in6_cksum() need it.
13842          */
13843 #ifdef INET6
13844         if (isipv6) {
13845                 /*
13846                  * we separately set hoplimit for every segment, since the
13847                  * user might want to change the value via setsockopt. Also,
13848                  * desired default hop limit might be changed via Neighbor
13849                  * Discovery.
13850                  */
13851                 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
13852
13853                 /*
13854                  * Set the packet size here for the benefit of DTrace
13855                  * probes. ip6_output() will set it properly; it's supposed
13856                  * to include the option header lengths as well.
13857                  */
13858                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
13859
13860                 if (V_path_mtu_discovery && maxseg > V_tcp_minmss)
13861                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13862                 else
13863                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13864
13865                 if (tp->t_state == TCPS_SYN_SENT)
13866                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
13867
13868                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
13869                 /* TODO: IPv6 IP6TOS_ECT bit on */
13870                 error = ip6_output(m, inp->in6p_outputopts,
13871                     &inp->inp_route6,
13872                     ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
13873                     NULL, NULL, inp);
13874
13875                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
13876                         mtu = inp->inp_route6.ro_nh->nh_mtu;
13877         }
13878 #endif                          /* INET6 */
13879 #if defined(INET) && defined(INET6)
13880         else
13881 #endif
13882 #ifdef INET
13883         {
13884                 ip->ip_len = htons(m->m_pkthdr.len);
13885 #ifdef INET6
13886                 if (isipv6)
13887                         ip->ip_ttl = in6_selecthlim(inp, NULL);
13888 #endif                          /* INET6 */
13889                 /*
13890                  * If we do path MTU discovery, then we set DF on every
13891                  * packet. This might not be the best thing to do according
13892                  * to RFC3390 Section 2. However the tcp hostcache migitates
13893                  * the problem so it affects only the first tcp connection
13894                  * with a host.
13895                  *
13896                  * NB: Don't set DF on small MTU/MSS to have a safe
13897                  * fallback.
13898                  */
13899                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
13900                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13901                         if (tp->t_port == 0 || len < V_tcp_minmss) {
13902                                 ip->ip_off |= htons(IP_DF);
13903                         }
13904                 } else {
13905                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13906                 }
13907
13908                 if (tp->t_state == TCPS_SYN_SENT)
13909                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
13910
13911                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
13912
13913                 error = ip_output(m, inp->inp_options, &inp->inp_route,
13914                     ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
13915                     inp);
13916                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
13917                         mtu = inp->inp_route.ro_nh->nh_mtu;
13918         }
13919 #endif                          /* INET */
13920 out:
13921
13922         if (lgb) {
13923                 lgb->tlb_errno = error;
13924                 lgb = NULL;
13925         }
13926         /*
13927          * In transmit state, time the transmission and arrange for the
13928          * retransmit.  In persist state, just set snd_max.
13929          */
13930         if (error == 0) {
13931                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13932                     (tp->t_flags & TF_SACK_PERMIT) &&
13933                     tp->rcv_numsacks > 0)
13934                         tcp_clean_dsack_blocks(tp);
13935                 /* We sent an ack clear the bbr_segs_rcvd count */
13936                 bbr->output_error_seen = 0;
13937                 bbr->oerror_cnt = 0;
13938                 bbr->bbr_segs_rcvd = 0;
13939                 if (len == 0)
13940                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1);
13941                 else if (hw_tls) {
13942                         if (filled_all ||
13943                             (len >= bbr->r_ctl.rc_pace_max_segs))
13944                                 BBR_STAT_INC(bbr_meets_tso_thresh);
13945                         else {
13946                                 if (doing_tlp) {
13947                                         BBR_STAT_INC(bbr_miss_tlp);
13948                                         bbr_log_type_hrdwtso(tp, bbr, len, 1, what_we_can);
13949
13950
13951                                 } else if (rsm) {
13952                                         BBR_STAT_INC(bbr_miss_retran);
13953                                         bbr_log_type_hrdwtso(tp, bbr, len, 2, what_we_can);
13954                                 } else if ((ctf_outstanding(tp) + bbr->r_ctl.rc_pace_max_segs) > sbavail(sb)) {
13955                                         BBR_STAT_INC(bbr_miss_tso_app);
13956                                         bbr_log_type_hrdwtso(tp, bbr, len, 3, what_we_can);
13957                                 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13958                                                                  bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_pace_max_segs) > tp->snd_cwnd) {
13959                                         BBR_STAT_INC(bbr_miss_tso_cwnd);
13960                                         bbr_log_type_hrdwtso(tp, bbr, len, 4, what_we_can);
13961                                 } else if ((ctf_outstanding(tp) + bbr->r_ctl.rc_pace_max_segs) > tp->snd_wnd) {
13962                                         BBR_STAT_INC(bbr_miss_tso_rwnd);
13963                                         bbr_log_type_hrdwtso(tp, bbr, len, 5, what_we_can);
13964                                 } else {
13965                                         BBR_STAT_INC(bbr_miss_unknown);
13966                                         bbr_log_type_hrdwtso(tp, bbr, len, 6, what_we_can);
13967                                 }
13968                         }
13969                 }
13970                 /* Do accounting for new sends */
13971                 if ((len > 0) && (rsm == NULL)) {
13972                         int idx;
13973                         if (tp->snd_una == tp->snd_max) {
13974                                 /*
13975                                  * Special case to match google, when
13976                                  * nothing is in flight the delivered
13977                                  * time does get updated to the current
13978                                  * time (see tcp_rate_bsd.c).
13979                                  */
13980                                 bbr->r_ctl.rc_del_time = cts;
13981                         }
13982                         if (len >= maxseg) {
13983                                 idx = (len / maxseg) + 3;
13984                                 if (idx >= TCP_MSS_ACCT_ATIMER)
13985                                         counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1);
13986                                 else
13987                                         counter_u64_add(bbr_out_size[idx], 1);
13988                         } else {
13989                                 /* smaller than a MSS */
13990                                 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options);
13991                                 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV)
13992                                         idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1);
13993                                 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1);
13994                         }
13995                 }
13996         }
13997         abandon = 0;
13998         /*
13999          * We must do the send accounting before we log the output,
14000          * otherwise the state of the rsm could change and we account to the
14001          * wrong bucket.
14002          */
14003         if (len > 0) {
14004                 bbr_do_send_accounting(tp, bbr, rsm, len, error);
14005                 if (error == 0) {
14006                         if (tp->snd_una == tp->snd_max)
14007                                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
14008                 }
14009         }
14010         bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error,
14011             cts, mb, &abandon, rsm, 0, sb);
14012         if (abandon) {
14013                 /*
14014                  * If bbr_log_output destroys the TCB or sees a TH_RST being
14015                  * sent we should hit this condition.
14016                  */
14017                 return (0);
14018         }
14019         if (bbr->rc_in_persist == 0) {
14020                 /*
14021                  * Advance snd_nxt over sequence space of this segment.
14022                  */
14023                 if (error)
14024                         /* We don't log or do anything with errors */
14025                         goto skip_upd;
14026
14027                 if (tp->snd_una == tp->snd_max &&
14028                     (len || (flags & (TH_SYN | TH_FIN)))) {
14029                         /*
14030                          * Update the time we just added data since none was
14031                          * outstanding.
14032                          */
14033                         bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
14034                         bbr->rc_tp->t_acktime  = ticks;
14035                 }
14036                 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) {
14037                         if (flags & TH_SYN) {
14038                                 /*
14039                                  * Smack the snd_max to iss + 1
14040                                  * if its a FO we will add len below.
14041                                  */
14042                                 tp->snd_max = tp->iss + 1;
14043                         }
14044                         if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
14045                                 tp->snd_max++;
14046                                 tp->t_flags |= TF_SENTFIN;
14047                         }
14048                 }
14049                 if (sack_rxmit == 0)
14050                         tp->snd_max += len;
14051 skip_upd:
14052                 if ((error == 0) && len)
14053                         tot_len += len;
14054         } else {
14055                 /* Persists case */
14056                 int32_t xlen = len;
14057
14058                 if (error)
14059                         goto nomore;
14060
14061                 if (flags & TH_SYN)
14062                         ++xlen;
14063                 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
14064                         ++xlen;
14065                         tp->t_flags |= TF_SENTFIN;
14066                 }
14067                 if (xlen && (tp->snd_una == tp->snd_max)) {
14068                         /*
14069                          * Update the time we just added data since none was
14070                          * outstanding.
14071                          */
14072                         bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
14073                         bbr->rc_tp->t_acktime = ticks;
14074                 }
14075                 if (sack_rxmit == 0)
14076                         tp->snd_max += xlen;
14077                 tot_len += (len + optlen + ipoptlen);
14078         }
14079 nomore:
14080         if (error) {
14081                 /*
14082                  * Failures do not advance the seq counter above. For the
14083                  * case of ENOBUFS we will fall out and become ack-clocked.
14084                  * capping the cwnd at the current flight.
14085                  * Everything else will just have to retransmit with the timer
14086                  * (no pacer).
14087                  */
14088                 SOCKBUF_UNLOCK_ASSERT(sb);
14089                 BBR_STAT_INC(bbr_saw_oerr);
14090                 /* Clear all delay/early tracks */
14091                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
14092                 bbr->r_ctl.rc_agg_early = 0;
14093                 bbr->r_agg_early_set = 0;
14094                 bbr->output_error_seen = 1;
14095                 if (bbr->oerror_cnt < 0xf)
14096                         bbr->oerror_cnt++;
14097                 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) {
14098                         /* drop the session */
14099                         tcp_set_inp_to_drop(inp, ENETDOWN);
14100                 }
14101                 switch (error) {
14102                 case ENOBUFS:
14103                         /*
14104                          * Make this guy have to get ack's to send
14105                          * more but lets make sure we don't
14106                          * slam him below a T-O (1MSS).
14107                          */
14108                         if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
14109                                 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
14110                                                                     bbr->r_ctl.rc_lost_bytes)) - maxseg;
14111                                 if (tp->snd_cwnd < maxseg)
14112                                         tp->snd_cwnd = maxseg;
14113                         }
14114                         slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt;
14115                         BBR_STAT_INC(bbr_saw_enobuf);
14116                         if (bbr->bbr_hdrw_pacing)
14117                                 counter_u64_add(bbr_hdwr_pacing_enobuf, 1);
14118                         else
14119                                 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1);
14120                         /*
14121                          * Here even in the enobuf's case we want to do our
14122                          * state update. The reason being we may have been
14123                          * called by the input function. If so we have had
14124                          * things change.
14125                          */
14126                         error = 0;
14127                         goto enobufs;
14128                 case EMSGSIZE:
14129                         /*
14130                          * For some reason the interface we used initially
14131                          * to send segments changed to another or lowered
14132                          * its MTU. If TSO was active we either got an
14133                          * interface without TSO capabilits or TSO was
14134                          * turned off. If we obtained mtu from ip_output()
14135                          * then update it and try again.
14136                          */
14137                         /* Turn on tracing (or try to) */
14138                         {
14139                                 int old_maxseg;
14140
14141                                 old_maxseg = tp->t_maxseg;
14142                                 BBR_STAT_INC(bbr_saw_emsgsiz);
14143                                 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts);
14144                                 if (mtu != 0)
14145                                         tcp_mss_update(tp, -1, mtu, NULL, NULL);
14146                                 if (old_maxseg <= tp->t_maxseg) {
14147                                         /* Huh it did not shrink? */
14148                                         tp->t_maxseg = old_maxseg - 40;
14149                                         bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts);
14150                                 }
14151                                 /*
14152                                  * Nuke all other things that can interfere
14153                                  * with slot
14154                                  */
14155                                 if ((tot_len + len) && (len >= tp->t_maxseg)) {
14156                                         slot = bbr_get_pacing_delay(bbr,
14157                                             bbr->r_ctl.rc_bbr_hptsi_gain,
14158                                             (tot_len + len), cts, 0);
14159                                         if (slot < bbr_error_base_paceout)
14160                                                 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
14161                                 } else
14162                                         slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
14163                                 bbr->rc_output_starts_timer = 1;
14164                                 bbr_start_hpts_timer(bbr, tp, cts, 10, slot,
14165                                     tot_len);
14166                                 return (error);
14167                         }
14168                 case EPERM:
14169                         tp->t_softerror = error;
14170                         /* Fall through */
14171                 case EHOSTDOWN:
14172                 case EHOSTUNREACH:
14173                 case ENETDOWN:
14174                 case ENETUNREACH:
14175                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
14176                                 tp->t_softerror = error;
14177                         }
14178                         /* FALLTHROUGH */
14179                 default:
14180                         slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt;
14181                         bbr->rc_output_starts_timer = 1;
14182                         bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0);
14183                         return (error);
14184                 }
14185 #ifdef STATS
14186         } else if (((tp->t_flags & TF_GPUTINPROG) == 0) &&
14187                     len &&
14188                     (rsm == NULL) &&
14189             (bbr->rc_in_persist == 0)) {
14190                 tp->gput_seq = bbr_seq;
14191                 tp->gput_ack = bbr_seq +
14192                     min(sbavail(&so->so_snd) - sb_offset, sendwin);
14193                 tp->gput_ts = cts;
14194                 tp->t_flags |= TF_GPUTINPROG;
14195 #endif
14196         }
14197         KMOD_TCPSTAT_INC(tcps_sndtotal);
14198         if ((bbr->bbr_hdw_pace_ena) &&
14199             (bbr->bbr_attempt_hdwr_pace == 0) &&
14200             (bbr->rc_past_init_win) &&
14201             (bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
14202             (get_filter_value(&bbr->r_ctl.rc_delrate)) &&
14203             (inp->inp_route.ro_nh &&
14204              inp->inp_route.ro_nh->nh_ifp)) {
14205                 /*
14206                  * We are past the initial window and
14207                  * have at least one measurement so we
14208                  * could use hardware pacing if its available.
14209                  * We have an interface and we have not attempted
14210                  * to setup hardware pacing, lets try to now.
14211                  */
14212                 uint64_t rate_wanted;
14213                 int err = 0;
14214
14215                 rate_wanted = bbr_get_hardware_rate(bbr);
14216                 bbr->bbr_attempt_hdwr_pace = 1;
14217                 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp,
14218                                                       inp->inp_route.ro_nh->nh_ifp,
14219                                                       rate_wanted,
14220                                                       (RS_PACING_GEQ|RS_PACING_SUB_OK),
14221                                                       &err);
14222                 if (bbr->r_ctl.crte) {
14223                         bbr_type_log_hdwr_pacing(bbr,
14224                                                  bbr->r_ctl.crte->ptbl->rs_ifp,
14225                                                  rate_wanted,
14226                                                  bbr->r_ctl.crte->rate,
14227                                                  __LINE__, cts, err);
14228                         BBR_STAT_INC(bbr_hdwr_rl_add_ok);
14229                         counter_u64_add(bbr_flows_nohdwr_pacing, -1);
14230                         counter_u64_add(bbr_flows_whdwr_pacing, 1);
14231                         bbr->bbr_hdrw_pacing = 1;
14232                         /* Now what is our gain status? */
14233                         if (bbr->r_ctl.crte->rate < rate_wanted) {
14234                                 /* We have a problem */
14235                                 bbr_setup_less_of_rate(bbr, cts,
14236                                                        bbr->r_ctl.crte->rate, rate_wanted);
14237                         } else {
14238                                 /* We are good */
14239                                 bbr->gain_is_limited = 0;
14240                                 bbr->skip_gain = 0;
14241                         }
14242                         tcp_bbr_tso_size_check(bbr, cts);
14243                 } else {
14244                         bbr_type_log_hdwr_pacing(bbr,
14245                                                  inp->inp_route.ro_nh->nh_ifp,
14246                                                  rate_wanted,
14247                                                  0,
14248                                                  __LINE__, cts, err);
14249                         BBR_STAT_INC(bbr_hdwr_rl_add_fail);
14250                 }
14251         }
14252         if (bbr->bbr_hdrw_pacing) {
14253                 /*
14254                  * Worry about cases where the route
14255                  * changes or something happened that we
14256                  * lost our hardware pacing possibly during
14257                  * the last ip_output call.
14258                  */
14259                 if (inp->inp_snd_tag == NULL) {
14260                         /* A change during ip output disabled hw pacing? */
14261                         bbr->bbr_hdrw_pacing = 0;
14262                 } else if ((inp->inp_route.ro_nh == NULL) ||
14263                     (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) {
14264                         /*
14265                          * We had an interface or route change,
14266                          * detach from the current hdwr pacing
14267                          * and setup to re-attempt next go
14268                          * round.
14269                          */
14270                         bbr->bbr_hdrw_pacing = 0;
14271                         bbr->bbr_attempt_hdwr_pace = 0;
14272                         tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
14273                         tcp_bbr_tso_size_check(bbr, cts);
14274                 }
14275         }
14276         /*
14277          * Data sent (as far as we can tell). If this advertises a larger
14278          * window than any other segment, then remember the size of the
14279          * advertised window. Any pending ACK has now been sent.
14280          */
14281         if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
14282                 tp->rcv_adv = tp->rcv_nxt + recwin;
14283
14284         tp->last_ack_sent = tp->rcv_nxt;
14285         if ((error == 0) &&
14286             (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) &&
14287             (doing_tlp == 0) &&
14288             (tso == 0) &&
14289             (hw_tls == 0) &&
14290             (len > 0) &&
14291             ((flags & TH_RST) == 0) &&
14292             ((flags & TH_SYN) == 0) &&
14293             (IN_RECOVERY(tp->t_flags) == 0) &&
14294             (bbr->rc_in_persist == 0) &&
14295             (tot_len < bbr->r_ctl.rc_pace_max_segs)) {
14296                 /*
14297                  * For non-tso we need to goto again until we have sent out
14298                  * enough data to match what we are hptsi out every hptsi
14299                  * interval.
14300                  */
14301                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14302                         /* Make sure snd_nxt is drug up */
14303                         tp->snd_nxt = tp->snd_max;
14304                 }
14305                 if (rsm != NULL) {
14306                         rsm = NULL;
14307                         goto skip_again;
14308                 }
14309                 rsm = NULL;
14310                 sack_rxmit = 0;
14311                 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14312                 goto again;
14313         }
14314 skip_again:
14315         if ((error == 0) && (flags & TH_FIN))
14316                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
14317         if ((error == 0) && (flags & TH_RST))
14318                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
14319         if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) {
14320                 /*
14321                  * Calculate/Re-Calculate the hptsi slot in usecs based on
14322                  * what we have sent so far
14323                  */
14324                 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
14325                 if (bbr->rc_no_pacing)
14326                         slot = 0;
14327         }
14328         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14329 enobufs:
14330         if (bbr->rc_use_google == 0)
14331                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
14332         bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
14333                                                         bbr->r_ctl.rc_lost_bytes)));
14334         bbr->rc_output_starts_timer = 1;
14335         if (bbr->bbr_use_rack_cheat &&
14336             (more_to_rxt ||
14337              ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) {
14338                 /* Rack cheats and shotguns out all rxt's 1ms apart */
14339                 if (slot > 1000)
14340                         slot = 1000;
14341         }
14342         if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) {
14343                 /*
14344                  * We don't change the tso size until some number of sends
14345                  * to give the hardware commands time to get down
14346                  * to the interface.
14347                  */
14348                 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++;
14349                 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) {
14350                         bbr->hw_pacing_set = 1;
14351                         tcp_bbr_tso_size_check(bbr, cts);
14352                 }
14353         }
14354         bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len);
14355         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14356                 /* Make sure snd_nxt is drug up */
14357                 tp->snd_nxt = tp->snd_max;
14358         }
14359         return (error);
14360
14361 }
14362
14363 /*
14364  * See bbr_output_wtime() for return values.
14365  */
14366 static int
14367 bbr_output(struct tcpcb *tp)
14368 {
14369         int32_t ret;
14370         struct timeval tv;
14371         struct tcp_bbr *bbr;
14372
14373         NET_EPOCH_ASSERT();
14374
14375         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14376         INP_WLOCK_ASSERT(tp->t_inpcb);
14377         (void)tcp_get_usecs(&tv);
14378         ret = bbr_output_wtime(tp, &tv);
14379         return (ret);
14380 }
14381
14382 static void
14383 bbr_mtu_chg(struct tcpcb *tp)
14384 {
14385         struct tcp_bbr *bbr;
14386         struct bbr_sendmap *rsm, *frsm = NULL;
14387         uint32_t maxseg;
14388
14389         /*
14390          * The MTU has changed. a) Clear the sack filter. b) Mark everything
14391          * over the current size as SACK_PASS so a retransmit will occur.
14392          */
14393
14394         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14395         maxseg = tp->t_maxseg - bbr->rc_last_options;
14396         sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
14397         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
14398                 /* Don't mess with ones acked (by sack?) */
14399                 if (rsm->r_flags & BBR_ACKED)
14400                         continue;
14401                 if ((rsm->r_end - rsm->r_start) > maxseg) {
14402                         /*
14403                          * We mark sack-passed on all the previous large
14404                          * sends we did. This will force them to retransmit.
14405                          */
14406                         rsm->r_flags |= BBR_SACK_PASSED;
14407                         if (((rsm->r_flags & BBR_MARKED_LOST) == 0) &&
14408                             bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) {
14409                                 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
14410                                 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
14411                                 rsm->r_flags |= BBR_MARKED_LOST;
14412                         }
14413                         if (frsm == NULL)
14414                                 frsm = rsm;
14415                 }
14416         }
14417         if (frsm) {
14418                 bbr->r_ctl.rc_resend = frsm;
14419         }
14420 }
14421
14422 /*
14423  * bbr_ctloutput() must drop the inpcb lock before performing copyin on
14424  * socket option arguments.  When it re-acquires the lock after the copy, it
14425  * has to revalidate that the connection is still valid for the socket
14426  * option.
14427  */
14428 static int
14429 bbr_set_sockopt(struct socket *so, struct sockopt *sopt,
14430                 struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14431 {
14432         struct epoch_tracker et;
14433         int32_t error = 0, optval;
14434
14435         switch (sopt->sopt_name) {
14436         case TCP_RACK_PACE_MAX_SEG:
14437         case TCP_RACK_MIN_TO:
14438         case TCP_RACK_REORD_THRESH:
14439         case TCP_RACK_REORD_FADE:
14440         case TCP_RACK_TLP_THRESH:
14441         case TCP_RACK_PKT_DELAY:
14442         case TCP_BBR_ALGORITHM:
14443         case TCP_BBR_TSLIMITS:
14444         case TCP_BBR_IWINTSO:
14445         case TCP_BBR_RECFORCE:
14446         case TCP_BBR_STARTUP_PG:
14447         case TCP_BBR_DRAIN_PG:
14448         case TCP_BBR_RWND_IS_APP:
14449         case TCP_BBR_PROBE_RTT_INT:
14450         case TCP_BBR_PROBE_RTT_GAIN:
14451         case TCP_BBR_PROBE_RTT_LEN:
14452         case TCP_BBR_STARTUP_LOSS_EXIT:
14453         case TCP_BBR_USEDEL_RATE:
14454         case TCP_BBR_MIN_RTO:
14455         case TCP_BBR_MAX_RTO:
14456         case TCP_BBR_PACE_PER_SEC:
14457         case TCP_DELACK:
14458         case TCP_BBR_PACE_DEL_TAR:
14459         case TCP_BBR_SEND_IWND_IN_TSO:
14460         case TCP_BBR_EXTRA_STATE:
14461         case TCP_BBR_UTTER_MAX_TSO:
14462         case TCP_BBR_MIN_TOPACEOUT:
14463         case TCP_BBR_FLOOR_MIN_TSO:
14464         case TCP_BBR_TSTMP_RAISES:
14465         case TCP_BBR_POLICER_DETECT:
14466         case TCP_BBR_USE_RACK_CHEAT:
14467         case TCP_DATA_AFTER_CLOSE:
14468         case TCP_BBR_HDWR_PACE:
14469         case TCP_BBR_PACE_SEG_MAX:
14470         case TCP_BBR_PACE_SEG_MIN:
14471         case TCP_BBR_PACE_CROSS:
14472         case TCP_BBR_PACE_OH:
14473 #ifdef NETFLIX_PEAKRATE
14474         case TCP_MAXPEAKRATE:
14475 #endif
14476         case TCP_BBR_TMR_PACE_OH:
14477         case TCP_BBR_RACK_RTT_USE:
14478         case TCP_BBR_RETRAN_WTSO:
14479                 break;
14480         default:
14481                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14482                 break;
14483         }
14484         INP_WUNLOCK(inp);
14485         error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
14486         if (error)
14487                 return (error);
14488         INP_WLOCK(inp);
14489         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
14490                 INP_WUNLOCK(inp);
14491                 return (ECONNRESET);
14492         }
14493         tp = intotcpcb(inp);
14494         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14495         switch (sopt->sopt_name) {
14496         case TCP_BBR_PACE_PER_SEC:
14497                 BBR_OPTS_INC(tcp_bbr_pace_per_sec);
14498                 bbr->r_ctl.bbr_hptsi_per_second = optval;
14499                 break;
14500         case TCP_BBR_PACE_DEL_TAR:
14501                 BBR_OPTS_INC(tcp_bbr_pace_del_tar);
14502                 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval;
14503                 break;
14504         case TCP_BBR_PACE_SEG_MAX:
14505                 BBR_OPTS_INC(tcp_bbr_pace_seg_max);
14506                 bbr->r_ctl.bbr_hptsi_segments_max = optval;
14507                 break;
14508         case TCP_BBR_PACE_SEG_MIN:
14509                 BBR_OPTS_INC(tcp_bbr_pace_seg_min);
14510                 bbr->r_ctl.bbr_hptsi_bytes_min = optval;
14511                 break;
14512         case TCP_BBR_PACE_CROSS:
14513                 BBR_OPTS_INC(tcp_bbr_pace_cross);
14514                 bbr->r_ctl.bbr_cross_over = optval;
14515                 break;
14516         case TCP_BBR_ALGORITHM:
14517                 BBR_OPTS_INC(tcp_bbr_algorithm);
14518                 if (optval && (bbr->rc_use_google == 0)) {
14519                         /* Turn on the google mode */
14520                         bbr_google_mode_on(bbr);
14521                         if ((optval > 3) && (optval < 500)) {
14522                                 /*
14523                                  * Must be at least greater than .3%
14524                                  * and must be less than 50.0%.
14525                                  */
14526                                 bbr->r_ctl.bbr_google_discount = optval;
14527                         }
14528                 } else if ((optval == 0) && (bbr->rc_use_google == 1)) {
14529                         /* Turn off the google mode */
14530                         bbr_google_mode_off(bbr);
14531                 }
14532                 break;
14533         case TCP_BBR_TSLIMITS:
14534                 BBR_OPTS_INC(tcp_bbr_tslimits);
14535                 if (optval == 1)
14536                         bbr->rc_use_ts_limit = 1;
14537                 else if (optval == 0)
14538                         bbr->rc_use_ts_limit = 0;
14539                 else
14540                         error = EINVAL;
14541                 break;
14542
14543         case TCP_BBR_IWINTSO:
14544                 BBR_OPTS_INC(tcp_bbr_iwintso);
14545                 if ((optval >= 0) && (optval < 128)) {
14546                         uint32_t twin;
14547
14548                         bbr->rc_init_win = optval;
14549                         twin = bbr_initial_cwnd(bbr, tp);
14550                         if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd))
14551                                 tp->snd_cwnd = twin;
14552                         else
14553                                 error = EBUSY;
14554                 } else
14555                         error = EINVAL;
14556                 break;
14557         case TCP_BBR_STARTUP_PG:
14558                 BBR_OPTS_INC(tcp_bbr_startup_pg);
14559                 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) {
14560                         bbr->r_ctl.rc_startup_pg = optval;
14561                         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
14562                                 bbr->r_ctl.rc_bbr_hptsi_gain = optval;
14563                         }
14564                 } else
14565                         error = EINVAL;
14566                 break;
14567         case TCP_BBR_DRAIN_PG:
14568                 BBR_OPTS_INC(tcp_bbr_drain_pg);
14569                 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE))
14570                         bbr->r_ctl.rc_drain_pg = optval;
14571                 else
14572                         error = EINVAL;
14573                 break;
14574         case TCP_BBR_PROBE_RTT_LEN:
14575                 BBR_OPTS_INC(tcp_bbr_probertt_len);
14576                 if (optval <= 1)
14577                         reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND));
14578                 else
14579                         error = EINVAL;
14580                 break;
14581         case TCP_BBR_PROBE_RTT_GAIN:
14582                 BBR_OPTS_INC(tcp_bbr_probertt_gain);
14583                 if (optval <= BBR_UNIT)
14584                         bbr->r_ctl.bbr_rttprobe_gain_val = optval;
14585                 else
14586                         error = EINVAL;
14587                 break;
14588         case TCP_BBR_PROBE_RTT_INT:
14589                 BBR_OPTS_INC(tcp_bbr_probe_rtt_int);
14590                 if (optval > 1000)
14591                         bbr->r_ctl.rc_probertt_int = optval;
14592                 else
14593                         error = EINVAL;
14594                 break;
14595         case TCP_BBR_MIN_TOPACEOUT:
14596                 BBR_OPTS_INC(tcp_bbr_topaceout);
14597                 if (optval == 0) {
14598                         bbr->no_pacing_until = 0;
14599                         bbr->rc_no_pacing = 0;
14600                 } else if (optval <= 0x00ff) {
14601                         bbr->no_pacing_until = optval;
14602                         if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) &&
14603                             (bbr->rc_bbr_state == BBR_STATE_STARTUP)){
14604                                 /* Turn on no pacing */
14605                                 bbr->rc_no_pacing = 1;
14606                         }
14607                 } else
14608                         error = EINVAL;
14609                 break;
14610         case TCP_BBR_STARTUP_LOSS_EXIT:
14611                 BBR_OPTS_INC(tcp_bbr_startup_loss_exit);
14612                 bbr->rc_loss_exit = optval;
14613                 break;
14614         case TCP_BBR_USEDEL_RATE:
14615                 error = EINVAL;
14616                 break;
14617         case TCP_BBR_MIN_RTO:
14618                 BBR_OPTS_INC(tcp_bbr_min_rto);
14619                 bbr->r_ctl.rc_min_rto_ms = optval;
14620                 break;
14621         case TCP_BBR_MAX_RTO:
14622                 BBR_OPTS_INC(tcp_bbr_max_rto);
14623                 bbr->rc_max_rto_sec = optval;
14624                 break;
14625         case TCP_RACK_MIN_TO:
14626                 /* Minimum time between rack t-o's in ms */
14627                 BBR_OPTS_INC(tcp_rack_min_to);
14628                 bbr->r_ctl.rc_min_to = optval;
14629                 break;
14630         case TCP_RACK_REORD_THRESH:
14631                 /* RACK reorder threshold (shift amount) */
14632                 BBR_OPTS_INC(tcp_rack_reord_thresh);
14633                 if ((optval > 0) && (optval < 31))
14634                         bbr->r_ctl.rc_reorder_shift = optval;
14635                 else
14636                         error = EINVAL;
14637                 break;
14638         case TCP_RACK_REORD_FADE:
14639                 /* Does reordering fade after ms time */
14640                 BBR_OPTS_INC(tcp_rack_reord_fade);
14641                 bbr->r_ctl.rc_reorder_fade = optval;
14642                 break;
14643         case TCP_RACK_TLP_THRESH:
14644                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
14645                 BBR_OPTS_INC(tcp_rack_tlp_thresh);
14646                 if (optval)
14647                         bbr->rc_tlp_threshold = optval;
14648                 else
14649                         error = EINVAL;
14650                 break;
14651         case TCP_BBR_USE_RACK_CHEAT:
14652                 BBR_OPTS_INC(tcp_use_rackcheat);
14653                 if (bbr->rc_use_google) {
14654                         error = EINVAL;
14655                         break;
14656                 }
14657                 BBR_OPTS_INC(tcp_rack_cheat);
14658                 if (optval)
14659                         bbr->bbr_use_rack_cheat = 1;
14660                 else
14661                         bbr->bbr_use_rack_cheat = 0;
14662                 break;
14663         case TCP_BBR_FLOOR_MIN_TSO:
14664                 BBR_OPTS_INC(tcp_utter_max_tso);
14665                 if ((optval >= 0) && (optval < 40))
14666                         bbr->r_ctl.bbr_hptsi_segments_floor = optval;
14667                 else
14668                         error = EINVAL;
14669                 break;
14670         case TCP_BBR_UTTER_MAX_TSO:
14671                 BBR_OPTS_INC(tcp_utter_max_tso);
14672                 if ((optval >= 0) && (optval < 0xffff))
14673                         bbr->r_ctl.bbr_utter_max = optval;
14674                 else
14675                         error = EINVAL;
14676                 break;
14677
14678         case TCP_BBR_EXTRA_STATE:
14679                 BBR_OPTS_INC(tcp_extra_state);
14680                 if (optval)
14681                         bbr->rc_use_idle_restart = 1;
14682                 else
14683                         bbr->rc_use_idle_restart = 0;
14684                 break;
14685         case TCP_BBR_SEND_IWND_IN_TSO:
14686                 BBR_OPTS_INC(tcp_iwnd_tso);
14687                 if (optval) {
14688                         bbr->bbr_init_win_cheat = 1;
14689                         if (bbr->rc_past_init_win == 0) {
14690                                 uint32_t cts;
14691                                 cts = tcp_get_usecs(&bbr->rc_tv);
14692                                 tcp_bbr_tso_size_check(bbr, cts);
14693                         }
14694                 } else
14695                         bbr->bbr_init_win_cheat = 0;
14696                 break;
14697         case TCP_BBR_HDWR_PACE:
14698                 BBR_OPTS_INC(tcp_hdwr_pacing);
14699                 if (optval){
14700                         bbr->bbr_hdw_pace_ena = 1;
14701                         bbr->bbr_attempt_hdwr_pace = 0;
14702                 } else {
14703                         bbr->bbr_hdw_pace_ena = 0;
14704 #ifdef RATELIMIT
14705                         if (bbr->bbr_hdrw_pacing) {
14706                                 bbr->bbr_hdrw_pacing = 0;
14707                                 in_pcbdetach_txrtlmt(bbr->rc_inp);
14708                         }
14709 #endif
14710                 }
14711                 break;
14712
14713         case TCP_DELACK:
14714                 BBR_OPTS_INC(tcp_delack);
14715                 if (optval < 100) {
14716                         if (optval == 0) /* off */
14717                                 tp->t_delayed_ack = 0;
14718                         else if (optval == 1) /* on which is 2 */
14719                                 tp->t_delayed_ack = 2;
14720                         else /* higher than 2 and less than 100 */
14721                                 tp->t_delayed_ack = optval;
14722                         if (tp->t_flags & TF_DELACK) {
14723                                 tp->t_flags &= ~TF_DELACK;
14724                                 tp->t_flags |= TF_ACKNOW;
14725                                 NET_EPOCH_ENTER(et);
14726                                 bbr_output(tp);
14727                                 NET_EPOCH_EXIT(et);
14728                         }
14729                 } else
14730                         error = EINVAL;
14731                 break;
14732         case TCP_RACK_PKT_DELAY:
14733                 /* RACK added ms i.e. rack-rtt + reord + N */
14734                 BBR_OPTS_INC(tcp_rack_pkt_delay);
14735                 bbr->r_ctl.rc_pkt_delay = optval;
14736                 break;
14737 #ifdef NETFLIX_PEAKRATE
14738         case TCP_MAXPEAKRATE:
14739                 BBR_OPTS_INC(tcp_maxpeak);
14740                 error = tcp_set_maxpeakrate(tp, optval);
14741                 if (!error)
14742                         tp->t_peakrate_thr = tp->t_maxpeakrate;
14743                 break;
14744 #endif
14745         case TCP_BBR_RETRAN_WTSO:
14746                 BBR_OPTS_INC(tcp_retran_wtso);
14747                 if (optval)
14748                         bbr->rc_resends_use_tso = 1;
14749                 else
14750                         bbr->rc_resends_use_tso = 0;
14751                 break;
14752         case TCP_DATA_AFTER_CLOSE:
14753                 BBR_OPTS_INC(tcp_data_ac);
14754                 if (optval)
14755                         bbr->rc_allow_data_af_clo = 1;
14756                 else
14757                         bbr->rc_allow_data_af_clo = 0;
14758                 break;
14759         case TCP_BBR_POLICER_DETECT:
14760                 BBR_OPTS_INC(tcp_policer_det);
14761                 if (bbr->rc_use_google == 0)
14762                         error = EINVAL;
14763                 else if (optval)
14764                         bbr->r_use_policer = 1;
14765                 else
14766                         bbr->r_use_policer = 0;
14767                 break;
14768
14769         case TCP_BBR_TSTMP_RAISES:
14770                 BBR_OPTS_INC(tcp_ts_raises);
14771                 if (optval)
14772                         bbr->ts_can_raise = 1;
14773                 else
14774                         bbr->ts_can_raise = 0;
14775                 break;
14776         case TCP_BBR_TMR_PACE_OH:
14777                 BBR_OPTS_INC(tcp_pacing_oh_tmr);
14778                 if (bbr->rc_use_google) {
14779                         error = EINVAL;
14780                 } else {
14781                         if (optval)
14782                                 bbr->r_ctl.rc_incr_tmrs = 1;
14783                         else
14784                                 bbr->r_ctl.rc_incr_tmrs = 0;
14785                 }
14786                 break;
14787         case TCP_BBR_PACE_OH:
14788                 BBR_OPTS_INC(tcp_pacing_oh);
14789                 if (bbr->rc_use_google) {
14790                         error = EINVAL;
14791                 } else {
14792                         if (optval > (BBR_INCL_TCP_OH|
14793                                       BBR_INCL_IP_OH|
14794                                       BBR_INCL_ENET_OH)) {
14795                                 error = EINVAL;
14796                                 break;
14797                         }
14798                         if (optval & BBR_INCL_TCP_OH)
14799                                 bbr->r_ctl.rc_inc_tcp_oh = 1;
14800                         else
14801                                 bbr->r_ctl.rc_inc_tcp_oh = 0;
14802                         if (optval & BBR_INCL_IP_OH)
14803                                 bbr->r_ctl.rc_inc_ip_oh = 1;
14804                         else
14805                                 bbr->r_ctl.rc_inc_ip_oh = 0;
14806                         if (optval & BBR_INCL_ENET_OH)
14807                                 bbr->r_ctl.rc_inc_enet_oh = 1;
14808                         else
14809                                 bbr->r_ctl.rc_inc_enet_oh = 0;
14810                 }
14811                 break;
14812         default:
14813                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14814                 break;
14815         }
14816 #ifdef NETFLIX_STATS
14817         tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
14818 #endif
14819         INP_WUNLOCK(inp);
14820         return (error);
14821 }
14822
14823 /*
14824  * return 0 on success, error-num on failure
14825  */
14826 static int
14827 bbr_get_sockopt(struct socket *so, struct sockopt *sopt,
14828     struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14829 {
14830         int32_t error, optval;
14831
14832         /*
14833          * Because all our options are either boolean or an int, we can just
14834          * pull everything into optval and then unlock and copy. If we ever
14835          * add a option that is not a int, then this will have quite an
14836          * impact to this routine.
14837          */
14838         switch (sopt->sopt_name) {
14839         case TCP_BBR_PACE_PER_SEC:
14840                 optval = bbr->r_ctl.bbr_hptsi_per_second;
14841                 break;
14842         case TCP_BBR_PACE_DEL_TAR:
14843                 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar;
14844                 break;
14845         case TCP_BBR_PACE_SEG_MAX:
14846                 optval = bbr->r_ctl.bbr_hptsi_segments_max;
14847                 break;
14848         case TCP_BBR_MIN_TOPACEOUT:
14849                 optval = bbr->no_pacing_until;
14850                 break;
14851         case TCP_BBR_PACE_SEG_MIN:
14852                 optval = bbr->r_ctl.bbr_hptsi_bytes_min;
14853                 break;
14854         case TCP_BBR_PACE_CROSS:
14855                 optval = bbr->r_ctl.bbr_cross_over;
14856                 break;
14857         case TCP_BBR_ALGORITHM:
14858                 optval = bbr->rc_use_google;
14859                 break;
14860         case TCP_BBR_TSLIMITS:
14861                 optval = bbr->rc_use_ts_limit;
14862                 break;
14863         case TCP_BBR_IWINTSO:
14864                 optval = bbr->rc_init_win;
14865                 break;
14866         case TCP_BBR_STARTUP_PG:
14867                 optval = bbr->r_ctl.rc_startup_pg;
14868                 break;
14869         case TCP_BBR_DRAIN_PG:
14870                 optval = bbr->r_ctl.rc_drain_pg;
14871                 break;
14872         case TCP_BBR_PROBE_RTT_INT:
14873                 optval = bbr->r_ctl.rc_probertt_int;
14874                 break;
14875         case TCP_BBR_PROBE_RTT_LEN:
14876                 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND);
14877                 break;
14878         case TCP_BBR_PROBE_RTT_GAIN:
14879                 optval = bbr->r_ctl.bbr_rttprobe_gain_val;
14880                 break;
14881         case TCP_BBR_STARTUP_LOSS_EXIT:
14882                 optval = bbr->rc_loss_exit;
14883                 break;
14884         case TCP_BBR_USEDEL_RATE:
14885                 error = EINVAL;
14886                 break;
14887         case TCP_BBR_MIN_RTO:
14888                 optval = bbr->r_ctl.rc_min_rto_ms;
14889                 break;
14890         case TCP_BBR_MAX_RTO:
14891                 optval = bbr->rc_max_rto_sec;
14892                 break;
14893         case TCP_RACK_PACE_MAX_SEG:
14894                 /* Max segments in a pace */
14895                 optval = bbr->r_ctl.rc_pace_max_segs;
14896                 break;
14897         case TCP_RACK_MIN_TO:
14898                 /* Minimum time between rack t-o's in ms */
14899                 optval = bbr->r_ctl.rc_min_to;
14900                 break;
14901         case TCP_RACK_REORD_THRESH:
14902                 /* RACK reorder threshold (shift amount) */
14903                 optval = bbr->r_ctl.rc_reorder_shift;
14904                 break;
14905         case TCP_RACK_REORD_FADE:
14906                 /* Does reordering fade after ms time */
14907                 optval = bbr->r_ctl.rc_reorder_fade;
14908                 break;
14909         case TCP_BBR_USE_RACK_CHEAT:
14910                 /* Do we use the rack cheat for rxt */
14911                 optval = bbr->bbr_use_rack_cheat;
14912                 break;
14913         case TCP_BBR_FLOOR_MIN_TSO:
14914                 optval = bbr->r_ctl.bbr_hptsi_segments_floor;
14915                 break;
14916         case TCP_BBR_UTTER_MAX_TSO:
14917                 optval = bbr->r_ctl.bbr_utter_max;
14918                 break;
14919         case TCP_BBR_SEND_IWND_IN_TSO:
14920                 /* Do we send TSO size segments initially */
14921                 optval = bbr->bbr_init_win_cheat;
14922                 break;
14923         case TCP_BBR_EXTRA_STATE:
14924                 optval = bbr->rc_use_idle_restart;
14925                 break;
14926         case TCP_RACK_TLP_THRESH:
14927                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
14928                 optval = bbr->rc_tlp_threshold;
14929                 break;
14930         case TCP_RACK_PKT_DELAY:
14931                 /* RACK added ms i.e. rack-rtt + reord + N */
14932                 optval = bbr->r_ctl.rc_pkt_delay;
14933                 break;
14934         case TCP_BBR_RETRAN_WTSO:
14935                 optval = bbr->rc_resends_use_tso;
14936                 break;
14937         case TCP_DATA_AFTER_CLOSE:
14938                 optval = bbr->rc_allow_data_af_clo;
14939                 break;
14940         case TCP_DELACK:
14941                 optval = tp->t_delayed_ack;
14942                 break;
14943         case TCP_BBR_HDWR_PACE:
14944                 optval = bbr->bbr_hdw_pace_ena;
14945                 break;
14946         case TCP_BBR_POLICER_DETECT:
14947                 optval = bbr->r_use_policer;
14948                 break;
14949         case TCP_BBR_TSTMP_RAISES:
14950                 optval = bbr->ts_can_raise;
14951                 break;
14952         case TCP_BBR_TMR_PACE_OH:
14953                 optval = bbr->r_ctl.rc_incr_tmrs;
14954                 break;
14955         case TCP_BBR_PACE_OH:
14956                 optval = 0;
14957                 if (bbr->r_ctl.rc_inc_tcp_oh)
14958                         optval |= BBR_INCL_TCP_OH;
14959                 if (bbr->r_ctl.rc_inc_ip_oh)
14960                         optval |= BBR_INCL_IP_OH;
14961                 if (bbr->r_ctl.rc_inc_enet_oh)
14962                         optval |= BBR_INCL_ENET_OH;
14963                 break;
14964         default:
14965                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14966                 break;
14967         }
14968         INP_WUNLOCK(inp);
14969         error = sooptcopyout(sopt, &optval, sizeof optval);
14970         return (error);
14971 }
14972
14973 /*
14974  * return 0 on success, error-num on failure
14975  */
14976 static int
14977 bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
14978 {
14979         int32_t error = EINVAL;
14980         struct tcp_bbr *bbr;
14981
14982         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14983         if (bbr == NULL) {
14984                 /* Huh? */
14985                 goto out;
14986         }
14987         if (sopt->sopt_dir == SOPT_SET) {
14988                 return (bbr_set_sockopt(so, sopt, inp, tp, bbr));
14989         } else if (sopt->sopt_dir == SOPT_GET) {
14990                 return (bbr_get_sockopt(so, sopt, inp, tp, bbr));
14991         }
14992 out:
14993         INP_WUNLOCK(inp);
14994         return (error);
14995 }
14996
14997 static int
14998 bbr_pru_options(struct tcpcb *tp, int flags)
14999 {
15000         if (flags & PRUS_OOB)
15001                 return (EOPNOTSUPP);
15002         return (0);
15003 }
15004
15005 struct tcp_function_block __tcp_bbr = {
15006         .tfb_tcp_block_name = __XSTRING(STACKNAME),
15007         .tfb_tcp_output = bbr_output,
15008         .tfb_do_queued_segments = ctf_do_queued_segments,
15009         .tfb_do_segment_nounlock = bbr_do_segment_nounlock,
15010         .tfb_tcp_do_segment = bbr_do_segment,
15011         .tfb_tcp_ctloutput = bbr_ctloutput,
15012         .tfb_tcp_fb_init = bbr_init,
15013         .tfb_tcp_fb_fini = bbr_fini,
15014         .tfb_tcp_timer_stop_all = bbr_stopall,
15015         .tfb_tcp_timer_activate = bbr_timer_activate,
15016         .tfb_tcp_timer_active = bbr_timer_active,
15017         .tfb_tcp_timer_stop = bbr_timer_stop,
15018         .tfb_tcp_rexmit_tmr = bbr_remxt_tmr,
15019         .tfb_tcp_handoff_ok = bbr_handoff_ok,
15020         .tfb_tcp_mtu_chg = bbr_mtu_chg,
15021         .tfb_pru_options = bbr_pru_options,
15022 };
15023
15024 static const char *bbr_stack_names[] = {
15025         __XSTRING(STACKNAME),
15026 #ifdef STACKALIAS
15027         __XSTRING(STACKALIAS),
15028 #endif
15029 };
15030
15031 static bool bbr_mod_inited = false;
15032
15033 static int
15034 tcp_addbbr(module_t mod, int32_t type, void *data)
15035 {
15036         int32_t err = 0;
15037         int num_stacks;
15038
15039         switch (type) {
15040         case MOD_LOAD:
15041                 printf("Attempting to load " __XSTRING(MODNAME) "\n");
15042                 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
15043                     sizeof(struct bbr_sendmap),
15044                     NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
15045                 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
15046                     sizeof(struct tcp_bbr),
15047                     NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
15048                 sysctl_ctx_init(&bbr_sysctl_ctx);
15049                 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
15050                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
15051                     OID_AUTO,
15052 #ifdef STACKALIAS
15053                     __XSTRING(STACKALIAS),
15054 #else
15055                     __XSTRING(STACKNAME),
15056 #endif
15057                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
15058                     "");
15059                 if (bbr_sysctl_root == NULL) {
15060                         printf("Failed to add sysctl node\n");
15061                         err = EFAULT;
15062                         goto free_uma;
15063                 }
15064                 bbr_init_sysctls();
15065                 num_stacks = nitems(bbr_stack_names);
15066                 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK,
15067                     bbr_stack_names, &num_stacks);
15068                 if (err) {
15069                         printf("Failed to register %s stack name for "
15070                             "%s module\n", bbr_stack_names[num_stacks],
15071                             __XSTRING(MODNAME));
15072                         sysctl_ctx_free(&bbr_sysctl_ctx);
15073         free_uma:
15074                         uma_zdestroy(bbr_zone);
15075                         uma_zdestroy(bbr_pcb_zone);
15076                         bbr_counter_destroy();
15077                         printf("Failed to register " __XSTRING(MODNAME)
15078                             " module err:%d\n", err);
15079                         return (err);
15080                 }
15081                 tcp_lro_reg_mbufq();
15082                 bbr_mod_inited = true;
15083                 printf(__XSTRING(MODNAME) " is now available\n");
15084                 break;
15085         case MOD_QUIESCE:
15086                 err = deregister_tcp_functions(&__tcp_bbr, true, false);
15087                 break;
15088         case MOD_UNLOAD:
15089                 err = deregister_tcp_functions(&__tcp_bbr, false, true);
15090                 if (err == EBUSY)
15091                         break;
15092                 if (bbr_mod_inited) {
15093                         uma_zdestroy(bbr_zone);
15094                         uma_zdestroy(bbr_pcb_zone);
15095                         sysctl_ctx_free(&bbr_sysctl_ctx);
15096                         bbr_counter_destroy();
15097                         printf(__XSTRING(MODNAME)
15098                             " is now no longer available\n");
15099                         bbr_mod_inited = false;
15100                 }
15101                 tcp_lro_dereg_mbufq();
15102                 err = 0;
15103                 break;
15104         default:
15105                 return (EOPNOTSUPP);
15106         }
15107         return (err);
15108 }
15109
15110 static moduledata_t tcp_bbr = {
15111         .name = __XSTRING(MODNAME),
15112             .evhand = tcp_addbbr,
15113             .priv = 0
15114 };
15115
15116 MODULE_VERSION(MODNAME, 1);
15117 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
15118 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);