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 <sys/param.h>
  42 #include <sys/arb.h>
  43 #include <sys/module.h>
  44 #include <sys/kernel.h>
  45 #include <sys/libkern.h>
  46 #ifdef TCP_HHOOK
  47 #include <sys/hhook.h>
  48 #endif
  49 #include <sys/malloc.h>
  50 #include <sys/mbuf.h>
  51 #include <sys/proc.h>
  52 #include <sys/socket.h>
  53 #include <sys/socketvar.h>
  54 #include <sys/sysctl.h>
  55 #include <sys/systm.h>
  56 #ifdef STATS
  57 #include <sys/qmath.h>
  58 #include <sys/tree.h>
  59 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
  60 #endif
  61 #include <sys/refcount.h>
  62 #include <sys/queue.h>
  63 #include <sys/eventhandler.h>
  64 #include <sys/smp.h>
  65 #include <sys/kthread.h>
  66 #include <sys/lock.h>
  67 #include <sys/mutex.h>
  68 #include <sys/tim_filter.h>
  69 #include <sys/time.h>
  70 #include <sys/protosw.h>
  71 #include <vm/uma.h>
  72 #include <sys/kern_prefetch.h>
  73
  74 #include <net/route.h>
  75 #include <net/route/nhop.h>
  76 #include <net/vnet.h>
  77
  78 #define TCPSTATES               /* for logging */
  79
  80 #include <netinet/in.h>
  81 #include <netinet/in_kdtrace.h>
  82 #include <netinet/in_pcb.h>
  83 #include <netinet/ip.h>
  84 #include <netinet/ip_icmp.h>    /* required for icmp_var.h */
  85 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
  86 #include <netinet/ip_var.h>
  87 #include <netinet/ip6.h>
  88 #include <netinet6/in6_pcb.h>
  89 #include <netinet6/ip6_var.h>
  90 #define TCPOUTFLAGS
  91 #include <netinet/tcp.h>
  92 #include <netinet/tcp_fsm.h>
  93 #include <netinet/tcp_seq.h>
  94 #include <netinet/tcp_timer.h>
  95 #include <netinet/tcp_var.h>
  96 #include <netinet/tcpip.h>
  97 #include <netinet/tcp_hpts.h>
  98 #include <netinet/cc/cc.h>
  99 #include <netinet/tcp_log_buf.h>
 100 #include <netinet/tcp_ratelimit.h>
 101 #include <netinet/tcp_lro.h>
 102 #ifdef TCPDEBUG
 103 #include <netinet/tcp_debug.h>
 104 #endif                          /* TCPDEBUG */
 105 #ifdef TCP_OFFLOAD
 106 #include <netinet/tcp_offload.h>
 107 #endif
 108 #ifdef INET6
 109 #include <netinet6/tcp6_var.h>
 110 #endif
 111 #include <netinet/tcp_fastopen.h>
 112
 113 #include <netipsec/ipsec_support.h>
 114 #include <net/if.h>
 115 #include <net/if_var.h>
 116 #include <net/ethernet.h>
 117
 118 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 119 #include <netipsec/ipsec.h>
 120 #include <netipsec/ipsec6.h>
 121 #endif                          /* IPSEC */
 122
 123 #include <netinet/udp.h>
 124 #include <netinet/udp_var.h>
 125 #include <machine/in_cksum.h>
 126
 127 #ifdef MAC
 128 #include <security/mac/mac_framework.h>
 129 #endif
 130
 131 #include "sack_filter.h"
 132 #include "tcp_bbr.h"
 133 #include "rack_bbr_common.h"
 134 uma_zone_t bbr_zone;
 135 uma_zone_t bbr_pcb_zone;
 136
 137 struct sysctl_ctx_list bbr_sysctl_ctx;
 138 struct sysctl_oid *bbr_sysctl_root;
 139
 140 #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \
 141         (tv) = (value); \
 142         if ((u_long)(tv) < (u_long)(tvmin)) \
 143                 (tv) = (tvmin); \
 144         if ((u_long)(tv) > (u_long)(tvmax)) \
 145                 (tv) = (tvmax); \
 146 } while(0)
 147
 148 /*#define BBR_INVARIANT 1*/
 149
 150 /*
 151  * initial window
 152  */
 153 static uint32_t bbr_def_init_win = 10;
 154 static int32_t bbr_persist_min = 250000;        /* 250ms */
 155 static int32_t bbr_persist_max = 1000000;       /* 1 Second */
 156 static int32_t bbr_cwnd_may_shrink = 0;
 157 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP;
 158 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT;
 159 static int32_t bbr_hardware_pacing_limit = 8000;
 160 static int32_t bbr_quanta = 3;  /* How much extra quanta do we get? */
 161 static int32_t bbr_no_retran = 0;
 162
 163 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */
 164 static int32_t bbr_max_net_error_cnt = 10;
 165 /* Should the following be dynamic too -- loss wise */
 166 static int32_t bbr_rtt_gain_thresh = 0;
 167 /* Measurement controls */
 168 static int32_t bbr_use_google_algo = 1;
 169 static int32_t bbr_ts_limiting = 1;
 170 static int32_t bbr_ts_can_raise = 0;
 171 static int32_t bbr_do_red = 600;
 172 static int32_t bbr_red_scale = 20000;
 173 static int32_t bbr_red_mul = 1;
 174 static int32_t bbr_red_div = 2;
 175 static int32_t bbr_red_growth_restrict = 1;
 176 static int32_t  bbr_target_is_bbunit = 0;
 177 static int32_t bbr_drop_limit = 0;
 178 /*
 179  * How much gain do we need to see to
 180  * stay in startup?
 181  */
 182 static int32_t bbr_marks_rxt_sack_passed = 0;
 183 static int32_t bbr_start_exit = 25;
 184 static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */
 185 static int32_t bbr_startup_loss_thresh = 2000;  /* 20.00% loss */
 186 static int32_t bbr_hptsi_max_mul = 1;   /* These two mul/div assure a min pacing */
 187 static int32_t bbr_hptsi_max_div = 2;   /* time, 0 means turned off. We need this
 188                                          * if we go back ever to where the pacer
 189                                          * has priority over timers.
 190                                          */
 191 static int32_t bbr_policer_call_from_rack_to = 0;
 192 static int32_t bbr_policer_detection_enabled = 1;
 193 static int32_t bbr_min_measurements_req = 1;    /* We need at least 2
 194                                                  * measurments before we are
 195                                                  * "good" note that 2 == 1.
 196                                                  * This is because we use a >
 197                                                  * comparison. This means if
 198                                                  * min_measure was 0, it takes
 199                                                  * num-measures > min(0) and
 200                                                  * you get 1 measurement and
 201                                                  * you are good. Set to 1, you
 202                                                  * have to have two
 203                                                  * measurements (this is done
 204                                                  * to prevent it from being ok
 205                                                  * to have no measurements). */
 206 static int32_t bbr_no_pacing_until = 4;
 207
 208 static int32_t bbr_min_usec_delta = 20000;      /* 20,000 usecs */
 209 static int32_t bbr_min_peer_delta = 20;         /* 20 units */
 210 static int32_t bbr_delta_percent = 150;         /* 15.0 % */
 211
 212 static int32_t bbr_target_cwnd_mult_limit = 8;
 213 /*
 214  * bbr_cwnd_min_val is the number of
 215  * segments we hold to in the RTT probe
 216  * state typically 4.
 217  */
 218 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS;
 219
 220 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS;
 221
 222 static int32_t bbr_gain_to_target = 1;
 223 static int32_t bbr_gain_gets_extra_too = 1;
 224 /*
 225  * bbr_high_gain is the 2/ln(2) value we need
 226  * to double the sending rate in startup. This
 227  * is used for both cwnd and hptsi gain's.
 228  */
 229 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1;
 230 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1;
 231 static int32_t bbr_use_lower_gain_in_startup = 1;
 232
 233 /* thresholds for reduction on drain in sub-states/drain */
 234 static int32_t bbr_drain_rtt = BBR_SRTT;
 235 static int32_t bbr_drain_floor = 88;
 236 static int32_t google_allow_early_out = 1;
 237 static int32_t google_consider_lost = 1;
 238 static int32_t bbr_drain_drop_mul = 4;
 239 static int32_t bbr_drain_drop_div = 5;
 240 static int32_t bbr_rand_ot = 50;
 241 static int32_t bbr_can_force_probertt = 0;
 242 static int32_t bbr_can_adjust_probertt = 1;
 243 static int32_t bbr_probertt_sets_rtt = 0;
 244 static int32_t bbr_can_use_ts_for_rtt = 1;
 245 static int32_t bbr_is_ratio = 0;
 246 static int32_t bbr_sub_drain_app_limit = 1;
 247 static int32_t bbr_prtt_slam_cwnd = 1;
 248 static int32_t bbr_sub_drain_slam_cwnd = 1;
 249 static int32_t bbr_slam_cwnd_in_main_drain = 1;
 250 static int32_t bbr_filter_len_sec = 6;  /* How long does the rttProp filter
 251                                          * hold */
 252 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4);
 253 /*
 254  * bbr_drain_gain is the reverse of the high_gain
 255  * designed to drain back out the standing queue
 256  * that is formed in startup by causing a larger
 257  * hptsi gain and thus drainging the packets
 258  * in flight.
 259  */
 260 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885;
 261 static int32_t bbr_rttprobe_gain = 192;
 262
 263 /*
 264  * The cwnd_gain is the default cwnd gain applied when
 265  * calculating a target cwnd. Note that the cwnd is
 266  * a secondary factor in the way BBR works (see the
 267  * paper and think about it, it will take some time).
 268  * Basically the hptsi_gain spreads the packets out
 269  * so you never get more than BDP to the peer even
 270  * if the cwnd is high. In our implemenation that
 271  * means in non-recovery/retransmission scenarios
 272  * cwnd will never be reached by the flight-size.
 273  */
 274 static int32_t bbr_cwnd_gain = BBR_UNIT * 2;
 275 static int32_t bbr_tlp_type_to_use = BBR_SRTT;
 276 static int32_t bbr_delack_time = 100000;        /* 100ms in useconds */
 277 static int32_t bbr_sack_not_required = 0;       /* set to one to allow non-sack to use bbr */
 278 static int32_t bbr_initial_bw_bps = 62500;      /* 500kbps in bytes ps */
 279 static int32_t bbr_ignore_data_after_close = 1;
 280 static int16_t bbr_hptsi_gain[] = {
 281         (BBR_UNIT *5 / 4),
 282         (BBR_UNIT * 3 / 4),
 283         BBR_UNIT,
 284         BBR_UNIT,
 285         BBR_UNIT,
 286         BBR_UNIT,
 287         BBR_UNIT,
 288         BBR_UNIT
 289 };
 290 int32_t bbr_use_rack_resend_cheat = 1;
 291 int32_t bbr_sends_full_iwnd = 1;
 292
 293 #define BBR_HPTSI_GAIN_MAX 8
 294 /*
 295  * The BBR module incorporates a number of
 296  * TCP ideas that have been put out into the IETF
 297  * over the last few years:
 298  * - Yuchung Cheng's RACK TCP (for which its named) that
 299  *    will stop us using the number of dup acks and instead
 300  *    use time as the gage of when we retransmit.
 301  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
 302  *    of Dukkipati et.al.
 303  * - Van Jacobson's et.al BBR.
 304  *
 305  * RACK depends on SACK, so if an endpoint arrives that
 306  * cannot do SACK the state machine below will shuttle the
 307  * connection back to using the "default" TCP stack that is
 308  * in FreeBSD.
 309  *
 310  * To implement BBR and RACK the original TCP stack was first decomposed
 311  * into a functional state machine with individual states
 312  * for each of the possible TCP connection states. The do_segement
 313  * functions role in life is to mandate the connection supports SACK
 314  * initially and then assure that the RACK state matches the conenction
 315  * state before calling the states do_segment function. Data processing
 316  * of inbound segments also now happens in the hpts_do_segment in general
 317  * with only one exception. This is so we can keep the connection on
 318  * a single CPU.
 319  *
 320  * Each state is simplified due to the fact that the original do_segment
 321  * has been decomposed and we *know* what state we are in (no
 322  * switches on the state) and all tests for SACK are gone. This
 323  * greatly simplifies what each state does.
 324  *
 325  * TCP output is also over-written with a new version since it
 326  * must maintain the new rack scoreboard and has had hptsi
 327  * integrated as a requirment. Still todo is to eliminate the
 328  * use of the callout_() system and use the hpts for all
 329  * timers as well.
 330  */
 331 static uint32_t bbr_rtt_probe_time = 200000;    /* 200ms in micro seconds */
 332 static uint32_t bbr_rtt_probe_cwndtarg = 4;     /* How many mss's outstanding */
 333 static const int32_t bbr_min_req_free = 2;      /* The min we must have on the
 334                                                  * free list */
 335 static int32_t bbr_tlp_thresh = 1;
 336 static int32_t bbr_reorder_thresh = 2;
 337 static int32_t bbr_reorder_fade = 60000000;     /* 0 - never fade, def
 338                                                  * 60,000,000 - 60 seconds */
 339 static int32_t bbr_pkt_delay = 1000;
 340 static int32_t bbr_min_to = 1000;       /* Number of usec's minimum timeout */
 341 static int32_t bbr_incr_timers = 1;
 342
 343 static int32_t bbr_tlp_min = 10000;     /* 10ms in usecs */
 344 static int32_t bbr_delayed_ack_time = 200000;   /* 200ms in usecs */
 345 static int32_t bbr_exit_startup_at_loss = 1;
 346
 347 /*
 348  * bbr_lt_bw_ratio is 1/8th
 349  * bbr_lt_bw_diff is  < 4 Kbit/sec
 350  */
 351 static uint64_t bbr_lt_bw_diff = 4000 / 8;      /* In bytes per second */
 352 static uint64_t bbr_lt_bw_ratio = 8;    /* For 1/8th */
 353 static uint32_t bbr_lt_bw_max_rtts = 48;        /* How many rtt's do we use
 354                                                  * the lt_bw for */
 355 static uint32_t bbr_lt_intvl_min_rtts = 4;      /* Min num of RTT's to measure
 356                                                  * lt_bw */
 357 static int32_t bbr_lt_intvl_fp = 0;             /* False positive epoch diff */
 358 static int32_t bbr_lt_loss_thresh = 196;        /* Lost vs delivered % */
 359 static int32_t bbr_lt_fd_thresh = 100;          /* false detection % */
 360
 361 static int32_t bbr_verbose_logging = 0;
 362 /*
 363  * Currently regular tcp has a rto_min of 30ms
 364  * the backoff goes 12 times so that ends up
 365  * being a total of 122.850 seconds before a
 366  * connection is killed.
 367  */
 368 static int32_t bbr_rto_min_ms = 30;     /* 30ms same as main freebsd */
 369 static int32_t bbr_rto_max_sec = 4;     /* 4 seconds */
 370
 371 /****************************************************/
 372 /* DEFAULT TSO SIZING  (cpu performance impacting)  */
 373 /****************************************************/
 374 /* What amount is our formula using to get TSO size */
 375 static int32_t bbr_hptsi_per_second = 1000;
 376
 377 /*
 378  * For hptsi under bbr_cross_over connections what is delay
 379  * target 7ms (in usec) combined with a seg_max of 2
 380  * gets us close to identical google behavior in
 381  * TSO size selection (possibly more 1MSS sends).
 382  */
 383 static int32_t bbr_hptsi_segments_delay_tar = 7000;
 384
 385 /* Does pacing delay include overhead's in its time calculations? */
 386 static int32_t bbr_include_enet_oh = 0;
 387 static int32_t bbr_include_ip_oh = 1;
 388 static int32_t bbr_include_tcp_oh = 1;
 389 static int32_t bbr_google_discount = 10;
 390
 391 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */
 392 static int32_t bbr_state_is_pkt_epoch = 0;
 393 static int32_t bbr_state_drain_2_tar = 1;
 394 /* What is the max the 0 - bbr_cross_over MBPS TSO target
 395  * can reach using our delay target. Note that this
 396  * value becomes the floor for the cross over
 397  * algorithm.
 398  */
 399 static int32_t bbr_hptsi_segments_max = 2;
 400 static int32_t bbr_hptsi_segments_floor = 1;
 401 static int32_t bbr_hptsi_utter_max = 0;
 402
 403 /* What is the min the 0 - bbr_cross-over MBPS  TSO target can be */
 404 static int32_t bbr_hptsi_bytes_min = 1460;
 405 static int32_t bbr_all_get_min = 0;
 406
 407 /* Cross over point from algo-a to algo-b */
 408 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS;
 409
 410 /* Do we deal with our restart state? */
 411 static int32_t bbr_uses_idle_restart = 0;
 412 static int32_t bbr_idle_restart_threshold = 100000;     /* 100ms in useconds */
 413
 414 /* Do we allow hardware pacing? */
 415 static int32_t bbr_allow_hdwr_pacing = 0;
 416 static int32_t bbr_hdwr_pace_adjust = 2;        /* multipler when we calc the tso size */
 417 static int32_t bbr_hdwr_pace_floor = 1;
 418 static int32_t bbr_hdwr_pacing_delay_cnt = 10;
 419
 420 /****************************************************/
 421 static int32_t bbr_resends_use_tso = 0;
 422 static int32_t bbr_tlp_max_resend = 2;
 423 static int32_t bbr_sack_block_limit = 128;
 424
 425 #define  BBR_MAX_STAT 19
 426 counter_u64_t bbr_state_time[BBR_MAX_STAT];
 427 counter_u64_t bbr_state_lost[BBR_MAX_STAT];
 428 counter_u64_t bbr_state_resend[BBR_MAX_STAT];
 429 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE];
 430 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE];
 431 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE];
 432 counter_u64_t bbr_flows_whdwr_pacing;
 433 counter_u64_t bbr_flows_nohdwr_pacing;
 434
 435 counter_u64_t bbr_nohdwr_pacing_enobuf;
 436 counter_u64_t bbr_hdwr_pacing_enobuf;
 437
 438 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr);
 439
 440 /*
 441  * Static defintions we need for forward declarations.
 442  */
 443 static uint32_t
 444 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain,
 445     uint32_t useconds_time, uint64_t bw);
 446 static uint32_t
 447 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain);
 448 static void
 449      bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win);
 450 static void
 451 bbr_set_probebw_gains(struct tcp_bbr *bbr,  uint32_t cts, uint32_t losses);
 452 static void
 453 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line,
 454                     int dolog);
 455 static uint32_t
 456 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain);
 457 static void
 458 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch,
 459                  int32_t pkt_epoch, uint32_t losses);
 460 static uint32_t
 461 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm);
 462 static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp);
 463 static uint32_t
 464 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
 465     struct bbr_sendmap *rsm, uint32_t srtt,
 466     uint32_t cts);
 467 static void
 468 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts,
 469     int32_t line);
 470 static void
 471      bbr_set_state_target(struct tcp_bbr *bbr, int line);
 472 static void
 473      bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line);
 474
 475 static void
 476      bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line);
 477
 478 static void
 479      tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts);
 480
 481 static void
 482      bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts);
 483
 484 static void
 485      bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt,
 486                          uint32_t line, uint8_t is_start, uint16_t set);
 487
 488 static struct bbr_sendmap *
 489             bbr_find_lowest_rsm(struct tcp_bbr *bbr);
 490 static __inline uint32_t
 491 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type);
 492 static void
 493      bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which);
 494
 495 static void
 496 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
 497     uint32_t thresh, uint32_t to);
 498 static void
 499      bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag);
 500
 501 static void
 502 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot,
 503     uint32_t del_by, uint32_t cts, uint32_t sloton, uint32_t prev_delay);
 504
 505 static void
 506 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr,
 507     uint32_t cts, int32_t line);
 508 static void
 509      bbr_stop_all_timers(struct tcpcb *tp);
 510 static void
 511      bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts);
 512 static void
 513      bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts);
 514 static void
 515      bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts);
 516
 517 static void
 518 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
 519     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod);
 520
 521 static int
 522 bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp,
 523     struct tcpcb *tp);
 524
 525 static inline uint8_t
 526 bbr_state_val(struct tcp_bbr *bbr)
 527 {
 528         return(bbr->rc_bbr_substate);
 529 }
 530
 531 static inline uint32_t
 532 get_min_cwnd(struct tcp_bbr *bbr)
 533 {
 534         int mss;
 535
 536         mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
 537         if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED)
 538                 return (bbr_cwnd_min_val_hs * mss);
 539         else
 540                 return (bbr_cwnd_min_val * mss);
 541 }
 542
 543 static uint32_t
 544 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr)
 545 {
 546         uint64_t srtt, var;
 547         uint64_t ret_val;
 548
 549         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
 550         if (tp->t_srtt == 0) {
 551                 srtt = (uint64_t)BBR_INITIAL_RTO;
 552                 var = 0;
 553         } else {
 554                 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
 555                 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT);
 556         }
 557         TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]),
 558             bbr_persist_min, bbr_persist_max);
 559         return ((uint32_t)ret_val);
 560 }
 561
 562 static uint32_t
 563 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
 564 {
 565         /*
 566          * Start the FR timer, we do this based on getting the first one in
 567          * the rc_tmap. Note that if its NULL we must stop the timer. in all
 568          * events we need to stop the running timer (if its running) before
 569          * starting the new one.
 570          */
 571         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
 572         int32_t idx;
 573         int32_t is_tlp_timer = 0;
 574         struct bbr_sendmap *rsm;
 575
 576         if (bbr->rc_all_timers_stopped) {
 577                 /* All timers have been stopped none are to run */
 578                 return (0);
 579         }
 580         if (bbr->rc_in_persist) {
 581                 /* We can't start any timer in persists */
 582                 return (bbr_get_persists_timer_val(tp, bbr));
 583         }
 584         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 585         if ((rsm == NULL) ||
 586             ((tp->t_flags & TF_SACK_PERMIT) == 0) ||
 587             (tp->t_state < TCPS_ESTABLISHED)) {
 588                 /* Nothing on the send map */
 589 activate_rxt:
 590                 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
 591                         uint64_t tov;
 592
 593                         time_since_sent = 0;
 594                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 595                         if (rsm) {
 596                                 idx = rsm->r_rtr_cnt - 1;
 597                                 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
 598                                         tstmp_touse = rsm->r_tim_lastsent[idx];
 599                                 else
 600                                         tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
 601                                 if (TSTMP_GT(tstmp_touse, cts))
 602                                     time_since_sent = cts - tstmp_touse;
 603                         }
 604                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
 605                         if (tp->t_srtt == 0)
 606                                 tov = BBR_INITIAL_RTO;
 607                         else
 608                                 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) +
 609                                     ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT);
 610                         if (tp->t_rxtshift)
 611                                 tov *= tcp_backoff[tp->t_rxtshift];
 612                         if (tov > time_since_sent)
 613                                 tov -= time_since_sent;
 614                         else
 615                                 tov = bbr->r_ctl.rc_min_to;
 616                         TCPT_RANGESET_NOSLOP(to, tov,
 617                             (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC),
 618                             (bbr->rc_max_rto_sec * USECS_IN_SECOND));
 619                         bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to);
 620                         return (to);
 621                 }
 622                 return (0);
 623         }
 624         if (rsm->r_flags & BBR_ACKED) {
 625                 rsm = bbr_find_lowest_rsm(bbr);
 626                 if (rsm == NULL) {
 627                         /* No lowest? */
 628                         goto activate_rxt;
 629                 }
 630         }
 631         /* Convert from ms to usecs */
 632         if (rsm->r_flags & BBR_SACK_PASSED) {
 633                 if ((tp->t_flags & TF_SENTFIN) &&
 634                     ((tp->snd_max - tp->snd_una) == 1) &&
 635                     (rsm->r_flags & BBR_HAS_FIN)) {
 636                         /*
 637                          * We don't start a bbr rack timer if all we have is
 638                          * a FIN outstanding.
 639                          */
 640                         goto activate_rxt;
 641                 }
 642                 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK);
 643                 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm);
 644                 idx = rsm->r_rtr_cnt - 1;
 645                 exp = rsm->r_tim_lastsent[idx] + thresh;
 646                 if (SEQ_GEQ(exp, cts)) {
 647                         to = exp - cts;
 648                         if (to < bbr->r_ctl.rc_min_to) {
 649                                 to = bbr->r_ctl.rc_min_to;
 650                         }
 651                 } else {
 652                         to = bbr->r_ctl.rc_min_to;
 653                 }
 654         } else {
 655                 /* Ok we need to do a TLP not RACK */
 656                 if (bbr->rc_tlp_in_progress != 0) {
 657                         /*
 658                          * The previous send was a TLP.
 659                          */
 660                         goto activate_rxt;
 661                 }
 662                 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
 663                 if (rsm == NULL) {
 664                         /* We found no rsm to TLP with. */
 665                         goto activate_rxt;
 666                 }
 667                 if (rsm->r_flags & BBR_HAS_FIN) {
 668                         /* If its a FIN we don't do TLP */
 669                         rsm = NULL;
 670                         goto activate_rxt;
 671                 }
 672                 time_since_sent = 0;
 673                 idx = rsm->r_rtr_cnt - 1;
 674                 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
 675                         tstmp_touse = rsm->r_tim_lastsent[idx];
 676                 else
 677                         tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
 678                 if (TSTMP_GT(tstmp_touse, cts))
 679                     time_since_sent = cts - tstmp_touse;
 680                 is_tlp_timer = 1;
 681                 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use);
 682                 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts);
 683                 if (thresh > time_since_sent)
 684                         to = thresh - time_since_sent;
 685                 else
 686                         to = bbr->r_ctl.rc_min_to;
 687                 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
 688                         /*
 689                          * If the TLP time works out to larger than the max
 690                          * RTO lets not do TLP.. just RTO.
 691                          */
 692                         goto activate_rxt;
 693                 }
 694                 if ((bbr->rc_tlp_rtx_out == 1) &&
 695                     (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) {
 696                         /*
 697                          * Second retransmit of the same TLP
 698                          * lets not.
 699                          */
 700                         bbr->rc_tlp_rtx_out = 0;
 701                         goto activate_rxt;
 702                 }
 703                 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) {
 704                         /*
 705                          * The tail is no longer the last one I did a probe
 706                          * on
 707                          */
 708                         bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
 709                         bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
 710                 }
 711         }
 712         if (is_tlp_timer == 0) {
 713                 BBR_STAT_INC(bbr_to_arm_rack);
 714                 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
 715         } else {
 716                 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to);
 717                 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
 718                         /*
 719                          * We have exceeded how many times we can retran the
 720                          * current TLP timer, switch to the RTO timer.
 721                          */
 722                         goto activate_rxt;
 723                 } else {
 724                         BBR_STAT_INC(bbr_to_arm_tlp);
 725                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
 726                 }
 727         }
 728         return (to);
 729 }
 730
 731 static inline int32_t
 732 bbr_minseg(struct tcp_bbr *bbr)
 733 {
 734         return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options);
 735 }
 736
 737 static void
 738 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len)
 739 {
 740         struct inpcb *inp;
 741         struct hpts_diag diag;
 742         uint32_t delayed_ack = 0;
 743         uint32_t left = 0;
 744         uint32_t hpts_timeout;
 745         uint8_t stopped;
 746         int32_t delay_calc = 0;
 747         uint32_t prev_delay = 0;
 748
 749         inp = tp->t_inpcb;
 750         if (inp->inp_in_hpts) {
 751                 /* A previous call is already set up */
 752                 return;
 753         }
 754         if ((tp->t_state == TCPS_CLOSED) ||
 755             (tp->t_state == TCPS_LISTEN)) {
 756                 return;
 757         }
 758         stopped = bbr->rc_tmr_stopped;
 759         if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) {
 760                 left = bbr->r_ctl.rc_timer_exp - cts;
 761         }
 762         bbr->r_ctl.rc_hpts_flags = 0;
 763         bbr->r_ctl.rc_timer_exp = 0;
 764         prev_delay = bbr->r_ctl.rc_last_delay_val;
 765         if (bbr->r_ctl.rc_last_delay_val &&
 766             (slot == 0)) {
 767                 /*
 768                  * If a previous pacer delay was in place we
 769                  * are not coming from the output side (where
 770                  * we calculate a delay, more likely a timer).
 771                  */
 772                 slot = bbr->r_ctl.rc_last_delay_val;
 773                 if (TSTMP_GT(cts, bbr->rc_pacer_started)) {
 774                         /* Compensate for time passed  */
 775                         delay_calc = cts - bbr->rc_pacer_started;
 776                         if (delay_calc <= slot)
 777                                 slot -= delay_calc;
 778                 }
 779         }
 780         /* Do we have early to make up for by pushing out the pacing time? */
 781         if (bbr->r_agg_early_set) {
 782                 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2);
 783                 slot += bbr->r_ctl.rc_agg_early;
 784                 bbr->r_ctl.rc_agg_early = 0;
 785                 bbr->r_agg_early_set = 0;
 786         }
 787         /* Are we running a total debt that needs to be compensated for? */
 788         if (bbr->r_ctl.rc_hptsi_agg_delay) {
 789                 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) {
 790                         /* We nuke the delay */
 791                         slot -= bbr->r_ctl.rc_hptsi_agg_delay;
 792                         bbr->r_ctl.rc_hptsi_agg_delay = 0;
 793                 } else {
 794                         /* We nuke some of the delay, put in a minimal 100usecs  */
 795                         bbr->r_ctl.rc_hptsi_agg_delay -= slot;
 796                         bbr->r_ctl.rc_last_delay_val = slot = 100;
 797                 }
 798         }
 799         bbr->r_ctl.rc_last_delay_val = slot;
 800         hpts_timeout = bbr_timer_start(tp, bbr, cts);
 801         if (tp->t_flags & TF_DELACK) {
 802                 if (bbr->rc_in_persist == 0) {
 803                         delayed_ack = bbr_delack_time;
 804                 } else {
 805                         /*
 806                          * We are in persists and have
 807                          * gotten a new data element.
 808                          */
 809                         if (hpts_timeout > bbr_delack_time) {
 810                                 /*
 811                                  * Lets make the persists timer (which acks)
 812                                  * be the smaller of hpts_timeout and bbr_delack_time.
 813                                  */
 814                                 hpts_timeout = bbr_delack_time;
 815                         }
 816                 }
 817         }
 818         if (delayed_ack &&
 819             ((hpts_timeout == 0) ||
 820              (delayed_ack < hpts_timeout))) {
 821                 /* We need a Delayed ack timer */
 822                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
 823                 hpts_timeout = delayed_ack;
 824         }
 825         if (slot) {
 826                 /* Mark that we have a pacing timer up */
 827                 BBR_STAT_INC(bbr_paced_segments);
 828                 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
 829         }
 830         /*
 831          * If no timers are going to run and we will fall off thfe hptsi
 832          * wheel, we resort to a keep-alive timer if its configured.
 833          */
 834         if ((hpts_timeout == 0) &&
 835             (slot == 0)) {
 836                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
 837                     (tp->t_state <= TCPS_CLOSING)) {
 838                         /*
 839                          * Ok we have no timer (persists, rack, tlp, rxt  or
 840                          * del-ack), we don't have segments being paced. So
 841                          * all that is left is the keepalive timer.
 842                          */
 843                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
 844                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
 845                         } else {
 846                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
 847                         }
 848                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
 849                 }
 850         }
 851         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
 852             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
 853                 /*
 854                  * RACK, TLP, persists and RXT timers all are restartable
 855                  * based on actions input .. i.e we received a packet (ack
 856                  * or sack) and that changes things (rw, or snd_una etc).
 857                  * Thus we can restart them with a new value. For
 858                  * keep-alive, delayed_ack we keep track of what was left
 859                  * and restart the timer with a smaller value.
 860                  */
 861                 if (left < hpts_timeout)
 862                         hpts_timeout = left;
 863         }
 864         if (bbr->r_ctl.rc_incr_tmrs && slot &&
 865             (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
 866                 /*
 867                  * If configured to do so, and the timer is either
 868                  * the TLP or RXT timer, we need to increase the timeout
 869                  * by the pacing time. Consider the bottleneck at my
 870                  * machine as an example, we are sending something
 871                  * to start a TLP on. The last packet won't be emitted
 872                  * fully until the pacing time (the bottleneck will hold
 873                  * the data in place). Once the packet is emitted that
 874                  * is when we want to start waiting for the TLP. This
 875                  * is most evident with hardware pacing (where the nic
 876                  * is holding the packet(s) before emitting). But it
 877                  * can also show up in the network so we do it for all
 878                  * cases. Technically we would take off one packet from
 879                  * this extra delay but this is easier and being more
 880                  * conservative is probably better.
 881                  */
 882                 hpts_timeout += slot;
 883         }
 884         if (hpts_timeout) {
 885                 /*
 886                  * Hack alert for now we can't time-out over 2147 seconds (a
 887                  * bit more than 35min)
 888                  */
 889                 if (hpts_timeout > 0x7ffffffe)
 890                         hpts_timeout = 0x7ffffffe;
 891                 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
 892         } else
 893                 bbr->r_ctl.rc_timer_exp = 0;
 894         if ((slot) &&
 895             (bbr->rc_use_google ||
 896              bbr->output_error_seen ||
 897              (slot <= hpts_timeout))  ) {
 898                 /*
 899                  * Tell LRO that it can queue packets while
 900                  * we pace.
 901                  */
 902                 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
 903                 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
 904                     (bbr->rc_cwnd_limited == 0)) {
 905                         /*
 906                          * If we are not cwnd limited and we
 907                          * are running a rack timer we put on
 908                          * the do not disturbe even for sack.
 909                          */
 910                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
 911                 } else
 912                         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
 913                 bbr->rc_pacer_started = cts;
 914
 915                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
 916                                            __LINE__, &diag);
 917                 bbr->rc_timer_first = 0;
 918                 bbr->bbr_timer_src = frm;
 919                 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1);
 920                 bbr_log_hpts_diag(bbr, cts, &diag);
 921         } else if (hpts_timeout) {
 922                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
 923                                            __LINE__, &diag);
 924                 /*
 925                  * We add the flag here as well if the slot is set,
 926                  * since hpts will call in to clear the queue first before
 927                  * calling the output routine (which does our timers).
 928                  * We don't want to set the flag if its just a timer
 929                  * else the arrival of data might (that causes us
 930                  * to send more) might get delayed. Imagine being
 931                  * on a keep-alive timer and a request comes in for
 932                  * more data.
 933                  */
 934                 if (slot)
 935                         bbr->rc_pacer_started = cts;
 936                 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
 937                     (bbr->rc_cwnd_limited == 0)) {
 938                         /*
 939                          * For a rack timer, don't wake us even
 940                          * if a sack arrives as long as we are
 941                          * not cwnd limited.
 942                          */
 943                         bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
 944                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
 945                 } else {
 946                         /* All other timers wake us up */
 947                         bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
 948                         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
 949                 }
 950                 bbr->bbr_timer_src = frm;
 951                 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0);
 952                 bbr_log_hpts_diag(bbr, cts, &diag);
 953                 bbr->rc_timer_first = 1;
 954         }
 955         bbr->rc_tmr_stopped = 0;
 956         bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay);
 957 }
 958
 959 static void
 960 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb)
 961 {
 962         /*
 963          * We received an ack, and then did not call send or were bounced
 964          * out due to the hpts was running. Now a timer is up as well, is it
 965          * the right timer?
 966          */
 967         struct inpcb *inp;
 968         struct bbr_sendmap *rsm;
 969         uint32_t hpts_timeout;
 970         int tmr_up;
 971
 972         tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
 973         if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
 974                 return;
 975         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 976         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
 977             (tmr_up == PACE_TMR_RXT)) {
 978                 /* Should be an RXT */
 979                 return;
 980         }
 981         inp = bbr->rc_inp;
 982         if (rsm == NULL) {
 983                 /* Nothing outstanding? */
 984                 if (tp->t_flags & TF_DELACK) {
 985                         if (tmr_up == PACE_TMR_DELACK)
 986                                 /*
 987                                  * We are supposed to have delayed ack up
 988                                  * and we do
 989                                  */
 990                                 return;
 991                 } else if (sbavail(&inp->inp_socket->so_snd) &&
 992                     (tmr_up == PACE_TMR_RXT)) {
 993                         /*
 994                          * if we hit enobufs then we would expect the
 995                          * possiblity of nothing outstanding and the RXT up
 996                          * (and the hptsi timer).
 997                          */
 998                         return;
 999                 } else if (((V_tcp_always_keepalive ||
1000                             inp->inp_socket->so_options & SO_KEEPALIVE) &&
1001                             (tp->t_state <= TCPS_CLOSING)) &&
1002                             (tmr_up == PACE_TMR_KEEP) &&
1003                     (tp->snd_max == tp->snd_una)) {
1004                         /* We should have keep alive up and we do */
1005                         return;
1006                 }
1007         }
1008         if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) {
1009                 if ((tp->t_flags & TF_SENTFIN) &&
1010                     ((tp->snd_max - tp->snd_una) == 1) &&
1011                     (rsm->r_flags & BBR_HAS_FIN)) {
1012                         /* needs to be a RXT */
1013                         if (tmr_up == PACE_TMR_RXT)
1014                                 return;
1015                         else
1016                                 goto wrong_timer;
1017                 } else if (tmr_up == PACE_TMR_RACK)
1018                         return;
1019                 else
1020                         goto wrong_timer;
1021         } else if (rsm && (tmr_up == PACE_TMR_RACK)) {
1022                 /* Rack timer has priority if we have data out */
1023                 return;
1024         } else if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1025                     ((tmr_up == PACE_TMR_TLP) ||
1026             (tmr_up == PACE_TMR_RXT))) {
1027                 /*
1028                  * Either a TLP or RXT is fine if no sack-passed is in place
1029                  * and data is outstanding.
1030                  */
1031                 return;
1032         } else if (tmr_up == PACE_TMR_DELACK) {
1033                 /*
1034                  * If the delayed ack was going to go off before the
1035                  * rtx/tlp/rack timer were going to expire, then that would
1036                  * be the timer in control. Note we don't check the time
1037                  * here trusting the code is correct.
1038                  */
1039                 return;
1040         }
1041         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1042             ((tmr_up == PACE_TMR_RXT) ||
1043              (tmr_up == PACE_TMR_TLP) ||
1044              (tmr_up == PACE_TMR_RACK))) {
1045                 /*
1046                  * We have outstanding data and
1047                  * we *do* have a RACK, TLP or RXT
1048                  * timer running. We won't restart
1049                  * anything here since thats probably ok we
1050                  * will get called with some timer here shortly.
1051                  */
1052                 return;
1053         }
1054         /*
1055          * Ok the timer originally started is not what we want now. We will
1056          * force the hpts to be stopped if any, and restart with the slot
1057          * set to what was in the saved slot.
1058          */
1059 wrong_timer:
1060         if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) {
1061                 if (inp->inp_in_hpts)
1062                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
1063                 bbr_timer_cancel(bbr, __LINE__, cts);
1064                 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val,
1065                     0);
1066         } else {
1067                 /*
1068                  * Output is hptsi so we just need to switch the type of
1069                  * timer. We don't bother with keep-alive, since when we
1070                  * jump through the output, it will start the keep-alive if
1071                  * nothing is sent.
1072                  *
1073                  * We only need a delayed-ack added and or the hpts_timeout.
1074                  */
1075                 hpts_timeout = bbr_timer_start(tp, bbr, cts);
1076                 if (tp->t_flags & TF_DELACK) {
1077                         if (hpts_timeout == 0) {
1078                                 hpts_timeout = bbr_delack_time;
1079                                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1080                         }
1081                         else if (hpts_timeout > bbr_delack_time) {
1082                                 hpts_timeout = bbr_delack_time;
1083                                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1084                         }
1085                 }
1086                 if (hpts_timeout) {
1087                         if (hpts_timeout > 0x7ffffffe)
1088                                 hpts_timeout = 0x7ffffffe;
1089                         bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
1090                 }
1091         }
1092 }
1093
1094 int32_t bbr_clear_lost = 0;
1095
1096 /*
1097  * Considers the two time values now (cts) and earlier.
1098  * If cts is smaller than earlier, we could have
1099  * had a sequence wrap (our counter wraps every
1100  * 70 min or so) or it could be just clock skew
1101  * getting us two differnt time values. Clock skew
1102  * will show up within 10ms or so. So in such
1103  * a case (where cts is behind earlier time by
1104  * less than 10ms) we return 0. Otherwise we
1105  * return the true difference between them.
1106  */
1107 static inline uint32_t
1108 bbr_calc_time(uint32_t cts, uint32_t earlier_time) {
1109         /*
1110          * Given two timestamps, the current time stamp cts, and some other
1111          * time-stamp taken in theory earlier return the difference. The
1112          * trick is here sometimes locking will get the other timestamp
1113          * after the cts. If this occurs we need to return 0.
1114          */
1115         if (TSTMP_GEQ(cts, earlier_time))
1116                 return (cts - earlier_time);
1117         /*
1118          * cts is behind earlier_time if its less than 10ms consider it 0.
1119          * If its more than 10ms difference then we had a time wrap. Else
1120          * its just the normal locking foo. I wonder if we should not go to
1121          * 64bit TS and get rid of this issue.
1122          */
1123         if (TSTMP_GEQ((cts + 10000), earlier_time))
1124                 return (0);
1125         /*
1126          * Ok the time must have wrapped. So we need to answer a large
1127          * amount of time, which the normal subtraction should do.
1128          */
1129         return (cts - earlier_time);
1130 }
1131
1132 static int
1133 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS)
1134 {
1135         uint32_t stat;
1136         int32_t error;
1137
1138         error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t));
1139         if (error || req->newptr == NULL)
1140                 return error;
1141
1142         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
1143         if (error)
1144                 return (error);
1145         if (stat == 1) {
1146 #ifdef BBR_INVARIANTS
1147                 printf("Clearing BBR lost counters\n");
1148 #endif
1149                 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT);
1150                 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT);
1151                 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT);
1152         } else if (stat == 2) {
1153 #ifdef BBR_INVARIANTS
1154                 printf("Clearing BBR option counters\n");
1155 #endif
1156                 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE);
1157         } else if (stat == 3) {
1158 #ifdef BBR_INVARIANTS
1159                 printf("Clearing BBR stats counters\n");
1160 #endif
1161                 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE);
1162         } else if (stat == 4) {
1163 #ifdef BBR_INVARIANTS
1164                 printf("Clearing BBR out-size counters\n");
1165 #endif
1166                 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE);
1167         }
1168         bbr_clear_lost = 0;
1169         return (0);
1170 }
1171
1172 static void
1173 bbr_init_sysctls(void)
1174 {
1175         struct sysctl_oid *bbr_probertt;
1176         struct sysctl_oid *bbr_hptsi;
1177         struct sysctl_oid *bbr_measure;
1178         struct sysctl_oid *bbr_cwnd;
1179         struct sysctl_oid *bbr_timeout;
1180         struct sysctl_oid *bbr_states;
1181         struct sysctl_oid *bbr_startup;
1182         struct sysctl_oid *bbr_policer;
1183
1184         /* Probe rtt controls */
1185         bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1186             SYSCTL_CHILDREN(bbr_sysctl_root),
1187             OID_AUTO,
1188             "probertt",
1189             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1190             "");
1191         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1192             SYSCTL_CHILDREN(bbr_probertt),
1193             OID_AUTO, "gain", CTLFLAG_RW,
1194             &bbr_rttprobe_gain, 192,
1195             "What is the filter gain drop in probe_rtt (0=disable)?");
1196         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1197             SYSCTL_CHILDREN(bbr_probertt),
1198             OID_AUTO, "cwnd", CTLFLAG_RW,
1199             &bbr_rtt_probe_cwndtarg, 4,
1200             "How many mss's are outstanding during probe-rtt");
1201         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1202             SYSCTL_CHILDREN(bbr_probertt),
1203             OID_AUTO, "int", CTLFLAG_RW,
1204             &bbr_rtt_probe_limit, 4000000,
1205             "If RTT has not shrank in this many micro-seconds enter probe-rtt");
1206         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1207             SYSCTL_CHILDREN(bbr_probertt),
1208             OID_AUTO, "mintime", CTLFLAG_RW,
1209             &bbr_rtt_probe_time, 200000,
1210             "How many microseconds in probe-rtt");
1211         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1212             SYSCTL_CHILDREN(bbr_probertt),
1213             OID_AUTO, "filter_len_sec", CTLFLAG_RW,
1214             &bbr_filter_len_sec, 6,
1215             "How long in seconds does the rttProp filter run?");
1216         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1217             SYSCTL_CHILDREN(bbr_probertt),
1218             OID_AUTO, "drain_rtt", CTLFLAG_RW,
1219             &bbr_drain_rtt, BBR_SRTT,
1220             "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?");
1221         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1222             SYSCTL_CHILDREN(bbr_probertt),
1223             OID_AUTO, "can_force", CTLFLAG_RW,
1224             &bbr_can_force_probertt, 0,
1225             "If we keep setting new low rtt's but delay going in probe-rtt can we force in??");
1226         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1227             SYSCTL_CHILDREN(bbr_probertt),
1228             OID_AUTO, "enter_sets_force", CTLFLAG_RW,
1229             &bbr_probertt_sets_rtt, 0,
1230             "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?");
1231         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1232             SYSCTL_CHILDREN(bbr_probertt),
1233             OID_AUTO, "can_adjust", CTLFLAG_RW,
1234             &bbr_can_adjust_probertt, 1,
1235             "Can we dynamically adjust the probe-rtt limits and times?");
1236         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1237             SYSCTL_CHILDREN(bbr_probertt),
1238             OID_AUTO, "is_ratio", CTLFLAG_RW,
1239             &bbr_is_ratio, 0,
1240             "is the limit to filter a ratio?");
1241         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1242             SYSCTL_CHILDREN(bbr_probertt),
1243             OID_AUTO, "use_cwnd", CTLFLAG_RW,
1244             &bbr_prtt_slam_cwnd, 0,
1245             "Should we set/recover cwnd?");
1246         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1247             SYSCTL_CHILDREN(bbr_probertt),
1248             OID_AUTO, "can_use_ts", CTLFLAG_RW,
1249             &bbr_can_use_ts_for_rtt, 1,
1250             "Can we use the ms timestamp if available for retransmistted rtt calculations?");
1251
1252         /* Pacing controls */
1253         bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1254             SYSCTL_CHILDREN(bbr_sysctl_root),
1255             OID_AUTO,
1256             "pacing",
1257             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1258             "");
1259         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1260             SYSCTL_CHILDREN(bbr_hptsi),
1261             OID_AUTO, "hw_pacing", CTLFLAG_RW,
1262             &bbr_allow_hdwr_pacing, 1,
1263             "Do we allow hardware pacing?");
1264         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1265             SYSCTL_CHILDREN(bbr_hptsi),
1266             OID_AUTO, "hw_pacing_limit", CTLFLAG_RW,
1267             &bbr_hardware_pacing_limit, 4000,
1268             "Do we have a limited number of connections for pacing chelsio (0=no limit)?");
1269         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1270             SYSCTL_CHILDREN(bbr_hptsi),
1271             OID_AUTO, "hw_pacing_adj", CTLFLAG_RW,
1272             &bbr_hdwr_pace_adjust, 2,
1273             "Multiplier to calculated tso size?");
1274         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1275             SYSCTL_CHILDREN(bbr_hptsi),
1276             OID_AUTO, "hw_pacing_floor", CTLFLAG_RW,
1277             &bbr_hdwr_pace_floor, 1,
1278             "Do we invoke the hardware pacing floor?");
1279         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1280             SYSCTL_CHILDREN(bbr_hptsi),
1281             OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW,
1282             &bbr_hdwr_pacing_delay_cnt, 10,
1283             "How many packets must be sent after hdwr pacing is enabled");
1284         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1285             SYSCTL_CHILDREN(bbr_hptsi),
1286             OID_AUTO, "bw_cross", CTLFLAG_RW,
1287             &bbr_cross_over, 3000000,
1288             "What is the point where we cross over to linux like TSO size set");
1289         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1290             SYSCTL_CHILDREN(bbr_hptsi),
1291             OID_AUTO, "seg_deltarg", CTLFLAG_RW,
1292             &bbr_hptsi_segments_delay_tar, 7000,
1293             "What is the worse case delay target for hptsi < 48Mbp connections");
1294         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1295             SYSCTL_CHILDREN(bbr_hptsi),
1296             OID_AUTO, "enet_oh", CTLFLAG_RW,
1297             &bbr_include_enet_oh, 0,
1298             "Do we include the ethernet overhead in calculating pacing delay?");
1299         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1300             SYSCTL_CHILDREN(bbr_hptsi),
1301             OID_AUTO, "ip_oh", CTLFLAG_RW,
1302             &bbr_include_ip_oh, 1,
1303             "Do we include the IP overhead in calculating pacing delay?");
1304         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1305             SYSCTL_CHILDREN(bbr_hptsi),
1306             OID_AUTO, "tcp_oh", CTLFLAG_RW,
1307             &bbr_include_tcp_oh, 0,
1308             "Do we include the TCP overhead in calculating pacing delay?");
1309         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1310             SYSCTL_CHILDREN(bbr_hptsi),
1311             OID_AUTO, "google_discount", CTLFLAG_RW,
1312             &bbr_google_discount, 10,
1313             "What is the default google discount percentage wise for pacing (11 = 1.1%%)?");
1314         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1315             SYSCTL_CHILDREN(bbr_hptsi),
1316             OID_AUTO, "all_get_min", CTLFLAG_RW,
1317             &bbr_all_get_min, 0,
1318             "If you are less than a MSS do you just get the min?");
1319         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1320             SYSCTL_CHILDREN(bbr_hptsi),
1321             OID_AUTO, "tso_min", CTLFLAG_RW,
1322             &bbr_hptsi_bytes_min, 1460,
1323             "For 0 -> 24Mbps what is floor number of segments for TSO");
1324         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1325             SYSCTL_CHILDREN(bbr_hptsi),
1326             OID_AUTO, "seg_tso_max", CTLFLAG_RW,
1327             &bbr_hptsi_segments_max, 6,
1328             "For 0 -> 24Mbps what is top number of segments for TSO");
1329         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1330             SYSCTL_CHILDREN(bbr_hptsi),
1331             OID_AUTO, "seg_floor", CTLFLAG_RW,
1332             &bbr_hptsi_segments_floor, 1,
1333             "Minimum TSO size we will fall too in segments");
1334         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1335             SYSCTL_CHILDREN(bbr_hptsi),
1336             OID_AUTO, "utter_max", CTLFLAG_RW,
1337             &bbr_hptsi_utter_max, 0,
1338             "The absolute maximum that any pacing (outside of hardware) can be");
1339         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1340             SYSCTL_CHILDREN(bbr_hptsi),
1341             OID_AUTO, "seg_divisor", CTLFLAG_RW,
1342             &bbr_hptsi_per_second, 100,
1343             "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps ");
1344         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1345             SYSCTL_CHILDREN(bbr_hptsi),
1346             OID_AUTO, "srtt_mul", CTLFLAG_RW,
1347             &bbr_hptsi_max_mul, 1,
1348             "The multiplier for pace len max");
1349         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1350             SYSCTL_CHILDREN(bbr_hptsi),
1351             OID_AUTO, "srtt_div", CTLFLAG_RW,
1352             &bbr_hptsi_max_div, 2,
1353             "The divisor for pace len max");
1354         /* Measurement controls */
1355         bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1356             SYSCTL_CHILDREN(bbr_sysctl_root),
1357             OID_AUTO,
1358             "measure",
1359             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1360             "Measurement controls");
1361         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1362             SYSCTL_CHILDREN(bbr_measure),
1363             OID_AUTO, "min_i_bw", CTLFLAG_RW,
1364             &bbr_initial_bw_bps, 62500,
1365             "Minimum initial b/w in bytes per second");
1366         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1367             SYSCTL_CHILDREN(bbr_measure),
1368             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1369             &bbr_sack_not_required, 0,
1370             "Do we allow bbr to run on connections not supporting SACK?");
1371         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1372             SYSCTL_CHILDREN(bbr_measure),
1373             OID_AUTO, "use_google", CTLFLAG_RW,
1374             &bbr_use_google_algo, 0,
1375             "Use has close to google V1.0 has possible?");
1376         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1377             SYSCTL_CHILDREN(bbr_measure),
1378             OID_AUTO, "ts_limiting", CTLFLAG_RW,
1379             &bbr_ts_limiting, 1,
1380             "Do we attempt to use the peers timestamp to limit b/w caculations?");
1381         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1382             SYSCTL_CHILDREN(bbr_measure),
1383             OID_AUTO, "ts_can_raise", CTLFLAG_RW,
1384             &bbr_ts_can_raise, 0,
1385             "Can we raise the b/w via timestamp b/w calculation?");
1386         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1387             SYSCTL_CHILDREN(bbr_measure),
1388             OID_AUTO, "ts_delta", CTLFLAG_RW,
1389             &bbr_min_usec_delta, 20000,
1390             "How long in usec between ts of our sends in ts validation code?");
1391         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1392             SYSCTL_CHILDREN(bbr_measure),
1393             OID_AUTO, "ts_peer_delta", CTLFLAG_RW,
1394             &bbr_min_peer_delta, 20,
1395             "What min numerical value should be between the peer deltas?");
1396         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1397             SYSCTL_CHILDREN(bbr_measure),
1398             OID_AUTO, "ts_delta_percent", CTLFLAG_RW,
1399             &bbr_delta_percent, 150,
1400             "What percentage (150 = 15.0) do we allow variance for?");
1401         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1402             SYSCTL_CHILDREN(bbr_measure),
1403             OID_AUTO, "min_measure_good_bw", CTLFLAG_RW,
1404             &bbr_min_measurements_req, 1,
1405             "What is the minimum measurment count we need before we switch to our b/w estimate");
1406         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1407             SYSCTL_CHILDREN(bbr_measure),
1408             OID_AUTO, "min_measure_before_pace", CTLFLAG_RW,
1409             &bbr_no_pacing_until, 4,
1410             "How many pkt-epoch's (0 is off) do we need before pacing is on?");
1411         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1412             SYSCTL_CHILDREN(bbr_measure),
1413             OID_AUTO, "quanta", CTLFLAG_RW,
1414             &bbr_quanta, 2,
1415             "Extra quanta to add when calculating the target (ID section 4.2.3.2).");
1416         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1417             SYSCTL_CHILDREN(bbr_measure),
1418             OID_AUTO, "noretran", CTLFLAG_RW,
1419             &bbr_no_retran, 0,
1420             "Should google mode not use retransmission measurements for the b/w estimation?");
1421         /* State controls */
1422         bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1423             SYSCTL_CHILDREN(bbr_sysctl_root),
1424             OID_AUTO,
1425             "states",
1426             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1427             "State controls");
1428         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1429             SYSCTL_CHILDREN(bbr_states),
1430             OID_AUTO, "idle_restart", CTLFLAG_RW,
1431             &bbr_uses_idle_restart, 0,
1432             "Do we use a new special idle_restart state to ramp back up quickly?");
1433         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1434             SYSCTL_CHILDREN(bbr_states),
1435             OID_AUTO, "idle_restart_threshold", CTLFLAG_RW,
1436             &bbr_idle_restart_threshold, 100000,
1437             "How long must we be idle before we restart??");
1438         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1439             SYSCTL_CHILDREN(bbr_states),
1440             OID_AUTO, "use_pkt_epoch", CTLFLAG_RW,
1441             &bbr_state_is_pkt_epoch, 0,
1442             "Do we use a pkt-epoch for substate if 0 rttProp?");
1443         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1444             SYSCTL_CHILDREN(bbr_states),
1445             OID_AUTO, "startup_rtt_gain", CTLFLAG_RW,
1446             &bbr_rtt_gain_thresh, 0,
1447             "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?");
1448         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1449             SYSCTL_CHILDREN(bbr_states),
1450             OID_AUTO, "drain_floor", CTLFLAG_RW,
1451             &bbr_drain_floor, 88,
1452             "What is the lowest we can drain (pg) too?");
1453         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1454             SYSCTL_CHILDREN(bbr_states),
1455             OID_AUTO, "drain_2_target", CTLFLAG_RW,
1456             &bbr_state_drain_2_tar, 1,
1457             "Do we drain to target in drain substate?");
1458         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1459             SYSCTL_CHILDREN(bbr_states),
1460             OID_AUTO, "gain_2_target", CTLFLAG_RW,
1461             &bbr_gain_to_target, 1,
1462             "Does probe bw gain to target??");
1463         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1464             SYSCTL_CHILDREN(bbr_states),
1465             OID_AUTO, "gain_extra_time", CTLFLAG_RW,
1466             &bbr_gain_gets_extra_too, 1,
1467             "Does probe bw gain get the extra time too?");
1468         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1469             SYSCTL_CHILDREN(bbr_states),
1470             OID_AUTO, "ld_div", CTLFLAG_RW,
1471             &bbr_drain_drop_div, 5,
1472             "Long drain drop divider?");
1473         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1474             SYSCTL_CHILDREN(bbr_states),
1475             OID_AUTO, "ld_mul", CTLFLAG_RW,
1476             &bbr_drain_drop_mul, 4,
1477             "Long drain drop multiplier?");
1478         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1479             SYSCTL_CHILDREN(bbr_states),
1480             OID_AUTO, "rand_ot_disc", CTLFLAG_RW,
1481             &bbr_rand_ot, 50,
1482             "Random discount of the ot?");
1483         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1484             SYSCTL_CHILDREN(bbr_states),
1485             OID_AUTO, "dr_filter_life", CTLFLAG_RW,
1486             &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT,
1487             "How many packet-epochs does the b/w delivery rate last?");
1488         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1489             SYSCTL_CHILDREN(bbr_states),
1490             OID_AUTO, "subdrain_applimited", CTLFLAG_RW,
1491             &bbr_sub_drain_app_limit, 0,
1492             "Does our sub-state drain invoke app limited if its long?");
1493         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1494             SYSCTL_CHILDREN(bbr_states),
1495             OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW,
1496             &bbr_sub_drain_slam_cwnd, 0,
1497             "Should we set/recover cwnd for sub-state drain?");
1498         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1499             SYSCTL_CHILDREN(bbr_states),
1500             OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW,
1501             &bbr_slam_cwnd_in_main_drain, 0,
1502             "Should we set/recover cwnd for main-state drain?");
1503         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1504             SYSCTL_CHILDREN(bbr_states),
1505             OID_AUTO, "google_gets_earlyout", CTLFLAG_RW,
1506             &google_allow_early_out, 1,
1507             "Should we allow google probe-bw/drain to exit early at flight target?");
1508         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1509             SYSCTL_CHILDREN(bbr_states),
1510             OID_AUTO, "google_exit_loss", CTLFLAG_RW,
1511             &google_consider_lost, 1,
1512             "Should we have losses exit gain of probebw in google mode??");
1513         /* Startup controls */
1514         bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1515             SYSCTL_CHILDREN(bbr_sysctl_root),
1516             OID_AUTO,
1517             "startup",
1518             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1519             "Startup controls");
1520         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1521             SYSCTL_CHILDREN(bbr_startup),
1522             OID_AUTO, "cheat_iwnd", CTLFLAG_RW,
1523             &bbr_sends_full_iwnd, 1,
1524             "Do we not pace but burst out initial windows has our TSO size?");
1525         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1526             SYSCTL_CHILDREN(bbr_startup),
1527             OID_AUTO, "loss_threshold", CTLFLAG_RW,
1528             &bbr_startup_loss_thresh, 2000,
1529             "In startup what is the loss threshold in a pe that will exit us from startup?");
1530         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1531             SYSCTL_CHILDREN(bbr_startup),
1532             OID_AUTO, "use_lowerpg", CTLFLAG_RW,
1533             &bbr_use_lower_gain_in_startup, 1,
1534             "Should we use a lower hptsi gain if we see loss in startup?");
1535         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1536             SYSCTL_CHILDREN(bbr_startup),
1537             OID_AUTO, "gain", CTLFLAG_RW,
1538             &bbr_start_exit, 25,
1539             "What gain percent do we need to see to stay in startup??");
1540         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1541             SYSCTL_CHILDREN(bbr_startup),
1542             OID_AUTO, "low_gain", CTLFLAG_RW,
1543             &bbr_low_start_exit, 15,
1544             "What gain percent do we need to see to stay in the lower gain startup??");
1545         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1546             SYSCTL_CHILDREN(bbr_startup),
1547             OID_AUTO, "loss_exit", CTLFLAG_RW,
1548             &bbr_exit_startup_at_loss, 1,
1549             "Should we exit startup at loss in an epoch if we are not gaining?");
1550         /* CWND controls */
1551         bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1552             SYSCTL_CHILDREN(bbr_sysctl_root),
1553             OID_AUTO,
1554             "cwnd",
1555             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1556             "Cwnd controls");
1557         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1558             SYSCTL_CHILDREN(bbr_cwnd),
1559             OID_AUTO, "tar_rtt", CTLFLAG_RW,
1560             &bbr_cwndtarget_rtt_touse, 0,
1561             "Target cwnd rtt measurment to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?");
1562         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1563             SYSCTL_CHILDREN(bbr_cwnd),
1564             OID_AUTO, "may_shrink", CTLFLAG_RW,
1565             &bbr_cwnd_may_shrink, 0,
1566             "Can the cwnd shrink if it would grow to more than the target?");
1567         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1568             SYSCTL_CHILDREN(bbr_cwnd),
1569             OID_AUTO, "max_target_limit", CTLFLAG_RW,
1570             &bbr_target_cwnd_mult_limit, 8,
1571             "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?");
1572         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1573             SYSCTL_CHILDREN(bbr_cwnd),
1574             OID_AUTO, "highspeed_min", CTLFLAG_RW,
1575             &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS,
1576             "What is the high-speed min cwnd (rttProp under 1ms)");
1577         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1578             SYSCTL_CHILDREN(bbr_cwnd),
1579             OID_AUTO, "lowspeed_min", CTLFLAG_RW,
1580             &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS,
1581             "What is the min cwnd (rttProp > 1ms)");
1582         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1583             SYSCTL_CHILDREN(bbr_cwnd),
1584             OID_AUTO, "initwin", CTLFLAG_RW,
1585             &bbr_def_init_win, 10,
1586             "What is the BBR initial window, if 0 use tcp version");
1587         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1588             SYSCTL_CHILDREN(bbr_cwnd),
1589             OID_AUTO, "do_loss_red", CTLFLAG_RW,
1590             &bbr_do_red, 600,
1591             "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?");
1592         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1593             SYSCTL_CHILDREN(bbr_cwnd),
1594             OID_AUTO, "red_scale", CTLFLAG_RW,
1595             &bbr_red_scale, 20000,
1596             "What RTT do we scale with?");
1597         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1598             SYSCTL_CHILDREN(bbr_cwnd),
1599             OID_AUTO, "red_growslow", CTLFLAG_RW,
1600             &bbr_red_growth_restrict, 1,
1601             "Do we restrict cwnd growth for whats in flight?");
1602         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1603             SYSCTL_CHILDREN(bbr_cwnd),
1604             OID_AUTO, "red_div", CTLFLAG_RW,
1605             &bbr_red_div, 2,
1606             "If we reduce whats the divisor?");
1607         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1608             SYSCTL_CHILDREN(bbr_cwnd),
1609             OID_AUTO, "red_mul", CTLFLAG_RW,
1610             &bbr_red_mul, 1,
1611             "If we reduce whats the mulitiplier?");
1612         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1613             SYSCTL_CHILDREN(bbr_cwnd),
1614             OID_AUTO, "target_is_unit", CTLFLAG_RW,
1615             &bbr_target_is_bbunit, 0,
1616             "Is the state target the pacing_gain or BBR_UNIT?");
1617         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1618             SYSCTL_CHILDREN(bbr_cwnd),
1619             OID_AUTO, "drop_limit", CTLFLAG_RW,
1620             &bbr_drop_limit, 0,
1621             "Number of segments limit for drop (0=use min_cwnd w/flight)?");
1622
1623         /* Timeout controls */
1624         bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1625             SYSCTL_CHILDREN(bbr_sysctl_root),
1626             OID_AUTO,
1627             "timeout",
1628             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1629             "Time out controls");
1630         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1631             SYSCTL_CHILDREN(bbr_timeout),
1632             OID_AUTO, "delack", CTLFLAG_RW,
1633             &bbr_delack_time, 100000,
1634             "BBR's delayed ack time");
1635         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1636             SYSCTL_CHILDREN(bbr_timeout),
1637             OID_AUTO, "tlp_uses", CTLFLAG_RW,
1638             &bbr_tlp_type_to_use, 3,
1639             "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt");
1640         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1641             SYSCTL_CHILDREN(bbr_timeout),
1642             OID_AUTO, "persmin", CTLFLAG_RW,
1643             &bbr_persist_min, 250000,
1644             "What is the minimum time in microseconds between persists");
1645         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1646             SYSCTL_CHILDREN(bbr_timeout),
1647             OID_AUTO, "persmax", CTLFLAG_RW,
1648             &bbr_persist_max, 1000000,
1649             "What is the largest delay in microseconds between persists");
1650         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1651             SYSCTL_CHILDREN(bbr_timeout),
1652             OID_AUTO, "tlp_minto", CTLFLAG_RW,
1653             &bbr_tlp_min, 10000,
1654             "TLP Min timeout in usecs");
1655         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1656             SYSCTL_CHILDREN(bbr_timeout),
1657             OID_AUTO, "tlp_dack_time", CTLFLAG_RW,
1658             &bbr_delayed_ack_time, 200000,
1659             "TLP delayed ack compensation value");
1660         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1661             SYSCTL_CHILDREN(bbr_sysctl_root),
1662             OID_AUTO, "minrto", CTLFLAG_RW,
1663             &bbr_rto_min_ms, 30,
1664             "Minimum RTO in ms");
1665         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1666             SYSCTL_CHILDREN(bbr_timeout),
1667             OID_AUTO, "maxrto", CTLFLAG_RW,
1668             &bbr_rto_max_sec, 4,
1669             "Maxiumum RTO in seconds -- should be at least as large as min_rto");
1670         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1671             SYSCTL_CHILDREN(bbr_timeout),
1672             OID_AUTO, "tlp_retry", CTLFLAG_RW,
1673             &bbr_tlp_max_resend, 2,
1674             "How many times does TLP retry a single segment or multiple with no ACK");
1675         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1676             SYSCTL_CHILDREN(bbr_timeout),
1677             OID_AUTO, "minto", CTLFLAG_RW,
1678             &bbr_min_to, 1000,
1679             "Minimum rack timeout in useconds");
1680         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1681             SYSCTL_CHILDREN(bbr_timeout),
1682             OID_AUTO, "pktdelay", CTLFLAG_RW,
1683             &bbr_pkt_delay, 1000,
1684             "Extra RACK time (in useconds) besides reordering thresh");
1685         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1686             SYSCTL_CHILDREN(bbr_timeout),
1687             OID_AUTO, "incr_tmrs", CTLFLAG_RW,
1688             &bbr_incr_timers, 1,
1689             "Increase the RXT/TLP timer by the pacing time used?");
1690         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1691             SYSCTL_CHILDREN(bbr_timeout),
1692             OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW,
1693             &bbr_marks_rxt_sack_passed, 0,
1694             "Mark sack passed on all those not ack'd when a RXT hits?");
1695         /* Policer controls */
1696         bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1697             SYSCTL_CHILDREN(bbr_sysctl_root),
1698             OID_AUTO,
1699             "policer",
1700             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1701             "Policer controls");
1702         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1703             SYSCTL_CHILDREN(bbr_policer),
1704             OID_AUTO, "detect_enable", CTLFLAG_RW,
1705             &bbr_policer_detection_enabled, 1,
1706             "Is policer detection enabled??");
1707         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1708             SYSCTL_CHILDREN(bbr_policer),
1709             OID_AUTO, "min_pes", CTLFLAG_RW,
1710             &bbr_lt_intvl_min_rtts, 4,
1711             "Minimum number of PE's?");
1712         SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1713             SYSCTL_CHILDREN(bbr_policer),
1714             OID_AUTO, "bwdiff", CTLFLAG_RW,
1715             &bbr_lt_bw_diff, (4000/8),
1716             "Minimal bw diff?");
1717         SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1718             SYSCTL_CHILDREN(bbr_policer),
1719             OID_AUTO, "bwratio", CTLFLAG_RW,
1720             &bbr_lt_bw_ratio, 8,
1721             "Minimal bw diff?");
1722         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1723             SYSCTL_CHILDREN(bbr_policer),
1724             OID_AUTO, "from_rack_rxt", CTLFLAG_RW,
1725             &bbr_policer_call_from_rack_to, 0,
1726             "Do we call the policer detection code from a rack-timeout?");
1727         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1728             SYSCTL_CHILDREN(bbr_policer),
1729             OID_AUTO, "false_postive", CTLFLAG_RW,
1730             &bbr_lt_intvl_fp, 0,
1731             "What packet epoch do we do false-postive detection at (0=no)?");
1732         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1733             SYSCTL_CHILDREN(bbr_policer),
1734             OID_AUTO, "loss_thresh", CTLFLAG_RW,
1735             &bbr_lt_loss_thresh, 196,
1736             "Loss threshold 196 = 19.6%?");
1737         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1738             SYSCTL_CHILDREN(bbr_policer),
1739             OID_AUTO, "false_postive_thresh", CTLFLAG_RW,
1740             &bbr_lt_fd_thresh, 100,
1741             "What percentage is the false detection threshold (150=15.0)?");
1742         /* All the rest */
1743         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1744             SYSCTL_CHILDREN(bbr_sysctl_root),
1745             OID_AUTO, "cheat_rxt", CTLFLAG_RW,
1746             &bbr_use_rack_resend_cheat, 0,
1747             "Do we burst 1ms between sends on retransmissions (like rack)?");
1748         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1749             SYSCTL_CHILDREN(bbr_sysctl_root),
1750             OID_AUTO, "error_paceout", CTLFLAG_RW,
1751             &bbr_error_base_paceout, 10000,
1752             "When we hit an error what is the min to pace out in usec's?");
1753         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1754             SYSCTL_CHILDREN(bbr_sysctl_root),
1755             OID_AUTO, "kill_paceout", CTLFLAG_RW,
1756             &bbr_max_net_error_cnt, 10,
1757             "When we hit this many errors in a row, kill the session?");
1758         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1759             SYSCTL_CHILDREN(bbr_sysctl_root),
1760             OID_AUTO, "data_after_close", CTLFLAG_RW,
1761             &bbr_ignore_data_after_close, 1,
1762             "Do we hold off sending a RST until all pending data is ack'd");
1763         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1764             SYSCTL_CHILDREN(bbr_sysctl_root),
1765             OID_AUTO, "resend_use_tso", CTLFLAG_RW,
1766             &bbr_resends_use_tso, 0,
1767             "Can resends use TSO?");
1768         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1769             SYSCTL_CHILDREN(bbr_sysctl_root),
1770             OID_AUTO, "sblklimit", CTLFLAG_RW,
1771             &bbr_sack_block_limit, 128,
1772             "When do we start ignoring small sack blocks");
1773         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1774             SYSCTL_CHILDREN(bbr_sysctl_root),
1775             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1776             &bbr_verbose_logging, 0,
1777             "Should BBR black box logging be verbose");
1778         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1779             SYSCTL_CHILDREN(bbr_sysctl_root),
1780             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1781             &bbr_reorder_thresh, 2,
1782             "What factor for rack will be added when seeing reordering (shift right)");
1783         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1784             SYSCTL_CHILDREN(bbr_sysctl_root),
1785             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1786             &bbr_reorder_fade, 0,
1787             "Does reorder detection fade, if so how many ms (0 means never)");
1788         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1789             SYSCTL_CHILDREN(bbr_sysctl_root),
1790             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1791             &bbr_tlp_thresh, 1,
1792             "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1793         /* Stats and counters */
1794         /* The pacing counters for hdwr/software can't be in the array */
1795         bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1796         bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1797         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1798             SYSCTL_CHILDREN(bbr_sysctl_root),
1799             OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD,
1800             &bbr_hdwr_pacing_enobuf,
1801             "Total number of enobufs for hardware paced flows");
1802         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1803             SYSCTL_CHILDREN(bbr_sysctl_root),
1804             OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD,
1805             &bbr_nohdwr_pacing_enobuf,
1806             "Total number of enobufs for non-hardware paced flows");
1807
1808         bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK);
1809         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1810             SYSCTL_CHILDREN(bbr_sysctl_root),
1811             OID_AUTO, "hdwr_pacing", CTLFLAG_RD,
1812             &bbr_flows_whdwr_pacing,
1813             "Total number of hardware paced flows");
1814         bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK);
1815         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1816             SYSCTL_CHILDREN(bbr_sysctl_root),
1817             OID_AUTO, "software_pacing", CTLFLAG_RD,
1818             &bbr_flows_nohdwr_pacing,
1819             "Total number of software paced flows");
1820         COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK);
1821         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1822             OID_AUTO, "stats", CTLFLAG_RD,
1823             bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats");
1824         COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK);
1825         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1826             OID_AUTO, "opts", CTLFLAG_RD,
1827             bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats");
1828         COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK);
1829         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1830             OID_AUTO, "lost", CTLFLAG_RD,
1831             bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur");
1832         COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK);
1833         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1834             OID_AUTO, "stateresend", CTLFLAG_RD,
1835             bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend");
1836         COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK);
1837         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1838             OID_AUTO, "statetime", CTLFLAG_RD,
1839             bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states");
1840         COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1841         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1842             OID_AUTO, "outsize", CTLFLAG_RD,
1843             bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls");
1844         SYSCTL_ADD_PROC(&bbr_sysctl_ctx,
1845             SYSCTL_CHILDREN(bbr_sysctl_root),
1846             OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1847             &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters");
1848 }
1849
1850 static void
1851 bbr_counter_destroy(void)
1852 {
1853         COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE);
1854         COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE);
1855         COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE);
1856         COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT);
1857         COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT);
1858         COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT);
1859         counter_u64_free(bbr_nohdwr_pacing_enobuf);
1860         counter_u64_free(bbr_hdwr_pacing_enobuf);
1861         counter_u64_free(bbr_flows_whdwr_pacing);
1862         counter_u64_free(bbr_flows_nohdwr_pacing);
1863
1864 }
1865
1866 static __inline void
1867 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts)
1868 {
1869         memset(l, 0, sizeof(union tcp_log_stackspecific));
1870         l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate;
1871         l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate);
1872         l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
1873         l->bw_inuse = bbr_get_bw(bbr);
1874         l->inflight = ctf_flight_size(bbr->rc_tp,
1875                           (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
1876         l->applimited = bbr->r_ctl.r_app_limited_until;
1877         l->delivered = bbr->r_ctl.rc_delivered;
1878         l->timeStamp = cts;
1879         l->lost = bbr->r_ctl.rc_lost;
1880         l->bbr_state = bbr->rc_bbr_state;
1881         l->bbr_substate = bbr_state_val(bbr);
1882         l->epoch = bbr->r_ctl.rc_rtt_epoch;
1883         l->lt_epoch = bbr->r_ctl.rc_lt_epoch;
1884         l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
1885         l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain;
1886         l->inhpts = bbr->rc_inp->inp_in_hpts;
1887         l->ininput = bbr->rc_inp->inp_in_input;
1888         l->use_lt_bw = bbr->rc_lt_use_bw;
1889         l->pkts_out = bbr->r_ctl.rc_flight_at_input;
1890         l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch;
1891 }
1892
1893 static void
1894 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason)
1895 {
1896         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1897                 union tcp_log_stackspecific log;
1898
1899                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1900                 log.u_bbr.flex1 = 0;
1901                 log.u_bbr.flex2 = 0;
1902                 log.u_bbr.flex5 = 0;
1903                 log.u_bbr.flex3 = 0;
1904                 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate;
1905                 log.u_bbr.flex7 = reason;
1906                 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt;
1907                 log.u_bbr.flex8 = 0;
1908                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1909                     &bbr->rc_inp->inp_socket->so_rcv,
1910                     &bbr->rc_inp->inp_socket->so_snd,
1911                     BBR_LOG_BW_RED_EV, 0,
1912                     0, &log, false, &bbr->rc_tv);
1913         }
1914 }
1915
1916 static void
1917 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count)
1918 {
1919         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1920                 union tcp_log_stackspecific log;
1921
1922                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1923                 log.u_bbr.flex1 = seq;
1924                 log.u_bbr.flex2 = count;
1925                 log.u_bbr.flex8 = mode;
1926                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1927                     &bbr->rc_inp->inp_socket->so_rcv,
1928                     &bbr->rc_inp->inp_socket->so_snd,
1929                     BBR_LOG_LOWGAIN, 0,
1930                     0, &log, false, &bbr->rc_tv);
1931         }
1932 }
1933
1934 static void
1935 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling,
1936     uint8_t reason, uint32_t p_maxseg, int len)
1937 {
1938         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1939                 union tcp_log_stackspecific log;
1940
1941                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1942                 log.u_bbr.flex1 = p_maxseg;
1943                 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags;
1944                 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
1945                 log.u_bbr.flex4 = reason;
1946                 log.u_bbr.flex5 = bbr->rc_in_persist;
1947                 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val;
1948                 log.u_bbr.flex7 = p_maxseg;
1949                 log.u_bbr.flex8 = bbr->rc_in_persist;
1950                 log.u_bbr.pkts_out = 0;
1951                 log.u_bbr.applimited = len;
1952                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1953                     &bbr->rc_inp->inp_socket->so_rcv,
1954                     &bbr->rc_inp->inp_socket->so_snd,
1955                     BBR_LOG_JUSTRET, 0,
1956                     tlen, &log, false, &bbr->rc_tv);
1957         }
1958 }
1959
1960 static void
1961 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq)
1962 {
1963         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1964                 union tcp_log_stackspecific log;
1965
1966                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1967                 log.u_bbr.flex1 = seq;
1968                 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
1969                 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start;
1970                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1971                     &bbr->rc_inp->inp_socket->so_rcv,
1972                     &bbr->rc_inp->inp_socket->so_snd,
1973                     BBR_LOG_ENTREC, 0,
1974                     0, &log, false, &bbr->rc_tv);
1975         }
1976 }
1977
1978 static void
1979 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)
1980 {
1981         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1982                 union tcp_log_stackspecific log;
1983
1984                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1985                 log.u_bbr.flex1 = tso;
1986                 log.u_bbr.flex2 = maxseg;
1987                 log.u_bbr.flex3 = mtu;
1988                 log.u_bbr.flex4 = csum_flags;
1989                 TCP_LOG_EVENTP(tp, NULL,
1990                     &bbr->rc_inp->inp_socket->so_rcv,
1991                     &bbr->rc_inp->inp_socket->so_snd,
1992                     BBR_LOG_MSGSIZE, 0,
1993                     0, &log, false, &bbr->rc_tv);
1994         }
1995 }
1996
1997 static void
1998 bbr_log_flowend(struct tcp_bbr *bbr)
1999 {
2000         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2001                 union tcp_log_stackspecific log;
2002                 struct sockbuf *r, *s;
2003                 struct timeval tv;
2004
2005                 if (bbr->rc_inp->inp_socket) {
2006                         r = &bbr->rc_inp->inp_socket->so_rcv;
2007                         s = &bbr->rc_inp->inp_socket->so_snd;
2008                 } else {
2009                         r = s = NULL;
2010                 }
2011                 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv));
2012                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2013                     r, s,
2014                     TCP_LOG_FLOWEND, 0,
2015                     0, &log, false, &tv);
2016         }
2017 }
2018
2019 static void
2020 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line,
2021     uint32_t lost, uint32_t del)
2022 {
2023         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2024                 union tcp_log_stackspecific log;
2025
2026                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2027                 log.u_bbr.flex1 = lost;
2028                 log.u_bbr.flex2 = del;
2029                 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw;
2030                 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt;
2031                 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2032                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2033                 log.u_bbr.flex7 = line;
2034                 log.u_bbr.flex8 = 0;
2035                 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count;
2036                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2037                     &bbr->rc_inp->inp_socket->so_rcv,
2038                     &bbr->rc_inp->inp_socket->so_snd,
2039                     BBR_LOG_PKT_EPOCH, 0,
2040                     0, &log, false, &bbr->rc_tv);
2041         }
2042 }
2043
2044 static void
2045 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time)
2046 {
2047         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2048                 union tcp_log_stackspecific log;
2049
2050                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2051                 log.u_bbr.flex1 = bbr->r_ctl.rc_lost;
2052                 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat;
2053                 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat;
2054                 log.u_bbr.flex7 = line;
2055                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2056                     &bbr->rc_inp->inp_socket->so_rcv,
2057                     &bbr->rc_inp->inp_socket->so_snd,
2058                     BBR_LOG_TIME_EPOCH, 0,
2059                     0, &log, false, &bbr->rc_tv);
2060         }
2061 }
2062
2063 static void
2064 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth)
2065 {
2066         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2067                 union tcp_log_stackspecific log;
2068
2069                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2070                 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2071                 log.u_bbr.flex2 = new_tar;
2072                 log.u_bbr.flex3 = line;
2073                 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2074                 log.u_bbr.flex5 = bbr_quanta;
2075                 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs;
2076                 log.u_bbr.flex7 = bbr->rc_last_options;
2077                 log.u_bbr.flex8 = meth;
2078                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2079                     &bbr->rc_inp->inp_socket->so_rcv,
2080                     &bbr->rc_inp->inp_socket->so_snd,
2081                     BBR_LOG_STATE_TARGET, 0,
2082                     0, &log, false, &bbr->rc_tv);
2083         }
2084
2085 }
2086
2087 static void
2088 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2089 {
2090         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2091                 union tcp_log_stackspecific log;
2092
2093                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2094                 log.u_bbr.flex1 = line;
2095                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2096                 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2097                 if (bbr_state_is_pkt_epoch)
2098                         log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
2099                 else
2100                         log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP);
2101                 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2102                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2103                 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000);
2104                 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra;
2105                 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state;
2106                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2107                     &bbr->rc_inp->inp_socket->so_rcv,
2108                     &bbr->rc_inp->inp_socket->so_snd,
2109                     BBR_LOG_STATE, 0,
2110                     0, &log, false, &bbr->rc_tv);
2111         }
2112 }
2113
2114 static void
2115 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied,
2116                     uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond)
2117 {
2118         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2119                 union tcp_log_stackspecific log;
2120
2121                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2122                 log.u_bbr.flex1 = line;
2123                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2124                 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt;
2125                 log.u_bbr.flex4 = applied;
2126                 log.u_bbr.flex5 = rtt;
2127                 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2128                 log.u_bbr.flex7 = cond;
2129                 log.u_bbr.flex8 = reas;
2130                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2131                     &bbr->rc_inp->inp_socket->so_rcv,
2132                     &bbr->rc_inp->inp_socket->so_snd,
2133                     BBR_LOG_RTT_SHRINKS, 0,
2134                     0, &log, false, &bbr->rc_tv);
2135         }
2136 }
2137
2138 static void
2139 bbr_log_type_exit_rec(struct tcp_bbr *bbr)
2140 {
2141         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2142                 union tcp_log_stackspecific log;
2143
2144                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2145                 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start;
2146                 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
2147                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2148                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2149                     &bbr->rc_inp->inp_socket->so_rcv,
2150                     &bbr->rc_inp->inp_socket->so_snd,
2151                     BBR_LOG_EXITREC, 0,
2152                     0, &log, false, &bbr->rc_tv);
2153         }
2154 }
2155
2156 static void
2157 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg,
2158     uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line)
2159 {
2160         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2161                 union tcp_log_stackspecific log;
2162
2163                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2164                 log.u_bbr.flex1 = line;
2165                 log.u_bbr.flex2 = prev_acked;
2166                 log.u_bbr.flex3 = bytes_this_ack;
2167                 log.u_bbr.flex4 = chg;
2168                 log.u_bbr.flex5 = th_ack;
2169                 log.u_bbr.flex6 = target;
2170                 log.u_bbr.flex8 = meth;
2171                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2172                     &bbr->rc_inp->inp_socket->so_rcv,
2173                     &bbr->rc_inp->inp_socket->so_snd,
2174                     BBR_LOG_CWND, 0,
2175                     0, &log, false, &bbr->rc_tv);
2176         }
2177 }
2178
2179 static void
2180 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin)
2181 {
2182         /*
2183          * Log the rtt sample we are applying to the srtt algorithm in
2184          * useconds.
2185          */
2186         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2187                 union tcp_log_stackspecific log;
2188
2189                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2190                 log.u_bbr.flex1 = rtt;
2191                 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time;
2192                 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay;
2193                 log.u_bbr.flex4 = bbr->rc_tp->ts_offset;
2194                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2195                 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv);
2196                 log.u_bbr.flex6 = tsin;
2197                 log.u_bbr.flex7 = 0;
2198                 log.u_bbr.flex8 = bbr->rc_ack_was_delayed;
2199                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2200                     &bbr->rc_inp->inp_socket->so_rcv,
2201                     &bbr->rc_inp->inp_socket->so_snd,
2202                     TCP_LOG_RTT, 0,
2203                     0, &log, false, &bbr->rc_tv);
2204         }
2205 }
2206
2207 static void
2208 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit)
2209 {
2210         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2211                 union tcp_log_stackspecific log;
2212
2213                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2214                 log.u_bbr.flex1 = time_in;
2215                 log.u_bbr.flex2 = line;
2216                 log.u_bbr.flex8 = enter_exit;
2217                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2218                     &bbr->rc_inp->inp_socket->so_rcv,
2219                     &bbr->rc_inp->inp_socket->so_snd,
2220                     BBR_LOG_PERSIST, 0,
2221                     0, &log, false, &bbr->rc_tv);
2222         }
2223 }
2224 static void
2225 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts)
2226 {
2227         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2228                 union tcp_log_stackspecific log;
2229
2230                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2231                 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age;
2232                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2233                 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2234                 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time;
2235                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2236                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2237                     &bbr->rc_inp->inp_socket->so_rcv,
2238                     &bbr->rc_inp->inp_socket->so_snd,
2239                     BBR_LOG_ACKCLEAR, 0,
2240                     0, &log, false, &bbr->rc_tv);
2241         }
2242 }
2243
2244 static void
2245 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen,
2246                   uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m)
2247 {
2248         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2249                 union tcp_log_stackspecific log;
2250                 struct timeval tv;
2251
2252                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2253                 log.u_bbr.flex1 = nsegs;
2254                 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes;
2255                 if (m) {
2256                         struct timespec ts;
2257
2258                         log.u_bbr.flex3 = m->m_flags;
2259                         if (m->m_flags & M_TSTMP) {
2260                                 mbuf_tstmp2timespec(m, &ts);
2261                                 tv.tv_sec = ts.tv_sec;
2262                                 tv.tv_usec = ts.tv_nsec / 1000;
2263                                 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv);
2264                         } else {
2265                                 log.u_bbr.lt_epoch = 0;
2266                         }
2267                         if (m->m_flags & M_TSTMP_LRO) {
2268                                 tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000;
2269                                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000;
2270                                 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv);
2271                         } else {
2272                                 /* No arrival timestamp */
2273                                 log.u_bbr.flex5 = 0;
2274                         }
2275
2276                         log.u_bbr.pkts_out = tcp_get_usecs(&tv);
2277                 } else {
2278                         log.u_bbr.flex3 = 0;
2279                         log.u_bbr.flex5 = 0;
2280                         log.u_bbr.flex6 = 0;
2281                         log.u_bbr.pkts_out = 0;
2282                 }
2283                 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state;
2284                 log.u_bbr.flex7 = bbr->r_wanted_output;
2285                 log.u_bbr.flex8 = bbr->rc_in_persist;
2286                 TCP_LOG_EVENTP(bbr->rc_tp, th,
2287                     &bbr->rc_inp->inp_socket->so_rcv,
2288                     &bbr->rc_inp->inp_socket->so_snd,
2289                     TCP_LOG_IN, 0,
2290                     tlen, &log, true, &bbr->rc_tv);
2291         }
2292 }
2293
2294 static void
2295 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out)
2296 {
2297         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2298                 union tcp_log_stackspecific log;
2299
2300                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2301                 log.u_bbr.flex1 = did_out;
2302                 log.u_bbr.flex2 = nxt_pkt;
2303                 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val;
2304                 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2305                 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp;
2306                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes;
2307                 log.u_bbr.flex7 = bbr->r_wanted_output;
2308                 log.u_bbr.flex8 = bbr->rc_in_persist;
2309                 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay;
2310                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2311                     &bbr->rc_inp->inp_socket->so_rcv,
2312                     &bbr->rc_inp->inp_socket->so_snd,
2313                     BBR_LOG_DOSEG_DONE, 0,
2314                     0, &log, true, &bbr->rc_tv);
2315         }
2316 }
2317
2318 static void
2319 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts,
2320     int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz)
2321 {
2322         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2323                 union tcp_log_stackspecific log;
2324
2325                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2326                 log.u_bbr.flex1 = line;
2327                 log.u_bbr.flex2 = o_len;
2328                 log.u_bbr.flex3 = segcnt;
2329                 log.u_bbr.flex4 = segsiz;
2330                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2331                     &bbr->rc_inp->inp_socket->so_rcv,
2332                     &bbr->rc_inp->inp_socket->so_snd,
2333                     BBR_LOG_ENOBUF_JMP, ENOBUFS,
2334                     len, &log, true, &bbr->rc_tv);
2335         }
2336 }
2337
2338 static void
2339 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling)
2340 {
2341         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2342                 union tcp_log_stackspecific log;
2343
2344                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2345                 log.u_bbr.flex1 = timers;
2346                 log.u_bbr.flex2 = ret;
2347                 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
2348                 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2349                 log.u_bbr.flex5 = cts;
2350                 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2351                 log.u_bbr.flex8 = hpts_calling;
2352                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2353                     &bbr->rc_inp->inp_socket->so_rcv,
2354                     &bbr->rc_inp->inp_socket->so_snd,
2355                     BBR_LOG_TO_PROCESS, 0,
2356                     0, &log, false, &bbr->rc_tv);
2357         }
2358 }
2359
2360 static void
2361 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num)
2362 {
2363         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2364                 union tcp_log_stackspecific log;
2365                 uint64_t ar;
2366
2367                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2368                 log.u_bbr.flex1 = bbr->bbr_timer_src;
2369                 log.u_bbr.flex2 = 0;
2370                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2371                 ar = (uint64_t)(bbr->r_ctl.rc_resend);
2372                 ar >>= 32;
2373                 ar &= 0x00000000ffffffff;
2374                 log.u_bbr.flex4 = (uint32_t)ar;
2375                 ar = (uint64_t)bbr->r_ctl.rc_resend;
2376                 ar &= 0x00000000ffffffff;
2377                 log.u_bbr.flex5 = (uint32_t)ar;
2378                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2379                 log.u_bbr.flex8 = to_num;
2380                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2381                     &bbr->rc_inp->inp_socket->so_rcv,
2382                     &bbr->rc_inp->inp_socket->so_snd,
2383                     BBR_LOG_RTO, 0,
2384                     0, &log, false, &bbr->rc_tv);
2385         }
2386 }
2387
2388 static void
2389 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason)
2390 {
2391         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2392                 union tcp_log_stackspecific log;
2393
2394                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2395                 log.u_bbr.flex1 = flex1;
2396                 log.u_bbr.flex2 = flex2;
2397                 log.u_bbr.flex3 = flex3;
2398                 log.u_bbr.flex4 = 0;
2399                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2400                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2401                 log.u_bbr.flex8 = reason;
2402                 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw;
2403                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2404                     &bbr->rc_inp->inp_socket->so_rcv,
2405                     &bbr->rc_inp->inp_socket->so_snd,
2406                     BBR_LOG_REDUCE, 0,
2407                     0, &log, false, &bbr->rc_tv);
2408         }
2409 }
2410
2411 static void
2412 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag)
2413 {
2414         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2415                 union tcp_log_stackspecific log;
2416
2417                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2418                 log.u_bbr.flex1 = diag->p_nxt_slot;
2419                 log.u_bbr.flex2 = diag->p_cur_slot;
2420                 log.u_bbr.flex3 = diag->slot_req;
2421                 log.u_bbr.flex4 = diag->inp_hptsslot;
2422                 log.u_bbr.flex5 = diag->slot_remaining;
2423                 log.u_bbr.flex6 = diag->need_new_to;
2424                 log.u_bbr.flex7 = diag->p_hpts_active;
2425                 log.u_bbr.flex8 = diag->p_on_min_sleep;
2426                 /* Hijack other fields as needed  */
2427                 log.u_bbr.epoch = diag->have_slept;
2428                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
2429                 log.u_bbr.pkts_out = diag->co_ret;
2430                 log.u_bbr.applimited = diag->hpts_sleep_time;
2431                 log.u_bbr.delivered = diag->p_prev_slot;
2432                 log.u_bbr.inflight = diag->p_runningslot;
2433                 log.u_bbr.bw_inuse = diag->wheel_slot;
2434                 log.u_bbr.rttProp = diag->wheel_cts;
2435                 log.u_bbr.delRate = diag->maxslots;
2436                 log.u_bbr.cur_del_rate = diag->p_curtick;
2437                 log.u_bbr.cur_del_rate <<= 32;
2438                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
2439                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2440                     &bbr->rc_inp->inp_socket->so_rcv,
2441                     &bbr->rc_inp->inp_socket->so_snd,
2442                     BBR_LOG_HPTSDIAG, 0,
2443                     0, &log, false, &bbr->rc_tv);
2444         }
2445 }
2446
2447 static void
2448 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
2449     uint32_t thresh, uint32_t to)
2450 {
2451         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2452                 union tcp_log_stackspecific log;
2453
2454                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2455                 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar;
2456                 log.u_bbr.flex2 = time_since_sent;
2457                 log.u_bbr.flex3 = srtt;
2458                 log.u_bbr.flex4 = thresh;
2459                 log.u_bbr.flex5 = to;
2460                 log.u_bbr.flex6 = bbr->rc_tp->t_srtt;
2461                 log.u_bbr.flex8 = mode;
2462                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2463                     &bbr->rc_inp->inp_socket->so_rcv,
2464                     &bbr->rc_inp->inp_socket->so_snd,
2465                     BBR_LOG_TIMERPREP, 0,
2466                     0, &log, false, &bbr->rc_tv);
2467         }
2468 }
2469
2470 static void
2471 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
2472     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod)
2473 {
2474         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2475                 union tcp_log_stackspecific log;
2476
2477                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2478                 log.u_bbr.flex1 = usecs;
2479                 log.u_bbr.flex2 = len;
2480                 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff);
2481                 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff);
2482                 if (override)
2483                         log.u_bbr.flex5 = (1 << 2);
2484                 else
2485                         log.u_bbr.flex5 = 0;
2486                 log.u_bbr.flex6 = override;
2487                 log.u_bbr.flex7 = gain;
2488                 log.u_bbr.flex8 = mod;
2489                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2490                     &bbr->rc_inp->inp_socket->so_rcv,
2491                     &bbr->rc_inp->inp_socket->so_snd,
2492                     BBR_LOG_HPTSI_CALC, 0,
2493                     len, &log, false, &bbr->rc_tv);
2494         }
2495 }
2496
2497 static void
2498 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2499 {
2500         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2501                 union tcp_log_stackspecific log;
2502
2503                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2504
2505                 log.u_bbr.flex1 = bbr->bbr_timer_src;
2506                 log.u_bbr.flex2 = to;
2507                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2508                 log.u_bbr.flex4 = slot;
2509                 log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot;
2510                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2511                 log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2;
2512                 log.u_bbr.flex8 = which;
2513                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2514                     &bbr->rc_inp->inp_socket->so_rcv,
2515                     &bbr->rc_inp->inp_socket->so_snd,
2516                     BBR_LOG_TIMERSTAR, 0,
2517                     0, &log, false, &bbr->rc_tv);
2518         }
2519 }
2520
2521 static void
2522 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)
2523 {
2524         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2525                 union tcp_log_stackspecific log;
2526
2527                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2528                 log.u_bbr.flex1 = thresh;
2529                 log.u_bbr.flex2 = lro;
2530                 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts;
2531                 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
2532                 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2533                 log.u_bbr.flex6 = srtt;
2534                 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift;
2535                 log.u_bbr.flex8 = frm;
2536                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2537                     &bbr->rc_inp->inp_socket->so_rcv,
2538                     &bbr->rc_inp->inp_socket->so_snd,
2539                     BBR_LOG_THRESH_CALC, 0,
2540                     0, &log, false, &bbr->rc_tv);
2541         }
2542 }
2543
2544 static void
2545 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed)
2546 {
2547         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2548                 union tcp_log_stackspecific log;
2549
2550                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2551                 log.u_bbr.flex1 = line;
2552                 log.u_bbr.flex2 = bbr->bbr_timer_src;
2553                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2554                 log.u_bbr.flex4 = bbr->rc_in_persist;
2555                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2556                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2557                 log.u_bbr.flex8 = hpts_removed;
2558                 log.u_bbr.pkts_out = bbr->rc_pacer_started;
2559                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2560                     &bbr->rc_inp->inp_socket->so_rcv,
2561                     &bbr->rc_inp->inp_socket->so_snd,
2562                     BBR_LOG_TIMERCANC, 0,
2563                     0, &log, false, &bbr->rc_tv);
2564         }
2565 }
2566
2567 static void
2568 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta)
2569 {
2570         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2571                 union tcp_log_stackspecific log;
2572
2573                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2574                 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio;
2575                 log.u_bbr.flex2 = (peer_delta >> 32);
2576                 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff);
2577                 log.u_bbr.flex4 = (delta >> 32);
2578                 log.u_bbr.flex5 = (delta & 0x00000000ffffffff);
2579                 log.u_bbr.flex7 = bbr->rc_ts_clock_set;
2580                 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used;
2581                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2582                     &bbr->rc_inp->inp_socket->so_rcv,
2583                     &bbr->rc_inp->inp_socket->so_snd,
2584                     BBR_LOG_TSTMP_VAL, 0,
2585                     0, &log, false, &bbr->rc_tv);
2586         }
2587 }
2588
2589 static void
2590 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)
2591 {
2592         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2593                 union tcp_log_stackspecific log;
2594
2595                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2596                 log.u_bbr.flex1 = tsosz;
2597                 log.u_bbr.flex2 = tls;
2598                 log.u_bbr.flex3 = tcp_min_hptsi_time;
2599                 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min;
2600                 log.u_bbr.flex5 = old_val;
2601                 log.u_bbr.flex6 = maxseg;
2602                 log.u_bbr.flex7 = bbr->rc_no_pacing;
2603                 log.u_bbr.flex7 <<= 1;
2604                 log.u_bbr.flex7 |= bbr->rc_past_init_win;
2605                 if (hdwr)
2606                         log.u_bbr.flex8 = 0x80 | bbr->rc_use_google;
2607                 else
2608                         log.u_bbr.flex8 = bbr->rc_use_google;
2609                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2610                     &bbr->rc_inp->inp_socket->so_rcv,
2611                     &bbr->rc_inp->inp_socket->so_snd,
2612                     BBR_LOG_BBRTSO, 0,
2613                     0, &log, false, &bbr->rc_tv);
2614         }
2615 }
2616
2617 static void
2618 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm,
2619                       uint32_t flags, uint32_t line)
2620 {
2621         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2622                 union tcp_log_stackspecific log;
2623
2624                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2625                 log.u_bbr.flex1 = line;
2626                 log.u_bbr.flex2 = rsm->r_start;
2627                 log.u_bbr.flex3 = rsm->r_end;
2628                 log.u_bbr.flex4 = rsm->r_delivered;
2629                 log.u_bbr.flex5 = rsm->r_rtr_cnt;
2630                 log.u_bbr.flex6 = rsm->r_dupack;
2631                 log.u_bbr.flex7 = rsm->r_tim_lastsent[0];
2632                 log.u_bbr.flex8 = rsm->r_flags;
2633                 /* Hijack the pkts_out fids */
2634                 log.u_bbr.applimited = flags;
2635                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2636                     &bbr->rc_inp->inp_socket->so_rcv,
2637                     &bbr->rc_inp->inp_socket->so_snd,
2638                     BBR_RSM_CLEARED, 0,
2639                     0, &log, false, &bbr->rc_tv);
2640         }
2641 }
2642
2643 static void
2644 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts,
2645     uint32_t flex3, uint32_t flex2, uint32_t flex5,
2646     uint32_t flex6, uint32_t pkts_out, int flex7,
2647     uint32_t flex4, uint32_t flex1)
2648 {
2649
2650         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2651                 union tcp_log_stackspecific log;
2652
2653                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2654                 log.u_bbr.flex1 = flex1;
2655                 log.u_bbr.flex2 = flex2;
2656                 log.u_bbr.flex3 = flex3;
2657                 log.u_bbr.flex4 = flex4;
2658                 log.u_bbr.flex5 = flex5;
2659                 log.u_bbr.flex6 = flex6;
2660                 log.u_bbr.flex7 = flex7;
2661                 /* Hijack the pkts_out fids */
2662                 log.u_bbr.pkts_out = pkts_out;
2663                 log.u_bbr.flex8 = flex8;
2664                 if (bbr->rc_ack_was_delayed)
2665                         log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay;
2666                 else
2667                         log.u_bbr.epoch = 0;
2668                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2669                     &bbr->rc_inp->inp_socket->so_rcv,
2670                     &bbr->rc_inp->inp_socket->so_snd,
2671                     BBR_LOG_BBRUPD, 0,
2672                     flex2, &log, false, &bbr->rc_tv);
2673         }
2674 }
2675
2676 static void
2677 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason,
2678         uint32_t newbw, uint32_t obw, uint32_t diff,
2679         uint32_t tim)
2680 {
2681         if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2682                 union tcp_log_stackspecific log;
2683
2684                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2685                 log.u_bbr.flex1 = reason;
2686                 log.u_bbr.flex2 = newbw;
2687                 log.u_bbr.flex3 = obw;
2688                 log.u_bbr.flex4 = diff;
2689                 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost;
2690                 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del;
2691                 log.u_bbr.flex7 = bbr->rc_lt_is_sampling;
2692                 log.u_bbr.pkts_out = tim;
2693                 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw;
2694                 if (bbr->rc_lt_use_bw == 0)
2695                         log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
2696                 else
2697                         log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
2698                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2699                     &bbr->rc_inp->inp_socket->so_rcv,
2700                     &bbr->rc_inp->inp_socket->so_snd,
2701                     BBR_LOG_BWSAMP, 0,
2702                     0, &log, false, &bbr->rc_tv);
2703         }
2704 }
2705
2706 static inline void
2707 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line)
2708 {
2709         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2710                 union tcp_log_stackspecific log;
2711
2712                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2713                 log.u_bbr.flex1 = line;
2714                 log.u_bbr.flex2 = tick;
2715                 log.u_bbr.flex3 = tp->t_maxunacktime;
2716                 log.u_bbr.flex4 = tp->t_acktime;
2717                 log.u_bbr.flex8 = event;
2718                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2719                     &bbr->rc_inp->inp_socket->so_rcv,
2720                     &bbr->rc_inp->inp_socket->so_snd,
2721                     BBR_LOG_PROGRESS, 0,
2722                     0, &log, false, &bbr->rc_tv);
2723         }
2724 }
2725
2726 static void
2727 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp,
2728                          uint64_t rate, uint64_t hw_rate, int line, uint32_t cts,
2729                          int error)
2730 {
2731         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2732                 union tcp_log_stackspecific log;
2733
2734                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2735                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2736                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2737                 log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2738                 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2739                 log.u_bbr.bw_inuse = rate;
2740                 log.u_bbr.flex5 = line;
2741                 log.u_bbr.flex6 = error;
2742                 log.u_bbr.flex8 = bbr->skip_gain;
2743                 log.u_bbr.flex8 <<= 1;
2744                 log.u_bbr.flex8 |= bbr->gain_is_limited;
2745                 log.u_bbr.flex8 <<= 1;
2746                 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing;
2747                 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
2748                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2749                     &bbr->rc_inp->inp_socket->so_rcv,
2750                     &bbr->rc_inp->inp_socket->so_snd,
2751                     BBR_LOG_HDWR_PACE, 0,
2752                     0, &log, false, &bbr->rc_tv);
2753         }
2754 }
2755
2756 static void
2757 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)
2758 {
2759         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2760                 union tcp_log_stackspecific log;
2761
2762                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2763                 log.u_bbr.flex1 = slot;
2764                 log.u_bbr.flex2 = del_by;
2765                 log.u_bbr.flex3 = prev_delay;
2766                 log.u_bbr.flex4 = line;
2767                 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val;
2768                 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay;
2769                 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags);
2770                 log.u_bbr.flex8 = bbr->rc_in_persist;
2771                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2772                     &bbr->rc_inp->inp_socket->so_rcv,
2773                     &bbr->rc_inp->inp_socket->so_snd,
2774                     BBR_LOG_BBRSND, 0,
2775                     len, &log, false, &bbr->rc_tv);
2776         }
2777 }
2778
2779 static void
2780 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)
2781 {
2782         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2783                 union tcp_log_stackspecific log;
2784
2785                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2786                 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered;
2787                 log.u_bbr.flex2 = 0;
2788                 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt;
2789                 log.u_bbr.flex4 = end;
2790                 log.u_bbr.flex5 = seq;
2791                 log.u_bbr.flex6 = t;
2792                 log.u_bbr.flex7 = match;
2793                 log.u_bbr.flex8 = flags;
2794                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2795                     &bbr->rc_inp->inp_socket->so_rcv,
2796                     &bbr->rc_inp->inp_socket->so_snd,
2797                     BBR_LOG_BBRRTT, 0,
2798                     0, &log, false, &bbr->rc_tv);
2799         }
2800 }
2801
2802 static void
2803 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method)
2804 {
2805         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2806                 union tcp_log_stackspecific log;
2807
2808                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2809                 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2810                 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
2811                 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch;
2812                 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2813                 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs;
2814                 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight;
2815                 log.u_bbr.flex7 = 0;
2816                 log.u_bbr.flex8 = entry_method;
2817                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2818                     &bbr->rc_inp->inp_socket->so_rcv,
2819                     &bbr->rc_inp->inp_socket->so_snd,
2820                     BBR_LOG_EXIT_GAIN, 0,
2821                     0, &log, false, &bbr->rc_tv);
2822         }
2823 }
2824
2825 static void
2826 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired)
2827 {
2828         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2829                 union tcp_log_stackspecific log;
2830
2831                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2832                 /* R-HU */
2833                 log.u_bbr.flex1 = 0;
2834                 log.u_bbr.flex2 = 0;
2835                 log.u_bbr.flex3 = 0;
2836                 log.u_bbr.flex4 = 0;
2837                 log.u_bbr.flex7 = 0;
2838                 log.u_bbr.flex8 = settings_desired;
2839
2840                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2841                     &bbr->rc_inp->inp_socket->so_rcv,
2842                     &bbr->rc_inp->inp_socket->so_snd,
2843                     BBR_LOG_SETTINGS_CHG, 0,
2844                     0, &log, false, &bbr->rc_tv);
2845         }
2846 }
2847
2848 /*
2849  * Returns the bw from the our filter.
2850  */
2851 static inline uint64_t
2852 bbr_get_full_bw(struct tcp_bbr *bbr)
2853 {
2854         uint64_t bw;
2855
2856         bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2857
2858         return (bw);
2859 }
2860
2861 static inline void
2862 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2863 {
2864         uint64_t calclr;
2865         uint32_t lost, del;
2866
2867         if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch)
2868                 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch;
2869         else
2870                 lost = 0;
2871         del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del;
2872         if (lost == 0)  {
2873                 calclr = 0;
2874         } else if (del) {
2875                 calclr = lost;
2876                 calclr *= (uint64_t)1000;
2877                 calclr /= (uint64_t)del;
2878         } else {
2879                 /* Nothing delivered? 100.0% loss */
2880                 calclr = 1000;
2881         }
2882         bbr->r_ctl.rc_pkt_epoch_loss_rate =  (uint32_t)calclr;
2883         if (IN_RECOVERY(bbr->rc_tp->t_flags))
2884                 bbr->r_ctl.recovery_lr += (uint32_t)calclr;
2885         bbr->r_ctl.rc_pkt_epoch++;
2886         if (bbr->rc_no_pacing &&
2887             (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) {
2888                 bbr->rc_no_pacing = 0;
2889                 tcp_bbr_tso_size_check(bbr, cts);
2890         }
2891         bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time);
2892         bbr->r_ctl.rc_pkt_epoch_time = cts;
2893         /* What was our loss rate */
2894         bbr_log_pkt_epoch(bbr, cts, line, lost, del);
2895         bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered;
2896         bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost;
2897 }
2898
2899 static inline void
2900 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2901 {
2902         uint32_t epoch_time;
2903
2904         /* Tick the RTT clock */
2905         bbr->r_ctl.rc_rtt_epoch++;
2906         epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start;
2907         bbr_log_time_epoch(bbr, cts, line, epoch_time);
2908         bbr->r_ctl.rc_rcv_epoch_start = cts;
2909 }
2910
2911 static inline void
2912 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked)
2913 {
2914         if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) {
2915                 bbr->rc_is_pkt_epoch_now = 1;
2916         }
2917 }
2918
2919 /*
2920  * Returns the bw from either the b/w filter
2921  * or from the lt_bw (if the connection is being
2922  * policed).
2923  */
2924 static inline uint64_t
2925 __bbr_get_bw(struct tcp_bbr *bbr)
2926 {
2927         uint64_t bw, min_bw;
2928         uint64_t rtt;
2929         int gm_measure_cnt = 1;
2930
2931         /*
2932          * For startup we make, like google, a
2933          * minimum b/w. This is generated from the
2934          * IW and the rttProp. We do fall back to srtt
2935          * if for some reason (initial handshake) we don't
2936          * have a rttProp. We, in the worst case, fall back
2937          * to the configured min_bw (rc_initial_hptsi_bw).
2938          */
2939         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
2940                 /* Attempt first to use rttProp */
2941                 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2942                 if (rtt && (rtt < 0xffffffff)) {
2943 measure:
2944                         min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2945                                 ((uint64_t)1000000);
2946                         min_bw /= rtt;
2947                         if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2948                                 min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2949                         }
2950
2951                 } else if (bbr->rc_tp->t_srtt != 0) {
2952                         /* No rttProp, use srtt? */
2953                         rtt = bbr_get_rtt(bbr, BBR_SRTT);
2954                         goto measure;
2955                 } else {
2956                         min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2957                 }
2958         } else
2959                 min_bw = 0;
2960
2961         if ((bbr->rc_past_init_win == 0) &&
2962             (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp)))
2963                 bbr->rc_past_init_win = 1;
2964         if ((bbr->rc_use_google)  && (bbr->r_ctl.r_measurement_count >= 1))
2965                 gm_measure_cnt = 0;
2966         if (gm_measure_cnt &&
2967             ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) ||
2968              (bbr->rc_past_init_win == 0))) {
2969                 /* For google we use our guess rate until we get 1 measurement */
2970
2971 use_initial_window:
2972                 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2973                 if (rtt && (rtt < 0xffffffff)) {
2974                         /*
2975                          * We have an RTT measurment. Use that in
2976                          * combination with our initial window to calculate
2977                          * a b/w.
2978                          */
2979                         bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2980                                 ((uint64_t)1000000);
2981                         bw /= rtt;
2982                         if (bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2983                                 bw = bbr->r_ctl.rc_initial_hptsi_bw;
2984                         }
2985                 } else {
2986                         /* Drop back to the 40 and punt to a default */
2987                         bw = bbr->r_ctl.rc_initial_hptsi_bw;
2988                 }
2989                 if (bw < 1)
2990                         /* Probably should panic */
2991                         bw = 1;
2992                 if (bw > min_bw)
2993                         return (bw);
2994                 else
2995                         return (min_bw);
2996         }
2997         if (bbr->rc_lt_use_bw)
2998                 bw = bbr->r_ctl.rc_lt_bw;
2999         else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0))
3000                 bw = bbr->r_ctl.red_bw;
3001         else
3002                 bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3003         if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) {
3004                 /*
3005                  * Enforce user set rate limit, keep in mind that
3006                  * t_peakrate_thr is in B/s already
3007                  */
3008                 bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw);
3009         }
3010         if (bw == 0) {
3011                 /* We should not be at 0, go to the initial window then  */
3012                 goto use_initial_window;
3013         }
3014         if (bw < 1)
3015                 /* Probably should panic */
3016                 bw = 1;
3017         if (bw < min_bw)
3018                 bw = min_bw;
3019         return (bw);
3020 }
3021
3022 static inline uint64_t
3023 bbr_get_bw(struct tcp_bbr *bbr)
3024 {
3025         uint64_t bw;
3026
3027         bw = __bbr_get_bw(bbr);
3028         return (bw);
3029 }
3030
3031 static inline void
3032 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts)
3033 {
3034         bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch;
3035         bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time;
3036         bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3037         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3038 }
3039
3040 static inline void
3041 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts)
3042 {
3043         bbr->rc_lt_is_sampling = 0;
3044         bbr->rc_lt_use_bw = 0;
3045         bbr->r_ctl.rc_lt_bw = 0;
3046         bbr_reset_lt_bw_interval(bbr, cts);
3047 }
3048
3049 static inline void
3050 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin)
3051 {
3052         uint64_t diff;
3053
3054         /* Do we have a previous sample? */
3055         if (bbr->r_ctl.rc_lt_bw) {
3056                 /* Get the diff in bytes per second */
3057                 if (bbr->r_ctl.rc_lt_bw > bw)
3058                         diff = bbr->r_ctl.rc_lt_bw - bw;
3059                 else
3060                         diff = bw - bbr->r_ctl.rc_lt_bw;
3061                 if ((diff <= bbr_lt_bw_diff) ||
3062                     (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) {
3063                         /* Consider us policed */
3064                         uint32_t saved_bw;
3065
3066                         saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw;
3067                         bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2;   /* average of two */
3068                         bbr->rc_lt_use_bw = 1;
3069                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
3070                         /*
3071                          * Use pkt based epoch for measuring length of
3072                          * policer up
3073                          */
3074                         bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch;
3075                         /*
3076                          * reason 4 is we need to start consider being
3077                          * policed
3078                          */
3079                         bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin);
3080                         return;
3081                 }
3082         }
3083         bbr->r_ctl.rc_lt_bw = bw;
3084         bbr_reset_lt_bw_interval(bbr, cts);
3085         bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin);
3086 }
3087
3088 static void
3089 bbr_randomize_extra_state_time(struct tcp_bbr *bbr)
3090 {
3091         uint32_t ran, deduct;
3092
3093         ran = arc4random_uniform(bbr_rand_ot);
3094         if (ran) {
3095                 deduct = bbr->r_ctl.rc_level_state_extra / ran;
3096                 bbr->r_ctl.rc_level_state_extra -= deduct;
3097         }
3098 }
3099 /*
3100  * Return randomly the starting state
3101  * to use in probebw.
3102  */
3103 static uint8_t
3104 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts)
3105 {
3106         uint32_t ran;
3107         uint8_t ret_val;
3108
3109         /* Initialize the offset to 0 */
3110         bbr->r_ctl.rc_exta_time_gd = 0;
3111         bbr->rc_hit_state_1 = 0;
3112         bbr->r_ctl.rc_level_state_extra = 0;
3113         ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1));
3114         /*
3115          * The math works funny here :) the return value is used to set the
3116          * substate and then the state change is called which increments by
3117          * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when
3118          * we fully enter the state. Note that the (8 - 1 - ran) assures that
3119          * we return 1 - 7, so we dont return 0 and end up starting in
3120          * state 1 (DRAIN).
3121          */
3122         ret_val = BBR_SUBSTATE_COUNT - 1 - ran;
3123         /* Set an epoch */
3124         if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP))
3125                 bbr_set_epoch(bbr, cts, __LINE__);
3126
3127         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
3128         return (ret_val);
3129 }
3130
3131 static void
3132 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected)
3133 {
3134         uint32_t diff, d_time;
3135         uint64_t del_time, bw, lost, delivered;
3136
3137         if (bbr->r_use_policer == 0)
3138                 return;
3139         if (bbr->rc_lt_use_bw) {
3140                 /* We are using lt bw do we stop yet? */
3141                 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
3142                 if (diff > bbr_lt_bw_max_rtts) {
3143                         /* Reset it all */
3144 reset_all:
3145                         bbr_reset_lt_bw_sampling(bbr, cts);
3146                         if (bbr->rc_filled_pipe) {
3147                                 bbr_set_epoch(bbr, cts, __LINE__);
3148                                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
3149                                 bbr_substate_change(bbr, cts, __LINE__, 0);
3150                                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
3151                                 bbr_log_type_statechange(bbr, cts, __LINE__);
3152                         } else {
3153                                 /*
3154                                  * This should not happen really
3155                                  * unless we remove the startup/drain
3156                                  * restrictions above.
3157                                  */
3158                                 bbr->rc_bbr_state = BBR_STATE_STARTUP;
3159                                 bbr_set_epoch(bbr, cts, __LINE__);
3160                                 bbr->r_ctl.rc_bbr_state_time = cts;
3161                                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
3162                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
3163                                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
3164                                 bbr_set_state_target(bbr, __LINE__);
3165                                 bbr_log_type_statechange(bbr, cts, __LINE__);
3166                         }
3167                         /* reason 0 is to stop using lt-bw */
3168                         bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0);
3169                         return;
3170                 }
3171                 if (bbr_lt_intvl_fp == 0) {
3172                         /* Not doing false-postive detection */
3173                         return;
3174                 }
3175                 /* False positive detection */
3176                 if (diff == bbr_lt_intvl_fp) {
3177                         /* At bbr_lt_intvl_fp we record the lost */
3178                         bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3179                         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3180                 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) {
3181                         /* Now is our loss rate still high? */
3182                         lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3183                         delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3184                         if ((delivered == 0) ||
3185                             (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) {
3186                                 /* No still below our threshold */
3187                                 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0);
3188                         } else {
3189                                 /* Yikes its still high, it must be a false positive */
3190                                 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0);
3191                                 goto reset_all;
3192                         }
3193                 }
3194                 return;
3195         }
3196         /*
3197          * Wait for the first loss before sampling, to let the policer
3198          * exhaust its tokens and estimate the steady-state rate allowed by
3199          * the policer. Starting samples earlier includes bursts that
3200          * over-estimate the bw.
3201          */
3202         if (bbr->rc_lt_is_sampling == 0) {
3203                 /* reason 1 is to begin doing the sampling  */
3204                 if (loss_detected == 0)
3205                         return;
3206                 bbr_reset_lt_bw_interval(bbr, cts);
3207                 bbr->rc_lt_is_sampling = 1;
3208                 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0);
3209                 return;
3210         }
3211         /* Now how long were we delivering long term last> */
3212         if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time))
3213                 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time;
3214         else
3215                 d_time = 0;
3216
3217         /* To avoid underestimates, reset sampling if we run out of data. */
3218         if (bbr->r_ctl.r_app_limited_until) {
3219                 /* Can not measure in app-limited state */
3220                 bbr_reset_lt_bw_sampling(bbr, cts);
3221                 /* reason 2 is to reset sampling due to app limits  */
3222                 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time);
3223                 return;
3224         }
3225         diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
3226         if (diff < bbr_lt_intvl_min_rtts) {
3227                 /*
3228                  * need more samples (we don't
3229                  * start on a round like linux so
3230                  * we need 1 more).
3231                  */
3232                 /* 6 is not_enough time or no-loss */
3233                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3234                 return;
3235         }
3236         if (diff > (4 * bbr_lt_intvl_min_rtts)) {
3237                 /*
3238                  * For now if we wait too long, reset all sampling. We need
3239                  * to do some research here, its possible that we should
3240                  * base this on how much loss as occurred.. something like
3241                  * if its under 10% (or some thresh) reset all otherwise
3242                  * don't.  Thats for phase II I guess.
3243                  */
3244                 bbr_reset_lt_bw_sampling(bbr, cts);
3245                 /* reason 3 is to reset sampling due too long of sampling */
3246                 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3247                 return;
3248         }
3249         /*
3250          * End sampling interval when a packet is lost, so we estimate the
3251          * policer tokens were exhausted. Stopping the sampling before the
3252          * tokens are exhausted under-estimates the policed rate.
3253          */
3254         if (loss_detected == 0) {
3255                 /* 6 is not_enough time or no-loss */
3256                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3257                 return;
3258         }
3259         /* Calculate packets lost and delivered in sampling interval. */
3260         lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3261         delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3262         if ((delivered == 0) ||
3263             (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) {
3264                 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time);
3265                 return;
3266         }
3267         if (d_time < 1000) {
3268                 /* Not enough time. wait */
3269                 /* 6 is not_enough time or no-loss */
3270                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3271                 return;
3272         }
3273         if (d_time >= (0xffffffff / USECS_IN_MSEC)) {
3274                 /* Too long */
3275                 bbr_reset_lt_bw_sampling(bbr, cts);
3276                 /* reason 3 is to reset sampling due too long of sampling */
3277                 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3278                 return;
3279         }
3280         del_time = d_time;
3281         bw = delivered;
3282         bw *= (uint64_t)USECS_IN_SECOND;
3283         bw /= del_time;
3284         bbr_lt_bw_samp_done(bbr, bw, cts, d_time);
3285 }
3286
3287 /*
3288  * Allocate a sendmap from our zone.
3289  */
3290 static struct bbr_sendmap *
3291 bbr_alloc(struct tcp_bbr *bbr)
3292 {
3293         struct bbr_sendmap *rsm;
3294
3295         BBR_STAT_INC(bbr_to_alloc);
3296         rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO));
3297         if (rsm) {
3298                 bbr->r_ctl.rc_num_maps_alloced++;
3299                 return (rsm);
3300         }
3301         if (bbr->r_ctl.rc_free_cnt) {
3302                 BBR_STAT_INC(bbr_to_alloc_emerg);
3303                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
3304                 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
3305                 bbr->r_ctl.rc_free_cnt--;
3306                 return (rsm);
3307         }
3308         BBR_STAT_INC(bbr_to_alloc_failed);
3309         return (NULL);
3310 }
3311
3312 static struct bbr_sendmap *
3313 bbr_alloc_full_limit(struct tcp_bbr *bbr)
3314 {
3315         if ((V_tcp_map_entries_limit > 0) &&
3316             (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3317                 BBR_STAT_INC(bbr_alloc_limited);
3318                 if (!bbr->alloc_limit_reported) {
3319                         bbr->alloc_limit_reported = 1;
3320                         BBR_STAT_INC(bbr_alloc_limited_conns);
3321                 }
3322                 return (NULL);
3323         }
3324         return (bbr_alloc(bbr));
3325 }
3326
3327 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3328 static struct bbr_sendmap *
3329 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type)
3330 {
3331         struct bbr_sendmap *rsm;
3332
3333         if (limit_type) {
3334                 /* currently there is only one limit type */
3335                 if (V_tcp_map_split_limit > 0 &&
3336                     bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
3337                         BBR_STAT_INC(bbr_split_limited);
3338                         if (!bbr->alloc_limit_reported) {
3339                                 bbr->alloc_limit_reported = 1;
3340                                 BBR_STAT_INC(bbr_alloc_limited_conns);
3341                         }
3342                         return (NULL);
3343                 }
3344         }
3345
3346         /* allocate and mark in the limit type, if set */
3347         rsm = bbr_alloc(bbr);
3348         if (rsm != NULL && limit_type) {
3349                 rsm->r_limit_type = limit_type;
3350                 bbr->r_ctl.rc_num_split_allocs++;
3351         }
3352         return (rsm);
3353 }
3354
3355 static void
3356 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
3357 {
3358         if (rsm->r_limit_type) {
3359                 /* currently there is only one limit type */
3360                 bbr->r_ctl.rc_num_split_allocs--;
3361         }
3362         if (rsm->r_is_smallmap)
3363                 bbr->r_ctl.rc_num_small_maps_alloced--;
3364         if (bbr->r_ctl.rc_tlp_send == rsm)
3365                 bbr->r_ctl.rc_tlp_send = NULL;
3366         if (bbr->r_ctl.rc_resend == rsm) {
3367                 bbr->r_ctl.rc_resend = NULL;
3368         }
3369         if (bbr->r_ctl.rc_next == rsm)
3370                 bbr->r_ctl.rc_next = NULL;
3371         if (bbr->r_ctl.rc_sacklast == rsm)
3372                 bbr->r_ctl.rc_sacklast = NULL;
3373         if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
3374                 memset(rsm, 0, sizeof(struct bbr_sendmap));
3375                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
3376                 rsm->r_limit_type = 0;
3377                 bbr->r_ctl.rc_free_cnt++;
3378                 return;
3379         }
3380         bbr->r_ctl.rc_num_maps_alloced--;
3381         uma_zfree(bbr_zone, rsm);
3382 }
3383
3384 /*
3385  * Returns the BDP.
3386  */
3387 static uint64_t
3388 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) {
3389         /*
3390          * Calculate the bytes in flight needed given the bw (in bytes per
3391          * second) and the specifyed rtt in useconds. We need to put out the
3392          * returned value per RTT to match that rate. Gain will normally
3393          * raise it up from there.
3394          *
3395          * This should not overflow as long as the bandwidth is below 1
3396          * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller
3397          * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30).
3398          */
3399         uint64_t usec_per_sec;
3400
3401         usec_per_sec = USECS_IN_SECOND;
3402         return ((rtt * bw) / usec_per_sec);
3403 }
3404
3405 /*
3406  * Return the initial cwnd.
3407  */
3408 static uint32_t
3409 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp)
3410 {
3411         uint32_t i_cwnd;
3412
3413         if (bbr->rc_init_win) {
3414                 i_cwnd = bbr->rc_init_win * tp->t_maxseg;
3415         } else if (V_tcp_initcwnd_segments)
3416                 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg),
3417                     max(2 * tp->t_maxseg, 14600));
3418         else if (V_tcp_do_rfc3390)
3419                 i_cwnd = min(4 * tp->t_maxseg,
3420                     max(2 * tp->t_maxseg, 4380));
3421         else {
3422                 /* Per RFC5681 Section 3.1 */
3423                 if (tp->t_maxseg > 2190)
3424                         i_cwnd = 2 * tp->t_maxseg;
3425                 else if (tp->t_maxseg > 1095)
3426                         i_cwnd = 3 * tp->t_maxseg;
3427                 else
3428                         i_cwnd = 4 * tp->t_maxseg;
3429         }
3430         return (i_cwnd);
3431 }
3432
3433 /*
3434  * Given a specified gain, return the target
3435  * cwnd based on that gain.
3436  */
3437 static uint32_t
3438 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw)
3439 {
3440         uint64_t bdp, rtt;
3441         uint32_t cwnd;
3442
3443         if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) ||
3444             (bbr_get_full_bw(bbr) == 0)) {
3445                 /* No measurements yet */
3446                 return (bbr_initial_cwnd(bbr, bbr->rc_tp));
3447         }
3448         /*
3449          * Get bytes per RTT needed (rttProp is normally in
3450          * bbr_cwndtarget_rtt_touse)
3451          */
3452         rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse);
3453         /* Get the bdp from the two values */
3454         bdp = bbr_get_bw_delay_prod(rtt, bw);
3455         /* Now apply the gain */
3456         cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT));
3457
3458         return (cwnd);
3459 }
3460
3461 static uint32_t
3462 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain)
3463 {
3464         uint32_t cwnd, mss;
3465
3466         mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
3467         /* Get the base cwnd with gain rounded to a mss */
3468         cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss);
3469         /*
3470          * Add in N (2 default since we do not have a
3471          * fq layer to trap packets in) quanta's per the I-D
3472          * section 4.2.3.2 quanta adjust.
3473          */
3474         cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs);
3475         if (bbr->rc_use_google) {
3476                 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3477                    (bbr_state_val(bbr) == BBR_SUB_GAIN)) {
3478                         /*
3479                          * The linux implementation adds
3480                          * an extra 2 x mss in gain cycle which
3481                          * is documented no-where except in the code.
3482                          * so we add more for Neal undocumented feature
3483                          */
3484                         cwnd += 2 * mss;
3485                 }
3486                 if ((cwnd / mss) & 0x1) {
3487                         /* Round up for odd num mss */
3488                         cwnd += mss;
3489                 }
3490         }
3491         /* Are we below the min cwnd? */
3492         if (cwnd < get_min_cwnd(bbr))
3493                 return (get_min_cwnd(bbr));
3494         return (cwnd);
3495 }
3496
3497 static uint16_t
3498 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain)
3499 {
3500         if (gain < 1)
3501                 gain = 1;
3502         return (gain);
3503 }
3504
3505 static uint32_t
3506 bbr_get_header_oh(struct tcp_bbr *bbr)
3507 {
3508         int seg_oh;
3509
3510         seg_oh = 0;
3511         if (bbr->r_ctl.rc_inc_tcp_oh) {
3512                 /* Do we include TCP overhead? */
3513                 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr));
3514         }
3515         if (bbr->r_ctl.rc_inc_ip_oh) {
3516                 /* Do we include IP overhead? */
3517 #ifdef INET6
3518                 if (bbr->r_is_v6) {
3519                         seg_oh += sizeof(struct ip6_hdr);
3520                 } else
3521 #endif
3522                 {
3523
3524 #ifdef INET
3525                         seg_oh += sizeof(struct ip);
3526 #endif
3527                 }
3528         }
3529         if (bbr->r_ctl.rc_inc_enet_oh) {
3530                 /* Do we include the ethernet overhead?  */
3531                 seg_oh += sizeof(struct ether_header);
3532         }
3533         return(seg_oh);
3534 }
3535
3536 static uint32_t
3537 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw)
3538 {
3539         uint64_t divor, res, tim;
3540
3541         if (useconds_time == 0)
3542                 return (0);
3543         gain = bbr_gain_adjust(bbr, gain);
3544         divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT;
3545         tim = useconds_time;
3546         res = (tim * bw * gain) / divor;
3547         if (res == 0)
3548                 res = 1;
3549         return ((uint32_t)res);
3550 }
3551
3552 /*
3553  * Given a gain and a length return the delay in useconds that
3554  * should be used to evenly space out packets
3555  * on the connection (based on the gain factor).
3556  */
3557 static uint32_t
3558 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog)
3559 {
3560         uint64_t bw, lentim, res;
3561         uint32_t usecs, srtt, over = 0;
3562         uint32_t seg_oh, num_segs, maxseg;
3563
3564         if (len == 0)
3565                 return (0);
3566
3567         maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
3568         num_segs = (len + maxseg - 1) / maxseg;
3569         if (bbr->rc_use_google == 0) {
3570                 seg_oh = bbr_get_header_oh(bbr);
3571                 len += (num_segs * seg_oh);
3572         }
3573         gain = bbr_gain_adjust(bbr, gain);
3574         bw = bbr_get_bw(bbr);
3575         if (bbr->rc_use_google) {
3576                 uint64_t cbw;
3577
3578                 /*
3579                  * Reduce the b/w by the google discount
3580                  * factor 10 = 1%.
3581                  */
3582                 cbw = bw *  (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount);
3583                 cbw /= (uint64_t)1000;
3584                 /* We don't apply a discount if it results in 0 */
3585                 if (cbw > 0)
3586                         bw = cbw;
3587         }
3588         lentim = ((uint64_t)len *
3589                   (uint64_t)USECS_IN_SECOND *
3590                   (uint64_t)BBR_UNIT);
3591         res = lentim / ((uint64_t)gain * bw);
3592         if (res == 0)
3593                 res = 1;
3594         usecs = (uint32_t)res;
3595         srtt = bbr_get_rtt(bbr, BBR_SRTT);
3596         if (bbr_hptsi_max_mul && bbr_hptsi_max_div &&
3597             (bbr->rc_use_google == 0) &&
3598             (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) {
3599                 /*
3600                  * We cannot let the delay be more than 1/2 the srtt time.
3601                  * Otherwise we cannot pace out or send properly.
3602                  */
3603                 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div;
3604                 BBR_STAT_INC(bbr_hpts_min_time);
3605         }
3606         if (!nolog)
3607                 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1);
3608         return (usecs);
3609 }
3610
3611 static void
3612 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack,
3613                  uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses)
3614 {
3615         INP_WLOCK_ASSERT(tp->t_inpcb);
3616         uint64_t bw;
3617         uint32_t cwnd, target_cwnd, saved_bytes, maxseg;
3618         int32_t meth;
3619
3620 #ifdef STATS
3621         if ((tp->t_flags & TF_GPUTINPROG) &&
3622             SEQ_GEQ(th->th_ack, tp->gput_ack)) {
3623                 /*
3624                  * Strech acks and compressed acks will cause this to
3625                  * oscillate but we are doing it the same way as the main
3626                  * stack so it will be compariable (though possibly not
3627                  * ideal).
3628                  */
3629                 int32_t cgput;
3630                 int64_t gput, time_stamp;
3631
3632                 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8;
3633                 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000));
3634                 cgput = gput / time_stamp;
3635                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
3636                                          cgput);
3637                 if (tp->t_stats_gput_prev > 0)
3638                         stats_voi_update_abs_s32(tp->t_stats,
3639                                                  VOI_TCP_GPUT_ND,
3640                                                  ((gput - tp->t_stats_gput_prev) * 100) /
3641                                                  tp->t_stats_gput_prev);
3642                 tp->t_flags &= ~TF_GPUTINPROG;
3643                 tp->t_stats_gput_prev = cgput;
3644         }
3645 #endif
3646         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3647             ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
3648                 /* We don't change anything in probe-rtt */
3649                 return;
3650         }
3651         maxseg = tp->t_maxseg - bbr->rc_last_options;
3652         saved_bytes = bytes_this_ack;
3653         bytes_this_ack += sack_changed;
3654         if (bytes_this_ack > prev_acked) {
3655                 bytes_this_ack -= prev_acked;
3656                 /*
3657                  * A byte ack'd gives us a full mss
3658                  * to be like linux i.e. they count packets.
3659                  */
3660                 if ((bytes_this_ack < maxseg) && bbr->rc_use_google)
3661                         bytes_this_ack = maxseg;
3662         } else {
3663                 /* Unlikely */
3664                 bytes_this_ack = 0;
3665         }
3666         cwnd = tp->snd_cwnd;
3667         bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3668         if (bw)
3669                 target_cwnd = bbr_get_target_cwnd(bbr,
3670                                                   bw,
3671                                                   (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain);
3672         else
3673                 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp);
3674         if (IN_RECOVERY(tp->t_flags) &&
3675             (bbr->bbr_prev_in_rec == 0)) {
3676                 /*
3677                  * We are entering recovery and
3678                  * thus packet conservation.
3679                  */
3680                 bbr->pkt_conservation = 1;
3681                 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime;
3682                 cwnd = ctf_flight_size(tp,
3683                                        (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3684                         bytes_this_ack;
3685         }
3686         if (IN_RECOVERY(tp->t_flags)) {
3687                 uint32_t flight;
3688
3689                 bbr->bbr_prev_in_rec = 1;
3690                 if (cwnd > losses) {
3691                         cwnd -= losses;
3692                         if (cwnd < maxseg)
3693                                 cwnd = maxseg;
3694                 } else
3695                         cwnd = maxseg;
3696                 flight = ctf_flight_size(tp,
3697                                          (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3698                 bbr_log_type_cwndupd(bbr, flight, 0,
3699                                      losses, 10, 0, 0, line);
3700                 if (bbr->pkt_conservation) {
3701                         uint32_t time_in;
3702
3703                         if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start))
3704                                 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start;
3705                         else
3706                                 time_in = 0;
3707
3708                         if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
3709                                 /* Clear packet conservation after an rttProp */
3710                                 bbr->pkt_conservation = 0;
3711                         } else {
3712                                 if ((flight + bytes_this_ack) > cwnd)
3713                                         cwnd = flight + bytes_this_ack;
3714                                 if (cwnd < get_min_cwnd(bbr))
3715                                         cwnd = get_min_cwnd(bbr);
3716                                 tp->snd_cwnd = cwnd;
3717                                 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed,
3718                                                      prev_acked, 1, target_cwnd, th->th_ack, line);
3719                                 return;
3720                         }
3721                 }
3722         } else
3723                 bbr->bbr_prev_in_rec = 0;
3724         if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) {
3725                 bbr->r_ctl.restrict_growth--;
3726                 if (bytes_this_ack > maxseg)
3727                         bytes_this_ack = maxseg;
3728         }
3729         if (bbr->rc_filled_pipe) {
3730                 /*
3731                  * Here we have exited startup and filled the pipe. We will
3732                  * thus allow the cwnd to shrink to the target. We hit here
3733                  * mostly.
3734                  */
3735                 uint32_t s_cwnd;
3736
3737                 meth = 2;
3738                 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd);
3739                 if (s_cwnd > cwnd)
3740                         cwnd = s_cwnd;
3741                 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing)
3742                         cwnd = s_cwnd;
3743         } else {
3744                 /*
3745                  * Here we are still in startup, we increase cwnd by what
3746                  * has been acked.
3747                  */
3748                 if ((cwnd < target_cwnd) ||
3749                     (bbr->rc_past_init_win == 0)) {
3750                         meth = 3;
3751                         cwnd += bytes_this_ack;
3752                 } else {
3753                         /*
3754                          * Method 4 means we are at target so no gain in
3755                          * startup and past the initial window.
3756                          */
3757                         meth = 4;
3758                 }
3759         }
3760         tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr));
3761         bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line);
3762 }
3763
3764 static void
3765 tcp_bbr_partialack(struct tcpcb *tp)
3766 {
3767         struct tcp_bbr *bbr;
3768
3769         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3770         INP_WLOCK_ASSERT(tp->t_inpcb);
3771         if (ctf_flight_size(tp,
3772                 (bbr->r_ctl.rc_sacked  + bbr->r_ctl.rc_lost_bytes)) <=
3773             tp->snd_cwnd) {
3774                 bbr->r_wanted_output = 1;
3775         }
3776 }
3777
3778 static void
3779 bbr_post_recovery(struct tcpcb *tp)
3780 {
3781         struct tcp_bbr *bbr;
3782         uint32_t  flight;
3783
3784         INP_WLOCK_ASSERT(tp->t_inpcb);
3785         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3786         /*
3787          * Here we just exit recovery.
3788          */
3789         EXIT_RECOVERY(tp->t_flags);
3790         /* Lock in our b/w reduction for the specified number of pkt-epochs */
3791         bbr->r_recovery_bw = 0;
3792         tp->snd_recover = tp->snd_una;
3793         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3794         bbr->pkt_conservation = 0;
3795         if (bbr->rc_use_google == 0) {
3796                 /*
3797                  * For non-google mode lets
3798                  * go ahead and make sure we clear
3799                  * the recovery state so if we
3800                  * bounce back in to recovery we
3801                  * will do PC.
3802                  */
3803                 bbr->bbr_prev_in_rec = 0;
3804         }
3805         bbr_log_type_exit_rec(bbr);
3806         if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
3807                 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
3808                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__);
3809         } else {
3810                 /* For probe-rtt case lets fix up its saved_cwnd */
3811                 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) {
3812                         bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent;
3813                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__);
3814                 }
3815         }
3816         flight = ctf_flight_size(tp,
3817                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3818         if ((bbr->rc_use_google == 0) &&
3819             bbr_do_red) {
3820                 uint64_t val, lr2use;
3821                 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd;
3822                 uint32_t *cwnd_p;
3823
3824                 if (bbr_get_rtt(bbr, BBR_SRTT)) {
3825                         val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000);
3826                         val /= bbr_get_rtt(bbr, BBR_SRTT);
3827                         ratio = (uint32_t)val;
3828                 } else
3829                         ratio = 1000;
3830
3831                 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div,
3832                                      bbr->r_ctl.recovery_lr, 21,
3833                                      ratio,
3834                                      bbr->r_ctl.rc_red_cwnd_pe,
3835                                      __LINE__);
3836                 if ((ratio < bbr_do_red) || (bbr_do_red == 0))
3837                         goto done;
3838                 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3839                      bbr_prtt_slam_cwnd) ||
3840                     (bbr_sub_drain_slam_cwnd &&
3841                      (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3842                      bbr->rc_hit_state_1 &&
3843                      (bbr_state_val(bbr) == BBR_SUB_DRAIN)) ||
3844                     ((bbr->rc_bbr_state == BBR_STATE_DRAIN) &&
3845                      bbr_slam_cwnd_in_main_drain)) {
3846                         /*
3847                          * Here we must poke at the saved cwnd
3848                          * as well as the cwnd.
3849                          */
3850                         cwnd = bbr->r_ctl.rc_saved_cwnd;
3851                         cwnd_p = &bbr->r_ctl.rc_saved_cwnd;
3852                 } else {
3853                         cwnd = tp->snd_cwnd;
3854                         cwnd_p = &tp->snd_cwnd;
3855                 }
3856                 maxseg = tp->t_maxseg - bbr->rc_last_options;
3857                 /* Add the overall lr with the recovery lr */
3858                 if (bbr->r_ctl.rc_lost == 0)
3859                         lr2use = 0;
3860                 else if (bbr->r_ctl.rc_delivered == 0)
3861                         lr2use = 1000;
3862                 else {
3863                         lr2use = bbr->r_ctl.rc_lost * 1000;
3864                         lr2use /= bbr->r_ctl.rc_delivered;
3865                 }
3866                 lr2use += bbr->r_ctl.recovery_lr;
3867                 acks_inflight = (flight / (maxseg * 2));
3868                 if (bbr_red_scale) {
3869                         lr2use *= bbr_get_rtt(bbr, BBR_SRTT);
3870                         lr2use /= bbr_red_scale;
3871                         if ((bbr_red_growth_restrict) &&
3872                             ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1))
3873                             bbr->r_ctl.restrict_growth += acks_inflight;
3874                 }
3875                 if (lr2use) {
3876                         val = (uint64_t)cwnd * lr2use;
3877                         val /= 1000;
3878                         if (cwnd > val)
3879                                 newcwnd = roundup((cwnd - val), maxseg);
3880                         else
3881                                 newcwnd = maxseg;
3882                 } else {
3883                         val = (uint64_t)cwnd * (uint64_t)bbr_red_mul;
3884                         val /= (uint64_t)bbr_red_div;
3885                         newcwnd = roundup((uint32_t)val, maxseg);
3886                 }
3887                 /* with standard delayed acks how many acks can I expect? */
3888                 if (bbr_drop_limit == 0) {
3889                         /*
3890                          * Anticpate how much we will
3891                          * raise the cwnd based on the acks.
3892                          */
3893                         if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) {
3894                                 /* We do enforce the min (with the acks) */
3895                                 newcwnd = (get_min_cwnd(bbr) - acks_inflight);
3896                         }
3897                 } else {
3898                         /*
3899                          * A strict drop limit of N is is inplace
3900                          */
3901                         if (newcwnd < (bbr_drop_limit * maxseg)) {
3902                                 newcwnd = bbr_drop_limit * maxseg;
3903                         }
3904                 }
3905                 /* For the next N acks do we restrict the growth */
3906                 *cwnd_p = newcwnd;
3907                 if (tp->snd_cwnd > newcwnd)
3908                         tp->snd_cwnd = newcwnd;
3909                 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22,
3910                                      (uint32_t)lr2use,
3911                                      bbr_get_rtt(bbr, BBR_SRTT), __LINE__);
3912                 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch;
3913         }
3914 done:
3915         bbr->r_ctl.recovery_lr = 0;
3916         if (flight <= tp->snd_cwnd) {
3917                 bbr->r_wanted_output = 1;
3918         }
3919         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3920 }
3921
3922 static void
3923 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts)
3924 {
3925         bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3926         /* Limit the drop in b/w to 1/2 our current filter. */
3927         if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate)
3928                 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate;
3929         if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2))
3930                 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2;
3931         tcp_bbr_tso_size_check(bbr, cts);
3932 }
3933
3934 static void
3935 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm)
3936 {
3937         struct tcp_bbr *bbr;
3938
3939         INP_WLOCK_ASSERT(tp->t_inpcb);
3940 #ifdef STATS
3941         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
3942 #endif
3943         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3944         switch (type) {
3945         case CC_NDUPACK:
3946                 if (!IN_RECOVERY(tp->t_flags)) {
3947                         tp->snd_recover = tp->snd_max;
3948                         /* Start a new epoch */
3949                         bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
3950                         if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) {
3951                                 /*
3952                                  * Move forward the lt epoch
3953                                  * so it won't count the truncated
3954                                  * epoch.
3955                                  */
3956                                 bbr->r_ctl.rc_lt_epoch++;
3957                         }
3958                         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
3959                                 /*
3960                                  * Just like the policer detection code
3961                                  * if we are in startup we must push
3962                                  * forward the last startup epoch
3963                                  * to hide the truncated PE.
3964                                  */
3965                                 bbr->r_ctl.rc_bbr_last_startup_epoch++;
3966                         }
3967                         bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd;
3968                         ENTER_RECOVERY(tp->t_flags);
3969                         bbr->rc_tlp_rtx_out = 0;
3970                         bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate;
3971                         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3972                         if (bbr->rc_inp->inp_in_hpts &&
3973                             ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) {
3974                                 /*
3975                                  * When we enter recovery, we need to restart
3976                                  * any timers. This may mean we gain an agg
3977                                  * early, which will be made up for at the last
3978                                  * rxt out.
3979                                  */
3980                                 bbr->rc_timer_first = 1;
3981                                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
3982                         }
3983                         /*
3984                          * Calculate a new cwnd based on to the current
3985                          * delivery rate with no gain. We get the bdp
3986                          * without gaining it up like we normally would and
3987                          * we use the last cur_del_rate.
3988                          */
3989                         if ((bbr->rc_use_google == 0) &&
3990                             (bbr->r_ctl.bbr_rttprobe_gain_val ||
3991                              (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) {
3992                                 tp->snd_cwnd = ctf_flight_size(tp,
3993                                                    (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3994                                         (tp->t_maxseg - bbr->rc_last_options);
3995                                 if (tp->snd_cwnd < get_min_cwnd(bbr)) {
3996                                         /* We always gate to min cwnd */
3997                                         tp->snd_cwnd = get_min_cwnd(bbr);
3998                                 }
3999                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__);
4000                         }
4001                         bbr_log_type_enter_rec(bbr, rsm->r_start);
4002                 }
4003                 break;
4004         case CC_RTO_ERR:
4005                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4006                 /* RTO was unnecessary, so reset everything. */
4007                 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime);
4008                 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
4009                         tp->snd_cwnd = tp->snd_cwnd_prev;
4010                         tp->snd_ssthresh = tp->snd_ssthresh_prev;
4011                         tp->snd_recover = tp->snd_recover_prev;
4012                         tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
4013                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__);
4014                 }
4015                 tp->t_badrxtwin = 0;
4016                 break;
4017         }
4018 }
4019
4020 /*
4021  * Indicate whether this ack should be delayed.  We can delay the ack if
4022  * following conditions are met:
4023  *      - There is no delayed ack timer in progress.
4024  *      - Our last ack wasn't a 0-sized window. We never want to delay
4025  *        the ack that opens up a 0-sized window.
4026  *      - LRO wasn't used for this segment. We make sure by checking that the
4027  *        segment size is not larger than the MSS.
4028  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4029  *        connection.
4030  *      - The data being acked is less than a full segment (a stretch ack
4031  *        of more than a segment we should ack.
4032  *      - nsegs is 1 (if its more than that we received more than 1 ack).
4033  */
4034 #define DELAY_ACK(tp, bbr, nsegs)                               \
4035         (((tp->t_flags & TF_RXWIN0SENT) == 0) &&                \
4036          ((tp->t_flags & TF_DELACK) == 0) &&                    \
4037          ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) &&  \
4038          (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4039
4040 /*
4041  * Return the lowest RSM in the map of
4042  * packets still in flight that is not acked.
4043  * This should normally find on the first one
4044  * since we remove packets from the send
4045  * map after they are marked ACKED.
4046  */
4047 static struct bbr_sendmap *
4048 bbr_find_lowest_rsm(struct tcp_bbr *bbr)
4049 {
4050         struct bbr_sendmap *rsm;
4051
4052         /*
4053          * Walk the time-order transmitted list looking for an rsm that is
4054          * not acked. This will be the one that was sent the longest time
4055          * ago that is still outstanding.
4056          */
4057         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) {
4058                 if (rsm->r_flags & BBR_ACKED) {
4059                         continue;
4060                 }
4061                 goto finish;
4062         }
4063 finish:
4064         return (rsm);
4065 }
4066
4067 static struct bbr_sendmap *
4068 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
4069 {
4070         struct bbr_sendmap *prsm;
4071
4072         /*
4073          * Walk the sequence order list backward until we hit and arrive at
4074          * the highest seq not acked. In theory when this is called it
4075          * should be the last segment (which it was not).
4076          */
4077         prsm = rsm;
4078         TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4079                 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) {
4080                         continue;
4081                 }
4082                 return (prsm);
4083         }
4084         return (NULL);
4085 }
4086
4087 /*
4088  * Returns to the caller the number of microseconds that
4089  * the packet can be outstanding before we think we
4090  * should have had an ack returned.
4091  */
4092 static uint32_t
4093 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm)
4094 {
4095         /*
4096          * lro is the flag we use to determine if we have seen reordering.
4097          * If it gets set we have seen reordering. The reorder logic either
4098          * works in one of two ways:
4099          *
4100          * If reorder-fade is configured, then we track the last time we saw
4101          * re-ordering occur. If we reach the point where enough time as
4102          * passed we no longer consider reordering has occuring.
4103          *
4104          * Or if reorder-face is 0, then once we see reordering we consider
4105          * the connection to alway be subject to reordering and just set lro
4106          * to 1.
4107          *
4108          * In the end if lro is non-zero we add the extra time for
4109          * reordering in.
4110          */
4111         int32_t lro;
4112         uint32_t thresh, t_rxtcur;
4113
4114         if (srtt == 0)
4115                 srtt = 1;
4116         if (bbr->r_ctl.rc_reorder_ts) {
4117                 if (bbr->r_ctl.rc_reorder_fade) {
4118                         if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) {
4119                                 lro = cts - bbr->r_ctl.rc_reorder_ts;
4120                                 if (lro == 0) {
4121                                         /*
4122                                          * No time as passed since the last
4123                                          * reorder, mark it as reordering.
4124                                          */
4125                                         lro = 1;
4126                                 }
4127                         } else {
4128                                 /* Negative time? */
4129                                 lro = 0;
4130                         }
4131                         if (lro > bbr->r_ctl.rc_reorder_fade) {
4132                                 /* Turn off reordering seen too */
4133                                 bbr->r_ctl.rc_reorder_ts = 0;
4134                                 lro = 0;
4135                         }
4136                 } else {
4137                         /* Reodering does not fade */
4138                         lro = 1;
4139                 }
4140         } else {
4141                 lro = 0;
4142         }
4143         thresh = srtt + bbr->r_ctl.rc_pkt_delay;
4144         if (lro) {
4145                 /* It must be set, if not you get 1/4 rtt */
4146                 if (bbr->r_ctl.rc_reorder_shift)
4147                         thresh += (srtt >> bbr->r_ctl.rc_reorder_shift);
4148                 else
4149                         thresh += (srtt >> 2);
4150         } else {
4151                 thresh += 1000;
4152         }
4153         /* We don't let the rack timeout be above a RTO */
4154         if ((bbr->rc_tp)->t_srtt == 0)
4155                 t_rxtcur = BBR_INITIAL_RTO;
4156         else
4157                 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
4158         if (thresh > t_rxtcur) {
4159                 thresh = t_rxtcur;
4160         }
4161         /* And we don't want it above the RTO max either */
4162         if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4163                 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4164         }
4165         bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK);
4166         return (thresh);
4167 }
4168
4169 /*
4170  * Return to the caller the amount of time in mico-seconds
4171  * that should be used for the TLP timer from the last
4172  * send time of this packet.
4173  */
4174 static uint32_t
4175 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
4176     struct bbr_sendmap *rsm, uint32_t srtt,
4177     uint32_t cts)
4178 {
4179         uint32_t thresh, len, maxseg, t_rxtcur;
4180         struct bbr_sendmap *prsm;
4181
4182         if (srtt == 0)
4183                 srtt = 1;
4184         if (bbr->rc_tlp_threshold)
4185                 thresh = srtt + (srtt / bbr->rc_tlp_threshold);
4186         else
4187                 thresh = (srtt * 2);
4188         maxseg = tp->t_maxseg - bbr->rc_last_options;
4189         /* Get the previous sent packet, if any  */
4190         len = rsm->r_end - rsm->r_start;
4191
4192         /* 2.1 behavior */
4193         prsm = TAILQ_PREV(rsm, bbr_head, r_tnext);
4194         if (prsm && (len <= maxseg)) {
4195                 /*
4196                  * Two packets outstanding, thresh should be (2*srtt) +
4197                  * possible inter-packet delay (if any).
4198                  */
4199                 uint32_t inter_gap = 0;
4200                 int idx, nidx;
4201
4202                 idx = rsm->r_rtr_cnt - 1;
4203                 nidx = prsm->r_rtr_cnt - 1;
4204                 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
4205                         /* Yes it was sent later (or at the same time) */
4206                         inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
4207                 }
4208                 thresh += inter_gap;
4209         } else if (len <= maxseg) {
4210                 /*
4211                  * Possibly compensate for delayed-ack.
4212                  */
4213                 uint32_t alt_thresh;
4214
4215                 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time;
4216                 if (alt_thresh > thresh)
4217                         thresh = alt_thresh;
4218         }
4219         /* Not above the current  RTO */
4220         if (tp->t_srtt == 0)
4221                 t_rxtcur = BBR_INITIAL_RTO;
4222         else
4223                 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur);
4224
4225         bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP);
4226         /* Not above an RTO */
4227         if (thresh > t_rxtcur) {
4228                 thresh = t_rxtcur;
4229         }
4230         /* Not above a RTO max */
4231         if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4232                 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4233         }
4234         /* And now apply the user TLP min */
4235         if (thresh < bbr_tlp_min) {
4236                 thresh = bbr_tlp_min;
4237         }
4238         return (thresh);
4239 }
4240
4241 /*
4242  * Return one of three RTTs to use (in microseconds).
4243  */
4244 static __inline uint32_t
4245 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type)
4246 {
4247         uint32_t f_rtt;
4248         uint32_t srtt;
4249
4250         f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
4251         if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) {
4252                 /* We have no rtt at all */
4253                 if (bbr->rc_tp->t_srtt == 0)
4254                         f_rtt = BBR_INITIAL_RTO;
4255                 else
4256                         f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4257                 /*
4258                  * Since we don't know how good the rtt is apply a
4259                  * delayed-ack min
4260                  */
4261                 if (f_rtt < bbr_delayed_ack_time) {
4262                         f_rtt = bbr_delayed_ack_time;
4263                 }
4264         }
4265         /* Take the filter version or last measured pkt-rtt */
4266         if (rtt_type == BBR_RTT_PROP) {
4267                 srtt = f_rtt;
4268         } else if (rtt_type == BBR_RTT_PKTRTT) {
4269                 if (bbr->r_ctl.rc_pkt_epoch_rtt) {
4270                         srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
4271                 } else {
4272                         /* No pkt rtt yet */
4273                         srtt = f_rtt;
4274                 }
4275         } else if (rtt_type == BBR_RTT_RACK) {
4276                 srtt = bbr->r_ctl.rc_last_rtt;
4277                 /* We need to add in any internal delay for our timer */
4278                 if (bbr->rc_ack_was_delayed)
4279                         srtt += bbr->r_ctl.rc_ack_hdwr_delay;
4280         } else if (rtt_type == BBR_SRTT) {
4281                 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4282         } else {
4283                 /* TSNH */
4284                 srtt = f_rtt;
4285 #ifdef BBR_INVARIANTS
4286                 panic("Unknown rtt request type %d", rtt_type);
4287 #endif
4288         }
4289         return (srtt);
4290 }
4291
4292 static int
4293 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts)
4294 {
4295         uint32_t thresh;
4296
4297         thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK),
4298                                       cts, rsm);
4299         if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) {
4300                 /* It is lost (past time) */
4301                 return (1);
4302         }
4303         return (0);
4304 }
4305
4306 /*
4307  * Return a sendmap if we need to retransmit something.
4308  */
4309 static struct bbr_sendmap *
4310 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4311 {
4312         /*
4313          * Check to see that we don't need to fall into recovery. We will
4314          * need to do so if our oldest transmit is past the time we should
4315          * have had an ack.
4316          */
4317
4318         struct bbr_sendmap *rsm;
4319         int32_t idx;
4320
4321         if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) {
4322                 /* Nothing outstanding that we know of */
4323                 return (NULL);
4324         }
4325         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
4326         if (rsm == NULL) {
4327                 /* Nothing in the transmit map */
4328                 return (NULL);
4329         }
4330         if (tp->t_flags & TF_SENTFIN) {
4331                 /* Fin restricted, don't find anything once a fin is sent */
4332                 return (NULL);
4333         }
4334         if (rsm->r_flags & BBR_ACKED) {
4335                 /*
4336                  * Ok the first one is acked (this really should not happen
4337                  * since we remove the from the tmap once they are acked)
4338                  */
4339                 rsm = bbr_find_lowest_rsm(bbr);
4340                 if (rsm == NULL)
4341                         return (NULL);
4342         }
4343         idx = rsm->r_rtr_cnt - 1;
4344         if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) {
4345                 /* Send timestamp is the same or less? can't be ready */
4346                 return (NULL);
4347         }
4348         /* Get our RTT time */
4349         if (bbr_is_lost(bbr, rsm, cts) &&
4350             ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
4351              (rsm->r_flags & BBR_SACK_PASSED))) {
4352                 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4353                         rsm->r_flags |= BBR_MARKED_LOST;
4354                         bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4355                         bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4356                 }
4357                 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm);
4358 #ifdef BBR_INVARIANTS
4359                 if ((rsm->r_end - rsm->r_start) == 0)
4360                         panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm);
4361 #endif
4362                 return (rsm);
4363         }
4364         return (NULL);
4365 }
4366
4367 /*
4368  * RACK Timer, here we simply do logging and house keeping.
4369  * the normal bbr_output_wtime() function will call the
4370  * appropriate thing to check if we need to do a RACK retransmit.
4371  * We return 1, saying don't proceed with bbr_output_wtime only
4372  * when all timers have been stopped (destroyed PCB?).
4373  */
4374 static int
4375 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4376 {
4377         /*
4378          * This timer simply provides an internal trigger to send out data.
4379          * The check_recovery_mode call will see if there are needed
4380          * retransmissions, if so we will enter fast-recovery. The output
4381          * call may or may not do the same thing depending on sysctl
4382          * settings.
4383          */
4384         uint32_t lost;
4385
4386         if (bbr->rc_all_timers_stopped) {
4387                 return (1);
4388         }
4389         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4390                 /* Its not time yet */
4391                 return (0);
4392         }
4393         BBR_STAT_INC(bbr_to_tot);
4394         lost = bbr->r_ctl.rc_lost;
4395         if (bbr->r_state && (bbr->r_state != tp->t_state))
4396                 bbr_set_state(tp, bbr, 0);
4397         bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK);
4398         if (bbr->r_ctl.rc_resend == NULL) {
4399                 /* Lets do the check here */
4400                 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
4401         }
4402         if (bbr_policer_call_from_rack_to)
4403                 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4404         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
4405         return (0);
4406 }
4407
4408 static __inline void
4409 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start)
4410 {
4411         int idx;
4412
4413         nrsm->r_start = start;
4414         nrsm->r_end = rsm->r_end;
4415         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
4416         nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed;
4417         nrsm->r_flags = rsm->r_flags;
4418         /* We don't transfer forward the SYN flag */
4419         nrsm->r_flags &= ~BBR_HAS_SYN;
4420         /* We move forward the FIN flag, not that this should happen */
4421         rsm->r_flags &= ~BBR_HAS_FIN;
4422         nrsm->r_dupack = rsm->r_dupack;
4423         nrsm->r_rtr_bytes = 0;
4424         nrsm->r_is_gain = rsm->r_is_gain;
4425         nrsm->r_is_drain = rsm->r_is_drain;
4426         nrsm->r_delivered = rsm->r_delivered;
4427         nrsm->r_ts_valid = rsm->r_ts_valid;
4428         nrsm->r_del_ack_ts = rsm->r_del_ack_ts;
4429         nrsm->r_del_time = rsm->r_del_time;
4430         nrsm->r_app_limited = rsm->r_app_limited;
4431         nrsm->r_first_sent_time = rsm->r_first_sent_time;
4432         nrsm->r_flight_at_send = rsm->r_flight_at_send;
4433         /* We split a piece the lower section looses any just_ret flag. */
4434         nrsm->r_bbr_state = rsm->r_bbr_state;
4435         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
4436                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
4437         }
4438         rsm->r_end = nrsm->r_start;
4439         idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
4440         idx /= 8;
4441         /* Check if we got too small */
4442         if ((rsm->r_is_smallmap == 0) &&
4443             ((rsm->r_end - rsm->r_start) <= idx)) {
4444                 bbr->r_ctl.rc_num_small_maps_alloced++;
4445                 rsm->r_is_smallmap = 1;
4446         }
4447         /* Check the new one as well */
4448         if ((nrsm->r_end - nrsm->r_start) <= idx) {
4449                 bbr->r_ctl.rc_num_small_maps_alloced++;
4450                 nrsm->r_is_smallmap = 1;
4451         }
4452 }
4453
4454 static int
4455 bbr_sack_mergable(struct bbr_sendmap *at,
4456                   uint32_t start, uint32_t end)
4457 {
4458         /*
4459          * Given a sack block defined by
4460          * start and end, and a current postion
4461          * at. Return 1 if either side of at
4462          * would show that the block is mergable
4463          * to that side. A block to be mergable
4464          * must have overlap with the start/end
4465          * and be in the SACK'd state.
4466          */
4467         struct bbr_sendmap *l_rsm;
4468         struct bbr_sendmap *r_rsm;
4469
4470         /* first get the either side blocks */
4471         l_rsm = TAILQ_PREV(at, bbr_head, r_next);
4472         r_rsm = TAILQ_NEXT(at, r_next);
4473         if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) {
4474                 /* Potentially mergeable */
4475                 if ((l_rsm->r_end == start) ||
4476                     (SEQ_LT(start, l_rsm->r_end) &&
4477                      SEQ_GT(end, l_rsm->r_end))) {
4478                             /*
4479                              * map blk   |------|
4480                              * sack blk         |------|
4481                              * <or>
4482                              * map blk   |------|
4483                              * sack blk      |------|
4484                              */
4485                             return (1);
4486                     }
4487         }
4488         if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) {
4489                 /* Potentially mergeable */
4490                 if ((r_rsm->r_start == end) ||
4491                     (SEQ_LT(start, r_rsm->r_start) &&
4492                      SEQ_GT(end, r_rsm->r_start))) {
4493                         /*
4494                          * map blk          |---------|
4495                          * sack blk    |----|
4496                          * <or>
4497                          * map blk          |---------|
4498                          * sack blk    |-------|
4499                          */
4500                         return (1);
4501                 }
4502         }
4503         return (0);
4504 }
4505
4506 static struct bbr_sendmap *
4507 bbr_merge_rsm(struct tcp_bbr *bbr,
4508               struct bbr_sendmap *l_rsm,
4509               struct bbr_sendmap *r_rsm)
4510 {
4511         /*
4512          * We are merging two ack'd RSM's,
4513          * the l_rsm is on the left (lower seq
4514          * values) and the r_rsm is on the right
4515          * (higher seq value). The simplest way
4516          * to merge these is to move the right
4517          * one into the left. I don't think there
4518          * is any reason we need to try to find
4519          * the oldest (or last oldest retransmitted).
4520          */
4521         l_rsm->r_end = r_rsm->r_end;
4522         if (l_rsm->r_dupack < r_rsm->r_dupack)
4523                 l_rsm->r_dupack = r_rsm->r_dupack;
4524         if (r_rsm->r_rtr_bytes)
4525                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
4526         if (r_rsm->r_in_tmap) {
4527                 /* This really should not happen */
4528                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext);
4529         }
4530         if (r_rsm->r_app_limited)
4531                 l_rsm->r_app_limited = r_rsm->r_app_limited;
4532         /* Now the flags */
4533         if (r_rsm->r_flags & BBR_HAS_FIN)
4534                 l_rsm->r_flags |= BBR_HAS_FIN;
4535         if (r_rsm->r_flags & BBR_TLP)
4536                 l_rsm->r_flags |= BBR_TLP;
4537         if (r_rsm->r_flags & BBR_RWND_COLLAPSED)
4538                 l_rsm->r_flags |= BBR_RWND_COLLAPSED;
4539         if (r_rsm->r_flags & BBR_MARKED_LOST) {
4540                 /* This really should not happen */
4541                 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start;
4542         }
4543         TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next);
4544         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
4545                 /* Transfer the split limit to the map we free */
4546                 r_rsm->r_limit_type = l_rsm->r_limit_type;
4547                 l_rsm->r_limit_type = 0;
4548         }
4549         bbr_free(bbr, r_rsm);
4550         return(l_rsm);
4551 }
4552
4553 /*
4554  * TLP Timer, here we simply setup what segment we want to
4555  * have the TLP expire on, the normal bbr_output_wtime() will then
4556  * send it out.
4557  *
4558  * We return 1, saying don't proceed with bbr_output_wtime only
4559  * when all timers have been stopped (destroyed PCB?).
4560  */
4561 static int
4562 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4563 {
4564         /*
4565          * Tail Loss Probe.
4566          */
4567         struct bbr_sendmap *rsm = NULL;
4568         struct socket *so;
4569         uint32_t amm;
4570         uint32_t out, avail;
4571         uint32_t maxseg;
4572         int collapsed_win = 0;
4573
4574         if (bbr->rc_all_timers_stopped) {
4575                 return (1);
4576         }
4577         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4578                 /* Its not time yet */
4579                 return (0);
4580         }
4581         if (ctf_progress_timeout_check(tp, true)) {
4582                 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4583                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4584                 return (1);
4585         }
4586         /* Did we somehow get into persists? */
4587         if (bbr->rc_in_persist) {
4588                 return (0);
4589         }
4590         if (bbr->r_state && (bbr->r_state != tp->t_state))
4591                 bbr_set_state(tp, bbr, 0);
4592         BBR_STAT_INC(bbr_tlp_tot);
4593         maxseg = tp->t_maxseg - bbr->rc_last_options;
4594         /*
4595          * A TLP timer has expired. We have been idle for 2 rtts. So we now
4596          * need to figure out how to force a full MSS segment out.
4597          */
4598         so = tp->t_inpcb->inp_socket;
4599         avail = sbavail(&so->so_snd);
4600         out = ctf_outstanding(tp);
4601         if (out > tp->snd_wnd) {
4602                 /* special case, we need a retransmission */
4603                 collapsed_win = 1;
4604                 goto need_retran;
4605         }
4606         if (avail > out) {
4607                 /* New data is available */
4608                 amm = avail - out;
4609                 if (amm > maxseg) {
4610                         amm = maxseg;
4611                 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) {
4612                         /* not enough to fill a MTU and no-delay is off */
4613                         goto need_retran;
4614                 }
4615                 /* Set the send-new override */
4616                 if ((out + amm) <= tp->snd_wnd) {
4617                         bbr->rc_tlp_new_data = 1;
4618                 } else {
4619                         goto need_retran;
4620                 }
4621                 bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4622                 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max;
4623                 bbr->r_ctl.rc_tlp_send = NULL;
4624                 /* cap any slots */
4625                 BBR_STAT_INC(bbr_tlp_newdata);
4626                 goto send;
4627         }
4628 need_retran:
4629         /*
4630          * Ok we need to arrange the last un-acked segment to be re-sent, or
4631          * optionally the first un-acked segment.
4632          */
4633         if (collapsed_win == 0) {
4634                 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
4635                 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) {
4636                         rsm = bbr_find_high_nonack(bbr, rsm);
4637                 }
4638                 if (rsm == NULL) {
4639                         goto restore;
4640                 }
4641         } else {
4642                 /*
4643                  * We must find the last segment
4644                  * that was acceptable by the client.
4645                  */
4646                 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4647                         if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) {
4648                                 /* Found one */
4649                                 break;
4650                         }
4651                 }
4652                 if (rsm == NULL) {
4653                         /* None? if so send the first */
4654                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4655                         if (rsm == NULL)
4656                                 goto restore;
4657                 }
4658         }
4659         if ((rsm->r_end - rsm->r_start) > maxseg) {
4660                 /*
4661                  * We need to split this the last segment in two.
4662                  */
4663                 struct bbr_sendmap *nrsm;
4664
4665                 nrsm = bbr_alloc_full_limit(bbr);
4666                 if (nrsm == NULL) {
4667                         /*
4668                          * We can't get memory to split, we can either just
4669                          * not split it. Or retransmit the whole piece, lets
4670                          * do the large send (BTLP :-) ).
4671                          */
4672                         goto go_for_it;
4673                 }
4674                 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg));
4675                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
4676                 if (rsm->r_in_tmap) {
4677                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
4678                         nrsm->r_in_tmap = 1;
4679                 }
4680                 rsm->r_flags &= (~BBR_HAS_FIN);
4681                 rsm = nrsm;
4682         }
4683 go_for_it:
4684         bbr->r_ctl.rc_tlp_send = rsm;
4685         bbr->rc_tlp_rtx_out = 1;
4686         if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) {
4687                 bbr->r_ctl.rc_tlp_seg_send_cnt++;
4688                 tp->t_rxtshift++;
4689         } else {
4690                 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
4691                 bbr->r_ctl.rc_tlp_seg_send_cnt = 1;
4692         }
4693 send:
4694         if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
4695                 /*
4696                  * Can't [re]/transmit a segment we have retranmitted the
4697                  * max times. We need the retransmit timer to take over.
4698                  */
4699 restore:
4700                 bbr->rc_tlp_new_data = 0;
4701                 bbr->r_ctl.rc_tlp_send = NULL;
4702                 if (rsm)
4703                         rsm->r_flags &= ~BBR_TLP;
4704                 BBR_STAT_INC(bbr_tlp_retran_fail);
4705                 return (0);
4706         } else if (rsm) {
4707                 rsm->r_flags |= BBR_TLP;
4708         }
4709         if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) &&
4710             (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) {
4711                 /*
4712                  * We have retransmitted to many times for TLP. Switch to
4713                  * the regular RTO timer
4714                  */
4715                 goto restore;
4716         }
4717         bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP);
4718         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
4719         return (0);
4720 }
4721
4722 /*
4723  * Delayed ack Timer, here we simply need to setup the
4724  * ACK_NOW flag and remove the DELACK flag. From there
4725  * the output routine will send the ack out.
4726  *
4727  * We only return 1, saying don't proceed, if all timers
4728  * are stopped (destroyed PCB?).
4729  */
4730 static int
4731 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4732 {
4733         if (bbr->rc_all_timers_stopped) {
4734                 return (1);
4735         }
4736         bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK);
4737         tp->t_flags &= ~TF_DELACK;
4738         tp->t_flags |= TF_ACKNOW;
4739         KMOD_TCPSTAT_INC(tcps_delack);
4740         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
4741         return (0);
4742 }
4743
4744 /*
4745  * Here we send a KEEP-ALIVE like probe to the
4746  * peer, we do not send data.
4747  *
4748  * We only return 1, saying don't proceed, if all timers
4749  * are stopped (destroyed PCB?).
4750  */
4751 static int
4752 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4753 {
4754         struct tcptemp *t_template;
4755         int32_t retval = 1;
4756
4757         if (bbr->rc_all_timers_stopped) {
4758                 return (1);
4759         }
4760         if (bbr->rc_in_persist == 0)
4761                 return (0);
4762         KASSERT(tp->t_inpcb != NULL,
4763             ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
4764         /*
4765          * Persistence timer into zero window. Force a byte to be output, if
4766          * possible.
4767          */
4768         bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST);
4769         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
4770         KMOD_TCPSTAT_INC(tcps_persisttimeo);
4771         /*
4772          * Have we exceeded the user specified progress time?
4773          */
4774         if (ctf_progress_timeout_check(tp, true)) {
4775                 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4776                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4777                 goto out;
4778         }
4779         /*
4780          * Hack: if the peer is dead/unreachable, we do not time out if the
4781          * window is closed.  After a full backoff, drop the connection if
4782          * the idle time (no responses to probes) reaches the maximum
4783          * backoff that we would use if retransmitting.
4784          */
4785         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
4786             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
4787             ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
4788                 KMOD_TCPSTAT_INC(tcps_persistdrop);
4789                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4790                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4791                 goto out;
4792         }
4793         if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) &&
4794             tp->snd_una == tp->snd_max) {
4795                 bbr_exit_persist(tp, bbr, cts, __LINE__);
4796                 retval = 0;
4797                 goto out;
4798         }
4799         /*
4800          * If the user has closed the socket then drop a persisting
4801          * connection after a much reduced timeout.
4802          */
4803         if (tp->t_state > TCPS_CLOSE_WAIT &&
4804             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
4805                 KMOD_TCPSTAT_INC(tcps_persistdrop);
4806                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4807                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4808                 goto out;
4809         }
4810         t_template = tcpip_maketemplate(bbr->rc_inp);
4811         if (t_template) {
4812                 tcp_respond(tp, t_template->tt_ipgen,
4813                             &t_template->tt_t, (struct mbuf *)NULL,
4814                             tp->rcv_nxt, tp->snd_una - 1, 0);
4815                 /* This sends an ack */
4816                 if (tp->t_flags & TF_DELACK)
4817                         tp->t_flags &= ~TF_DELACK;
4818                 free(t_template, M_TEMP);
4819         }
4820         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
4821                 tp->t_rxtshift++;
4822         bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0);
4823 out:
4824         return (retval);
4825 }
4826
4827 /*
4828  * If a keepalive goes off, we had no other timers
4829  * happening. We always return 1 here since this
4830  * routine either drops the connection or sends
4831  * out a segment with respond.
4832  */
4833 static int
4834 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4835 {
4836         struct tcptemp *t_template;
4837         struct inpcb *inp;
4838
4839         if (bbr->rc_all_timers_stopped) {
4840                 return (1);
4841         }
4842         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
4843         inp = tp->t_inpcb;
4844         bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP);
4845         /*
4846          * Keep-alive timer went off; send something or drop connection if
4847          * idle for too long.
4848          */
4849         KMOD_TCPSTAT_INC(tcps_keeptimeo);
4850         if (tp->t_state < TCPS_ESTABLISHED)
4851                 goto dropit;
4852         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
4853             tp->t_state <= TCPS_CLOSING) {
4854                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
4855                         goto dropit;
4856                 /*
4857                  * Send a packet designed to force a response if the peer is
4858                  * up and reachable: either an ACK if the connection is
4859                  * still alive, or an RST if the peer has closed the
4860                  * connection due to timeout or reboot. Using sequence
4861                  * number tp->snd_una-1 causes the transmitted zero-length
4862                  * segment to lie outside the receive window; by the
4863                  * protocol spec, this requires the correspondent TCP to
4864                  * respond.
4865                  */
4866                 KMOD_TCPSTAT_INC(tcps_keepprobe);
4867                 t_template = tcpip_maketemplate(inp);
4868                 if (t_template) {
4869                         tcp_respond(tp, t_template->tt_ipgen,
4870                             &t_template->tt_t, (struct mbuf *)NULL,
4871                             tp->rcv_nxt, tp->snd_una - 1, 0);
4872                         free(t_template, M_TEMP);
4873                 }
4874         }
4875         bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0);
4876         return (1);
4877 dropit:
4878         KMOD_TCPSTAT_INC(tcps_keepdrops);
4879         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
4880         tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4881         return (1);
4882 }
4883
4884 /*
4885  * Retransmit helper function, clear up all the ack
4886  * flags and take care of important book keeping.
4887  */
4888 static void
4889 bbr_remxt_tmr(struct tcpcb *tp)
4890 {
4891         /*
4892          * The retransmit timer went off, all sack'd blocks must be
4893          * un-acked.
4894          */
4895         struct bbr_sendmap *rsm, *trsm = NULL;
4896         struct tcp_bbr *bbr;
4897         uint32_t cts, lost;
4898
4899         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
4900         cts = tcp_get_usecs(&bbr->rc_tv);
4901         lost = bbr->r_ctl.rc_lost;
4902         if (bbr->r_state && (bbr->r_state != tp->t_state))
4903                 bbr_set_state(tp, bbr, 0);
4904
4905         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
4906                 if (rsm->r_flags & BBR_ACKED) {
4907                         uint32_t old_flags;
4908
4909                         rsm->r_dupack = 0;
4910                         if (rsm->r_in_tmap == 0) {
4911                                 /* We must re-add it back to the tlist */
4912                                 if (trsm == NULL) {
4913                                         TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
4914                                 } else {
4915                                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext);
4916                                 }
4917                                 rsm->r_in_tmap = 1;
4918                         }
4919                         old_flags = rsm->r_flags;
4920                         rsm->r_flags |= BBR_RXT_CLEARED;
4921                         rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS);
4922                         bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
4923                 } else {
4924                         if ((tp->t_state < TCPS_ESTABLISHED) &&
4925                             (rsm->r_start == tp->snd_una)) {
4926                                 /*
4927                                  * Special case for TCP FO. Where
4928                                  * we sent more data beyond the snd_max.
4929                                  * We don't mark that as lost and stop here.
4930                                  */
4931                                 break;
4932                         }
4933                         if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4934                                 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4935                                 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4936                         }
4937                         if (bbr_marks_rxt_sack_passed) {
4938                                 /*
4939                                  * With this option, we will rack out
4940                                  * in 1ms increments the rest of the packets.
4941                                  */
4942                                 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST;
4943                                 rsm->r_flags &= ~BBR_WAS_SACKPASS;
4944                         } else {
4945                                 /*
4946                                  * With this option we only mark them lost
4947                                  * and remove all sack'd markings. We will run
4948                                  * another RXT or a TLP. This will cause
4949                                  * us to eventually send more based on what
4950                                  * ack's come in.
4951                                  */
4952                                 rsm->r_flags |= BBR_MARKED_LOST;
4953                                 rsm->r_flags &= ~BBR_WAS_SACKPASS;
4954                                 rsm->r_flags &= ~BBR_SACK_PASSED;
4955                         }
4956                 }
4957                 trsm = rsm;
4958         }
4959         bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4960         /* Clear the count (we just un-acked them) */
4961         bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR);
4962         bbr->rc_tlp_new_data = 0;
4963         bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4964         /* zap the behindness on a rxt */
4965         bbr->r_ctl.rc_hptsi_agg_delay = 0;
4966         bbr->r_agg_early_set = 0;
4967         bbr->r_ctl.rc_agg_early = 0;
4968         bbr->rc_tlp_rtx_out = 0;
4969         bbr->r_ctl.rc_sacked = 0;
4970         bbr->r_ctl.rc_sacklast = NULL;
4971         bbr->r_timer_override = 1;
4972         bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4973 }
4974
4975 /*
4976  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
4977  * we will setup to retransmit the lowest seq number outstanding.
4978  */
4979 static int
4980 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4981 {
4982         int32_t rexmt;
4983         int32_t retval = 0;
4984         bool isipv6;
4985
4986         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
4987         if (bbr->rc_all_timers_stopped) {
4988                 return (1);
4989         }
4990         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
4991             (tp->snd_una == tp->snd_max)) {
4992                 /* Nothing outstanding .. nothing to do */
4993                 return (0);
4994         }
4995         /*
4996          * Retransmission timer went off.  Message has not been acked within
4997          * retransmit interval.  Back off to a longer retransmit interval
4998          * and retransmit one segment.
4999          */
5000         if (ctf_progress_timeout_check(tp, true)) {
5001                 retval = 1;
5002                 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
5003                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
5004                 goto out;
5005         }
5006         bbr_remxt_tmr(tp);
5007         if ((bbr->r_ctl.rc_resend == NULL) ||
5008             ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) {
5009                 /*
5010                  * If the rwnd collapsed on
5011                  * the one we are retransmitting
5012                  * it does not count against the
5013                  * rxt count.
5014                  */
5015                 tp->t_rxtshift++;
5016         }
5017         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
5018                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
5019                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
5020                 retval = 1;
5021                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
5022                 tcp_set_inp_to_drop(bbr->rc_inp,
5023                     (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
5024                 goto out;
5025         }
5026         if (tp->t_state == TCPS_SYN_SENT) {
5027                 /*
5028                  * If the SYN was retransmitted, indicate CWND to be limited
5029                  * to 1 segment in cc_conn_init().
5030                  */
5031                 tp->snd_cwnd = 1;
5032         } else if (tp->t_rxtshift == 1) {
5033                 /*
5034                  * first retransmit; record ssthresh and cwnd so they can be
5035                  * recovered if this turns out to be a "bad" retransmit. A
5036                  * retransmit is considered "bad" if an ACK for this segment
5037                  * is received within RTT/2 interval; the assumption here is
5038                  * that the ACK was already in flight.  See "On Estimating
5039                  * End-to-End Network Path Properties" by Allman and Paxson
5040                  * for more details.
5041                  */
5042                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5043                 if (!IN_RECOVERY(tp->t_flags)) {
5044                         tp->snd_cwnd_prev = tp->snd_cwnd;
5045                         tp->snd_ssthresh_prev = tp->snd_ssthresh;
5046                         tp->snd_recover_prev = tp->snd_recover;
5047                         tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
5048                         tp->t_flags |= TF_PREVVALID;
5049                 } else {
5050                         tp->t_flags &= ~TF_PREVVALID;
5051                 }
5052                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5053         } else {
5054                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5055                 tp->t_flags &= ~TF_PREVVALID;
5056         }
5057         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
5058         if ((tp->t_state == TCPS_SYN_SENT) ||
5059             (tp->t_state == TCPS_SYN_RECEIVED))
5060                 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift];
5061         else
5062                 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
5063         TCPT_RANGESET(tp->t_rxtcur, rexmt,
5064             MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms),
5065             MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
5066         /*
5067          * We enter the path for PLMTUD if connection is established or, if
5068          * connection is FIN_WAIT_1 status, reason for the last is that if
5069          * amount of data we send is very small, we could send it in couple
5070          * of packets and process straight to FIN. In that case we won't
5071          * catch ESTABLISHED state.
5072          */
5073 #ifdef INET6
5074         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
5075 #else
5076         isipv6 = false;
5077 #endif
5078         if (((V_tcp_pmtud_blackhole_detect == 1) ||
5079             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
5080             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
5081             ((tp->t_state == TCPS_ESTABLISHED) ||
5082             (tp->t_state == TCPS_FIN_WAIT_1))) {
5083                 /*
5084                  * Idea here is that at each stage of mtu probe (usually,
5085                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
5086                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
5087                  * should take care of that.
5088                  */
5089                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
5090                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
5091                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
5092                     tp->t_rxtshift % 2 == 0)) {
5093                         /*
5094                          * Enter Path MTU Black-hole Detection mechanism: -
5095                          * Disable Path MTU Discovery (IP "DF" bit). -
5096                          * Reduce MTU to lower value than what we negotiated
5097                          * with peer.
5098                          */
5099                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
5100                                 /*
5101                                  * Record that we may have found a black
5102                                  * hole.
5103                                  */
5104                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
5105                                 /* Keep track of previous MSS. */
5106                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
5107                         }
5108                         /*
5109                          * Reduce the MSS to blackhole value or to the
5110                          * default in an attempt to retransmit.
5111                          */
5112 #ifdef INET6
5113                         isipv6 = bbr->r_is_v6;
5114                         if (isipv6 &&
5115                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
5116                                 /* Use the sysctl tuneable blackhole MSS. */
5117                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
5118                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5119                         } else if (isipv6) {
5120                                 /* Use the default MSS. */
5121                                 tp->t_maxseg = V_tcp_v6mssdflt;
5122                                 /*
5123                                  * Disable Path MTU Discovery when we switch
5124                                  * to minmss.
5125                                  */
5126                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5127                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5128                         }
5129 #endif
5130 #if defined(INET6) && defined(INET)
5131                         else
5132 #endif
5133 #ifdef INET
5134                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
5135                                 /* Use the sysctl tuneable blackhole MSS. */
5136                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
5137                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5138                         } else {
5139                                 /* Use the default MSS. */
5140                                 tp->t_maxseg = V_tcp_mssdflt;
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                 } else {
5150                         /*
5151                          * If further retransmissions are still unsuccessful
5152                          * with a lowered MTU, maybe this isn't a blackhole
5153                          * and we restore the previous MSS and blackhole
5154                          * detection flags. The limit '6' is determined by
5155                          * giving each probe stage (1448, 1188, 524) 2
5156                          * chances to recover.
5157                          */
5158                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
5159                             (tp->t_rxtshift >= 6)) {
5160                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
5161                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
5162                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
5163                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
5164                         }
5165                 }
5166         }
5167         /*
5168          * Disable RFC1323 and SACK if we haven't got any response to our
5169          * third SYN to work-around some broken terminal servers (most of
5170          * which have hopefully been retired) that have bad VJ header
5171          * compression code which trashes TCP segments containing
5172          * unknown-to-them TCP options.
5173          */
5174         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
5175             (tp->t_rxtshift == 3))
5176                 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT);
5177         /*
5178          * If we backed off this far, our srtt estimate is probably bogus.
5179          * Clobber it so we'll take the next rtt measurement as our srtt;
5180          * move the current srtt into rttvar to keep the current retransmit
5181          * times until then.
5182          */
5183         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
5184 #ifdef INET6
5185                 if (bbr->r_is_v6)
5186                         in6_losing(tp->t_inpcb);
5187                 else
5188 #endif
5189                         in_losing(tp->t_inpcb);
5190                 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
5191                 tp->t_srtt = 0;
5192         }
5193         sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
5194         tp->snd_recover = tp->snd_max;
5195         tp->t_flags |= TF_ACKNOW;
5196         tp->t_rtttime = 0;
5197 out:
5198         return (retval);
5199 }
5200
5201 static int
5202 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling)
5203 {
5204         int32_t ret = 0;
5205         int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
5206
5207         if (timers == 0) {
5208                 return (0);
5209         }
5210         if (tp->t_state == TCPS_LISTEN) {
5211                 /* no timers on listen sockets */
5212                 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
5213                         return (0);
5214                 return (1);
5215         }
5216         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
5217                 uint32_t left;
5218
5219                 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
5220                         ret = -1;
5221                         bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5222                         return (0);
5223                 }
5224                 if (hpts_calling == 0) {
5225                         ret = -2;
5226                         bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5227                         return (0);
5228                 }
5229                 /*
5230                  * Ok our timer went off early and we are not paced false
5231                  * alarm, go back to sleep.
5232                  */
5233                 left = bbr->r_ctl.rc_timer_exp - cts;
5234                 ret = -3;
5235                 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling);
5236                 tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left));
5237                 return (1);
5238         }
5239         bbr->rc_tmr_stopped = 0;
5240         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
5241         if (timers & PACE_TMR_DELACK) {
5242                 ret = bbr_timeout_delack(tp, bbr, cts);
5243         } else if (timers & PACE_TMR_PERSIT) {
5244                 ret = bbr_timeout_persist(tp, bbr, cts);
5245         } else if (timers & PACE_TMR_RACK) {
5246                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5247                 ret = bbr_timeout_rack(tp, bbr, cts);
5248         } else if (timers & PACE_TMR_TLP) {
5249                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5250                 ret = bbr_timeout_tlp(tp, bbr, cts);
5251         } else if (timers & PACE_TMR_RXT) {
5252                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5253                 ret = bbr_timeout_rxt(tp, bbr, cts);
5254         } else if (timers & PACE_TMR_KEEP) {
5255                 ret = bbr_timeout_keepalive(tp, bbr, cts);
5256         }
5257         bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling);
5258         return (ret);
5259 }
5260
5261 static void
5262 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts)
5263 {
5264         if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
5265                 uint8_t hpts_removed = 0;
5266
5267                 if (bbr->rc_inp->inp_in_hpts &&
5268                     (bbr->rc_timer_first == 1)) {
5269                         /*
5270                          * If we are canceling timer's when we have the
5271                          * timer ahead of the output being paced. We also
5272                          * must remove ourselves from the hpts.
5273                          */
5274                         hpts_removed = 1;
5275                         tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
5276                         if (bbr->r_ctl.rc_last_delay_val) {
5277                                 /* Update the last hptsi delay too */
5278                                 uint32_t time_since_send;
5279
5280                                 if (TSTMP_GT(cts, bbr->rc_pacer_started))
5281                                         time_since_send = cts - bbr->rc_pacer_started;
5282                                 else
5283                                         time_since_send = 0;
5284                                 if (bbr->r_ctl.rc_last_delay_val > time_since_send) {
5285                                         /* Cut down our slot time */
5286                                         bbr->r_ctl.rc_last_delay_val -= time_since_send;
5287                                 } else {
5288                                         bbr->r_ctl.rc_last_delay_val = 0;
5289                                 }
5290                                 bbr->rc_pacer_started = cts;
5291                         }
5292                 }
5293                 bbr->rc_timer_first = 0;
5294                 bbr_log_to_cancel(bbr, line, cts, hpts_removed);
5295                 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
5296                 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
5297         }
5298 }
5299
5300 static void
5301 bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type)
5302 {
5303         struct tcp_bbr *bbr;
5304
5305         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
5306         bbr->rc_all_timers_stopped = 1;
5307         return;
5308 }
5309
5310 /*
5311  * stop all timers always returning 0.
5312  */
5313 static int
5314 bbr_stopall(struct tcpcb *tp)
5315 {
5316         return (0);
5317 }
5318
5319 static void
5320 bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
5321 {
5322         return;
5323 }
5324
5325 /*
5326  * return true if a bbr timer (rack or tlp) is active.
5327  */
5328 static int
5329 bbr_timer_active(struct tcpcb *tp, uint32_t timer_type)
5330 {
5331         return (0);
5332 }
5333
5334 static uint32_t
5335 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts)
5336 {
5337         struct bbr_sendmap *rsm;
5338
5339         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
5340         if ((rsm == NULL) || (u_rsm == rsm))
5341                 return (cts);
5342         return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
5343 }
5344
5345 static void
5346 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr,
5347      struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time)
5348 {
5349         int32_t idx;
5350
5351         rsm->r_rtr_cnt++;
5352         rsm->r_dupack = 0;
5353         if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) {
5354                 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS;
5355                 rsm->r_flags |= BBR_OVERMAX;
5356         }
5357         if (rsm->r_flags & BBR_RWND_COLLAPSED) {
5358                 /* Take off the collapsed flag at rxt */
5359                 rsm->r_flags &= ~BBR_RWND_COLLAPSED;
5360         }
5361         if (rsm->r_flags & BBR_MARKED_LOST) {
5362                 /* We have retransmitted, its no longer lost */
5363                 rsm->r_flags &= ~BBR_MARKED_LOST;
5364                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
5365         }
5366         if (rsm->r_flags & BBR_RXT_CLEARED) {
5367                 /*
5368                  * We hit a RXT timer on it and
5369                  * we cleared the "acked" flag.
5370                  * We now have it going back into
5371                  * flight, we can remove the cleared
5372                  * flag and possibly do accounting on
5373                  * this piece.
5374                  */
5375                 rsm->r_flags &= ~BBR_RXT_CLEARED;
5376         }
5377         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) {
5378                 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
5379                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
5380         }
5381         idx = rsm->r_rtr_cnt - 1;
5382         rsm->r_tim_lastsent[idx] = cts;
5383         rsm->r_pacing_delay = pacing_time;
5384         rsm->r_delivered = bbr->r_ctl.rc_delivered;
5385         rsm->r_ts_valid = bbr->rc_ts_valid;
5386         if (bbr->rc_ts_valid)
5387                 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
5388         if (bbr->r_ctl.r_app_limited_until)
5389                 rsm->r_app_limited = 1;
5390         else
5391                 rsm->r_app_limited = 0;
5392         if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
5393                 rsm->r_bbr_state = bbr_state_val(bbr);
5394         else
5395                 rsm->r_bbr_state = 8;
5396         if (rsm->r_flags & BBR_ACKED) {
5397                 /* Problably MTU discovery messing with us */
5398                 uint32_t old_flags;
5399
5400                 old_flags = rsm->r_flags;
5401                 rsm->r_flags &= ~BBR_ACKED;
5402                 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
5403                 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
5404                 if (bbr->r_ctl.rc_sacked == 0)
5405                         bbr->r_ctl.rc_sacklast = NULL;
5406         }
5407         if (rsm->r_in_tmap) {
5408                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5409         }
5410         TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5411         rsm->r_in_tmap = 1;
5412         if (rsm->r_flags & BBR_SACK_PASSED) {
5413                 /* We have retransmitted due to the SACK pass */
5414                 rsm->r_flags &= ~BBR_SACK_PASSED;
5415                 rsm->r_flags |= BBR_WAS_SACKPASS;
5416         }
5417         rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
5418         rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
5419                                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
5420         bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
5421         if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
5422                 rsm->r_is_gain = 1;
5423                 rsm->r_is_drain = 0;
5424         } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
5425                 rsm->r_is_drain = 1;
5426                 rsm->r_is_gain = 0;
5427         } else {
5428                 rsm->r_is_drain = 0;
5429                 rsm->r_is_gain = 0;
5430         }
5431         rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */
5432 }
5433
5434 /*
5435  * Returns 0, or the sequence where we stopped
5436  * updating. We also update the lenp to be the amount
5437  * of data left.
5438  */
5439
5440 static uint32_t
5441 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr,
5442     struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time)
5443 {
5444         /*
5445          * We (re-)transmitted starting at rsm->r_start for some length
5446          * (possibly less than r_end.
5447          */
5448         struct bbr_sendmap *nrsm;
5449         uint32_t c_end;
5450         int32_t len;
5451
5452         len = *lenp;
5453         c_end = rsm->r_start + len;
5454         if (SEQ_GEQ(c_end, rsm->r_end)) {
5455                 /*
5456                  * We retransmitted the whole piece or more than the whole
5457                  * slopping into the next rsm.
5458                  */
5459                 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5460                 if (c_end == rsm->r_end) {
5461                         *lenp = 0;
5462                         return (0);
5463                 } else {
5464                         int32_t act_len;
5465
5466                         /* Hangs over the end return whats left */
5467                         act_len = rsm->r_end - rsm->r_start;
5468                         *lenp = (len - act_len);
5469                         return (rsm->r_end);
5470                 }
5471                 /* We don't get out of this block. */
5472         }
5473         /*
5474          * Here we retransmitted less than the whole thing which means we
5475          * have to split this into what was transmitted and what was not.
5476          */
5477         nrsm = bbr_alloc_full_limit(bbr);
5478         if (nrsm == NULL) {
5479                 *lenp = 0;
5480                 return (0);
5481         }
5482         /*
5483          * So here we are going to take the original rsm and make it what we
5484          * retransmitted. nrsm will be the tail portion we did not
5485          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
5486          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
5487          * 1, 6 and the new piece will be 6, 11.
5488          */
5489         bbr_clone_rsm(bbr, nrsm, rsm, c_end);
5490         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
5491         nrsm->r_dupack = 0;
5492         if (rsm->r_in_tmap) {
5493                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
5494                 nrsm->r_in_tmap = 1;
5495         }
5496         rsm->r_flags &= (~BBR_HAS_FIN);
5497         bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5498         *lenp = 0;
5499         return (0);
5500 }
5501
5502 static uint64_t
5503 bbr_get_hardware_rate(struct tcp_bbr *bbr)
5504 {
5505         uint64_t bw;
5506
5507         bw = bbr_get_bw(bbr);
5508         bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN];
5509         bw /= (uint64_t)BBR_UNIT;
5510         return(bw);
5511 }
5512
5513 static void
5514 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts,
5515                        uint64_t act_rate, uint64_t rate_wanted)
5516 {
5517         /*
5518          * We could not get a full gains worth
5519          * of rate.
5520          */
5521         if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) {
5522                 /* we can't even get the real rate */
5523                 uint64_t red;
5524
5525                 bbr->skip_gain = 1;
5526                 bbr->gain_is_limited = 0;
5527                 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate;
5528                 if (red)
5529                         filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts);
5530         } else {
5531                 /* We can use a lower gain */
5532                 bbr->skip_gain = 0;
5533                 bbr->gain_is_limited = 1;
5534         }
5535 }
5536
5537 static void
5538 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts)
5539 {
5540         const struct tcp_hwrate_limit_table *nrte;
5541         int error, rate = -1;
5542
5543         if (bbr->r_ctl.crte == NULL)
5544                 return;
5545         if ((bbr->rc_inp->inp_route.ro_nh == NULL) ||
5546             (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) {
5547                 /* Lost our routes? */
5548                 /* Clear the way for a re-attempt */
5549                 bbr->bbr_attempt_hdwr_pace = 0;
5550 lost_rate:
5551                 bbr->gain_is_limited = 0;
5552                 bbr->skip_gain = 0;
5553                 bbr->bbr_hdrw_pacing = 0;
5554                 counter_u64_add(bbr_flows_whdwr_pacing, -1);
5555                 counter_u64_add(bbr_flows_nohdwr_pacing, 1);
5556                 tcp_bbr_tso_size_check(bbr, cts);
5557                 return;
5558         }
5559         rate = bbr_get_hardware_rate(bbr);
5560         nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte,
5561                                    bbr->rc_tp,
5562                                    bbr->rc_inp->inp_route.ro_nh->nh_ifp,
5563                                    rate,
5564                                    (RS_PACING_GEQ|RS_PACING_SUB_OK),
5565                                    &error, NULL);
5566         if (nrte == NULL) {
5567                 goto lost_rate;
5568         }
5569         if (nrte != bbr->r_ctl.crte) {
5570                 bbr->r_ctl.crte = nrte;
5571                 if (error == 0)  {
5572                         BBR_STAT_INC(bbr_hdwr_rl_mod_ok);
5573                         if (bbr->r_ctl.crte->rate < rate) {
5574                                 /* We have a problem */
5575                                 bbr_setup_less_of_rate(bbr, cts,
5576                                                        bbr->r_ctl.crte->rate, rate);
5577                         } else {
5578                                 /* We are good */
5579                                 bbr->gain_is_limited = 0;
5580                                 bbr->skip_gain = 0;
5581                         }
5582                 } else {
5583                         /* A failure should release the tag */
5584                         BBR_STAT_INC(bbr_hdwr_rl_mod_fail);
5585                         bbr->gain_is_limited = 0;
5586                         bbr->skip_gain = 0;
5587                         bbr->bbr_hdrw_pacing = 0;
5588                 }
5589                 bbr_type_log_hdwr_pacing(bbr,
5590                                          bbr->r_ctl.crte->ptbl->rs_ifp,
5591                                          rate,
5592                                          ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate),
5593                                          __LINE__,
5594                                          cts,
5595                                          error);
5596         }
5597 }
5598
5599 static void
5600 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts)
5601 {
5602         /*
5603          * If we have hardware pacing support
5604          * we need to factor that in for our
5605          * TSO size.
5606          */
5607         const struct tcp_hwrate_limit_table *rlp;
5608         uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay;
5609
5610         if ((bbr->bbr_hdrw_pacing == 0) ||
5611             (IN_RECOVERY(bbr->rc_tp->t_flags)) ||
5612             (bbr->r_ctl.crte == NULL))
5613                 return;
5614         if (bbr->hw_pacing_set == 0) {
5615                 /* Not yet by the hdwr pacing count delay */
5616                 return;
5617         }
5618         if (bbr_hdwr_pace_adjust == 0) {
5619                 /* No adjustment */
5620                 return;
5621         }
5622         rlp = bbr->r_ctl.crte;
5623         if (bbr->rc_tp->t_maxseg > bbr->rc_last_options)
5624                 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5625         else
5626                 maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5627         /*
5628          * So lets first get the
5629          * time we will take between
5630          * TSO sized sends currently without
5631          * hardware help.
5632          */
5633         cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT,
5634                         bbr->r_ctl.rc_pace_max_segs, cts, 1);
5635         hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg;
5636         hdwr_delay *= rlp->time_between;
5637         if (cur_delay > hdwr_delay)
5638                 delta = cur_delay - hdwr_delay;
5639         else
5640                 delta = 0;
5641         bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay,
5642                              (bbr->r_ctl.rc_pace_max_segs / maxseg),
5643                              1);
5644         if (delta &&
5645             (delta < (max(rlp->time_between,
5646                           bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) {
5647                 /*
5648                  * Now lets divide by the pacing
5649                  * time between each segment the
5650                  * hardware sends rounding up and
5651                  * derive a bytes from that. We multiply
5652                  * that by bbr_hdwr_pace_adjust to get
5653                  * more bang for our buck.
5654                  *
5655                  * The goal is to have the software pacer
5656                  * waiting no more than an additional
5657                  * pacing delay if we can (without the
5658                  * compensation i.e. x bbr_hdwr_pace_adjust).
5659                  */
5660                 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between),
5661                              (bbr->r_ctl.rc_pace_max_segs/maxseg));
5662                 seg_sz *= bbr_hdwr_pace_adjust;
5663                 if (bbr_hdwr_pace_floor &&
5664                     (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5665                         /* Currently hardware paces
5666                          * out rs_min_seg segments at a time.
5667                          * We need to make sure we always send at least
5668                          * a full burst of bbr_hdwr_pace_floor down.
5669                          */
5670                         seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5671                 }
5672                 seg_sz *= maxseg;
5673         } else if (delta == 0) {
5674                 /*
5675                  * The highest pacing rate is
5676                  * above our b/w gained. This means
5677                  * we probably are going quite fast at
5678                  * the hardware highest rate. Lets just multiply
5679                  * the calculated TSO size by the
5680                  * multiplier factor (its probably
5681                  * 4 segments in the default config for
5682                  * mlx).
5683                  */
5684                 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust;
5685                 if (bbr_hdwr_pace_floor &&
5686                     (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5687                         /* Currently hardware paces
5688                          * out rs_min_seg segments at a time.
5689                          * We need to make sure we always send at least
5690                          * a full burst of bbr_hdwr_pace_floor down.
5691                          */
5692                         seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5693                 }
5694         } else {
5695                 /*
5696                  * The pacing time difference is so
5697                  * big that the hardware will
5698                  * pace out more rapidly then we
5699                  * really want and then we
5700                  * will have a long delay. Lets just keep
5701                  * the same TSO size so its as if
5702                  * we were not using hdwr pacing (we
5703                  * just gain a bit of spacing from the
5704                  * hardware if seg_sz > 1).
5705                  */
5706                 seg_sz = bbr->r_ctl.rc_pace_max_segs;
5707         }
5708         if (seg_sz > bbr->r_ctl.rc_pace_max_segs)
5709                 new_tso = seg_sz;
5710         else
5711                 new_tso = bbr->r_ctl.rc_pace_max_segs;
5712         if (new_tso >= (PACE_MAX_IP_BYTES-maxseg))
5713                 new_tso = PACE_MAX_IP_BYTES - maxseg;
5714
5715         if (new_tso != bbr->r_ctl.rc_pace_max_segs) {
5716                 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0);
5717                 bbr->r_ctl.rc_pace_max_segs = new_tso;
5718         }
5719 }
5720
5721 static void
5722 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts)
5723 {
5724         uint64_t bw;
5725         uint32_t old_tso = 0, new_tso;
5726         uint32_t maxseg, bytes;
5727         uint32_t tls_seg=0;
5728         /*
5729          * Google/linux uses the following algorithm to determine
5730          * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18):
5731          *
5732          *  bytes = bw_in_bytes_per_second / 1000
5733          *  bytes = min(bytes, 64k)
5734          *  tso_segs = bytes / MSS
5735          *  if (bw < 1.2Mbs)
5736          *      min_tso_segs = 1
5737          *  else
5738          *      min_tso_segs = 2
5739          * tso_segs = max(tso_segs, min_tso_segs)
5740          *
5741          * * Note apply a device specific limit (we apply this in the
5742          *   tcp_m_copym).
5743          * Note that before the initial measurement is made google bursts out
5744          * a full iwnd just like new-reno/cubic.
5745          *
5746          * We do not use this algorithm. Instead we
5747          * use a two phased approach:
5748          *
5749          *  if ( bw <= per-tcb-cross-over)
5750          *     goal_tso =  calculate how much with this bw we
5751          *                 can send in goal-time seconds.
5752          *     if (goal_tso > mss)
5753          *         seg = goal_tso / mss
5754          *         tso = seg * mss
5755          *     else
5756          *         tso = mss
5757          *     if (tso > per-tcb-max)
5758          *         tso = per-tcb-max
5759          *  else if ( bw > 512Mbps)
5760          *     tso = max-tso (64k/mss)
5761          *  else
5762          *     goal_tso = bw / per-tcb-divsor
5763          *     seg = (goal_tso + mss-1)/mss
5764          *     tso = seg * mss
5765          *
5766          * if (tso < per-tcb-floor)
5767          *    tso = per-tcb-floor
5768          * if (tso > per-tcb-utter_max)
5769          *    tso = per-tcb-utter_max
5770          *
5771          * Note the default per-tcb-divisor is 1000 (same as google).
5772          * the goal cross over is 30Mbps however. To recreate googles
5773          * algorithm you need to set:
5774          *
5775          * cross-over = 23,168,000 bps
5776          * goal-time = 18000
5777          * per-tcb-max = 2
5778          * per-tcb-divisor = 1000
5779          * per-tcb-floor = 1
5780          *
5781          * This will get you "google bbr" behavior with respect to tso size.
5782          *
5783          * Note we do set anything TSO size until we are past the initial
5784          * window. Before that we gnerally use either a single MSS
5785          * or we use the full IW size (so we burst a IW at a time)
5786          */
5787
5788         if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) {
5789                 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5790         } else {
5791                 maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5792         }
5793         old_tso = bbr->r_ctl.rc_pace_max_segs;
5794         if (bbr->rc_past_init_win == 0) {
5795                 /*
5796                  * Not enough data has been acknowledged to make a
5797                  * judgement. Set up the initial TSO based on if we
5798                  * are sending a full IW at once or not.
5799                  */
5800                 if (bbr->rc_use_google)
5801                         bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2);
5802                 else if (bbr->bbr_init_win_cheat)
5803                         bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp);
5804                 else
5805                         bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5806                 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg)
5807                         bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg;
5808                 if (bbr->r_ctl.rc_pace_max_segs == 0) {
5809                         bbr->r_ctl.rc_pace_max_segs = maxseg;
5810                 }
5811                 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0);
5812                         bbr_adjust_for_hw_pacing(bbr, cts);
5813                 return;
5814         }
5815         /**
5816          * Now lets set the TSO goal based on our delivery rate in
5817          * bytes per second. Note we only do this if
5818          * we have acked at least the initial cwnd worth of data.
5819          */
5820         bw = bbr_get_bw(bbr);
5821         if (IN_RECOVERY(bbr->rc_tp->t_flags) &&
5822              (bbr->rc_use_google == 0)) {
5823                 /* We clamp to one MSS in recovery */
5824                 new_tso = maxseg;
5825         } else if (bbr->rc_use_google) {
5826                 int min_tso_segs;
5827
5828                 /* Google considers the gain too */
5829                 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) {
5830                         bw *= bbr->r_ctl.rc_bbr_hptsi_gain;
5831                         bw /= BBR_UNIT;
5832                 }
5833                 bytes = bw / 1024;
5834                 if (bytes > (64 * 1024))
5835                         bytes = 64 * 1024;
5836                 new_tso = bytes / maxseg;
5837                 if (bw < ONE_POINT_TWO_MEG)
5838                         min_tso_segs = 1;
5839                 else
5840                         min_tso_segs = 2;
5841                 if (new_tso < min_tso_segs)
5842                         new_tso = min_tso_segs;
5843                 new_tso *= maxseg;
5844         } else if (bbr->rc_no_pacing) {
5845                 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg;
5846         } else if (bw <= bbr->r_ctl.bbr_cross_over) {
5847                 /*
5848                  * Calculate the worse case b/w TSO if we are inserting no
5849                  * more than a delay_target number of TSO's.
5850                  */
5851                 uint32_t tso_len, min_tso;
5852
5853                 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw);
5854                 if (tso_len > maxseg) {
5855                         new_tso = tso_len / maxseg;
5856                         if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max)
5857                                 new_tso = bbr->r_ctl.bbr_hptsi_segments_max;
5858                         new_tso *= maxseg;
5859                 } else {
5860                         /*
5861                          * less than a full sized frame yikes.. long rtt or
5862                          * low bw?
5863                          */
5864                         min_tso = bbr_minseg(bbr);
5865                         if ((tso_len > min_tso) && (bbr_all_get_min == 0))
5866                                 new_tso = rounddown(tso_len, min_tso);
5867                         else
5868                                 new_tso = min_tso;
5869                 }
5870         } else if (bw > FIVETWELVE_MBPS) {
5871                 /*
5872                  * This guy is so fast b/w wise that we can TSO as large as
5873                  * possible of segments that the NIC will allow.
5874                  */
5875                 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5876         } else {
5877                 /*
5878                  * This formula is based on attempting to send a segment or
5879                  * more every bbr_hptsi_per_second. The default is 1000
5880                  * which means you are targeting what you can send every 1ms
5881                  * based on the peers bw.
5882                  *
5883                  * If the number drops to say 500, then you are looking more
5884                  * at 2ms and you will raise how much we send in a single
5885                  * TSO thus saving CPU (less bbr_output_wtime() calls). The
5886                  * trade off of course is you will send more at once and
5887                  * thus tend to clump up the sends into larger "bursts"
5888                  * building a queue.
5889                  */
5890                 bw /= bbr->r_ctl.bbr_hptsi_per_second;
5891                 new_tso = roundup(bw, (uint64_t)maxseg);
5892                 /*
5893                  * Gate the floor to match what our lower than 48Mbps
5894                  * algorithm does. The ceiling (bbr_hptsi_segments_max) thus
5895                  * becomes the floor for this calculation.
5896                  */
5897                 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg))
5898                         new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg);
5899         }
5900         if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor)))
5901                 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor;
5902         if (new_tso > PACE_MAX_IP_BYTES)
5903                 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5904         /* Enforce an utter maximum. */
5905         if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) {
5906                 new_tso = bbr->r_ctl.bbr_utter_max * maxseg;
5907         }
5908         if (old_tso != new_tso) {
5909                 /* Only log changes */
5910                 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0);
5911                 bbr->r_ctl.rc_pace_max_segs = new_tso;
5912         }
5913         /* We have hardware pacing! */
5914         bbr_adjust_for_hw_pacing(bbr, cts);
5915 }
5916
5917 static void
5918 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len,
5919     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t cts,
5920     struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc,
5921     struct sockbuf *sb)
5922 {
5923
5924         struct bbr_sendmap *rsm, *nrsm;
5925         register uint32_t snd_max, snd_una;
5926         uint32_t pacing_time;
5927         /*
5928          * Add to the RACK log of packets in flight or retransmitted. If
5929          * there is a TS option we will use the TS echoed, if not we will
5930          * grab a TS.
5931          *
5932          * Retransmissions will increment the count and move the ts to its
5933          * proper place. Note that if options do not include TS's then we
5934          * won't be able to effectively use the ACK for an RTT on a retran.
5935          *
5936          * Notes about r_start and r_end. Lets consider a send starting at
5937          * sequence 1 for 10 bytes. In such an example the r_start would be
5938          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
5939          * This means that r_end is actually the first sequence for the next
5940          * slot (11).
5941          *
5942          */
5943         INP_WLOCK_ASSERT(tp->t_inpcb);
5944         if (err) {
5945                 /*
5946                  * We don't log errors -- we could but snd_max does not
5947                  * advance in this case either.
5948                  */
5949                 return;
5950         }
5951         if (th_flags & TH_RST) {
5952                 /*
5953                  * We don't log resets and we return immediately from
5954                  * sending
5955                  */
5956                 *abandon = 1;
5957                 return;
5958         }
5959         snd_una = tp->snd_una;
5960         if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) {
5961                 /*
5962                  * The call to bbr_log_output is made before bumping
5963                  * snd_max. This means we can record one extra byte on a SYN
5964                  * or FIN if seq_out is adding more on and a FIN is present
5965                  * (and we are not resending).
5966                  */
5967                 if ((th_flags & TH_SYN) && (tp->iss == seq_out))
5968                         len++;
5969                 if (th_flags & TH_FIN)
5970                         len++;
5971         }
5972         if (SEQ_LEQ((seq_out + len), snd_una)) {
5973                 /* Are sending an old segment to induce an ack (keep-alive)? */
5974                 return;
5975         }
5976         if (SEQ_LT(seq_out, snd_una)) {
5977                 /* huh? should we panic? */
5978                 uint32_t end;
5979
5980                 end = seq_out + len;
5981                 seq_out = snd_una;
5982                 len = end - seq_out;
5983         }
5984         snd_max = tp->snd_max;
5985         if (len == 0) {
5986                 /* We don't log zero window probes */
5987                 return;
5988         }
5989         pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1);
5990         /* First question is it a retransmission? */
5991         if (seq_out == snd_max) {
5992 again:
5993                 rsm = bbr_alloc(bbr);
5994                 if (rsm == NULL) {
5995                         return;
5996                 }
5997                 rsm->r_flags = 0;
5998                 if (th_flags & TH_SYN)
5999                         rsm->r_flags |= BBR_HAS_SYN;
6000                 if (th_flags & TH_FIN)
6001                         rsm->r_flags |= BBR_HAS_FIN;
6002                 rsm->r_tim_lastsent[0] = cts;
6003                 rsm->r_rtr_cnt = 1;
6004                 rsm->r_rtr_bytes = 0;
6005                 rsm->r_start = seq_out;
6006                 rsm->r_end = rsm->r_start + len;
6007                 rsm->r_dupack = 0;
6008                 rsm->r_delivered = bbr->r_ctl.rc_delivered;
6009                 rsm->r_pacing_delay = pacing_time;
6010                 rsm->r_ts_valid = bbr->rc_ts_valid;
6011                 if (bbr->rc_ts_valid)
6012                         rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
6013                 rsm->r_del_time = bbr->r_ctl.rc_del_time;
6014                 if (bbr->r_ctl.r_app_limited_until)
6015                         rsm->r_app_limited = 1;
6016                 else
6017                         rsm->r_app_limited = 0;
6018                 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
6019                 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
6020                                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
6021                 /*
6022                  * Here we must also add in this rsm since snd_max
6023                  * is updated after we return from a new send.
6024                  */
6025                 rsm->r_flight_at_send += len;
6026                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
6027                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
6028                 rsm->r_in_tmap = 1;
6029                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
6030                         rsm->r_bbr_state = bbr_state_val(bbr);
6031                 else
6032                         rsm->r_bbr_state = 8;
6033                 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
6034                         rsm->r_is_gain = 1;
6035                         rsm->r_is_drain = 0;
6036                 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
6037                         rsm->r_is_drain = 1;
6038                         rsm->r_is_gain = 0;
6039                 } else {
6040                         rsm->r_is_drain = 0;
6041                         rsm->r_is_gain = 0;
6042                 }
6043                 return;
6044         }
6045         /*
6046          * If we reach here its a retransmission and we need to find it.
6047          */
6048 more:
6049         if (hintrsm && (hintrsm->r_start == seq_out)) {
6050                 rsm = hintrsm;
6051                 hintrsm = NULL;
6052         } else if (bbr->r_ctl.rc_next) {
6053                 /* We have a hint from a previous run */
6054                 rsm = bbr->r_ctl.rc_next;
6055         } else {
6056                 /* No hints sorry */
6057                 rsm = NULL;
6058         }
6059         if ((rsm) && (rsm->r_start == seq_out)) {
6060                 /*
6061                  * We used rc_next or hintrsm  to retransmit, hopefully the
6062                  * likely case.
6063                  */
6064                 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6065                 if (len == 0) {
6066                         return;
6067                 } else {
6068                         goto more;
6069                 }
6070         }
6071         /* Ok it was not the last pointer go through it the hard way. */
6072         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6073                 if (rsm->r_start == seq_out) {
6074                         seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6075                         bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
6076                         if (len == 0) {
6077                                 return;
6078                         } else {
6079                                 continue;
6080                         }
6081                 }
6082                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
6083                         /* Transmitted within this piece */
6084                         /*
6085                          * Ok we must split off the front and then let the
6086                          * update do the rest
6087                          */
6088                         nrsm = bbr_alloc_full_limit(bbr);
6089                         if (nrsm == NULL) {
6090                                 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
6091                                 return;
6092                         }
6093                         /*
6094                          * copy rsm to nrsm and then trim the front of rsm
6095                          * to not include this part.
6096                          */
6097                         bbr_clone_rsm(bbr, nrsm, rsm, seq_out);
6098                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
6099                         if (rsm->r_in_tmap) {
6100                                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6101                                 nrsm->r_in_tmap = 1;
6102                         }
6103                         rsm->r_flags &= (~BBR_HAS_FIN);
6104                         seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time);
6105                         if (len == 0) {
6106                                 return;
6107                         }
6108                 }
6109         }
6110         /*
6111          * Hmm not found in map did they retransmit both old and on into the
6112          * new?
6113          */
6114         if (seq_out == tp->snd_max) {
6115                 goto again;
6116         } else if (SEQ_LT(seq_out, tp->snd_max)) {
6117 #ifdef BBR_INVARIANTS
6118                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
6119                     seq_out, len, tp->snd_una, tp->snd_max);
6120                 printf("Starting Dump of all rack entries\n");
6121                 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6122                         printf("rsm:%p start:%u end:%u\n",
6123                             rsm, rsm->r_start, rsm->r_end);
6124                 }
6125                 printf("Dump complete\n");
6126                 panic("seq_out not found rack:%p tp:%p",
6127                     bbr, tp);
6128 #endif
6129         } else {
6130 #ifdef BBR_INVARIANTS
6131                 /*
6132                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
6133                  * flag)
6134                  */
6135                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
6136                     seq_out, len, tp->snd_max, tp);
6137 #endif
6138         }
6139 }
6140
6141 static void
6142 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt)
6143 {
6144         /*
6145          * Collapse timeout back the cum-ack moved.
6146          */
6147         tp->t_rxtshift = 0;
6148         tp->t_softerror = 0;
6149 }
6150
6151 static void
6152 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin)
6153 {
6154         bbr->rtt_valid = 1;
6155         bbr->r_ctl.cur_rtt = rtt_usecs;
6156         bbr->r_ctl.ts_in = tsin;
6157         if (rsm_send_time)
6158                 bbr->r_ctl.cur_rtt_send_time = rsm_send_time;
6159 }
6160
6161 static void
6162 bbr_make_timestamp_determination(struct tcp_bbr *bbr)
6163 {
6164         /**
6165          * We have in our bbr control:
6166          * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp).
6167          * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts).
6168          * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts)
6169          * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time)
6170          *
6171          * Now we can calculate the time between the sends by doing:
6172          *
6173          * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts
6174          *
6175          * And the peer's time between receiving them by doing:
6176          *
6177          * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp
6178          *
6179          * We want to figure out if the timestamp values are in msec, 10msec or usec.
6180          * We also may find that we can't use the timestamps if say we see
6181          * that the peer_delta indicates that though we may have taken 10ms to
6182          * pace out the data, it only saw 1ms between the two packets. This would
6183          * indicate that somewhere on the path is a batching entity that is giving
6184          * out time-slices of the actual b/w. This would mean we could not use
6185          * reliably the peers timestamps.
6186          *
6187          * We expect delta > peer_delta initially. Until we figure out the
6188          * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio.
6189          * If we place 1000 there then its a ms vs our usec. If we place 10000 there
6190          * then its 10ms vs our usec. If the peer is running a usec clock we would
6191          * put a 1 there. If the value is faster then ours, we will disable the
6192          * use of timestamps (though we could revist this later if we find it to be not
6193          * just an isolated one or two flows)).
6194          *
6195          * To detect the batching middle boxes we will come up with our compensation and
6196          * if with it in place, we find the peer is drastically off (by some margin) in
6197          * the smaller direction, then we will assume the worst case and disable use of timestamps.
6198          *
6199          */
6200         uint64_t delta, peer_delta, delta_up;
6201
6202         delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts;
6203         if (delta < bbr_min_usec_delta) {
6204                 /*
6205                  * Have not seen a min amount of time
6206                  * between our send times so we can
6207                  * make a determination of the timestamp
6208                  * yet.
6209                  */
6210                 return;
6211         }
6212         peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp;
6213         if (peer_delta < bbr_min_peer_delta) {
6214                 /*
6215                  * We may have enough in the form of
6216                  * our delta but the peers number
6217                  * has not changed that much. It could
6218                  * be its clock ratio is such that
6219                  * we need more data (10ms tick) or
6220                  * there may be other compression scenarios
6221                  * going on. In any event we need the
6222                  * spread to be larger.
6223                  */
6224                 return;
6225         }
6226         /* Ok lets first see which way our delta is going */
6227         if (peer_delta > delta) {
6228                 /* Very unlikely, the peer without
6229                  * compensation shows that it saw
6230                  * the two sends arrive further apart
6231                  * then we saw then in micro-seconds.
6232                  */
6233                 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) {
6234                         /* well it looks like the peer is a micro-second clock. */
6235                         bbr->rc_ts_clock_set = 1;
6236                         bbr->r_ctl.bbr_peer_tsratio = 1;
6237                 } else {
6238                         bbr->rc_ts_cant_be_used = 1;
6239                         bbr->rc_ts_clock_set = 1;
6240                 }
6241                 return;
6242         }
6243         /* Ok we know that the peer_delta is smaller than our send distance */
6244         bbr->rc_ts_clock_set = 1;
6245         /* First question is it within the percentage that they are using usec time? */
6246         delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent;
6247         if ((peer_delta + delta_up) >= delta) {
6248                 /* Its a usec clock */
6249                 bbr->r_ctl.bbr_peer_tsratio = 1;
6250                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6251                 return;
6252         }
6253         /* Ok if not usec, what about 10usec (though unlikely)? */
6254         delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent;
6255         if (((peer_delta * 10) + delta_up) >= delta) {
6256                 bbr->r_ctl.bbr_peer_tsratio = 10;
6257                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6258                 return;
6259         }
6260         /* And what about 100usec (though again unlikely)? */
6261         delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent;
6262         if (((peer_delta * 100) + delta_up) >= delta) {
6263                 bbr->r_ctl.bbr_peer_tsratio = 100;
6264                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6265                 return;
6266         }
6267         /* And how about 1 msec (the most likely one)? */
6268         delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent;
6269         if (((peer_delta * 1000) + delta_up) >= delta) {
6270                 bbr->r_ctl.bbr_peer_tsratio = 1000;
6271                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6272                 return;
6273         }
6274         /* Ok if not msec could it be 10 msec? */
6275         delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent;
6276         if (((peer_delta * 10000) + delta_up) >= delta) {
6277                 bbr->r_ctl.bbr_peer_tsratio = 10000;
6278                 return;
6279         }
6280         /* If we fall down here the clock tick so slowly we can't use it */
6281         bbr->rc_ts_cant_be_used = 1;
6282         bbr->r_ctl.bbr_peer_tsratio = 0;
6283         bbr_log_tstmp_validation(bbr, peer_delta, delta);
6284 }
6285
6286 /*
6287  * Collect new round-trip time estimate
6288  * and update averages and current timeout.
6289  */
6290 static void
6291 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts)
6292 {
6293         int32_t delta;
6294         uint32_t rtt, tsin;
6295         int32_t rtt_ticks;
6296
6297         if (bbr->rtt_valid == 0)
6298                 /* No valid sample */
6299                 return;
6300
6301         rtt = bbr->r_ctl.cur_rtt;
6302         tsin = bbr->r_ctl.ts_in;
6303         if (bbr->rc_prtt_set_ts) {
6304                 /*
6305                  * We are to force feed the rttProp filter due
6306                  * to an entry into PROBE_RTT. This assures
6307                  * that the times are sync'd between when we
6308                  * go into PROBE_RTT and the filter expiration.
6309                  *
6310                  * Google does not use a true filter, so they do
6311                  * this implicitly since they only keep one value
6312                  * and when they enter probe-rtt they update the
6313                  * value to the newest rtt.
6314                  */
6315                 uint32_t rtt_prop;
6316
6317                 bbr->rc_prtt_set_ts = 0;
6318                 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
6319                 if (rtt > rtt_prop)
6320                         filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts);
6321                 else
6322                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6323         }
6324         if (bbr->rc_ack_was_delayed)
6325                 rtt += bbr->r_ctl.rc_ack_hdwr_delay;
6326
6327         if (rtt < bbr->r_ctl.rc_lowest_rtt)
6328                 bbr->r_ctl.rc_lowest_rtt = rtt;
6329         bbr_log_rtt_sample(bbr, rtt, tsin);
6330         if (bbr->r_init_rtt) {
6331                 /*
6332                  * The initial rtt is not-trusted, nuke it and lets get
6333                  * our first valid measurement in.
6334                  */
6335                 bbr->r_init_rtt = 0;
6336                 tp->t_srtt = 0;
6337         }
6338         if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) {
6339                 /*
6340                  * So we have not yet figured out
6341                  * what the peers TSTMP value is
6342                  * in (most likely ms). We need a
6343                  * series of cum-ack's to determine
6344                  * this reliably.
6345                  */
6346                 if (bbr->rc_ack_is_cumack) {
6347                         if (bbr->rc_ts_data_set) {
6348                                 /* Lets attempt to determine the timestamp granularity. */
6349                                 bbr_make_timestamp_determination(bbr);
6350                         } else {
6351                                 bbr->rc_ts_data_set = 1;
6352                                 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts;
6353                                 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time;
6354                         }
6355                 } else {
6356                         /*
6357                          * We have to have consecutive acks
6358                          * reset any "filled" state to none.
6359                          */
6360                         bbr->rc_ts_data_set = 0;
6361                 }
6362         }
6363         /* Round it up */
6364         rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1)));
6365         if (rtt_ticks == 0)
6366                 rtt_ticks = 1;
6367         if (tp->t_srtt != 0) {
6368                 /*
6369                  * srtt is stored as fixed point with 5 bits after the
6370                  * binary point (i.e., scaled by 8).  The following magic is
6371                  * equivalent to the smoothing algorithm in rfc793 with an
6372                  * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
6373                  * Adjust rtt to origin 0.
6374                  */
6375
6376                 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT)
6377                     - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
6378
6379                 tp->t_srtt += delta;
6380                 if (tp->t_srtt <= 0)
6381                         tp->t_srtt = 1;
6382
6383                 /*
6384                  * We accumulate a smoothed rtt variance (actually, a
6385                  * smoothed mean difference), then set the retransmit timer
6386                  * to smoothed rtt + 4 times the smoothed variance. rttvar
6387                  * is stored as fixed point with 4 bits after the binary
6388                  * point (scaled by 16).  The following is equivalent to
6389                  * rfc793 smoothing with an alpha of .75 (rttvar =
6390                  * rttvar*3/4 + |delta| / 4).  This replaces rfc793's
6391                  * wired-in beta.
6392                  */
6393                 if (delta < 0)
6394                         delta = -delta;
6395                 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
6396                 tp->t_rttvar += delta;
6397                 if (tp->t_rttvar <= 0)
6398                         tp->t_rttvar = 1;
6399                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
6400                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6401         } else {
6402                 /*
6403                  * No rtt measurement yet - use the unsmoothed rtt. Set the
6404                  * variance to half the rtt (so our first retransmit happens
6405                  * at 3*rtt).
6406                  */
6407                 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT;
6408                 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1);
6409                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6410         }
6411         KMOD_TCPSTAT_INC(tcps_rttupdated);
6412         tp->t_rttupdated++;
6413 #ifdef STATS
6414         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks));
6415 #endif
6416         /*
6417          * the retransmit should happen at rtt + 4 * rttvar. Because of the
6418          * way we do the smoothing, srtt and rttvar will each average +1/2
6419          * tick of bias.  When we compute the retransmit timer, we want 1/2
6420          * tick of rounding and 1 extra tick because of +-1/2 tick
6421          * uncertainty in the firing of the timer.  The bias will give us
6422          * exactly the 1.5 tick we need.  But, because the bias is
6423          * statistical, we have to test that we don't drop below the minimum
6424          * feasible timer (which is 2 ticks).
6425          */
6426         TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
6427             max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2),
6428             MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
6429
6430         /*
6431          * We received an ack for a packet that wasn't retransmitted; it is
6432          * probably safe to discard any error indications we've received
6433          * recently.  This isn't quite right, but close enough for now (a
6434          * route might have failed after we sent a segment, and the return
6435          * path might not be symmetrical).
6436          */
6437         tp->t_softerror = 0;
6438         rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
6439         if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt)
6440                 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt;
6441 }
6442
6443 static void
6444 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line)
6445 {
6446         bbr->r_ctl.rc_rtt_shrinks = cts;
6447         if (bbr_can_force_probertt &&
6448             (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
6449             ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
6450                 /*
6451                  * We should enter probe-rtt its been too long
6452                  * since we have been there.
6453                  */
6454                 bbr_enter_probe_rtt(bbr, cts, __LINE__);
6455         } else
6456                 bbr_check_probe_rtt_limits(bbr, cts);
6457 }
6458
6459 static void
6460 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts)
6461 {
6462         uint64_t orig_bw;
6463
6464         if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) {
6465                 /* We never apply a zero measurment */
6466                 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0,
6467                                     0, 0, 0, 0, 0, 0);
6468                 return;
6469         }
6470         if (bbr->r_ctl.r_measurement_count < 0xffffffff)
6471                 bbr->r_ctl.r_measurement_count++;
6472         orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
6473         apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch);
6474         bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw,
6475                             (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate),
6476                             0, 0, 0, 0, 0, 0);
6477         if (orig_bw &&
6478             (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) {
6479                 if (bbr->bbr_hdrw_pacing) {
6480                         /*
6481                          * Apply a new rate to the hardware
6482                          * possibly.
6483                          */
6484                         bbr_update_hardware_pacing_rate(bbr, cts);
6485                 }
6486                 bbr_set_state_target(bbr, __LINE__);
6487                 tcp_bbr_tso_size_check(bbr, cts);
6488                 if (bbr->r_recovery_bw)  {
6489                         bbr_setup_red_bw(bbr, cts);
6490                         bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW);
6491                 }
6492         } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate))
6493                 tcp_bbr_tso_size_check(bbr, cts);
6494 }
6495
6496 static void
6497 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6498 {
6499         if (bbr->rc_in_persist == 0) {
6500                 /* We log only when not in persist */
6501                 /* Translate to a Bytes Per Second */
6502                 uint64_t tim, bw, ts_diff, ts_bw;
6503                 uint32_t upper, lower, delivered;
6504
6505                 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6506                         tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6507                 else
6508                         tim = 1;
6509                 /*
6510                  * Now that we have processed the tim (skipping the sample
6511                  * or possibly updating the time, go ahead and
6512                  * calculate the cdr.
6513                  */
6514                 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6515                 bw = (uint64_t)delivered;
6516                 bw *= (uint64_t)USECS_IN_SECOND;
6517                 bw /= tim;
6518                 if (bw == 0) {
6519                         /* We must have a calculatable amount */
6520                         return;
6521                 }
6522                 upper = (bw >> 32) & 0x00000000ffffffff;
6523                 lower = bw & 0x00000000ffffffff;
6524                 /*
6525                  * If we are using this b/w shove it in now so we
6526                  * can see in the trace viewer if it gets over-ridden.
6527                  */
6528                 if (rsm->r_ts_valid &&
6529                     bbr->rc_ts_valid &&
6530                     bbr->rc_ts_clock_set &&
6531                     (bbr->rc_ts_cant_be_used == 0) &&
6532                     bbr->rc_use_ts_limit) {
6533                         ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1);
6534                         ts_diff *= bbr->r_ctl.bbr_peer_tsratio;
6535                         if ((delivered == 0) ||
6536                             (rtt < 1000)) {
6537                                 /* Can't use the ts */
6538                                 bbr_log_type_bbrupd(bbr, 61, cts,
6539                                                     ts_diff,
6540                                                     bbr->r_ctl.last_inbound_ts,
6541                                                     rsm->r_del_ack_ts, 0,
6542                                                     0, 0, 0, delivered);
6543                         } else {
6544                                 ts_bw = (uint64_t)delivered;
6545                                 ts_bw *= (uint64_t)USECS_IN_SECOND;
6546                                 ts_bw /= ts_diff;
6547                                 bbr_log_type_bbrupd(bbr, 62, cts,
6548                                                     (ts_bw >> 32),
6549                                                     (ts_bw & 0xffffffff), 0, 0,
6550                                                     0, 0, ts_diff, delivered);
6551                                 if ((bbr->ts_can_raise) &&
6552                                     (ts_bw > bw)) {
6553                                         bbr_log_type_bbrupd(bbr, 8, cts,
6554                                                             delivered,
6555                                                             ts_diff,
6556                                                             (bw >> 32),
6557                                                             (bw & 0x00000000ffffffff),
6558                                                             0, 0, 0, 0);
6559                                         bw = ts_bw;
6560                                 } else if (ts_bw && (ts_bw < bw)) {
6561                                         bbr_log_type_bbrupd(bbr, 7, cts,
6562                                                             delivered,
6563                                                             ts_diff,
6564                                                             (bw >> 32),
6565                                                             (bw & 0x00000000ffffffff),
6566                                                             0, 0, 0, 0);
6567                                         bw = ts_bw;
6568                                 }
6569                         }
6570                 }
6571                 if (rsm->r_first_sent_time &&
6572                     TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6573                         uint64_t sbw, sti;
6574                         /*
6575                          * We use what was in flight at the time of our
6576                          * send  and the size of this send to figure
6577                          * out what we have been sending at (amount).
6578                          * For the time we take from the time of
6579                          * the send of the first send outstanding
6580                          * until this send plus this sends pacing
6581                          * time. This gives us a good calculation
6582                          * as to the rate we have been sending at.
6583                          */
6584
6585                         sbw = (uint64_t)(rsm->r_flight_at_send);
6586                         sbw *= (uint64_t)USECS_IN_SECOND;
6587                         sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6588                         sti += rsm->r_pacing_delay;
6589                         sbw /= sti;
6590                         if (sbw < bw) {
6591                                 bbr_log_type_bbrupd(bbr, 6, cts,
6592                                                     delivered,
6593                                                     (uint32_t)sti,
6594                                                     (bw >> 32),
6595                                                     (uint32_t)bw,
6596                                                     rsm->r_first_sent_time, 0, (sbw >> 32),
6597                                                     (uint32_t)sbw);
6598                                 bw = sbw;
6599                         }
6600                 }
6601                 /* Use the google algorithm for b/w measurements */
6602                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6603                 if ((rsm->r_app_limited == 0) ||
6604                     (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) {
6605                         tcp_bbr_commit_bw(bbr, cts);
6606                         bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6607                                             0, 0, 0, 0,  bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6608                 }
6609         }
6610 }
6611
6612 static void
6613 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6614 {
6615         if (bbr->rc_in_persist == 0) {
6616                 /* We log only when not in persist */
6617                 /* Translate to a Bytes Per Second */
6618                 uint64_t tim, bw;
6619                 uint32_t upper, lower, delivered;
6620                 int no_apply = 0;
6621
6622                 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6623                         tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6624                 else
6625                         tim = 1;
6626                 /*
6627                  * Now that we have processed the tim (skipping the sample
6628                  * or possibly updating the time, go ahead and
6629                  * calculate the cdr.
6630                  */
6631                 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6632                 bw = (uint64_t)delivered;
6633                 bw *= (uint64_t)USECS_IN_SECOND;
6634                 bw /= tim;
6635                 if (tim < bbr->r_ctl.rc_lowest_rtt) {
6636                         bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6637                                             tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6638
6639                         no_apply = 1;
6640                 }
6641                 upper = (bw >> 32) & 0x00000000ffffffff;
6642                 lower = bw & 0x00000000ffffffff;
6643                 /*
6644                  * If we are using this b/w shove it in now so we
6645                  * can see in the trace viewer if it gets over-ridden.
6646                  */
6647                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6648                 /* Gate by the sending rate */
6649                 if (rsm->r_first_sent_time &&
6650                     TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6651                         uint64_t sbw, sti;
6652                         /*
6653                          * We use what was in flight at the time of our
6654                          * send  and the size of this send to figure
6655                          * out what we have been sending at (amount).
6656                          * For the time we take from the time of
6657                          * the send of the first send outstanding
6658                          * until this send plus this sends pacing
6659                          * time. This gives us a good calculation
6660                          * as to the rate we have been sending at.
6661                          */
6662
6663                         sbw = (uint64_t)(rsm->r_flight_at_send);
6664                         sbw *= (uint64_t)USECS_IN_SECOND;
6665                         sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6666                         sti += rsm->r_pacing_delay;
6667                         sbw /= sti;
6668                         if (sbw < bw) {
6669                                 bbr_log_type_bbrupd(bbr, 6, cts,
6670                                                     delivered,
6671                                                     (uint32_t)sti,
6672                                                     (bw >> 32),
6673                                                     (uint32_t)bw,
6674                                                     rsm->r_first_sent_time, 0, (sbw >> 32),
6675                                                     (uint32_t)sbw);
6676                                 bw = sbw;
6677                         }
6678                         if ((sti > tim) &&
6679                             (sti < bbr->r_ctl.rc_lowest_rtt)) {
6680                                 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6681                                                     (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6682                                 no_apply = 1;
6683                         } else
6684                                 no_apply = 0;
6685                 }
6686                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6687                 if ((no_apply == 0) &&
6688                     ((rsm->r_app_limited == 0) ||
6689                      (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) {
6690                         tcp_bbr_commit_bw(bbr, cts);
6691                         bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6692                                             0, 0, 0, 0, bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6693                 }
6694         }
6695 }
6696
6697 static void
6698 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin,
6699     uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to)
6700 {
6701         uint64_t old_rttprop;
6702
6703         /* Update our delivery time and amount */
6704         bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start);
6705         bbr->r_ctl.rc_del_time = cts;
6706         if (rtt == 0) {
6707                 /*
6708                  * 0 means its a retransmit, for now we don't use these for
6709                  * the rest of BBR.
6710                  */
6711                 return;
6712         }
6713         if ((bbr->rc_use_google == 0) &&
6714             (match != BBR_RTT_BY_EXACTMATCH) &&
6715             (match != BBR_RTT_BY_TIMESTAMP)){
6716                 /*
6717                  * We get a lot of rtt updates, lets not pay attention to
6718                  * any that are not an exact match. That way we don't have
6719                  * to worry about timestamps and the whole nonsense of
6720                  * unsure if its a retransmission etc (if we ever had the
6721                  * timestamp fixed to always have the last thing sent this
6722                  * would not be a issue).
6723                  */
6724                 return;
6725         }
6726         if ((bbr_no_retran && bbr->rc_use_google) &&
6727             (match != BBR_RTT_BY_EXACTMATCH) &&
6728             (match != BBR_RTT_BY_TIMESTAMP)){
6729                 /*
6730                  * We only do measurements in google mode
6731                  * with bbr_no_retran on for sure things.
6732                  */
6733                 return;
6734         }
6735         /* Only update srtt if we know by exact match */
6736         tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin);
6737         if (ack_type == BBR_CUM_ACKED)
6738                 bbr->rc_ack_is_cumack = 1;
6739         else
6740                 bbr->rc_ack_is_cumack = 0;
6741         old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP);
6742         /*
6743          * Note the following code differs to the original
6744          * BBR spec. It calls for <= not <. However after a
6745          * long discussion in email with Neal, he acknowledged
6746          * that it should be < than so that we will have flows
6747          * going into probe-rtt (we were seeing cases where that
6748          * did not happen and caused ugly things to occur). We
6749          * have added this agreed upon fix to our code base.
6750          */
6751         if (rtt < old_rttprop) {
6752                 /* Update when we last saw a rtt drop */
6753                 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0);
6754                 bbr_set_reduced_rtt(bbr, cts, __LINE__);
6755         }
6756         bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts,
6757             match, rsm->r_start, rsm->r_flags);
6758         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6759         if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) {
6760                 /*
6761                  * The RTT-prop moved, reset the target (may be a
6762                  * nop for some states).
6763                  */
6764                 bbr_set_state_target(bbr, __LINE__);
6765                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)
6766                         bbr_log_rtt_shrinks(bbr, cts, 0, 0,
6767                                             __LINE__, BBR_RTTS_NEW_TARGET, 0);
6768                 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP))
6769                         /* It went up */
6770                         bbr_check_probe_rtt_limits(bbr, cts);
6771         }
6772         if ((bbr->rc_use_google == 0) &&
6773             (match == BBR_RTT_BY_TIMESTAMP)) {
6774                 /*
6775                  * We don't do b/w update with
6776                  * these since they are not really
6777                  * reliable.
6778                  */
6779                 return;
6780         }
6781         if (bbr->r_ctl.r_app_limited_until &&
6782             (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) {
6783                 /* We are no longer app-limited */
6784                 bbr->r_ctl.r_app_limited_until = 0;
6785         }
6786         if (bbr->rc_use_google) {
6787                 bbr_google_measurement(bbr, rsm, rtt, cts);
6788         } else {
6789                 bbr_nf_measurement(bbr, rsm, rtt, cts);
6790         }
6791 }
6792
6793 /*
6794  * Convert a timestamp that the main stack
6795  * uses (milliseconds) into one that bbr uses
6796  * (microseconds). Return that converted timestamp.
6797  */
6798 static uint32_t
6799 bbr_ts_convert(uint32_t cts) {
6800         uint32_t sec, msec;
6801
6802         sec = cts / MS_IN_USEC;
6803         msec = cts - (MS_IN_USEC * sec);
6804         return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC));
6805 }
6806
6807 /*
6808  * Return 0 if we did not update the RTT time, return
6809  * 1 if we did.
6810  */
6811 static int
6812 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr,
6813     struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack)
6814 {
6815         int32_t i;
6816         uint32_t t, uts = 0;
6817
6818         if ((rsm->r_flags & BBR_ACKED) ||
6819             (rsm->r_flags & BBR_WAS_RENEGED) ||
6820             (rsm->r_flags & BBR_RXT_CLEARED)) {
6821                 /* Already done */
6822                 return (0);
6823         }
6824         if (rsm->r_rtt_not_allowed) {
6825                 /* Not allowed */
6826                 return (0);
6827         }
6828         if (rsm->r_rtr_cnt == 1) {
6829                 /*
6830                  * Only one transmit. Hopefully the normal case.
6831                  */
6832                 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0]))
6833                         t = cts - rsm->r_tim_lastsent[0];
6834                 else
6835                         t = 1;
6836                 if ((int)t <= 0)
6837                         t = 1;
6838                 bbr->r_ctl.rc_last_rtt = t;
6839                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6840                                     BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to);
6841                 return (1);
6842         }
6843         /* Convert to usecs */
6844         if ((bbr_can_use_ts_for_rtt == 1) &&
6845             (bbr->rc_use_google == 1) &&
6846             (ack_type == BBR_CUM_ACKED) &&
6847             (to->to_flags & TOF_TS) &&
6848             (to->to_tsecr != 0)) {
6849                 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr;
6850                 if (t < 1)
6851                         t = 1;
6852                 t *= MS_IN_USEC;
6853                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6854                                     BBR_RTT_BY_TIMESTAMP,
6855                                     rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)],
6856                                     ack_type, to);
6857                 return (1);
6858         }
6859         uts = bbr_ts_convert(to->to_tsecr);
6860         if ((to->to_flags & TOF_TS) &&
6861             (to->to_tsecr != 0) &&
6862             (ack_type == BBR_CUM_ACKED) &&
6863             ((rsm->r_flags & BBR_OVERMAX) == 0)) {
6864                 /*
6865                  * Now which timestamp does it match? In this block the ACK
6866                  * may be coming from a previous transmission.
6867                  */
6868                 uint32_t fudge;
6869
6870                 fudge = BBR_TIMER_FUDGE;
6871                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
6872                         if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) &&
6873                             (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) {
6874                                 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6875                                         t = cts - rsm->r_tim_lastsent[i];
6876                                 else
6877                                         t = 1;
6878                                 if ((int)t <= 0)
6879                                         t = 1;
6880                                 bbr->r_ctl.rc_last_rtt = t;
6881                                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING,
6882                                                     rsm->r_tim_lastsent[i], ack_type, to);
6883                                 if ((i + 1) < rsm->r_rtr_cnt) {
6884                                         /* Likely */
6885                                         return (0);
6886                                 } else if (rsm->r_flags & BBR_TLP) {
6887                                         bbr->rc_tlp_rtx_out = 0;
6888                                 }
6889                                 return (1);
6890                         }
6891                 }
6892                 /* Fall through if we can't find a matching timestamp */
6893         }
6894         /*
6895          * Ok its a SACK block that we retransmitted. or a windows
6896          * machine without timestamps. We can tell nothing from the
6897          * time-stamp since its not there or the time the peer last
6898          * recieved a segment that moved forward its cum-ack point.
6899          *
6900          * Lets look at the last retransmit and see what we can tell
6901          * (with BBR for space we only keep 2 note we have to keep
6902          * at least 2 so the map can not be condensed more).
6903          */
6904         i = rsm->r_rtr_cnt - 1;
6905         if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6906                 t = cts - rsm->r_tim_lastsent[i];
6907         else
6908                 goto not_sure;
6909         if (t < bbr->r_ctl.rc_lowest_rtt) {
6910                 /*
6911                  * We retransmitted and the ack came back in less
6912                  * than the smallest rtt we have observed in the
6913                  * windowed rtt. We most likey did an improper
6914                  * retransmit as outlined in 4.2 Step 3 point 2 in
6915                  * the rack-draft.
6916                  *
6917                  * Use the prior transmission to update all the
6918                  * information as long as there is only one prior
6919                  * transmission.
6920                  */
6921                 if ((rsm->r_flags & BBR_OVERMAX) == 0) {
6922 #ifdef BBR_INVARIANTS
6923                         if (rsm->r_rtr_cnt == 1)
6924                                 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags);
6925 #endif
6926                         i = rsm->r_rtr_cnt - 2;
6927                         if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6928                                 t = cts - rsm->r_tim_lastsent[i];
6929                         else
6930                                 t = 1;
6931                         bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET,
6932                                             rsm->r_tim_lastsent[i], ack_type, to);
6933                         return (0);
6934                 } else {
6935                         /*
6936                          * Too many prior transmissions, just
6937                          * updated BBR delivered
6938                          */
6939 not_sure:
6940                         bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
6941                                             BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
6942                 }
6943         } else {
6944                 /*
6945                  * We retransmitted it and the retransmit did the
6946                  * job.
6947                  */
6948                 if (rsm->r_flags & BBR_TLP)
6949                         bbr->rc_tlp_rtx_out = 0;
6950                 if ((rsm->r_flags & BBR_OVERMAX) == 0)
6951                         bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts,
6952                                             BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to);
6953                 else
6954                         bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
6955                                             BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
6956                 return (1);
6957         }
6958         return (0);
6959 }
6960
6961 /*
6962  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
6963  */
6964 static void
6965 bbr_log_sack_passed(struct tcpcb *tp,
6966     struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
6967 {
6968         struct bbr_sendmap *nrsm;
6969
6970         nrsm = rsm;
6971         TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap,
6972             bbr_head, r_tnext) {
6973                 if (nrsm == rsm) {
6974                         /* Skip orginal segment he is acked */
6975                         continue;
6976                 }
6977                 if (nrsm->r_flags & BBR_ACKED) {
6978                         /* Skip ack'd segments */
6979                         continue;
6980                 }
6981                 if (nrsm->r_flags & BBR_SACK_PASSED) {
6982                         /*
6983                          * We found one that is already marked
6984                          * passed, we have been here before and
6985                          * so all others below this are marked.
6986                          */
6987                         break;
6988                 }
6989                 BBR_STAT_INC(bbr_sack_passed);
6990                 nrsm->r_flags |= BBR_SACK_PASSED;
6991                 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) &&
6992                     bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) {
6993                         bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start;
6994                         bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start;
6995                         nrsm->r_flags |= BBR_MARKED_LOST;
6996                 }
6997                 nrsm->r_flags &= ~BBR_WAS_SACKPASS;
6998         }
6999 }
7000
7001 /*
7002  * Returns the number of bytes that were
7003  * newly ack'd by sack blocks.
7004  */
7005 static uint32_t
7006 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack,
7007     struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts)
7008 {
7009         int32_t times = 0;
7010         uint32_t start, end, maxseg, changed = 0;
7011         struct bbr_sendmap *rsm, *nrsm;
7012         int32_t used_ref = 1;
7013         uint8_t went_back = 0, went_fwd = 0;
7014
7015         maxseg = tp->t_maxseg - bbr->rc_last_options;
7016         start = sack->start;
7017         end = sack->end;
7018         rsm = *prsm;
7019         if (rsm == NULL)
7020                 used_ref = 0;
7021
7022         /* Do we locate the block behind where we last were? */
7023         if (rsm && SEQ_LT(start, rsm->r_start)) {
7024                 went_back = 1;
7025                 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
7026                         if (SEQ_GEQ(start, rsm->r_start) &&
7027                             SEQ_LT(start, rsm->r_end)) {
7028                                 goto do_rest_ofb;
7029                         }
7030                 }
7031         }
7032 start_at_beginning:
7033         went_fwd = 1;
7034         /*
7035          * Ok lets locate the block where this guy is fwd from rsm (if its
7036          * set)
7037          */
7038         TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) {
7039                 if (SEQ_GEQ(start, rsm->r_start) &&
7040                     SEQ_LT(start, rsm->r_end)) {
7041                         break;
7042                 }
7043         }
7044 do_rest_ofb:
7045         if (rsm == NULL) {
7046                 /*
7047                  * This happens when we get duplicate sack blocks with the
7048                  * same end. For example SACK 4: 100 SACK 3: 100 The sort
7049                  * will not change there location so we would just start at
7050                  * the end of the first one and get lost.
7051                  */
7052                 if (tp->t_flags & TF_SENTFIN) {
7053                         /*
7054                          * Check to see if we have not logged the FIN that
7055                          * went out.
7056                          */
7057                         nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7058                         if (nrsm && (nrsm->r_end + 1) == tp->snd_max) {
7059                                 /*
7060                                  * Ok we did not get the FIN logged.
7061                                  */
7062                                 nrsm->r_end++;
7063                                 rsm = nrsm;
7064                                 goto do_rest_ofb;
7065                         }
7066                 }
7067                 if (times == 1) {
7068 #ifdef BBR_INVARIANTS
7069                         panic("tp:%p bbr:%p sack:%p to:%p prsm:%p",
7070                             tp, bbr, sack, to, prsm);
7071 #else
7072                         goto out;
7073 #endif
7074                 }
7075                 times++;
7076                 BBR_STAT_INC(bbr_sack_proc_restart);
7077                 rsm = NULL;
7078                 goto start_at_beginning;
7079         }
7080         /* Ok we have an ACK for some piece of rsm */
7081         if (rsm->r_start != start) {
7082                 /*
7083                  * Need to split this in two pieces the before and after.
7084                  */
7085                 if (bbr_sack_mergable(rsm, start, end))
7086                         nrsm = bbr_alloc_full_limit(bbr);
7087                 else
7088                         nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7089                 if (nrsm == NULL) {
7090                         /* We could not allocate ignore the sack */
7091                         struct sackblk blk;
7092
7093                         blk.start = start;
7094                         blk.end = end;
7095                         sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7096                         goto out;
7097                 }
7098                 bbr_clone_rsm(bbr, nrsm, rsm, start);
7099                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7100                 if (rsm->r_in_tmap) {
7101                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7102                         nrsm->r_in_tmap = 1;
7103                 }
7104                 rsm->r_flags &= (~BBR_HAS_FIN);
7105                 rsm = nrsm;
7106         }
7107         if (SEQ_GEQ(end, rsm->r_end)) {
7108                 /*
7109                  * The end of this block is either beyond this guy or right
7110                  * at this guy.
7111                  */
7112                 if ((rsm->r_flags & BBR_ACKED) == 0) {
7113                         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7114                         changed += (rsm->r_end - rsm->r_start);
7115                         bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7116                         bbr_log_sack_passed(tp, bbr, rsm);
7117                         if (rsm->r_flags & BBR_MARKED_LOST) {
7118                                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7119                         }
7120                         /* Is Reordering occuring? */
7121                         if (rsm->r_flags & BBR_SACK_PASSED) {
7122                                 BBR_STAT_INC(bbr_reorder_seen);
7123                                 bbr->r_ctl.rc_reorder_ts = cts;
7124                                 if (rsm->r_flags & BBR_MARKED_LOST) {
7125                                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7126                                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7127                                                 /* LT sampling also needs adjustment */
7128                                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7129                                 }
7130                         }
7131                         rsm->r_flags |= BBR_ACKED;
7132                         rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7133                         if (rsm->r_in_tmap) {
7134                                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7135                                 rsm->r_in_tmap = 0;
7136                         }
7137                 }
7138                 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7139                 if (end == rsm->r_end) {
7140                         /* This block only - done */
7141                         goto out;
7142                 }
7143                 /* There is more not coverend by this rsm move on */
7144                 start = rsm->r_end;
7145                 nrsm = TAILQ_NEXT(rsm, r_next);
7146                 rsm = nrsm;
7147                 times = 0;
7148                 goto do_rest_ofb;
7149         }
7150         if (rsm->r_flags & BBR_ACKED) {
7151                 /* Been here done that */
7152                 goto out;
7153         }
7154         /* Ok we need to split off this one at the tail */
7155         if (bbr_sack_mergable(rsm, start, end))
7156                 nrsm = bbr_alloc_full_limit(bbr);
7157         else
7158                 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7159         if (nrsm == NULL) {
7160                 /* failed XXXrrs what can we do but loose the sack info? */
7161                 struct sackblk blk;
7162
7163                 blk.start = start;
7164                 blk.end = end;
7165                 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7166                 goto out;
7167         }
7168         /* Clone it */
7169         bbr_clone_rsm(bbr, nrsm, rsm, end);
7170         /* The sack block does not cover this guy fully */
7171         rsm->r_flags &= (~BBR_HAS_FIN);
7172         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7173         if (rsm->r_in_tmap) {
7174                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7175                 nrsm->r_in_tmap = 1;
7176         }
7177         nrsm->r_dupack = 0;
7178         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7179         bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7180         changed += (rsm->r_end - rsm->r_start);
7181         bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7182         bbr_log_sack_passed(tp, bbr, rsm);
7183         /* Is Reordering occuring? */
7184         if (rsm->r_flags & BBR_MARKED_LOST) {
7185                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7186         }
7187         if (rsm->r_flags & BBR_SACK_PASSED) {
7188                 BBR_STAT_INC(bbr_reorder_seen);
7189                 bbr->r_ctl.rc_reorder_ts = cts;
7190                 if (rsm->r_flags & BBR_MARKED_LOST) {
7191                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7192                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7193                                 /* LT sampling also needs adjustment */
7194                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7195                 }
7196         }
7197         rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7198         rsm->r_flags |= BBR_ACKED;
7199         if (rsm->r_in_tmap) {
7200                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7201                 rsm->r_in_tmap = 0;
7202         }
7203 out:
7204         if (rsm && (rsm->r_flags & BBR_ACKED)) {
7205                 /*
7206                  * Now can we merge this newly acked
7207                  * block with either the previous or
7208                  * next block?
7209                  */
7210                 nrsm = TAILQ_NEXT(rsm, r_next);
7211                 if (nrsm &&
7212                     (nrsm->r_flags & BBR_ACKED)) {
7213                         /* yep this and next can be merged */
7214                         rsm = bbr_merge_rsm(bbr, rsm, nrsm);
7215                 }
7216                 /* Now what about the previous? */
7217                 nrsm = TAILQ_PREV(rsm, bbr_head, r_next);
7218                 if (nrsm &&
7219                     (nrsm->r_flags & BBR_ACKED)) {
7220                         /* yep the previous and this can be merged */
7221                         rsm = bbr_merge_rsm(bbr, nrsm, rsm);
7222                 }
7223         }
7224         if (used_ref == 0) {
7225                 BBR_STAT_INC(bbr_sack_proc_all);
7226         } else {
7227                 BBR_STAT_INC(bbr_sack_proc_short);
7228         }
7229         if (went_fwd && went_back) {
7230                 BBR_STAT_INC(bbr_sack_search_both);
7231         } else if (went_fwd) {
7232                 BBR_STAT_INC(bbr_sack_search_fwd);
7233         } else if (went_back) {
7234                 BBR_STAT_INC(bbr_sack_search_back);
7235         }
7236         /* Save off where the next seq is */
7237         if (rsm)
7238                 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next);
7239         else
7240                 bbr->r_ctl.rc_sacklast = NULL;
7241         *prsm = rsm;
7242         return (changed);
7243 }
7244
7245 static void inline
7246 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack)
7247 {
7248         struct bbr_sendmap *tmap;
7249
7250         BBR_STAT_INC(bbr_reneges_seen);
7251         tmap = NULL;
7252         while (rsm && (rsm->r_flags & BBR_ACKED)) {
7253                 /* Its no longer sacked, mark it so */
7254                 uint32_t oflags;
7255                 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7256 #ifdef BBR_INVARIANTS
7257                 if (rsm->r_in_tmap) {
7258                         panic("bbr:%p rsm:%p flags:0x%x in tmap?",
7259                             bbr, rsm, rsm->r_flags);
7260                 }
7261 #endif
7262                 oflags = rsm->r_flags;
7263                 if (rsm->r_flags & BBR_MARKED_LOST) {
7264                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7265                         bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7266                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7267                                 /* LT sampling also needs adjustment */
7268                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7269                 }
7270                 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST);
7271                 rsm->r_flags |= BBR_WAS_RENEGED;
7272                 rsm->r_flags |= BBR_RXT_CLEARED;
7273                 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__);
7274                 /* Rebuild it into our tmap */
7275                 if (tmap == NULL) {
7276                         TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7277                         tmap = rsm;
7278                 } else {
7279                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext);
7280                         tmap = rsm;
7281                 }
7282                 tmap->r_in_tmap = 1;
7283                 /*
7284                  * XXXrrs Delivered? Should we do anything here?
7285                  *
7286                  * Of course we don't on a rxt timeout so maybe its ok that
7287                  * we don't?
7288                  *
7289                  * For now lets not.
7290                  */
7291                 rsm = TAILQ_NEXT(rsm, r_next);
7292         }
7293         /*
7294          * Now lets possibly clear the sack filter so we start recognizing
7295          * sacks that cover this area.
7296          */
7297         sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack);
7298 }
7299
7300 static void
7301 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to)
7302 {
7303         struct tcp_bbr *bbr;
7304         struct bbr_sendmap *rsm;
7305         uint32_t cts;
7306
7307         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7308         cts = bbr->r_ctl.rc_rcvtime;
7309         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7310         if (rsm && (rsm->r_flags & BBR_HAS_SYN)) {
7311                 if ((rsm->r_end - rsm->r_start) <= 1) {
7312                         /* Log out the SYN completely */
7313                         bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7314                         rsm->r_rtr_bytes = 0;
7315                         TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7316                         if (rsm->r_in_tmap) {
7317                                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7318                                 rsm->r_in_tmap = 0;
7319                         }
7320                         if (bbr->r_ctl.rc_next == rsm) {
7321                                 /* scoot along the marker */
7322                                 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7323                         }
7324                         if (to != NULL)
7325                                 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0);
7326                         bbr_free(bbr, rsm);
7327                 } else {
7328                         /* There is more (Fast open)? strip out SYN. */
7329                         rsm->r_flags &= ~BBR_HAS_SYN;
7330                         rsm->r_start++;
7331                 }
7332         }
7333 }
7334
7335 /*
7336  * Returns the number of bytes that were
7337  * acknowledged by SACK blocks.
7338  */
7339
7340 static uint32_t
7341 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th,
7342     uint32_t *prev_acked)
7343 {
7344         uint32_t changed, last_seq, entered_recovery = 0;
7345         struct tcp_bbr *bbr;
7346         struct bbr_sendmap *rsm;
7347         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
7348         register uint32_t th_ack;
7349         int32_t i, j, k, new_sb, num_sack_blks = 0;
7350         uint32_t cts, acked, ack_point, sack_changed = 0;
7351         uint32_t p_maxseg, maxseg, p_acked = 0;
7352
7353         INP_WLOCK_ASSERT(tp->t_inpcb);
7354         if (th->th_flags & TH_RST) {
7355                 /* We don't log resets */
7356                 return (0);
7357         }
7358         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7359         cts = bbr->r_ctl.rc_rcvtime;
7360
7361         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7362         changed = 0;
7363         maxseg = tp->t_maxseg - bbr->rc_last_options;
7364         p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg);
7365         th_ack = th->th_ack;
7366         if (SEQ_GT(th_ack, tp->snd_una)) {
7367                 acked = th_ack - tp->snd_una;
7368                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__);
7369                 bbr->rc_tp->t_acktime = ticks;
7370         } else
7371                 acked = 0;
7372         if (SEQ_LEQ(th_ack, tp->snd_una)) {
7373                 /* Only sent here for sack processing */
7374                 goto proc_sack;
7375         }
7376         if (rsm && SEQ_GT(th_ack, rsm->r_start)) {
7377                 changed = th_ack - rsm->r_start;
7378         } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) {
7379                 /*
7380                  * For the SYN incoming case we will not have called
7381                  * tcp_output for the sending of the SYN, so there will be
7382                  * no map. All other cases should probably be a panic.
7383                  */
7384                 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) {
7385                         /*
7386                          * We have a timestamp that can be used to generate
7387                          * an initial RTT.
7388                          */
7389                         uint32_t ts, now, rtt;
7390
7391                         ts = bbr_ts_convert(to->to_tsecr);
7392                         now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv));
7393                         rtt = now - ts;
7394                         if (rtt < 1)
7395                                 rtt = 1;
7396                         bbr_log_type_bbrrttprop(bbr, rtt,
7397                                                 tp->iss, 0, cts,
7398                                                 BBR_RTT_BY_TIMESTAMP, tp->iss, 0);
7399                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
7400                         changed = 1;
7401                         bbr->r_wanted_output = 1;
7402                         goto out;
7403                 }
7404                 goto proc_sack;
7405         } else if (rsm == NULL) {
7406                 goto out;
7407         }
7408         if (changed) {
7409                 /*
7410                  * The ACK point is advancing to th_ack, we must drop off
7411                  * the packets in the rack log and calculate any eligble
7412                  * RTT's.
7413                  */
7414                 bbr->r_wanted_output = 1;
7415 more:
7416                 if (rsm == NULL) {
7417                         if (tp->t_flags & TF_SENTFIN) {
7418                                 /* if we send a FIN we will not hav a map */
7419                                 goto proc_sack;
7420                         }
7421 #ifdef BBR_INVARIANTS
7422                         panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n",
7423                             tp,
7424                             th, tp->t_state, bbr,
7425                             tp->snd_una, tp->snd_max, changed);
7426 #endif
7427                         goto proc_sack;
7428                 }
7429         }
7430         if (SEQ_LT(th_ack, rsm->r_start)) {
7431                 /* Huh map is missing this */
7432 #ifdef BBR_INVARIANTS
7433                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n",
7434                     rsm->r_start,
7435                     th_ack, tp->t_state,
7436                     bbr->r_state, bbr);
7437                 panic("th-ack is bad bbr:%p tp:%p", bbr, tp);
7438 #endif
7439                 goto proc_sack;
7440         } else if (th_ack == rsm->r_start) {
7441                 /* None here to ack */
7442                 goto proc_sack;
7443         }
7444         /*
7445          * Clear the dup ack counter, it will
7446          * either be freed or if there is some
7447          * remaining we need to start it at zero.
7448          */
7449         rsm->r_dupack = 0;
7450         /* Now do we consume the whole thing? */
7451         if (SEQ_GEQ(th_ack, rsm->r_end)) {
7452                 /* Its all consumed. */
7453                 uint32_t left;
7454
7455                 if (rsm->r_flags & BBR_ACKED) {
7456                         /*
7457                          * It was acked on the scoreboard -- remove it from
7458                          * total
7459                          */
7460                         p_acked += (rsm->r_end - rsm->r_start);
7461                         bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7462                         if (bbr->r_ctl.rc_sacked == 0)
7463                                 bbr->r_ctl.rc_sacklast = NULL;
7464                 } else {
7465                         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack);
7466                         if (rsm->r_flags & BBR_MARKED_LOST) {
7467                                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7468                         }
7469                         if (rsm->r_flags & BBR_SACK_PASSED) {
7470                                 /*
7471                                  * There are acked segments ACKED on the
7472                                  * scoreboard further up. We are seeing
7473                                  * reordering.
7474                                  */
7475                                 BBR_STAT_INC(bbr_reorder_seen);
7476                                 bbr->r_ctl.rc_reorder_ts = cts;
7477                                 if (rsm->r_flags & BBR_MARKED_LOST) {
7478                                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7479                                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7480                                                 /* LT sampling also needs adjustment */
7481                                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7482                                 }
7483                         }
7484                         rsm->r_flags &= ~BBR_MARKED_LOST;
7485                 }
7486                 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7487                 rsm->r_rtr_bytes = 0;
7488                 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7489                 if (rsm->r_in_tmap) {
7490                         TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7491                         rsm->r_in_tmap = 0;
7492                 }
7493                 if (bbr->r_ctl.rc_next == rsm) {
7494                         /* scoot along the marker */
7495                         bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7496                 }
7497                 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7498                 /* Adjust the packet counts */
7499                 left = th_ack - rsm->r_end;
7500                 /* Free back to zone */
7501                 bbr_free(bbr, rsm);
7502                 if (left) {
7503                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7504                         goto more;
7505                 }
7506                 goto proc_sack;
7507         }
7508         if (rsm->r_flags & BBR_ACKED) {
7509                 /*
7510                  * It was acked on the scoreboard -- remove it from total
7511                  * for the part being cum-acked.
7512                  */
7513                 p_acked += (rsm->r_end - rsm->r_start);
7514                 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
7515                 if (bbr->r_ctl.rc_sacked == 0)
7516                         bbr->r_ctl.rc_sacklast = NULL;
7517         } else {
7518                 /*
7519                  * It was acked up to th_ack point for the first time
7520                  */
7521                 struct bbr_sendmap lrsm;
7522
7523                 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap));
7524                 lrsm.r_end = th_ack;
7525                 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack);
7526         }
7527         if ((rsm->r_flags & BBR_MARKED_LOST) &&
7528             ((rsm->r_flags & BBR_ACKED) == 0)) {
7529                 /*
7530                  * It was marked lost and partly ack'd now
7531                  * for the first time. We lower the rc_lost_bytes
7532                  * and still leave it MARKED.
7533                  */
7534                 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start;
7535         }
7536         bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7537         bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7538         rsm->r_rtr_bytes = 0;
7539         /* adjust packet count */
7540         rsm->r_start = th_ack;
7541 proc_sack:
7542         /* Check for reneging */
7543         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7544         if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) {
7545                 /*
7546                  * The peer has moved snd_una up to the edge of this send,
7547                  * i.e. one that it had previously acked. The only way that
7548                  * can be true if the peer threw away data (space issues)
7549                  * that it had previously sacked (else it would have given
7550                  * us snd_una up to (rsm->r_end). We need to undo the acked
7551                  * markings here.
7552                  *
7553                  * Note we have to look to make sure th_ack is our
7554                  * rsm->r_start in case we get an old ack where th_ack is
7555                  * behind snd_una.
7556                  */
7557                 bbr_peer_reneges(bbr, rsm, th->th_ack);
7558         }
7559         if ((to->to_flags & TOF_SACK) == 0) {
7560                 /* We are done nothing left to log */
7561                 goto out;
7562         }
7563         rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7564         if (rsm) {
7565                 last_seq = rsm->r_end;
7566         } else {
7567                 last_seq = tp->snd_max;
7568         }
7569         /* Sack block processing */
7570         if (SEQ_GT(th_ack, tp->snd_una))
7571                 ack_point = th_ack;
7572         else
7573                 ack_point = tp->snd_una;
7574         for (i = 0; i < to->to_nsacks; i++) {
7575                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
7576                     &sack, sizeof(sack));
7577                 sack.start = ntohl(sack.start);
7578                 sack.end = ntohl(sack.end);
7579                 if (SEQ_GT(sack.end, sack.start) &&
7580                     SEQ_GT(sack.start, ack_point) &&
7581                     SEQ_LT(sack.start, tp->snd_max) &&
7582                     SEQ_GT(sack.end, ack_point) &&
7583                     SEQ_LEQ(sack.end, tp->snd_max)) {
7584                         if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) &&
7585                             (SEQ_LT(sack.end, last_seq)) &&
7586                             ((sack.end - sack.start) < (p_maxseg / 8))) {
7587                                 /*
7588                                  * Not the last piece and its smaller than
7589                                  * 1/8th of a p_maxseg. We ignore this.
7590                                  */
7591                                 BBR_STAT_INC(bbr_runt_sacks);
7592                                 continue;
7593                         }
7594                         sack_blocks[num_sack_blks] = sack;
7595                         num_sack_blks++;
7596 #ifdef NETFLIX_STATS
7597                 } else if (SEQ_LEQ(sack.start, th_ack) &&
7598                     SEQ_LEQ(sack.end, th_ack)) {
7599                         /*
7600                          * Its a D-SACK block.
7601                          */
7602                         tcp_record_dsack(sack.start, sack.end);
7603 #endif
7604                 }
7605         }
7606         if (num_sack_blks == 0)
7607                 goto out;
7608         /*
7609          * Sort the SACK blocks so we can update the rack scoreboard with
7610          * just one pass.
7611          */
7612         new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks,
7613                                   num_sack_blks, th->th_ack);
7614         ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks);
7615         BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks);
7616         BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb));
7617         num_sack_blks = new_sb;
7618         if (num_sack_blks < 2) {
7619                 goto do_sack_work;
7620         }
7621         /* Sort the sacks */
7622         for (i = 0; i < num_sack_blks; i++) {
7623                 for (j = i + 1; j < num_sack_blks; j++) {
7624                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
7625                                 sack = sack_blocks[i];
7626                                 sack_blocks[i] = sack_blocks[j];
7627                                 sack_blocks[j] = sack;
7628                         }
7629                 }
7630         }
7631         /*
7632          * Now are any of the sack block ends the same (yes some
7633          * implememtations send these)?
7634          */
7635 again:
7636         if (num_sack_blks > 1) {
7637                 for (i = 0; i < num_sack_blks; i++) {
7638                         for (j = i + 1; j < num_sack_blks; j++) {
7639                                 if (sack_blocks[i].end == sack_blocks[j].end) {
7640                                         /*
7641                                          * Ok these two have the same end we
7642                                          * want the smallest end and then
7643                                          * throw away the larger and start
7644                                          * again.
7645                                          */
7646                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
7647                                                 /*
7648                                                  * The second block covers
7649                                                  * more area use that
7650                                                  */
7651                                                 sack_blocks[i].start = sack_blocks[j].start;
7652                                         }
7653                                         /*
7654                                          * Now collapse out the dup-sack and
7655                                          * lower the count
7656                                          */
7657                                         for (k = (j + 1); k < num_sack_blks; k++) {
7658                                                 sack_blocks[j].start = sack_blocks[k].start;
7659                                                 sack_blocks[j].end = sack_blocks[k].end;
7660                                                 j++;
7661                                         }
7662                                         num_sack_blks--;
7663                                         goto again;
7664                                 }
7665                         }
7666                 }
7667         }
7668 do_sack_work:
7669         rsm = bbr->r_ctl.rc_sacklast;
7670         for (i = 0; i < num_sack_blks; i++) {
7671                 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts);
7672                 if (acked) {
7673                         bbr->r_wanted_output = 1;
7674                         changed += acked;
7675                         sack_changed += acked;
7676                 }
7677         }
7678 out:
7679         *prev_acked = p_acked;
7680         if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) {
7681                 /*
7682                  * Ok we have a high probability that we need to go in to
7683                  * recovery since we have data sack'd
7684                  */
7685                 struct bbr_sendmap *rsm;
7686
7687                 rsm = bbr_check_recovery_mode(tp, bbr, cts);
7688                 if (rsm) {
7689                         /* Enter recovery */
7690                         entered_recovery = 1;
7691                         bbr->r_wanted_output = 1;
7692                         /*
7693                          * When we enter recovery we need to assure we send
7694                          * one packet.
7695                          */
7696                         if (bbr->r_ctl.rc_resend == NULL) {
7697                                 bbr->r_ctl.rc_resend = rsm;
7698                         }
7699                 }
7700         }
7701         if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) {
7702                 /*
7703                  * See if we need to rack-retransmit anything if so set it
7704                  * up as the thing to resend assuming something else is not
7705                  * already in that position.
7706                  */
7707                 if (bbr->r_ctl.rc_resend == NULL) {
7708                         bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
7709                 }
7710         }
7711         /*
7712          * We return the amount that changed via sack, this is used by the
7713          * ack-received code to augment what was changed between th_ack <->
7714          * snd_una.
7715          */
7716         return (sack_changed);
7717 }
7718
7719 static void
7720 bbr_strike_dupack(struct tcp_bbr *bbr)
7721 {
7722         struct bbr_sendmap *rsm;
7723
7724         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
7725         if (rsm && (rsm->r_dupack < 0xff)) {
7726                 rsm->r_dupack++;
7727                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD)
7728                         bbr->r_wanted_output = 1;
7729         }
7730 }
7731
7732 /*
7733  * Return value of 1, we do not need to call bbr_process_data().
7734  * return value of 0, bbr_process_data can be called.
7735  * For ret_val if its 0 the TCB is locked and valid, if its non-zero
7736  * its unlocked and probably unsafe to touch the TCB.
7737  */
7738 static int
7739 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
7740     struct tcpcb *tp, struct tcpopt *to,
7741     uint32_t tiwin, int32_t tlen,
7742     int32_t * ofia, int32_t thflags, int32_t * ret_val)
7743 {
7744         int32_t ourfinisacked = 0;
7745         int32_t acked_amount;
7746         uint16_t nsegs;
7747         int32_t acked;
7748         uint32_t lost, sack_changed = 0;
7749         struct mbuf *mfree;
7750         struct tcp_bbr *bbr;
7751         uint32_t prev_acked = 0;
7752
7753         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7754         lost = bbr->r_ctl.rc_lost;
7755         nsegs = max(1, m->m_pkthdr.lro_nsegs);
7756         if (SEQ_GT(th->th_ack, tp->snd_max)) {
7757                 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
7758                 bbr->r_wanted_output = 1;
7759                 return (1);
7760         }
7761         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
7762                 /* Process the ack */
7763                 if (bbr->rc_in_persist)
7764                         tp->t_rxtshift = 0;
7765                 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd))
7766                         bbr_strike_dupack(bbr);
7767                 sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
7768         }
7769         bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost));
7770         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
7771                 /*
7772                  * Old ack, behind the last one rcv'd or a duplicate ack
7773                  * with SACK info.
7774                  */
7775                 if (th->th_ack == tp->snd_una) {
7776                         bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0);
7777                         if (bbr->r_state == TCPS_SYN_SENT) {
7778                                 /*
7779                                  * Special case on where we sent SYN. When
7780                                  * the SYN-ACK is processed in syn_sent
7781                                  * state it bumps the snd_una. This causes
7782                                  * us to hit here even though we did ack 1
7783                                  * byte.
7784                                  *
7785                                  * Go through the nothing left case so we
7786                                  * send data.
7787                                  */
7788                                 goto nothing_left;
7789                         }
7790                 }
7791                 return (0);
7792         }
7793         /*
7794          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
7795          * something we sent.
7796          */
7797         if (tp->t_flags & TF_NEEDSYN) {
7798                 /*
7799                  * T/TCP: Connection was half-synchronized, and our SYN has
7800                  * been ACK'd (so connection is now fully synchronized).  Go
7801                  * to non-starred state, increment snd_una for ACK of SYN,
7802                  * and check if we can do window scaling.
7803                  */
7804                 tp->t_flags &= ~TF_NEEDSYN;
7805                 tp->snd_una++;
7806                 /* Do window scaling? */
7807                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
7808                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
7809                         tp->rcv_scale = tp->request_r_scale;
7810                         /* Send window already scaled. */
7811                 }
7812         }
7813         INP_WLOCK_ASSERT(tp->t_inpcb);
7814
7815         acked = BYTES_THIS_ACK(tp, th);
7816         KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
7817         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
7818
7819         /*
7820          * If we just performed our first retransmit, and the ACK arrives
7821          * within our recovery window, then it was a mistake to do the
7822          * retransmit in the first place.  Recover our original cwnd and
7823          * ssthresh, and proceed to transmit where we left off.
7824          */
7825         if (tp->t_flags & TF_PREVVALID) {
7826                 tp->t_flags &= ~TF_PREVVALID;
7827                 if (tp->t_rxtshift == 1 &&
7828                     (int)(ticks - tp->t_badrxtwin) < 0)
7829                         bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
7830         }
7831         SOCKBUF_LOCK(&so->so_snd);
7832         acked_amount = min(acked, (int)sbavail(&so->so_snd));
7833         tp->snd_wnd -= acked_amount;
7834         mfree = sbcut_locked(&so->so_snd, acked_amount);
7835         /* NB: sowwakeup_locked() does an implicit unlock. */
7836         sowwakeup_locked(so);
7837         m_freem(mfree);
7838         if (SEQ_GT(th->th_ack, tp->snd_una)) {
7839                 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
7840         }
7841         tp->snd_una = th->th_ack;
7842         bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost));
7843         if (IN_RECOVERY(tp->t_flags)) {
7844                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
7845                     (SEQ_LT(th->th_ack, tp->snd_max))) {
7846                         tcp_bbr_partialack(tp);
7847                 } else {
7848                         bbr_post_recovery(tp);
7849                 }
7850         }
7851         if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
7852                 tp->snd_recover = tp->snd_una;
7853         }
7854         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
7855                 tp->snd_nxt = tp->snd_max;
7856         }
7857         if (tp->snd_una == tp->snd_max) {
7858                 /* Nothing left outstanding */
7859 nothing_left:
7860                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
7861                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
7862                         bbr->rc_tp->t_acktime = 0;
7863                 if ((sbused(&so->so_snd) == 0) &&
7864                     (tp->t_flags & TF_SENTFIN)) {
7865                         ourfinisacked = 1;
7866                 }
7867                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
7868                 if (bbr->rc_in_persist == 0) {
7869                         bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
7870                 }
7871                 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
7872                 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
7873                 /*
7874                  * We invalidate the last ack here since we
7875                  * don't want to transfer forward the time
7876                  * for our sum's calculations.
7877                  */
7878                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7879                     (sbavail(&so->so_snd) == 0) &&
7880                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
7881                         /*
7882                          * The socket was gone and the peer sent data, time
7883                          * to reset him.
7884                          */
7885                         *ret_val = 1;
7886                         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
7887                         /* tcp_close will kill the inp pre-log the Reset */
7888                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
7889                         tp = tcp_close(tp);
7890                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
7891                         BBR_STAT_INC(bbr_dropped_af_data);
7892                         return (1);
7893                 }
7894                 /* Set need output so persist might get set */
7895                 bbr->r_wanted_output = 1;
7896         }
7897         if (ofia)
7898                 *ofia = ourfinisacked;
7899         return (0);
7900 }
7901
7902 static void
7903 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
7904 {
7905         if (bbr->rc_in_persist == 0) {
7906                 bbr_timer_cancel(bbr, __LINE__, cts);
7907                 bbr->r_ctl.rc_last_delay_val = 0;
7908                 tp->t_rxtshift = 0;
7909                 bbr->rc_in_persist = 1;
7910                 bbr->r_ctl.rc_went_idle_time = cts;
7911                 /* We should be capped when rw went to 0 but just in case */
7912                 bbr_log_type_pesist(bbr, cts, 0, line, 1);
7913                 /* Time freezes for the state, so do the accounting now */
7914                 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
7915                         uint32_t time_in;
7916
7917                         time_in = cts - bbr->r_ctl.rc_bbr_state_time;
7918                         if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
7919                                 int32_t idx;
7920
7921                                 idx = bbr_state_val(bbr);
7922                                 counter_u64_add(bbr_state_time[(idx + 5)], time_in);
7923                         } else {
7924                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
7925                         }
7926                 }
7927                 bbr->r_ctl.rc_bbr_state_time = cts;
7928         }
7929 }
7930
7931 static void
7932 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time)
7933 {
7934         /*
7935          * Note that if idle time does not exceed our
7936          * threshold, we do nothing continuing the state
7937          * transitions we were last walking through.
7938          */
7939         if (idle_time >= bbr_idle_restart_threshold) {
7940                 if (bbr->rc_use_idle_restart) {
7941                         bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT;
7942                         /*
7943                          * Set our target using BBR_UNIT, so
7944                          * we increase at a dramatic rate but
7945                          * we stop when we get the pipe
7946                          * full again for our current b/w estimate.
7947                          */
7948                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
7949                         bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
7950                         bbr_set_state_target(bbr, __LINE__);
7951                         /* Now setup our gains to ramp up */
7952                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
7953                         bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
7954                         bbr_log_type_statechange(bbr, cts, __LINE__);
7955                 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
7956                         bbr_substate_change(bbr, cts, __LINE__, 1);
7957                 }
7958         }
7959 }
7960
7961 static void
7962 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
7963 {
7964         uint32_t idle_time;
7965
7966         if (bbr->rc_in_persist == 0)
7967                 return;
7968         idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time);
7969         bbr->rc_in_persist = 0;
7970         bbr->rc_hit_state_1 = 0;
7971         bbr->r_ctl.rc_del_time = cts;
7972         /*
7973          * We invalidate the last ack here since we
7974          * don't want to transfer forward the time
7975          * for our sum's calculations.
7976          */
7977         if (bbr->rc_inp->inp_in_hpts) {
7978                 tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
7979                 bbr->rc_timer_first = 0;
7980                 bbr->r_ctl.rc_hpts_flags = 0;
7981                 bbr->r_ctl.rc_last_delay_val = 0;
7982                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
7983                 bbr->r_agg_early_set = 0;
7984                 bbr->r_ctl.rc_agg_early = 0;
7985         }
7986         bbr_log_type_pesist(bbr, cts, idle_time, line, 0);
7987         if (idle_time >= bbr_rtt_probe_time) {
7988                 /*
7989                  * This qualifies as a RTT_PROBE session since we drop the
7990                  * data outstanding to nothing and waited more than
7991                  * bbr_rtt_probe_time.
7992                  */
7993                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0);
7994                 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts;
7995         }
7996         tp->t_rxtshift = 0;
7997         /*
7998          * If in probeBW and we have persisted more than an RTT lets do
7999          * special handling.
8000          */
8001         /* Force a time based epoch */
8002         bbr_set_epoch(bbr, cts, __LINE__);
8003         /*
8004          * Setup the lost so we don't count anything against the guy
8005          * we have been stuck with during persists.
8006          */
8007         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
8008         /* Time un-freezes for the state */
8009         bbr->r_ctl.rc_bbr_state_time = cts;
8010         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) ||
8011             (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) {
8012                 /*
8013                  * If we are going back to probe-bw
8014                  * or probe_rtt, we may need to possibly
8015                  * do a fast restart.
8016                  */
8017                 bbr_restart_after_idle(bbr, cts, idle_time);
8018         }
8019 }
8020
8021 static void
8022 bbr_collapsed_window(struct tcp_bbr *bbr)
8023 {
8024         /*
8025          * Now we must walk the
8026          * send map and divide the
8027          * ones left stranded. These
8028          * guys can't cause us to abort
8029          * the connection and are really
8030          * "unsent". However if a buggy
8031          * client actually did keep some
8032          * of the data i.e. collapsed the win
8033          * and refused to ack and then opened
8034          * the win and acked that data. We would
8035          * get into an ack war, the simplier
8036          * method then of just pretending we
8037          * did not send those segments something
8038          * won't work.
8039          */
8040         struct bbr_sendmap *rsm, *nrsm;
8041         tcp_seq max_seq;
8042         uint32_t maxseg;
8043         int can_split = 0;
8044         int fnd = 0;
8045
8046         maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
8047         max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd;
8048         bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0);
8049         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
8050                 /* Find the first seq past or at maxseq */
8051                 if (rsm->r_flags & BBR_RWND_COLLAPSED)
8052                         rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8053                 if (SEQ_GEQ(max_seq, rsm->r_start) &&
8054                     SEQ_GEQ(rsm->r_end, max_seq)) {
8055                         fnd = 1;
8056                         break;
8057                 }
8058         }
8059         bbr->rc_has_collapsed = 0;
8060         if (!fnd) {
8061                 /* Nothing to do strange */
8062                 return;
8063         }
8064         /*
8065          * Now can we split?
8066          *
8067          * We don't want to split if splitting
8068          * would generate too many small segments
8069          * less we let an attacker fragment our
8070          * send_map and leave us out of memory.
8071          */
8072         if ((max_seq != rsm->r_start) &&
8073             (max_seq != rsm->r_end)){
8074                 /* can we split? */
8075                 int res1, res2;
8076
8077                 res1 = max_seq - rsm->r_start;
8078                 res2 = rsm->r_end - max_seq;
8079                 if ((res1 >= (maxseg/8)) &&
8080                     (res2 >= (maxseg/8))) {
8081                         /* No small pieces here */
8082                         can_split = 1;
8083                 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) {
8084                         /* We are under the limit */
8085                         can_split = 1;
8086                 }
8087         }
8088         /* Ok do we need to split this rsm? */
8089         if (max_seq == rsm->r_start) {
8090                 /* It's this guy no split required */
8091                 nrsm = rsm;
8092         } else if (max_seq == rsm->r_end) {
8093                 /* It's the next one no split required. */
8094                 nrsm = TAILQ_NEXT(rsm, r_next);
8095                 if (nrsm == NULL) {
8096                         /* Huh? */
8097                         return;
8098                 }
8099         } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) {
8100                 /* yep we need to split it */
8101                 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
8102                 if (nrsm == NULL) {
8103                         /* failed XXXrrs what can we do mark the whole? */
8104                         nrsm = rsm;
8105                         goto no_split;
8106                 }
8107                 /* Clone it */
8108                 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0);
8109                 bbr_clone_rsm(bbr, nrsm, rsm, max_seq);
8110                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
8111                 if (rsm->r_in_tmap) {
8112                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8113                         nrsm->r_in_tmap = 1;
8114                 }
8115         } else {
8116                 /*
8117                  * Split not allowed just start here just
8118                  * use this guy.
8119                  */
8120                 nrsm = rsm;
8121         }
8122 no_split:
8123         BBR_STAT_INC(bbr_collapsed_win);
8124         /* reuse fnd as a count */
8125         fnd = 0;
8126         TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) {
8127                 nrsm->r_flags |= BBR_RWND_COLLAPSED;
8128                 fnd++;
8129                 bbr->rc_has_collapsed = 1;
8130         }
8131         bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd);
8132 }
8133
8134 static void
8135 bbr_un_collapse_window(struct tcp_bbr *bbr)
8136 {
8137         struct bbr_sendmap *rsm;
8138         int cleared = 0;
8139
8140         TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
8141                 if (rsm->r_flags & BBR_RWND_COLLAPSED) {
8142                         /* Clear the flag */
8143                         rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8144                         cleared++;
8145                 } else
8146                         break;
8147         }
8148         bbr_log_type_rwnd_collapse(bbr,
8149                                    (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared);
8150         bbr->rc_has_collapsed = 0;
8151 }
8152
8153 /*
8154  * Return value of 1, the TCB is unlocked and most
8155  * likely gone, return value of 0, the TCB is still
8156  * locked.
8157  */
8158 static int
8159 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
8160     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
8161     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8162 {
8163         /*
8164          * Update window information. Don't look at window if no ACK: TAC's
8165          * send garbage on first SYN.
8166          */
8167         uint16_t nsegs;
8168         int32_t tfo_syn;
8169         struct tcp_bbr *bbr;
8170
8171         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8172         INP_WLOCK_ASSERT(tp->t_inpcb);
8173         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8174         if ((thflags & TH_ACK) &&
8175             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
8176             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
8177             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
8178                 /* keep track of pure window updates */
8179                 if (tlen == 0 &&
8180                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
8181                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
8182                 tp->snd_wnd = tiwin;
8183                 tp->snd_wl1 = th->th_seq;
8184                 tp->snd_wl2 = th->th_ack;
8185                 if (tp->snd_wnd > tp->max_sndwnd)
8186                         tp->max_sndwnd = tp->snd_wnd;
8187                 bbr->r_wanted_output = 1;
8188         } else if (thflags & TH_ACK) {
8189                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
8190                         tp->snd_wnd = tiwin;
8191                         tp->snd_wl1 = th->th_seq;
8192                         tp->snd_wl2 = th->th_ack;
8193                 }
8194         }
8195         if (tp->snd_wnd < ctf_outstanding(tp))
8196                 /* The peer collapsed its window on us */
8197                 bbr_collapsed_window(bbr);
8198         else if (bbr->rc_has_collapsed)
8199                 bbr_un_collapse_window(bbr);
8200         /* Was persist timer active and now we have window space? */
8201         if ((bbr->rc_in_persist != 0) &&
8202             (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8203                                 bbr_minseg(bbr)))) {
8204                 /*
8205                  * Make the rate persist at end of persist mode if idle long
8206                  * enough
8207                  */
8208                 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8209
8210                 /* Make sure we output to start the timer */
8211                 bbr->r_wanted_output = 1;
8212         }
8213         /* Do we need to enter persist? */
8214         if ((bbr->rc_in_persist == 0) &&
8215             (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8216             TCPS_HAVEESTABLISHED(tp->t_state) &&
8217             (tp->snd_max == tp->snd_una) &&
8218             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8219             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8220                 /* No send window.. we must enter persist */
8221                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8222         }
8223         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
8224                 m_freem(m);
8225                 return (0);
8226         }
8227         /*
8228          * We don't support urgent data but
8229          * drag along the up just to make sure
8230          * if there is a stack switch no one
8231          * is surprised.
8232          */
8233         tp->rcv_up = tp->rcv_nxt;
8234         INP_WLOCK_ASSERT(tp->t_inpcb);
8235
8236         /*
8237          * Process the segment text, merging it into the TCP sequencing
8238          * queue, and arranging for acknowledgment of receipt if necessary.
8239          * This process logically involves adjusting tp->rcv_wnd as data is
8240          * presented to the user (this happens in tcp_usrreq.c, case
8241          * PRU_RCVD).  If a FIN has already been received on this connection
8242          * then we just ignore the text.
8243          */
8244         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
8245                    IS_FASTOPEN(tp->t_flags));
8246         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
8247             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8248                 tcp_seq save_start = th->th_seq;
8249                 tcp_seq save_rnxt  = tp->rcv_nxt;
8250                 int     save_tlen  = tlen;
8251
8252                 m_adj(m, drop_hdrlen);  /* delayed header drop */
8253                 /*
8254                  * Insert segment which includes th into TCP reassembly
8255                  * queue with control block tp.  Set thflags to whether
8256                  * reassembly now includes a segment with FIN.  This handles
8257                  * the common case inline (segment is the next to be
8258                  * received on an established connection, and the queue is
8259                  * empty), avoiding linkage into and removal from the queue
8260                  * and repetition of various conversions. Set DELACK for
8261                  * segments received in order, but ack immediately when
8262                  * segments are out of order (so fast retransmit can work).
8263                  */
8264                 if (th->th_seq == tp->rcv_nxt &&
8265                     SEGQ_EMPTY(tp) &&
8266                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
8267                     tfo_syn)) {
8268 #ifdef NETFLIX_SB_LIMITS
8269                         u_int mcnt, appended;
8270
8271                         if (so->so_rcv.sb_shlim) {
8272                                 mcnt = m_memcnt(m);
8273                                 appended = 0;
8274                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8275                                     CFO_NOSLEEP, NULL) == false) {
8276                                         counter_u64_add(tcp_sb_shlim_fails, 1);
8277                                         m_freem(m);
8278                                         return (0);
8279                                 }
8280                         }
8281
8282 #endif
8283                         if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) {
8284                                 bbr->bbr_segs_rcvd += max(1, nsegs);
8285                                 tp->t_flags |= TF_DELACK;
8286                                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8287                         } else {
8288                                 bbr->r_wanted_output = 1;
8289                                 tp->t_flags |= TF_ACKNOW;
8290                         }
8291                         tp->rcv_nxt += tlen;
8292                         if (tlen &&
8293                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8294                             (tp->t_fbyte_in == 0)) {
8295                                 tp->t_fbyte_in = ticks;
8296                                 if (tp->t_fbyte_in == 0)
8297                                         tp->t_fbyte_in = 1;
8298                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
8299                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8300                         }
8301                         thflags = th->th_flags & TH_FIN;
8302                         KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8303                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8304                         SOCKBUF_LOCK(&so->so_rcv);
8305                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
8306                                 m_freem(m);
8307                         else
8308 #ifdef NETFLIX_SB_LIMITS
8309                                 appended =
8310 #endif
8311                                         sbappendstream_locked(&so->so_rcv, m, 0);
8312                         /* NB: sorwakeup_locked() does an implicit unlock. */
8313                         sorwakeup_locked(so);
8314 #ifdef NETFLIX_SB_LIMITS
8315                         if (so->so_rcv.sb_shlim && appended != mcnt)
8316                                 counter_fo_release(so->so_rcv.sb_shlim,
8317                                     mcnt - appended);
8318 #endif
8319
8320                 } else {
8321                         /*
8322                          * XXX: Due to the header drop above "th" is
8323                          * theoretically invalid by now.  Fortunately
8324                          * m_adj() doesn't actually frees any mbufs when
8325                          * trimming from the head.
8326                          */
8327                         tcp_seq temp = save_start;
8328
8329                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
8330                         tp->t_flags |= TF_ACKNOW;
8331                         if (tp->t_flags & TF_WAKESOR) {
8332                                 tp->t_flags &= ~TF_WAKESOR;
8333                                 /* NB: sorwakeup_locked() does an implicit unlock. */
8334                                 sorwakeup_locked(so);
8335                         }
8336                 }
8337                 if ((tp->t_flags & TF_SACK_PERMIT) &&
8338                     (save_tlen > 0) &&
8339                     TCPS_HAVEESTABLISHED(tp->t_state)) {
8340                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
8341                                 /*
8342                                  * DSACK actually handled in the fastpath
8343                                  * above.
8344                                  */
8345                                 tcp_update_sack_list(tp, save_start,
8346                                     save_start + save_tlen);
8347                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
8348                                 if ((tp->rcv_numsacks >= 1) &&
8349                                     (tp->sackblks[0].end == save_start)) {
8350                                         /*
8351                                          * Partial overlap, recorded at todrop
8352                                          * above.
8353                                          */
8354                                         tcp_update_sack_list(tp,
8355                                             tp->sackblks[0].start,
8356                                             tp->sackblks[0].end);
8357                                 } else {
8358                                         tcp_update_dsack_list(tp, save_start,
8359                                             save_start + save_tlen);
8360                                 }
8361                         } else if (tlen >= save_tlen) {
8362                                 /* Update of sackblks. */
8363                                 tcp_update_dsack_list(tp, save_start,
8364                                     save_start + save_tlen);
8365                         } else if (tlen > 0) {
8366                                 tcp_update_dsack_list(tp, save_start,
8367                                     save_start + tlen);
8368                         }
8369                 }
8370         } else {
8371                 m_freem(m);
8372                 thflags &= ~TH_FIN;
8373         }
8374
8375         /*
8376          * If FIN is received ACK the FIN and let the user know that the
8377          * connection is closing.
8378          */
8379         if (thflags & TH_FIN) {
8380                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8381                         /* The socket upcall is handled by socantrcvmore. */
8382                         socantrcvmore(so);
8383                         /*
8384                          * If connection is half-synchronized (ie NEEDSYN
8385                          * flag on) then delay ACK, so it may be piggybacked
8386                          * when SYN is sent. Otherwise, since we received a
8387                          * FIN then no more input can be expected, send ACK
8388                          * now.
8389                          */
8390                         if (tp->t_flags & TF_NEEDSYN) {
8391                                 tp->t_flags |= TF_DELACK;
8392                                 bbr_timer_cancel(bbr,
8393                                     __LINE__, bbr->r_ctl.rc_rcvtime);
8394                         } else {
8395                                 tp->t_flags |= TF_ACKNOW;
8396                         }
8397                         tp->rcv_nxt++;
8398                 }
8399                 switch (tp->t_state) {
8400                         /*
8401                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
8402                          * CLOSE_WAIT state.
8403                          */
8404                 case TCPS_SYN_RECEIVED:
8405                         tp->t_starttime = ticks;
8406                         /* FALLTHROUGH */
8407                 case TCPS_ESTABLISHED:
8408                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
8409                         break;
8410
8411                         /*
8412                          * If still in FIN_WAIT_1 STATE FIN has not been
8413                          * acked so enter the CLOSING state.
8414                          */
8415                 case TCPS_FIN_WAIT_1:
8416                         tcp_state_change(tp, TCPS_CLOSING);
8417                         break;
8418
8419                         /*
8420                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
8421                          * starting the time-wait timer, turning off the
8422                          * other standard timers.
8423                          */
8424                 case TCPS_FIN_WAIT_2:
8425                         bbr->rc_timer_first = 1;
8426                         bbr_timer_cancel(bbr,
8427                             __LINE__, bbr->r_ctl.rc_rcvtime);
8428                         INP_WLOCK_ASSERT(tp->t_inpcb);
8429                         tcp_twstart(tp);
8430                         return (1);
8431                 }
8432         }
8433         /*
8434          * Return any desired output.
8435          */
8436         if ((tp->t_flags & TF_ACKNOW) ||
8437             (sbavail(&so->so_snd) > ctf_outstanding(tp))) {
8438                 bbr->r_wanted_output = 1;
8439         }
8440         INP_WLOCK_ASSERT(tp->t_inpcb);
8441         return (0);
8442 }
8443
8444 /*
8445  * Here nothing is really faster, its just that we
8446  * have broken out the fast-data path also just like
8447  * the fast-ack. Return 1 if we processed the packet
8448  * return 0 if you need to take the "slow-path".
8449  */
8450 static int
8451 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
8452     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8453     uint32_t tiwin, int32_t nxt_pkt)
8454 {
8455         uint16_t nsegs;
8456         int32_t newsize = 0;    /* automatic sockbuf scaling */
8457         struct tcp_bbr *bbr;
8458 #ifdef NETFLIX_SB_LIMITS
8459         u_int mcnt, appended;
8460 #endif
8461 #ifdef TCPDEBUG
8462         /*
8463          * The size of tcp_saveipgen must be the size of the max ip header,
8464          * now IPv6.
8465          */
8466         u_char tcp_saveipgen[IP6_HDR_LEN];
8467         struct tcphdr tcp_savetcp;
8468         short ostate = 0;
8469
8470 #endif
8471         /* On the hpts and we would have called output */
8472         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8473
8474         /*
8475          * If last ACK falls within this segment's sequence numbers, record
8476          * the timestamp. NOTE that the test is modified according to the
8477          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8478          */
8479         if (bbr->r_ctl.rc_resend != NULL) {
8480                 return (0);
8481         }
8482         if (tiwin && tiwin != tp->snd_wnd) {
8483                 return (0);
8484         }
8485         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
8486                 return (0);
8487         }
8488         if (__predict_false((to->to_flags & TOF_TS) &&
8489             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
8490                 return (0);
8491         }
8492         if (__predict_false((th->th_ack != tp->snd_una))) {
8493                 return (0);
8494         }
8495         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
8496                 return (0);
8497         }
8498         if ((to->to_flags & TOF_TS) != 0 &&
8499             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8500                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
8501                 tp->ts_recent = to->to_tsval;
8502         }
8503         /*
8504          * This is a pure, in-sequence data packet with nothing on the
8505          * reassembly queue and we have enough buffer space to take it.
8506          */
8507         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8508
8509 #ifdef NETFLIX_SB_LIMITS
8510         if (so->so_rcv.sb_shlim) {
8511                 mcnt = m_memcnt(m);
8512                 appended = 0;
8513                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8514                     CFO_NOSLEEP, NULL) == false) {
8515                         counter_u64_add(tcp_sb_shlim_fails, 1);
8516                         m_freem(m);
8517                         return (1);
8518                 }
8519         }
8520 #endif
8521         /* Clean receiver SACK report if present */
8522         if (tp->rcv_numsacks)
8523                 tcp_clean_sackreport(tp);
8524         KMOD_TCPSTAT_INC(tcps_preddat);
8525         tp->rcv_nxt += tlen;
8526         if (tlen &&
8527             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8528             (tp->t_fbyte_in == 0)) {
8529                 tp->t_fbyte_in = ticks;
8530                 if (tp->t_fbyte_in == 0)
8531                         tp->t_fbyte_in = 1;
8532                 if (tp->t_fbyte_out && tp->t_fbyte_in)
8533                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8534         }
8535         /*
8536          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
8537          */
8538         tp->snd_wl1 = th->th_seq;
8539         /*
8540          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
8541          */
8542         tp->rcv_up = tp->rcv_nxt;
8543         KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8544         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8545 #ifdef TCPDEBUG
8546         if (so->so_options & SO_DEBUG)
8547                 tcp_trace(TA_INPUT, ostate, tp,
8548                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
8549 #endif
8550         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
8551
8552         /* Add data to socket buffer. */
8553         SOCKBUF_LOCK(&so->so_rcv);
8554         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8555                 m_freem(m);
8556         } else {
8557                 /*
8558                  * Set new socket buffer size. Give up when limit is
8559                  * reached.
8560                  */
8561                 if (newsize)
8562                         if (!sbreserve_locked(&so->so_rcv,
8563                             newsize, so, NULL))
8564                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
8565                 m_adj(m, drop_hdrlen);  /* delayed header drop */
8566
8567 #ifdef NETFLIX_SB_LIMITS
8568                 appended =
8569 #endif
8570                         sbappendstream_locked(&so->so_rcv, m, 0);
8571                 ctf_calc_rwin(so, tp);
8572         }
8573         /* NB: sorwakeup_locked() does an implicit unlock. */
8574         sorwakeup_locked(so);
8575 #ifdef NETFLIX_SB_LIMITS
8576         if (so->so_rcv.sb_shlim && mcnt != appended)
8577                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
8578 #endif
8579         if (DELAY_ACK(tp, bbr, nsegs)) {
8580                 bbr->bbr_segs_rcvd += max(1, nsegs);
8581                 tp->t_flags |= TF_DELACK;
8582                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8583         } else {
8584                 bbr->r_wanted_output = 1;
8585                 tp->t_flags |= TF_ACKNOW;
8586         }
8587         return (1);
8588 }
8589
8590 /*
8591  * This subfunction is used to try to highly optimize the
8592  * fast path. We again allow window updates that are
8593  * in sequence to remain in the fast-path. We also add
8594  * in the __predict's to attempt to help the compiler.
8595  * Note that if we return a 0, then we can *not* process
8596  * it and the caller should push the packet into the
8597  * slow-path. If we return 1, then all is well and
8598  * the packet is fully processed.
8599  */
8600 static int
8601 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8602     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8603     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
8604 {
8605         int32_t acked;
8606         uint16_t nsegs;
8607         uint32_t sack_changed;
8608 #ifdef TCPDEBUG
8609         /*
8610          * The size of tcp_saveipgen must be the size of the max ip header,
8611          * now IPv6.
8612          */
8613         u_char tcp_saveipgen[IP6_HDR_LEN];
8614         struct tcphdr tcp_savetcp;
8615         short ostate = 0;
8616
8617 #endif
8618         uint32_t prev_acked = 0;
8619         struct tcp_bbr *bbr;
8620
8621         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8622                 /* Old ack, behind (or duplicate to) the last one rcv'd */
8623                 return (0);
8624         }
8625         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
8626                 /* Above what we have sent? */
8627                 return (0);
8628         }
8629         if (__predict_false(tiwin == 0)) {
8630                 /* zero window */
8631                 return (0);
8632         }
8633         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
8634                 /* We need a SYN or a FIN, unlikely.. */
8635                 return (0);
8636         }
8637         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
8638                 /* Timestamp is behind .. old ack with seq wrap? */
8639                 return (0);
8640         }
8641         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
8642                 /* Still recovering */
8643                 return (0);
8644         }
8645         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8646         if (__predict_false(bbr->r_ctl.rc_resend != NULL)) {
8647                 /* We are retransmitting */
8648                 return (0);
8649         }
8650         if (__predict_false(bbr->rc_in_persist != 0)) {
8651                 /* In persist mode */
8652                 return (0);
8653         }
8654         if (bbr->r_ctl.rc_sacked) {
8655                 /* We have sack holes on our scoreboard */
8656                 return (0);
8657         }
8658         /* Ok if we reach here, we can process a fast-ack */
8659         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8660         sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
8661         /*
8662          * We never detect loss in fast ack [we can't
8663          * have a sack and can't be in recovery so
8664          * we always pass 0 (nothing detected)].
8665          */
8666         bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0);
8667         /* Did the window get updated? */
8668         if (tiwin != tp->snd_wnd) {
8669                 tp->snd_wnd = tiwin;
8670                 tp->snd_wl1 = th->th_seq;
8671                 if (tp->snd_wnd > tp->max_sndwnd)
8672                         tp->max_sndwnd = tp->snd_wnd;
8673         }
8674         /* Do we need to exit persists? */
8675         if ((bbr->rc_in_persist != 0) &&
8676             (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8677                                bbr_minseg(bbr)))) {
8678                 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8679                 bbr->r_wanted_output = 1;
8680         }
8681         /* Do we need to enter persists? */
8682         if ((bbr->rc_in_persist == 0) &&
8683             (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8684             TCPS_HAVEESTABLISHED(tp->t_state) &&
8685             (tp->snd_max == tp->snd_una) &&
8686             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8687             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8688                 /* No send window.. we must enter persist */
8689                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8690         }
8691         /*
8692          * If last ACK falls within this segment's sequence numbers, record
8693          * the timestamp. NOTE that the test is modified according to the
8694          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8695          */
8696         if ((to->to_flags & TOF_TS) != 0 &&
8697             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8698                 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime;
8699                 tp->ts_recent = to->to_tsval;
8700         }
8701         /*
8702          * This is a pure ack for outstanding data.
8703          */
8704         KMOD_TCPSTAT_INC(tcps_predack);
8705
8706         /*
8707          * "bad retransmit" recovery.
8708          */
8709         if (tp->t_flags & TF_PREVVALID) {
8710                 tp->t_flags &= ~TF_PREVVALID;
8711                 if (tp->t_rxtshift == 1 &&
8712                     (int)(ticks - tp->t_badrxtwin) < 0)
8713                         bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
8714         }
8715         /*
8716          * Recalculate the transmit timer / rtt.
8717          *
8718          * Some boxes send broken timestamp replies during the SYN+ACK
8719          * phase, ignore timestamps of 0 or we could calculate a huge RTT
8720          * and blow up the retransmit timer.
8721          */
8722         acked = BYTES_THIS_ACK(tp, th);
8723
8724 #ifdef TCP_HHOOK
8725         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
8726         hhook_run_tcp_est_in(tp, th, to);
8727 #endif
8728
8729         KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
8730         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
8731         sbdrop(&so->so_snd, acked);
8732
8733         if (SEQ_GT(th->th_ack, tp->snd_una))
8734                 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
8735         tp->snd_una = th->th_ack;
8736         if (tp->snd_wnd < ctf_outstanding(tp))
8737                 /* The peer collapsed its window on us */
8738                 bbr_collapsed_window(bbr);
8739         else if (bbr->rc_has_collapsed)
8740                 bbr_un_collapse_window(bbr);
8741
8742         if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
8743                 tp->snd_recover = tp->snd_una;
8744         }
8745         bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0);
8746         /*
8747          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
8748          */
8749         tp->snd_wl2 = th->th_ack;
8750         m_freem(m);
8751         /*
8752          * If all outstanding data are acked, stop retransmit timer,
8753          * otherwise restart timer using current (possibly backed-off)
8754          * value. If process is waiting for space, wakeup/selwakeup/signal.
8755          * If data are ready to send, let tcp_output decide between more
8756          * output or persist.
8757          */
8758 #ifdef TCPDEBUG
8759         if (so->so_options & SO_DEBUG)
8760                 tcp_trace(TA_INPUT, ostate, tp,
8761                     (void *)tcp_saveipgen,
8762                     &tcp_savetcp, 0);
8763 #endif
8764         /* Wake up the socket if we have room to write more */
8765         sowwakeup(so);
8766         if (tp->snd_una == tp->snd_max) {
8767                 /* Nothing left outstanding */
8768                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
8769                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
8770                         bbr->rc_tp->t_acktime = 0;
8771                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8772                 if (bbr->rc_in_persist == 0) {
8773                         bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
8774                 }
8775                 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
8776                 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
8777                 /*
8778                  * We invalidate the last ack here since we
8779                  * don't want to transfer forward the time
8780                  * for our sum's calculations.
8781                  */
8782                 bbr->r_wanted_output = 1;
8783         }
8784         if (sbavail(&so->so_snd)) {
8785                 bbr->r_wanted_output = 1;
8786         }
8787         return (1);
8788 }
8789
8790 /*
8791  * Return value of 1, the TCB is unlocked and most
8792  * likely gone, return value of 0, the TCB is still
8793  * locked.
8794  */
8795 static int
8796 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
8797     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8798     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
8799 {
8800         int32_t todrop;
8801         int32_t ourfinisacked = 0;
8802         struct tcp_bbr *bbr;
8803         int32_t ret_val = 0;
8804
8805         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8806         ctf_calc_rwin(so, tp);
8807         /*
8808          * If the state is SYN_SENT: if seg contains an ACK, but not for our
8809          * SYN, drop the input. if seg contains a RST, then drop the
8810          * connection. if seg does not contain SYN, then drop it. Otherwise
8811          * this is an acceptable SYN segment initialize tp->rcv_nxt and
8812          * tp->irs if seg contains ack then advance tp->snd_una. BRR does
8813          * not support ECN so we will not say we are capable. if SYN has
8814          * been acked change to ESTABLISHED else SYN_RCVD state arrange for
8815          * segment to be acked (eventually) continue processing rest of
8816          * data/controls, beginning with URG
8817          */
8818         if ((thflags & TH_ACK) &&
8819             (SEQ_LEQ(th->th_ack, tp->iss) ||
8820             SEQ_GT(th->th_ack, tp->snd_max))) {
8821                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8822                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
8823                 return (1);
8824         }
8825         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
8826                 TCP_PROBE5(connect__refused, NULL, tp,
8827                     mtod(m, const char *), tp, th);
8828                 tp = tcp_drop(tp, ECONNREFUSED);
8829                 ctf_do_drop(m, tp);
8830                 return (1);
8831         }
8832         if (thflags & TH_RST) {
8833                 ctf_do_drop(m, tp);
8834                 return (1);
8835         }
8836         if (!(thflags & TH_SYN)) {
8837                 ctf_do_drop(m, tp);
8838                 return (1);
8839         }
8840         tp->irs = th->th_seq;
8841         tcp_rcvseqinit(tp);
8842         if (thflags & TH_ACK) {
8843                 int tfo_partial = 0;
8844
8845                 KMOD_TCPSTAT_INC(tcps_connects);
8846                 soisconnected(so);
8847 #ifdef MAC
8848                 mac_socketpeer_set_from_mbuf(m, so);
8849 #endif
8850                 /* Do window scaling on this connection? */
8851                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
8852                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
8853                         tp->rcv_scale = tp->request_r_scale;
8854                 }
8855                 tp->rcv_adv += min(tp->rcv_wnd,
8856                     TCP_MAXWIN << tp->rcv_scale);
8857                 /*
8858                  * If not all the data that was sent in the TFO SYN
8859                  * has been acked, resend the remainder right away.
8860                  */
8861                 if (IS_FASTOPEN(tp->t_flags) &&
8862                     (tp->snd_una != tp->snd_max)) {
8863                         tp->snd_nxt = th->th_ack;
8864                         tfo_partial = 1;
8865                 }
8866                 /*
8867                  * If there's data, delay ACK; if there's also a FIN ACKNOW
8868                  * will be turned on later.
8869                  */
8870                 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) {
8871                         bbr->bbr_segs_rcvd += 1;
8872                         tp->t_flags |= TF_DELACK;
8873                         bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8874                 } else {
8875                         bbr->r_wanted_output = 1;
8876                         tp->t_flags |= TF_ACKNOW;
8877                 }
8878                 if (SEQ_GT(th->th_ack, tp->iss)) {
8879                         /*
8880                          * The SYN is acked
8881                          * handle it specially.
8882                          */
8883                         bbr_log_syn(tp, to);
8884                 }
8885                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
8886                         /*
8887                          * We advance snd_una for the
8888                          * fast open case. If th_ack is
8889                          * acknowledging data beyond
8890                          * snd_una we can't just call
8891                          * ack-processing since the
8892                          * data stream in our send-map
8893                          * will start at snd_una + 1 (one
8894                          * beyond the SYN). If its just
8895                          * equal we don't need to do that
8896                          * and there is no send_map.
8897                          */
8898                         tp->snd_una++;
8899                 }
8900                 /*
8901                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
8902                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
8903                  */
8904                 tp->t_starttime = ticks;
8905                 if (tp->t_flags & TF_NEEDFIN) {
8906                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
8907                         tp->t_flags &= ~TF_NEEDFIN;
8908                         thflags &= ~TH_SYN;
8909                 } else {
8910                         tcp_state_change(tp, TCPS_ESTABLISHED);
8911                         TCP_PROBE5(connect__established, NULL, tp,
8912                             mtod(m, const char *), tp, th);
8913                         cc_conn_init(tp);
8914                 }
8915         } else {
8916                 /*
8917                  * Received initial SYN in SYN-SENT[*] state => simultaneous
8918                  * open.  If segment contains CC option and there is a
8919                  * cached CC, apply TAO test. If it succeeds, connection is *
8920                  * half-synchronized. Otherwise, do 3-way handshake:
8921                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
8922                  * there was no CC option, clear cached CC value.
8923                  */
8924                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
8925                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
8926         }
8927         INP_WLOCK_ASSERT(tp->t_inpcb);
8928         /*
8929          * Advance th->th_seq to correspond to first data byte. If data,
8930          * trim to stay within window, dropping FIN if necessary.
8931          */
8932         th->th_seq++;
8933         if (tlen > tp->rcv_wnd) {
8934                 todrop = tlen - tp->rcv_wnd;
8935                 m_adj(m, -todrop);
8936                 tlen = tp->rcv_wnd;
8937                 thflags &= ~TH_FIN;
8938                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
8939                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
8940         }
8941         tp->snd_wl1 = th->th_seq - 1;
8942         tp->rcv_up = th->th_seq;
8943         /*
8944          * Client side of transaction: already sent SYN and data. If the
8945          * remote host used T/TCP to validate the SYN, our data will be
8946          * ACK'd; if so, enter normal data segment processing in the middle
8947          * of step 5, ack processing. Otherwise, goto step 6.
8948          */
8949         if (thflags & TH_ACK) {
8950                 if ((to->to_flags & TOF_TS) != 0) {
8951                         uint32_t t, rtt;
8952
8953                         t = tcp_tv_to_mssectick(&bbr->rc_tv);
8954                         if (TSTMP_GEQ(t, to->to_tsecr)) {
8955                                 rtt = t - to->to_tsecr;
8956                                 if (rtt == 0) {
8957                                         rtt = 1;
8958                                 }
8959                                 rtt *= MS_IN_USEC;
8960                                 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
8961                                 apply_filter_min_small(&bbr->r_ctl.rc_rttprop,
8962                                                        rtt, bbr->r_ctl.rc_rcvtime);
8963                         }
8964                 }
8965                 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
8966                         return (ret_val);
8967                 /* We may have changed to FIN_WAIT_1 above */
8968                 if (tp->t_state == TCPS_FIN_WAIT_1) {
8969                         /*
8970                          * In FIN_WAIT_1 STATE in addition to the processing
8971                          * for the ESTABLISHED state if our FIN is now
8972                          * acknowledged then enter FIN_WAIT_2.
8973                          */
8974                         if (ourfinisacked) {
8975                                 /*
8976                                  * If we can't receive any more data, then
8977                                  * closing user can proceed. Starting the
8978                                  * timer is contrary to the specification,
8979                                  * but if we don't get a FIN we'll hang
8980                                  * forever.
8981                                  *
8982                                  * XXXjl: we should release the tp also, and
8983                                  * use a compressed state.
8984                                  */
8985                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8986                                         soisdisconnected(so);
8987                                         tcp_timer_activate(tp, TT_2MSL,
8988                                             (tcp_fast_finwait2_recycle ?
8989                                             tcp_finwait2_timeout :
8990                                             TP_MAXIDLE(tp)));
8991                                 }
8992                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
8993                         }
8994                 }
8995         }
8996         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
8997             tiwin, thflags, nxt_pkt));
8998 }
8999
9000 /*
9001  * Return value of 1, the TCB is unlocked and most
9002  * likely gone, return value of 0, the TCB is still
9003  * locked.
9004  */
9005 static int
9006 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
9007                 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9008                 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9009 {
9010         int32_t ourfinisacked = 0;
9011         int32_t ret_val;
9012         struct tcp_bbr *bbr;
9013
9014         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9015         ctf_calc_rwin(so, tp);
9016         if ((thflags & TH_ACK) &&
9017             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
9018              SEQ_GT(th->th_ack, tp->snd_max))) {
9019                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9020                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9021                 return (1);
9022         }
9023         if (IS_FASTOPEN(tp->t_flags)) {
9024                 /*
9025                  * When a TFO connection is in SYN_RECEIVED, the only valid
9026                  * packets are the initial SYN, a retransmit/copy of the
9027                  * initial SYN (possibly with a subset of the original
9028                  * data), a valid ACK, a FIN, or a RST.
9029                  */
9030                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
9031                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9032                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9033                         return (1);
9034                 } else if (thflags & TH_SYN) {
9035                         /* non-initial SYN is ignored */
9036                         if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
9037                             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
9038                             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
9039                                 ctf_do_drop(m, NULL);
9040                                 return (0);
9041                         }
9042                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
9043                         ctf_do_drop(m, NULL);
9044                         return (0);
9045                 }
9046         }
9047         if ((thflags & TH_RST) ||
9048             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9049                 return (ctf_process_rst(m, th, so, tp));
9050         /*
9051          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9052          * it's less than ts_recent, drop it.
9053          */
9054         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9055             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9056                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9057                         return (ret_val);
9058         }
9059         /*
9060          * In the SYN-RECEIVED state, validate that the packet belongs to
9061          * this connection before trimming the data to fit the receive
9062          * window.  Check the sequence number versus IRS since we know the
9063          * sequence numbers haven't wrapped.  This is a partial fix for the
9064          * "LAND" DoS attack.
9065          */
9066         if (SEQ_LT(th->th_seq, tp->irs)) {
9067                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9068                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9069                 return (1);
9070         }
9071         INP_WLOCK_ASSERT(tp->t_inpcb);
9072         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9073                 return (ret_val);
9074         }
9075         /*
9076          * If last ACK falls within this segment's sequence numbers, record
9077          * its timestamp. NOTE: 1) That the test incorporates suggestions
9078          * from the latest proposal of the tcplw@cray.com list (Braden
9079          * 1993/04/26). 2) That updating only on newer timestamps interferes
9080          * with our earlier PAWS tests, so this check should be solely
9081          * predicated on the sequence space of this segment. 3) That we
9082          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9083          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9084          * SEG.Len, This modified check allows us to overcome RFC1323's
9085          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9086          * p.869. In such cases, we can still calculate the RTT correctly
9087          * when RCV.NXT == Last.ACK.Sent.
9088          */
9089         if ((to->to_flags & TOF_TS) != 0 &&
9090             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9091             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9092                     ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9093                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9094                 tp->ts_recent = to->to_tsval;
9095         }
9096         tp->snd_wnd = tiwin;
9097         /*
9098          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9099          * is on (half-synchronized state), then queue data for later
9100          * processing; else drop segment and return.
9101          */
9102         if ((thflags & TH_ACK) == 0) {
9103                 if (IS_FASTOPEN(tp->t_flags)) {
9104                         cc_conn_init(tp);
9105                 }
9106                 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9107                                          tiwin, thflags, nxt_pkt));
9108         }
9109         KMOD_TCPSTAT_INC(tcps_connects);
9110         soisconnected(so);
9111         /* Do window scaling? */
9112         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9113             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9114                 tp->rcv_scale = tp->request_r_scale;
9115         }
9116         /*
9117          * ok for the first time in lets see if we can use the ts to figure
9118          * out what the initial RTT was.
9119          */
9120         if ((to->to_flags & TOF_TS) != 0) {
9121                 uint32_t t, rtt;
9122
9123                 t = tcp_tv_to_mssectick(&bbr->rc_tv);
9124                 if (TSTMP_GEQ(t, to->to_tsecr)) {
9125                         rtt = t - to->to_tsecr;
9126                         if (rtt == 0) {
9127                                 rtt = 1;
9128                         }
9129                         rtt *= MS_IN_USEC;
9130                         tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9131                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime);
9132                 }
9133         }
9134         /* Drop off any SYN in the send map (probably not there)  */
9135         if (thflags & TH_ACK)
9136                 bbr_log_syn(tp, to);
9137         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
9138                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
9139                 tp->t_tfo_pending = NULL;
9140         }
9141         /*
9142          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
9143          * FIN-WAIT-1
9144          */
9145         tp->t_starttime = ticks;
9146         if (tp->t_flags & TF_NEEDFIN) {
9147                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
9148                 tp->t_flags &= ~TF_NEEDFIN;
9149         } else {
9150                 tcp_state_change(tp, TCPS_ESTABLISHED);
9151                 TCP_PROBE5(accept__established, NULL, tp,
9152                            mtod(m, const char *), tp, th);
9153                 /*
9154                  * TFO connections call cc_conn_init() during SYN
9155                  * processing.  Calling it again here for such connections
9156                  * is not harmless as it would undo the snd_cwnd reduction
9157                  * that occurs when a TFO SYN|ACK is retransmitted.
9158                  */
9159                 if (!IS_FASTOPEN(tp->t_flags))
9160                         cc_conn_init(tp);
9161         }
9162         /*
9163          * Account for the ACK of our SYN prior to
9164          * regular ACK processing below, except for
9165          * simultaneous SYN, which is handled later.
9166          */
9167         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
9168                 tp->snd_una++;
9169         /*
9170          * If segment contains data or ACK, will call tcp_reass() later; if
9171          * not, do so now to pass queued data to user.
9172          */
9173         if (tlen == 0 && (thflags & TH_FIN) == 0) {
9174                 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
9175                         (struct mbuf *)0);
9176                 if (tp->t_flags & TF_WAKESOR) {
9177                         tp->t_flags &= ~TF_WAKESOR;
9178                         /* NB: sorwakeup_locked() does an implicit unlock. */
9179                         sorwakeup_locked(so);
9180                 }
9181         }
9182         tp->snd_wl1 = th->th_seq - 1;
9183         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9184                 return (ret_val);
9185         }
9186         if (tp->t_state == TCPS_FIN_WAIT_1) {
9187                 /* We could have went to FIN_WAIT_1 (or EST) above */
9188                 /*
9189                  * In FIN_WAIT_1 STATE in addition to the processing for the
9190                  * ESTABLISHED state if our FIN is now acknowledged then
9191                  * enter FIN_WAIT_2.
9192                  */
9193                 if (ourfinisacked) {
9194                         /*
9195                          * If we can't receive any more data, then closing
9196                          * user can proceed. Starting the timer is contrary
9197                          * to the specification, but if we don't get a FIN
9198                          * we'll hang forever.
9199                          *
9200                          * XXXjl: we should release the tp also, and use a
9201                          * compressed state.
9202                          */
9203                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9204                                 soisdisconnected(so);
9205                                 tcp_timer_activate(tp, TT_2MSL,
9206                                                    (tcp_fast_finwait2_recycle ?
9207                                                     tcp_finwait2_timeout :
9208                                                     TP_MAXIDLE(tp)));
9209                         }
9210                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
9211                 }
9212         }
9213         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9214                                  tiwin, thflags, nxt_pkt));
9215 }
9216
9217 /*
9218  * Return value of 1, the TCB is unlocked and most
9219  * likely gone, return value of 0, the TCB is still
9220  * locked.
9221  */
9222 static int
9223 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
9224     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9225     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9226 {
9227         struct tcp_bbr *bbr;
9228         int32_t ret_val;
9229
9230         /*
9231          * Header prediction: check for the two common cases of a
9232          * uni-directional data xfer.  If the packet has no control flags,
9233          * is in-sequence, the window didn't change and we're not
9234          * retransmitting, it's a candidate.  If the length is zero and the
9235          * ack moved forward, we're the sender side of the xfer.  Just free
9236          * the data acked & wake any higher level process that was blocked
9237          * waiting for space.  If the length is non-zero and the ack didn't
9238          * move, we're the receiver side.  If we're getting packets in-order
9239          * (the reassembly queue is empty), add the data toc The socket
9240          * buffer and note that we need a delayed ack. Make sure that the
9241          * hidden state-flags are also off. Since we check for
9242          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
9243          */
9244         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9245         if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) {
9246                 /*
9247                  * If we have delived under 4 segments increase the initial
9248                  * window if raised by the peer. We use this to determine
9249                  * dynamic and static rwnd's at the end of a connection.
9250                  */
9251                 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd);
9252         }
9253         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
9254             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) &&
9255             __predict_true(SEGQ_EMPTY(tp)) &&
9256             __predict_true(th->th_seq == tp->rcv_nxt)) {
9257                 if (tlen == 0) {
9258                         if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
9259                             tiwin, nxt_pkt, iptos)) {
9260                                 return (0);
9261                         }
9262                 } else {
9263                         if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
9264                             tiwin, nxt_pkt)) {
9265                                 return (0);
9266                         }
9267                 }
9268         }
9269         ctf_calc_rwin(so, tp);
9270
9271         if ((thflags & TH_RST) ||
9272             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9273                 return (ctf_process_rst(m, th, so, tp));
9274         /*
9275          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9276          * synchronized state.
9277          */
9278         if (thflags & TH_SYN) {
9279                 ctf_challenge_ack(m, th, tp, &ret_val);
9280                 return (ret_val);
9281         }
9282         /*
9283          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9284          * it's less than ts_recent, drop it.
9285          */
9286         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9287             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9288                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9289                         return (ret_val);
9290         }
9291         INP_WLOCK_ASSERT(tp->t_inpcb);
9292         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9293                 return (ret_val);
9294         }
9295         /*
9296          * If last ACK falls within this segment's sequence numbers, record
9297          * its timestamp. NOTE: 1) That the test incorporates suggestions
9298          * from the latest proposal of the tcplw@cray.com list (Braden
9299          * 1993/04/26). 2) That updating only on newer timestamps interferes
9300          * with our earlier PAWS tests, so this check should be solely
9301          * predicated on the sequence space of this segment. 3) That we
9302          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9303          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9304          * SEG.Len, This modified check allows us to overcome RFC1323's
9305          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9306          * p.869. In such cases, we can still calculate the RTT correctly
9307          * when RCV.NXT == Last.ACK.Sent.
9308          */
9309         if ((to->to_flags & TOF_TS) != 0 &&
9310             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9311             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9312             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9313                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9314                 tp->ts_recent = to->to_tsval;
9315         }
9316         /*
9317          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9318          * is on (half-synchronized state), then queue data for later
9319          * processing; else drop segment and return.
9320          */
9321         if ((thflags & TH_ACK) == 0) {
9322                 if (tp->t_flags & TF_NEEDSYN) {
9323                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9324                             tiwin, thflags, nxt_pkt));
9325                 } else if (tp->t_flags & TF_ACKNOW) {
9326                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9327                         bbr->r_wanted_output = 1;
9328                         return (ret_val);
9329                 } else {
9330                         ctf_do_drop(m, NULL);
9331                         return (0);
9332                 }
9333         }
9334         /*
9335          * Ack processing.
9336          */
9337         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9338                 return (ret_val);
9339         }
9340         if (sbavail(&so->so_snd)) {
9341                 if (ctf_progress_timeout_check(tp, true)) {
9342                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9343                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9344                         return (1);
9345                 }
9346         }
9347         /* State changes only happen in bbr_process_data() */
9348         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9349             tiwin, thflags, nxt_pkt));
9350 }
9351
9352 /*
9353  * Return value of 1, the TCB is unlocked and most
9354  * likely gone, return value of 0, the TCB is still
9355  * locked.
9356  */
9357 static int
9358 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
9359     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9360     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9361 {
9362         struct tcp_bbr *bbr;
9363         int32_t ret_val;
9364
9365         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9366         ctf_calc_rwin(so, tp);
9367         if ((thflags & TH_RST) ||
9368             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9369                 return (ctf_process_rst(m, th, so, tp));
9370         /*
9371          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9372          * synchronized state.
9373          */
9374         if (thflags & TH_SYN) {
9375                 ctf_challenge_ack(m, th, tp, &ret_val);
9376                 return (ret_val);
9377         }
9378         /*
9379          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9380          * it's less than ts_recent, drop it.
9381          */
9382         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9383             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9384                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9385                         return (ret_val);
9386         }
9387         INP_WLOCK_ASSERT(tp->t_inpcb);
9388         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9389                 return (ret_val);
9390         }
9391         /*
9392          * If last ACK falls within this segment's sequence numbers, record
9393          * its timestamp. NOTE: 1) That the test incorporates suggestions
9394          * from the latest proposal of the tcplw@cray.com list (Braden
9395          * 1993/04/26). 2) That updating only on newer timestamps interferes
9396          * with our earlier PAWS tests, so this check should be solely
9397          * predicated on the sequence space of this segment. 3) That we
9398          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9399          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9400          * SEG.Len, This modified check allows us to overcome RFC1323's
9401          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9402          * p.869. In such cases, we can still calculate the RTT correctly
9403          * when RCV.NXT == Last.ACK.Sent.
9404          */
9405         if ((to->to_flags & TOF_TS) != 0 &&
9406             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9407             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9408             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9409                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9410                 tp->ts_recent = to->to_tsval;
9411         }
9412         /*
9413          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9414          * is on (half-synchronized state), then queue data for later
9415          * processing; else drop segment and return.
9416          */
9417         if ((thflags & TH_ACK) == 0) {
9418                 if (tp->t_flags & TF_NEEDSYN) {
9419                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9420                             tiwin, thflags, nxt_pkt));
9421                 } else if (tp->t_flags & TF_ACKNOW) {
9422                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9423                         bbr->r_wanted_output = 1;
9424                         return (ret_val);
9425                 } else {
9426                         ctf_do_drop(m, NULL);
9427                         return (0);
9428                 }
9429         }
9430         /*
9431          * Ack processing.
9432          */
9433         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9434                 return (ret_val);
9435         }
9436         if (sbavail(&so->so_snd)) {
9437                 if (ctf_progress_timeout_check(tp, true)) {
9438                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9439                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9440                         return (1);
9441                 }
9442         }
9443         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9444             tiwin, thflags, nxt_pkt));
9445 }
9446
9447 static int
9448 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr,
9449     struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so)
9450 {
9451
9452         if (bbr->rc_allow_data_af_clo == 0) {
9453 close_now:
9454                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9455                 /* tcp_close will kill the inp pre-log the Reset */
9456                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9457                 tp = tcp_close(tp);
9458                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
9459                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
9460                 return (1);
9461         }
9462         if (sbavail(&so->so_snd) == 0)
9463                 goto close_now;
9464         /* Ok we allow data that is ignored and a followup reset */
9465         tp->rcv_nxt = th->th_seq + *tlen;
9466         tp->t_flags2 |= TF2_DROP_AF_DATA;
9467         bbr->r_wanted_output = 1;
9468         *tlen = 0;
9469         return (0);
9470 }
9471
9472 /*
9473  * Return value of 1, the TCB is unlocked and most
9474  * likely gone, return value of 0, the TCB is still
9475  * locked.
9476  */
9477 static int
9478 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
9479     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9480     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9481 {
9482         int32_t ourfinisacked = 0;
9483         int32_t ret_val;
9484         struct tcp_bbr *bbr;
9485
9486         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9487         ctf_calc_rwin(so, tp);
9488         if ((thflags & TH_RST) ||
9489             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9490                 return (ctf_process_rst(m, th, so, tp));
9491         /*
9492          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9493          * synchronized state.
9494          */
9495         if (thflags & TH_SYN) {
9496                 ctf_challenge_ack(m, th, tp, &ret_val);
9497                 return (ret_val);
9498         }
9499         /*
9500          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9501          * it's less than ts_recent, drop it.
9502          */
9503         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9504             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9505                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9506                         return (ret_val);
9507         }
9508         INP_WLOCK_ASSERT(tp->t_inpcb);
9509         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9510                 return (ret_val);
9511         }
9512         /*
9513          * If new data are received on a connection after the user processes
9514          * are gone, then RST the other end.
9515          */
9516         if ((so->so_state & SS_NOFDREF) && tlen) {
9517                 /*
9518                  * We call a new function now so we might continue and setup
9519                  * to reset at all data being ack'd.
9520                  */
9521                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9522                         return (1);
9523         }
9524         /*
9525          * If last ACK falls within this segment's sequence numbers, record
9526          * its timestamp. NOTE: 1) That the test incorporates suggestions
9527          * from the latest proposal of the tcplw@cray.com list (Braden
9528          * 1993/04/26). 2) That updating only on newer timestamps interferes
9529          * with our earlier PAWS tests, so this check should be solely
9530          * predicated on the sequence space of this segment. 3) That we
9531          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9532          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9533          * SEG.Len, This modified check allows us to overcome RFC1323's
9534          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9535          * p.869. In such cases, we can still calculate the RTT correctly
9536          * when RCV.NXT == Last.ACK.Sent.
9537          */
9538         if ((to->to_flags & TOF_TS) != 0 &&
9539             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9540             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9541             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9542                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9543                 tp->ts_recent = to->to_tsval;
9544         }
9545         /*
9546          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9547          * is on (half-synchronized state), then queue data for later
9548          * processing; else drop segment and return.
9549          */
9550         if ((thflags & TH_ACK) == 0) {
9551                 if (tp->t_flags & TF_NEEDSYN) {
9552                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9553                             tiwin, thflags, nxt_pkt));
9554                 } else if (tp->t_flags & TF_ACKNOW) {
9555                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9556                         bbr->r_wanted_output = 1;
9557                         return (ret_val);
9558                 } else {
9559                         ctf_do_drop(m, NULL);
9560                         return (0);
9561                 }
9562         }
9563         /*
9564          * Ack processing.
9565          */
9566         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9567                 return (ret_val);
9568         }
9569         if (ourfinisacked) {
9570                 /*
9571                  * If we can't receive any more data, then closing user can
9572                  * proceed. Starting the timer is contrary to the
9573                  * specification, but if we don't get a FIN we'll hang
9574                  * forever.
9575                  *
9576                  * XXXjl: we should release the tp also, and use a
9577                  * compressed state.
9578                  */
9579                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9580                         soisdisconnected(so);
9581                         tcp_timer_activate(tp, TT_2MSL,
9582                             (tcp_fast_finwait2_recycle ?
9583                             tcp_finwait2_timeout :
9584                             TP_MAXIDLE(tp)));
9585                 }
9586                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
9587         }
9588         if (sbavail(&so->so_snd)) {
9589                 if (ctf_progress_timeout_check(tp, true)) {
9590                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9591                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9592                         return (1);
9593                 }
9594         }
9595         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9596             tiwin, thflags, nxt_pkt));
9597 }
9598
9599 /*
9600  * Return value of 1, the TCB is unlocked and most
9601  * likely gone, return value of 0, the TCB is still
9602  * locked.
9603  */
9604 static int
9605 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
9606     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9607     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9608 {
9609         int32_t ourfinisacked = 0;
9610         int32_t ret_val;
9611         struct tcp_bbr *bbr;
9612
9613         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9614         ctf_calc_rwin(so, tp);
9615         if ((thflags & TH_RST) ||
9616             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9617                 return (ctf_process_rst(m, th, so, tp));
9618         /*
9619          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9620          * synchronized state.
9621          */
9622         if (thflags & TH_SYN) {
9623                 ctf_challenge_ack(m, th, tp, &ret_val);
9624                 return (ret_val);
9625         }
9626         /*
9627          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9628          * it's less than ts_recent, drop it.
9629          */
9630         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9631             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9632                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9633                         return (ret_val);
9634         }
9635         INP_WLOCK_ASSERT(tp->t_inpcb);
9636         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9637                 return (ret_val);
9638         }
9639         /*
9640          * If new data are received on a connection after the user processes
9641          * are gone, then RST the other end.
9642          */
9643         if ((so->so_state & SS_NOFDREF) && tlen) {
9644                 /*
9645                  * We call a new function now so we might continue and setup
9646                  * to reset at all data being ack'd.
9647                  */
9648                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9649                         return (1);
9650         }
9651         /*
9652          * If last ACK falls within this segment's sequence numbers, record
9653          * its timestamp. NOTE: 1) That the test incorporates suggestions
9654          * from the latest proposal of the tcplw@cray.com list (Braden
9655          * 1993/04/26). 2) That updating only on newer timestamps interferes
9656          * with our earlier PAWS tests, so this check should be solely
9657          * predicated on the sequence space of this segment. 3) That we
9658          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9659          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9660          * SEG.Len, This modified check allows us to overcome RFC1323's
9661          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9662          * p.869. In such cases, we can still calculate the RTT correctly
9663          * when RCV.NXT == Last.ACK.Sent.
9664          */
9665         if ((to->to_flags & TOF_TS) != 0 &&
9666             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9667             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9668             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9669                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9670                 tp->ts_recent = to->to_tsval;
9671         }
9672         /*
9673          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9674          * is on (half-synchronized state), then queue data for later
9675          * processing; else drop segment and return.
9676          */
9677         if ((thflags & TH_ACK) == 0) {
9678                 if (tp->t_flags & TF_NEEDSYN) {
9679                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9680                             tiwin, thflags, nxt_pkt));
9681                 } else if (tp->t_flags & TF_ACKNOW) {
9682                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9683                         bbr->r_wanted_output = 1;
9684                         return (ret_val);
9685                 } else {
9686                         ctf_do_drop(m, NULL);
9687                         return (0);
9688                 }
9689         }
9690         /*
9691          * Ack processing.
9692          */
9693         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9694                 return (ret_val);
9695         }
9696         if (ourfinisacked) {
9697                 tcp_twstart(tp);
9698                 m_freem(m);
9699                 return (1);
9700         }
9701         if (sbavail(&so->so_snd)) {
9702                 if (ctf_progress_timeout_check(tp, true)) {
9703                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9704                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9705                         return (1);
9706                 }
9707         }
9708         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9709             tiwin, thflags, nxt_pkt));
9710 }
9711
9712 /*
9713  * Return value of 1, the TCB is unlocked and most
9714  * likely gone, return value of 0, the TCB is still
9715  * locked.
9716  */
9717 static int
9718 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9719     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9720     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9721 {
9722         int32_t ourfinisacked = 0;
9723         int32_t ret_val;
9724         struct tcp_bbr *bbr;
9725
9726         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9727         ctf_calc_rwin(so, tp);
9728         if ((thflags & TH_RST) ||
9729             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9730                 return (ctf_process_rst(m, th, so, tp));
9731         /*
9732          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9733          * synchronized state.
9734          */
9735         if (thflags & TH_SYN) {
9736                 ctf_challenge_ack(m, th, tp, &ret_val);
9737                 return (ret_val);
9738         }
9739         /*
9740          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9741          * it's less than ts_recent, drop it.
9742          */
9743         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9744             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9745                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9746                         return (ret_val);
9747         }
9748         INP_WLOCK_ASSERT(tp->t_inpcb);
9749         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9750                 return (ret_val);
9751         }
9752         /*
9753          * If new data are received on a connection after the user processes
9754          * are gone, then RST the other end.
9755          */
9756         if ((so->so_state & SS_NOFDREF) && tlen) {
9757                 /*
9758                  * We call a new function now so we might continue and setup
9759                  * to reset at all data being ack'd.
9760                  */
9761                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9762                         return (1);
9763         }
9764         /*
9765          * If last ACK falls within this segment's sequence numbers, record
9766          * its timestamp. NOTE: 1) That the test incorporates suggestions
9767          * from the latest proposal of the tcplw@cray.com list (Braden
9768          * 1993/04/26). 2) That updating only on newer timestamps interferes
9769          * with our earlier PAWS tests, so this check should be solely
9770          * predicated on the sequence space of this segment. 3) That we
9771          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9772          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9773          * SEG.Len, This modified check allows us to overcome RFC1323's
9774          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9775          * p.869. In such cases, we can still calculate the RTT correctly
9776          * when RCV.NXT == Last.ACK.Sent.
9777          */
9778         if ((to->to_flags & TOF_TS) != 0 &&
9779             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9780             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9781             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9782                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9783                 tp->ts_recent = to->to_tsval;
9784         }
9785         /*
9786          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9787          * is on (half-synchronized state), then queue data for later
9788          * processing; else drop segment and return.
9789          */
9790         if ((thflags & TH_ACK) == 0) {
9791                 if (tp->t_flags & TF_NEEDSYN) {
9792                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9793                             tiwin, thflags, nxt_pkt));
9794                 } else if (tp->t_flags & TF_ACKNOW) {
9795                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9796                         bbr->r_wanted_output = 1;
9797                         return (ret_val);
9798                 } else {
9799                         ctf_do_drop(m, NULL);
9800                         return (0);
9801                 }
9802         }
9803         /*
9804          * case TCPS_LAST_ACK: Ack processing.
9805          */
9806         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9807                 return (ret_val);
9808         }
9809         if (ourfinisacked) {
9810                 tp = tcp_close(tp);
9811                 ctf_do_drop(m, tp);
9812                 return (1);
9813         }
9814         if (sbavail(&so->so_snd)) {
9815                 if (ctf_progress_timeout_check(tp, true)) {
9816                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9817                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9818                         return (1);
9819                 }
9820         }
9821         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9822             tiwin, thflags, nxt_pkt));
9823 }
9824
9825 /*
9826  * Return value of 1, the TCB is unlocked and most
9827  * likely gone, return value of 0, the TCB is still
9828  * locked.
9829  */
9830 static int
9831 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
9832     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9833     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9834 {
9835         int32_t ourfinisacked = 0;
9836         int32_t ret_val;
9837         struct tcp_bbr *bbr;
9838
9839         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9840         ctf_calc_rwin(so, tp);
9841         /* Reset receive buffer auto scaling when not in bulk receive mode. */
9842         if ((thflags & TH_RST) ||
9843             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9844                 return (ctf_process_rst(m, th, so, tp));
9845
9846         /*
9847          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9848          * synchronized state.
9849          */
9850         if (thflags & TH_SYN) {
9851                 ctf_challenge_ack(m, th, tp, &ret_val);
9852                 return (ret_val);
9853         }
9854         INP_WLOCK_ASSERT(tp->t_inpcb);
9855         /*
9856          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9857          * it's less than ts_recent, drop it.
9858          */
9859         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9860             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9861                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9862                         return (ret_val);
9863         }
9864         INP_WLOCK_ASSERT(tp->t_inpcb);
9865         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9866                 return (ret_val);
9867         }
9868         /*
9869          * If new data are received on a connection after the user processes
9870          * are gone, then we may RST the other end depending on the outcome
9871          * of bbr_check_data_after_close.
9872          */
9873         if ((so->so_state & SS_NOFDREF) &&
9874             tlen) {
9875                 /*
9876                  * We call a new function now so we might continue and setup
9877                  * to reset at all data being ack'd.
9878                  */
9879                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9880                         return (1);
9881         }
9882         INP_WLOCK_ASSERT(tp->t_inpcb);
9883         /*
9884          * If last ACK falls within this segment's sequence numbers, record
9885          * its timestamp. NOTE: 1) That the test incorporates suggestions
9886          * from the latest proposal of the tcplw@cray.com list (Braden
9887          * 1993/04/26). 2) That updating only on newer timestamps interferes
9888          * with our earlier PAWS tests, so this check should be solely
9889          * predicated on the sequence space of this segment. 3) That we
9890          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9891          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9892          * SEG.Len, This modified check allows us to overcome RFC1323's
9893          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9894          * p.869. In such cases, we can still calculate the RTT correctly
9895          * when RCV.NXT == Last.ACK.Sent.
9896          */
9897         INP_WLOCK_ASSERT(tp->t_inpcb);
9898         if ((to->to_flags & TOF_TS) != 0 &&
9899             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9900             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9901             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9902                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9903                 tp->ts_recent = to->to_tsval;
9904         }
9905         /*
9906          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9907          * is on (half-synchronized state), then queue data for later
9908          * processing; else drop segment and return.
9909          */
9910         if ((thflags & TH_ACK) == 0) {
9911                 if (tp->t_flags & TF_NEEDSYN) {
9912                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9913                             tiwin, thflags, nxt_pkt));
9914                 } else if (tp->t_flags & TF_ACKNOW) {
9915                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9916                         bbr->r_wanted_output = 1;
9917                         return (ret_val);
9918                 } else {
9919                         ctf_do_drop(m, NULL);
9920                         return (0);
9921                 }
9922         }
9923         /*
9924          * Ack processing.
9925          */
9926         INP_WLOCK_ASSERT(tp->t_inpcb);
9927         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9928                 return (ret_val);
9929         }
9930         if (sbavail(&so->so_snd)) {
9931                 if (ctf_progress_timeout_check(tp, true)) {
9932                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9933                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9934                         return (1);
9935                 }
9936         }
9937         INP_WLOCK_ASSERT(tp->t_inpcb);
9938         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9939             tiwin, thflags, nxt_pkt));
9940 }
9941
9942 static void
9943 bbr_stop_all_timers(struct tcpcb *tp)
9944 {
9945         struct tcp_bbr *bbr;
9946
9947         /*
9948          * Assure no timers are running.
9949          */
9950         if (tcp_timer_active(tp, TT_PERSIST)) {
9951                 /* We enter in persists, set the flag appropriately */
9952                 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9953                 bbr->rc_in_persist = 1;
9954         }
9955         tcp_timer_suspend(tp, TT_PERSIST);
9956         tcp_timer_suspend(tp, TT_REXMT);
9957         tcp_timer_suspend(tp, TT_KEEP);
9958         tcp_timer_suspend(tp, TT_DELACK);
9959 }
9960
9961 static void
9962 bbr_google_mode_on(struct tcp_bbr *bbr)
9963 {
9964         bbr->rc_use_google = 1;
9965         bbr->rc_no_pacing = 0;
9966         bbr->r_ctl.bbr_google_discount = bbr_google_discount;
9967         bbr->r_use_policer = bbr_policer_detection_enabled;
9968         bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
9969         bbr->bbr_use_rack_cheat = 0;
9970         bbr->r_ctl.rc_incr_tmrs = 0;
9971         bbr->r_ctl.rc_inc_tcp_oh = 0;
9972         bbr->r_ctl.rc_inc_ip_oh = 0;
9973         bbr->r_ctl.rc_inc_enet_oh = 0;
9974         reset_time(&bbr->r_ctl.rc_delrate,
9975                    BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
9976         reset_time_small(&bbr->r_ctl.rc_rttprop,
9977                          (11 * USECS_IN_SECOND));
9978         tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
9979 }
9980
9981 static void
9982 bbr_google_mode_off(struct tcp_bbr *bbr)
9983 {
9984         bbr->rc_use_google = 0;
9985         bbr->r_ctl.bbr_google_discount = 0;
9986         bbr->no_pacing_until = bbr_no_pacing_until;
9987         bbr->r_use_policer = 0;
9988         if (bbr->no_pacing_until)
9989                 bbr->rc_no_pacing = 1;
9990         else
9991                 bbr->rc_no_pacing = 0;
9992         if (bbr_use_rack_resend_cheat)
9993                 bbr->bbr_use_rack_cheat = 1;
9994         else
9995                 bbr->bbr_use_rack_cheat = 0;
9996         if (bbr_incr_timers)
9997                 bbr->r_ctl.rc_incr_tmrs = 1;
9998         else
9999                 bbr->r_ctl.rc_incr_tmrs = 0;
10000         if (bbr_include_tcp_oh)
10001                 bbr->r_ctl.rc_inc_tcp_oh = 1;
10002         else
10003                 bbr->r_ctl.rc_inc_tcp_oh = 0;
10004         if (bbr_include_ip_oh)
10005                 bbr->r_ctl.rc_inc_ip_oh = 1;
10006         else
10007                 bbr->r_ctl.rc_inc_ip_oh = 0;
10008         if (bbr_include_enet_oh)
10009                 bbr->r_ctl.rc_inc_enet_oh = 1;
10010         else
10011                 bbr->r_ctl.rc_inc_enet_oh = 0;
10012         bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10013         reset_time(&bbr->r_ctl.rc_delrate,
10014                    bbr_num_pktepo_for_del_limit);
10015         reset_time_small(&bbr->r_ctl.rc_rttprop,
10016                          (bbr_filter_len_sec * USECS_IN_SECOND));
10017         tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
10018 }
10019 /*
10020  * Return 0 on success, non-zero on failure
10021  * which indicates the error (usually no memory).
10022  */
10023 static int
10024 bbr_init(struct tcpcb *tp)
10025 {
10026         struct tcp_bbr *bbr = NULL;
10027         struct inpcb *inp;
10028         uint32_t cts;
10029
10030         tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO));
10031         if (tp->t_fb_ptr == NULL) {
10032                 /*
10033                  * We need to allocate memory but cant. The INP and INP_INFO
10034                  * locks and they are recusive (happens during setup. So a
10035                  * scheme to drop the locks fails :(
10036                  *
10037                  */
10038                 return (ENOMEM);
10039         }
10040         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10041         bbr->rtt_valid = 0;
10042         inp = tp->t_inpcb;
10043         inp->inp_flags2 |= INP_CANNOT_DO_ECN;
10044         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
10045         TAILQ_INIT(&bbr->r_ctl.rc_map);
10046         TAILQ_INIT(&bbr->r_ctl.rc_free);
10047         TAILQ_INIT(&bbr->r_ctl.rc_tmap);
10048         bbr->rc_tp = tp;
10049         if (tp->t_inpcb) {
10050                 bbr->rc_inp = tp->t_inpcb;
10051         }
10052         cts = tcp_get_usecs(&bbr->rc_tv);
10053         tp->t_acktime = 0;
10054         bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close;
10055         bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade;
10056         bbr->rc_tlp_threshold = bbr_tlp_thresh;
10057         bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh;
10058         bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay;
10059         bbr->r_ctl.rc_min_to = bbr_min_to;
10060         bbr->rc_bbr_state = BBR_STATE_STARTUP;
10061         bbr->r_ctl.bbr_lost_at_state = 0;
10062         bbr->r_ctl.rc_lost_at_startup = 0;
10063         bbr->rc_all_timers_stopped = 0;
10064         bbr->r_ctl.rc_bbr_lastbtlbw = 0;
10065         bbr->r_ctl.rc_pkt_epoch_del = 0;
10066         bbr->r_ctl.rc_pkt_epoch = 0;
10067         bbr->r_ctl.rc_lowest_rtt = 0xffffffff;
10068         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain;
10069         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
10070         bbr->r_ctl.rc_went_idle_time = cts;
10071         bbr->rc_pacer_started = cts;
10072         bbr->r_ctl.rc_pkt_epoch_time = cts;
10073         bbr->r_ctl.rc_rcvtime = cts;
10074         bbr->r_ctl.rc_bbr_state_time = cts;
10075         bbr->r_ctl.rc_del_time = cts;
10076         bbr->r_ctl.rc_tlp_rxt_last_time = cts;
10077         bbr->r_ctl.last_in_probertt = cts;
10078         bbr->skip_gain = 0;
10079         bbr->gain_is_limited = 0;
10080         bbr->no_pacing_until = bbr_no_pacing_until;
10081         if (bbr->no_pacing_until)
10082                 bbr->rc_no_pacing = 1;
10083         if (bbr_use_google_algo) {
10084                 bbr->rc_no_pacing = 0;
10085                 bbr->rc_use_google = 1;
10086                 bbr->r_ctl.bbr_google_discount = bbr_google_discount;
10087                 bbr->r_use_policer = bbr_policer_detection_enabled;
10088         } else {
10089                 bbr->rc_use_google = 0;
10090                 bbr->r_ctl.bbr_google_discount = 0;
10091                 bbr->r_use_policer = 0;
10092         }
10093         if (bbr_ts_limiting)
10094                 bbr->rc_use_ts_limit = 1;
10095         else
10096                 bbr->rc_use_ts_limit = 0;
10097         if (bbr_ts_can_raise)
10098                 bbr->ts_can_raise = 1;
10099         else
10100                 bbr->ts_can_raise = 0;
10101         if (V_tcp_delack_enabled == 1)
10102                 tp->t_delayed_ack = 2;
10103         else if (V_tcp_delack_enabled == 0)
10104                 tp->t_delayed_ack = 0;
10105         else if (V_tcp_delack_enabled < 100)
10106                 tp->t_delayed_ack = V_tcp_delack_enabled;
10107         else
10108                 tp->t_delayed_ack = 2;
10109         if (bbr->rc_use_google == 0)
10110                 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10111         else
10112                 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10113         bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms;
10114         bbr->rc_max_rto_sec = bbr_rto_max_sec;
10115         bbr->rc_init_win = bbr_def_init_win;
10116         if (tp->t_flags & TF_REQ_TSTMP)
10117                 bbr->rc_last_options = TCP_TS_OVERHEAD;
10118         bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options;
10119         bbr->r_ctl.rc_high_rwnd = tp->snd_wnd;
10120         bbr->r_init_rtt = 1;
10121
10122         counter_u64_add(bbr_flows_nohdwr_pacing, 1);
10123         if (bbr_allow_hdwr_pacing)
10124                 bbr->bbr_hdw_pace_ena = 1;
10125         else
10126                 bbr->bbr_hdw_pace_ena = 0;
10127         if (bbr_sends_full_iwnd)
10128                 bbr->bbr_init_win_cheat = 1;
10129         else
10130                 bbr->bbr_init_win_cheat = 0;
10131         bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max;
10132         bbr->r_ctl.rc_drain_pg = bbr_drain_gain;
10133         bbr->r_ctl.rc_startup_pg = bbr_high_gain;
10134         bbr->rc_loss_exit = bbr_exit_startup_at_loss;
10135         bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain;
10136         bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second;
10137         bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar;
10138         bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max;
10139         bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor;
10140         bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min;
10141         bbr->r_ctl.bbr_cross_over = bbr_cross_over;
10142         bbr->r_ctl.rc_rtt_shrinks = cts;
10143         if (bbr->rc_use_google) {
10144                 setup_time_filter(&bbr->r_ctl.rc_delrate,
10145                                   FILTER_TYPE_MAX,
10146                                   BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10147                 setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10148                                         FILTER_TYPE_MIN, (11 * USECS_IN_SECOND));
10149         } else {
10150                 setup_time_filter(&bbr->r_ctl.rc_delrate,
10151                                   FILTER_TYPE_MAX,
10152                                   bbr_num_pktepo_for_del_limit);
10153                 setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10154                                         FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND));
10155         }
10156         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0);
10157         if (bbr_uses_idle_restart)
10158                 bbr->rc_use_idle_restart = 1;
10159         else
10160                 bbr->rc_use_idle_restart = 0;
10161         bbr->r_ctl.rc_bbr_cur_del_rate = 0;
10162         bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps;
10163         if (bbr_resends_use_tso)
10164                 bbr->rc_resends_use_tso = 1;
10165 #ifdef NETFLIX_PEAKRATE
10166         tp->t_peakrate_thr = tp->t_maxpeakrate;
10167 #endif
10168         if (tp->snd_una != tp->snd_max) {
10169                 /* Create a send map for the current outstanding data */
10170                 struct bbr_sendmap *rsm;
10171
10172                 rsm = bbr_alloc(bbr);
10173                 if (rsm == NULL) {
10174                         uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10175                         tp->t_fb_ptr = NULL;
10176                         return (ENOMEM);
10177                 }
10178                 rsm->r_rtt_not_allowed = 1;
10179                 rsm->r_tim_lastsent[0] = cts;
10180                 rsm->r_rtr_cnt = 1;
10181                 rsm->r_rtr_bytes = 0;
10182                 rsm->r_start = tp->snd_una;
10183                 rsm->r_end = tp->snd_max;
10184                 rsm->r_dupack = 0;
10185                 rsm->r_delivered = bbr->r_ctl.rc_delivered;
10186                 rsm->r_ts_valid = 0;
10187                 rsm->r_del_ack_ts = tp->ts_recent;
10188                 rsm->r_del_time = cts;
10189                 if (bbr->r_ctl.r_app_limited_until)
10190                         rsm->r_app_limited = 1;
10191                 else
10192                         rsm->r_app_limited = 0;
10193                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
10194                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
10195                 rsm->r_in_tmap = 1;
10196                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
10197                         rsm->r_bbr_state = bbr_state_val(bbr);
10198                 else
10199                         rsm->r_bbr_state = 8;
10200         }
10201         if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0))
10202                 bbr->bbr_use_rack_cheat = 1;
10203         if (bbr_incr_timers && (bbr->rc_use_google == 0))
10204                 bbr->r_ctl.rc_incr_tmrs = 1;
10205         if (bbr_include_tcp_oh && (bbr->rc_use_google == 0))
10206                 bbr->r_ctl.rc_inc_tcp_oh = 1;
10207         if (bbr_include_ip_oh && (bbr->rc_use_google == 0))
10208                 bbr->r_ctl.rc_inc_ip_oh = 1;
10209         if (bbr_include_enet_oh && (bbr->rc_use_google == 0))
10210                 bbr->r_ctl.rc_inc_enet_oh = 1;
10211
10212         bbr_log_type_statechange(bbr, cts, __LINE__);
10213         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
10214             (tp->t_srtt)) {
10215                 uint32_t rtt;
10216
10217                 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
10218                 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
10219         }
10220         /* announce the settings and state */
10221         bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT);
10222         tcp_bbr_tso_size_check(bbr, cts);
10223         /*
10224          * Now call the generic function to start a timer. This will place
10225          * the TCB on the hptsi wheel if a timer is needed with appropriate
10226          * flags.
10227          */
10228         bbr_stop_all_timers(tp);
10229         bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0);
10230         return (0);
10231 }
10232
10233 /*
10234  * Return 0 if we can accept the connection. Return
10235  * non-zero if we can't handle the connection. A EAGAIN
10236  * means you need to wait until the connection is up.
10237  * a EADDRNOTAVAIL means we can never handle the connection
10238  * (no SACK).
10239  */
10240 static int
10241 bbr_handoff_ok(struct tcpcb *tp)
10242 {
10243         if ((tp->t_state == TCPS_CLOSED) ||
10244             (tp->t_state == TCPS_LISTEN)) {
10245                 /* Sure no problem though it may not stick */
10246                 return (0);
10247         }
10248         if ((tp->t_state == TCPS_SYN_SENT) ||
10249             (tp->t_state == TCPS_SYN_RECEIVED)) {
10250                 /*
10251                  * We really don't know you have to get to ESTAB or beyond
10252                  * to tell.
10253                  */
10254                 return (EAGAIN);
10255         }
10256         if (tp->t_flags & TF_SENTFIN)
10257                 return (EINVAL);
10258         if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) {
10259                 return (0);
10260         }
10261         /*
10262          * If we reach here we don't do SACK on this connection so we can
10263          * never do rack.
10264          */
10265         return (EINVAL);
10266 }
10267
10268 static void
10269 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged)
10270 {
10271         if (tp->t_fb_ptr) {
10272                 uint32_t calc;
10273                 struct tcp_bbr *bbr;
10274                 struct bbr_sendmap *rsm;
10275
10276                 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10277                 if (bbr->r_ctl.crte)
10278                         tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
10279                 bbr_log_flowend(bbr);
10280                 bbr->rc_tp = NULL;
10281                 if (tp->t_inpcb) {
10282                         /* Backout any flags2 we applied */
10283                         tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN;
10284                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
10285                         tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
10286                 }
10287                 if (bbr->bbr_hdrw_pacing)
10288                         counter_u64_add(bbr_flows_whdwr_pacing, -1);
10289                 else
10290                         counter_u64_add(bbr_flows_nohdwr_pacing, -1);
10291                 if (bbr->r_ctl.crte != NULL) {
10292                         tcp_rel_pacing_rate(bbr->r_ctl.crte, tp);
10293                         bbr->r_ctl.crte = NULL;
10294                 }
10295                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10296                 while (rsm) {
10297                         TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
10298                         uma_zfree(bbr_zone, rsm);
10299                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10300                 }
10301                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10302                 while (rsm) {
10303                         TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
10304                         uma_zfree(bbr_zone, rsm);
10305                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10306                 }
10307                 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd;
10308                 if (calc > (bbr->r_ctl.rc_init_rwnd / 10))
10309                         BBR_STAT_INC(bbr_dynamic_rwnd);
10310                 else
10311                         BBR_STAT_INC(bbr_static_rwnd);
10312                 bbr->r_ctl.rc_free_cnt = 0;
10313                 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10314                 tp->t_fb_ptr = NULL;
10315         }
10316         /* Make sure snd_nxt is correctly set */
10317         tp->snd_nxt = tp->snd_max;
10318 }
10319
10320 static void
10321 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win)
10322 {
10323         switch (tp->t_state) {
10324         case TCPS_SYN_SENT:
10325                 bbr->r_state = TCPS_SYN_SENT;
10326                 bbr->r_substate = bbr_do_syn_sent;
10327                 break;
10328         case TCPS_SYN_RECEIVED:
10329                 bbr->r_state = TCPS_SYN_RECEIVED;
10330                 bbr->r_substate = bbr_do_syn_recv;
10331                 break;
10332         case TCPS_ESTABLISHED:
10333                 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd);
10334                 bbr->r_state = TCPS_ESTABLISHED;
10335                 bbr->r_substate = bbr_do_established;
10336                 break;
10337         case TCPS_CLOSE_WAIT:
10338                 bbr->r_state = TCPS_CLOSE_WAIT;
10339                 bbr->r_substate = bbr_do_close_wait;
10340                 break;
10341         case TCPS_FIN_WAIT_1:
10342                 bbr->r_state = TCPS_FIN_WAIT_1;
10343                 bbr->r_substate = bbr_do_fin_wait_1;
10344                 break;
10345         case TCPS_CLOSING:
10346                 bbr->r_state = TCPS_CLOSING;
10347                 bbr->r_substate = bbr_do_closing;
10348                 break;
10349         case TCPS_LAST_ACK:
10350                 bbr->r_state = TCPS_LAST_ACK;
10351                 bbr->r_substate = bbr_do_lastack;
10352                 break;
10353         case TCPS_FIN_WAIT_2:
10354                 bbr->r_state = TCPS_FIN_WAIT_2;
10355                 bbr->r_substate = bbr_do_fin_wait_2;
10356                 break;
10357         case TCPS_LISTEN:
10358         case TCPS_CLOSED:
10359         case TCPS_TIME_WAIT:
10360         default:
10361                 break;
10362         };
10363 }
10364
10365 static void
10366 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog)
10367 {
10368         /*
10369          * Now what state are we going into now? Is there adjustments
10370          * needed?
10371          */
10372         int32_t old_state, old_gain;
10373
10374         old_state = bbr_state_val(bbr);
10375         old_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
10376         if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) {
10377                 /* Save the lowest srtt we saw in our end of the sub-state */
10378                 bbr->rc_hit_state_1 = 0;
10379                 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff)
10380                         bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state;
10381         }
10382         bbr->rc_bbr_substate++;
10383         if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) {
10384                 /* Cycle back to first state-> gain */
10385                 bbr->rc_bbr_substate = 0;
10386         }
10387         if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10388                 /*
10389                  * We enter the gain(5/4) cycle (possibly less if
10390                  * shallow buffer detection is enabled)
10391                  */
10392                 if (bbr->skip_gain) {
10393                         /*
10394                          * Hardware pacing has set our rate to
10395                          * the max and limited our b/w just
10396                          * do level i.e. no gain.
10397                          */
10398                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1];
10399                 } else if (bbr->gain_is_limited &&
10400                            bbr->bbr_hdrw_pacing &&
10401                            bbr->r_ctl.crte) {
10402                         /*
10403                          * We can't gain above the hardware pacing
10404                          * rate which is less than our rate + the gain
10405                          * calculate the gain needed to reach the hardware
10406                          * pacing rate..
10407                          */
10408                         uint64_t bw, rate, gain_calc;
10409
10410                         bw = bbr_get_bw(bbr);
10411                         rate = bbr->r_ctl.crte->rate;
10412                         if ((rate > bw) &&
10413                             (((bw *  (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) {
10414                                 gain_calc = (rate * BBR_UNIT) / bw;
10415                                 if (gain_calc < BBR_UNIT)
10416                                         gain_calc = BBR_UNIT;
10417                                 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc;
10418                         } else {
10419                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10420                         }
10421                 } else
10422                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10423                 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) {
10424                         bbr->r_ctl.rc_bbr_state_atflight = cts;
10425                 } else
10426                         bbr->r_ctl.rc_bbr_state_atflight = 0;
10427         } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10428                 bbr->rc_hit_state_1 = 1;
10429                 bbr->r_ctl.rc_exta_time_gd = 0;
10430                 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10431                                                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10432                 if (bbr_state_drain_2_tar) {
10433                         bbr->r_ctl.rc_bbr_state_atflight = 0;
10434                 } else
10435                         bbr->r_ctl.rc_bbr_state_atflight = cts;
10436                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN];
10437         } else {
10438                 /* All other cycles hit here 2-7 */
10439                 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) {
10440                         if (bbr_sub_drain_slam_cwnd &&
10441                             (bbr->rc_use_google == 0) &&
10442                             (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10443                                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10444                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10445                         }
10446                         if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP))
10447                                 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) -
10448                                                                bbr_get_rtt(bbr, BBR_RTT_PROP));
10449                         else
10450                                 bbr->r_ctl.rc_exta_time_gd = 0;
10451                         if (bbr->r_ctl.rc_exta_time_gd) {
10452                                 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd;
10453                                 /* Now chop up the time for each state (div by 7) */
10454                                 bbr->r_ctl.rc_level_state_extra /= 7;
10455                                 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) {
10456                                         /* Add a randomization */
10457                                         bbr_randomize_extra_state_time(bbr);
10458                                 }
10459                         }
10460                 }
10461                 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10462                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)];
10463         }
10464         if (bbr->rc_use_google) {
10465                 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10466         }
10467         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10468         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10469         if (dolog)
10470                 bbr_log_type_statechange(bbr, cts, line);
10471
10472         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10473                 uint32_t time_in;
10474
10475                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10476                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
10477                         counter_u64_add(bbr_state_time[(old_state + 5)], time_in);
10478                 } else {
10479                         counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10480                 }
10481         }
10482         bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
10483         bbr_set_state_target(bbr, __LINE__);
10484         if (bbr_sub_drain_slam_cwnd &&
10485             (bbr->rc_use_google == 0) &&
10486             (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10487                 /* Slam down the cwnd */
10488                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10489                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10490                 if (bbr_sub_drain_app_limit) {
10491                         /* Go app limited if we are on a long drain */
10492                         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered +
10493                                                           ctf_flight_size(bbr->rc_tp,
10494                                                               (bbr->r_ctl.rc_sacked +
10495                                                                bbr->r_ctl.rc_lost_bytes)));
10496                 }
10497                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10498         }
10499         if (bbr->rc_lt_use_bw) {
10500                 /* In policed mode we clamp pacing_gain to BBR_UNIT */
10501                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10502         }
10503         /* Google changes TSO size every cycle */
10504         if (bbr->rc_use_google)
10505                 tcp_bbr_tso_size_check(bbr, cts);
10506         bbr->r_ctl.gain_epoch = cts;
10507         bbr->r_ctl.rc_bbr_state_time = cts;
10508         bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch;
10509 }
10510
10511 static void
10512 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10513 {
10514         if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) &&
10515             (google_allow_early_out == 1) &&
10516             (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) {
10517                 /* We have reached out target flight size possibly early */
10518                 goto change_state;
10519         }
10520         if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10521                 return;
10522         }
10523         if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) {
10524                 /*
10525                  * Must be a rttProp movement forward before
10526                  * we can change states.
10527                  */
10528                 return;
10529         }
10530         if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10531                 /*
10532                  * The needed time has passed but for
10533                  * the gain cycle extra rules apply:
10534                  * 1) If we have seen loss, we exit
10535                  * 2) If we have not reached the target
10536                  *    we stay in GAIN (gain-to-target).
10537                  */
10538                 if (google_consider_lost && losses)
10539                         goto change_state;
10540                 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) {
10541                         return;
10542                 }
10543         }
10544 change_state:
10545         /* For gain we must reach our target, all others last 1 rttProp */
10546         bbr_substate_change(bbr, cts, __LINE__, 1);
10547 }
10548
10549 static void
10550 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10551 {
10552         uint32_t flight, bbr_cur_cycle_time;
10553
10554         if (bbr->rc_use_google) {
10555                 bbr_set_probebw_google_gains(bbr, cts, losses);
10556                 return;
10557         }
10558         if (cts == 0) {
10559                 /*
10560                  * Never alow cts to be 0 we
10561                  * do this so we can judge if
10562                  * we have set a timestamp.
10563                  */
10564                 cts = 1;
10565         }
10566         if (bbr_state_is_pkt_epoch)
10567                 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
10568         else
10569                 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP);
10570
10571         if (bbr->r_ctl.rc_bbr_state_atflight == 0) {
10572                 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10573                         flight = ctf_flight_size(bbr->rc_tp,
10574                                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10575                         if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) {
10576                                 /* Keep it slam down */
10577                                 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) {
10578                                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10579                                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10580                                 }
10581                                 if (bbr_sub_drain_app_limit) {
10582                                         /* Go app limited if we are on a long drain */
10583                                         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight);
10584                                 }
10585                         }
10586                         if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) &&
10587                             (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) ||
10588                              (flight >= bbr->r_ctl.flightsize_at_drain))) {
10589                                 /*
10590                                  * Still here after the same time as
10591                                  * the gain. We need to drain harder
10592                                  * for the next srtt. Reduce by a set amount
10593                                  * the gain drop is capped at DRAIN states
10594                                  * value (88).
10595                                  */
10596                                 bbr->r_ctl.flightsize_at_drain = flight;
10597                                 if (bbr_drain_drop_mul &&
10598                                     bbr_drain_drop_div &&
10599                                     (bbr_drain_drop_mul < bbr_drain_drop_div)) {
10600                                         /* Use your specific drop value (def 4/5 = 20%) */
10601                                         bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul;
10602                                         bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div;
10603                                 } else {
10604                                         /* You get drop of 20% */
10605                                         bbr->r_ctl.rc_bbr_hptsi_gain *= 4;
10606                                         bbr->r_ctl.rc_bbr_hptsi_gain /= 5;
10607                                 }
10608                                 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) {
10609                                         /* Reduce our gain again to the bottom  */
10610                                         bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
10611                                 }
10612                                 bbr_log_exit_gain(bbr, cts, 4);
10613                                 /*
10614                                  * Extend out so we wait another
10615                                  * epoch before dropping again.
10616                                  */
10617                                 bbr->r_ctl.gain_epoch = cts;
10618                         }
10619                         if (flight <= bbr->r_ctl.rc_target_at_state) {
10620                                 if (bbr_sub_drain_slam_cwnd &&
10621                                     (bbr->rc_use_google == 0) &&
10622                                     (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10623                                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10624                                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10625                                 }
10626                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10627                                 bbr_log_exit_gain(bbr, cts, 3);
10628                         }
10629                 } else {
10630                         /* Its a gain  */
10631                         if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) {
10632                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10633                                 goto change_state;
10634                         }
10635                         if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) ||
10636                             ((ctf_outstanding(bbr->rc_tp) +  bbr->rc_tp->t_maxseg - 1) >=
10637                              bbr->rc_tp->snd_wnd)) {
10638                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10639                                 bbr_log_exit_gain(bbr, cts, 2);
10640                         }
10641                 }
10642                 /**
10643                  * We fall through and return always one of two things has
10644                  * occurred.
10645                  * 1) We are still not at target
10646                  *    <or>
10647                  * 2) We reached the target and set rc_bbr_state_atflight
10648                  *    which means we no longer hit this block
10649                  *    next time we are called.
10650                  */
10651                 return;
10652         }
10653 change_state:
10654         if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time))
10655                 return;
10656         if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) {
10657                 /* Less than a full time-period has passed */
10658                 return;
10659         }
10660         if (bbr->r_ctl.rc_level_state_extra &&
10661             (bbr_state_val(bbr) > BBR_SUB_DRAIN) &&
10662             ((cts - bbr->r_ctl.rc_bbr_state_time) <
10663              (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10664                 /* Less than a full time-period + extra has passed */
10665                 return;
10666         }
10667         if (bbr_gain_gets_extra_too &&
10668             bbr->r_ctl.rc_level_state_extra &&
10669             (bbr_state_val(bbr) == BBR_SUB_GAIN) &&
10670             ((cts - bbr->r_ctl.rc_bbr_state_time) <
10671              (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10672                 /* Less than a full time-period + extra has passed */
10673                 return;
10674         }
10675         bbr_substate_change(bbr, cts, __LINE__, 1);
10676 }
10677
10678 static uint32_t
10679 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain)
10680 {
10681         uint32_t mss, tar;
10682
10683         if (bbr->rc_use_google) {
10684                 /* Google just uses the cwnd target */
10685                 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain);
10686         } else {
10687                 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options),
10688                           bbr->r_ctl.rc_pace_max_segs);
10689                 /* Get the base cwnd with gain rounded to a mss */
10690                 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr),
10691                                                       gain), mss);
10692                 /* Make sure it is within our min */
10693                 if (tar < get_min_cwnd(bbr))
10694                         return (get_min_cwnd(bbr));
10695         }
10696         return (tar);
10697 }
10698
10699 static void
10700 bbr_set_state_target(struct tcp_bbr *bbr, int line)
10701 {
10702         uint32_t tar, meth;
10703
10704         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
10705             ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
10706                 /* Special case using old probe-rtt method */
10707                 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10708                 meth = 1;
10709         } else {
10710                 /* Non-probe-rtt case and reduced probe-rtt  */
10711                 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
10712                     (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) {
10713                         /* For gain cycle we use the hptsi gain */
10714                         tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10715                         meth = 2;
10716                 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) {
10717                         /*
10718                          * If configured, or for google all other states
10719                          * get BBR_UNIT.
10720                          */
10721                         tar = bbr_get_a_state_target(bbr, BBR_UNIT);
10722                         meth = 3;
10723                 } else {
10724                         /*
10725                          * Or we set a target based on the pacing gain
10726                          * for non-google mode and default (non-configured).
10727                          * Note we don't set a target goal below drain (192).
10728                          */
10729                         if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN])  {
10730                                 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]);
10731                                 meth = 4;
10732                         } else {
10733                                 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10734                                 meth = 5;
10735                         }
10736                 }
10737         }
10738         bbr_log_set_of_state_target(bbr, tar, line, meth);
10739         bbr->r_ctl.rc_target_at_state = tar;
10740 }
10741
10742 static void
10743 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
10744 {
10745         /* Change to probe_rtt */
10746         uint32_t time_in;
10747
10748         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10749         bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10750                                              (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10751         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain
10752                                           + bbr->r_ctl.rc_delivered);
10753         /* Setup so we force feed the filter */
10754         if (bbr->rc_use_google || bbr_probertt_sets_rtt)
10755                 bbr->rc_prtt_set_ts = 1;
10756         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10757                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10758                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10759         }
10760         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0);
10761         bbr->r_ctl.rc_rtt_shrinks = cts;
10762         bbr->r_ctl.last_in_probertt = cts;
10763         bbr->r_ctl.rc_probertt_srttchktim = cts;
10764         bbr->r_ctl.rc_bbr_state_time = cts;
10765         bbr->rc_bbr_state = BBR_STATE_PROBE_RTT;
10766         /* We need to force the filter to update */
10767
10768         if ((bbr_sub_drain_slam_cwnd) &&
10769             bbr->rc_hit_state_1 &&
10770             (bbr->rc_use_google == 0) &&
10771             (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10772                 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd)
10773                         bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10774         } else
10775                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10776         /* Update the lost */
10777         bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10778         if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){
10779                 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN)  */
10780                 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10781                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10782                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10783                 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10784                 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6);
10785                 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd;
10786         } else {
10787                 /*
10788                  * We bring it down slowly by using a hptsi gain that is
10789                  * probably 75%. This will slowly float down our outstanding
10790                  * without tampering with the cwnd.
10791                  */
10792                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
10793                 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10794                 bbr_set_state_target(bbr, __LINE__);
10795                 if (bbr_prtt_slam_cwnd &&
10796                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
10797                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10798                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10799                 }
10800         }
10801         if (ctf_flight_size(bbr->rc_tp,
10802                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
10803             bbr->r_ctl.rc_target_at_state) {
10804                 /* We are at target */
10805                 bbr->r_ctl.rc_bbr_enters_probertt = cts;
10806         } else {
10807                 /* We need to come down to reach target before our time begins */
10808                 bbr->r_ctl.rc_bbr_enters_probertt = 0;
10809         }
10810         bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch;
10811         BBR_STAT_INC(bbr_enter_probertt);
10812         bbr_log_exit_gain(bbr, cts, 0);
10813         bbr_log_type_statechange(bbr, cts, line);
10814 }
10815
10816 static void
10817 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts)
10818 {
10819         /*
10820          * Sanity check on probe-rtt intervals.
10821          * In crazy situations where we are competing
10822          * against new-reno flows with huge buffers
10823          * our rtt-prop interval could come to dominate
10824          * things if we can't get through a full set
10825          * of cycles, we need to adjust it.
10826          */
10827         if (bbr_can_adjust_probertt &&
10828             (bbr->rc_use_google == 0)) {
10829                 uint16_t val = 0;
10830                 uint32_t cur_rttp, fval, newval, baseval;
10831
10832                 /* Are we to small and go into probe-rtt to often? */
10833                 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1));
10834                 cur_rttp = roundup(baseval, USECS_IN_SECOND);
10835                 fval = bbr_filter_len_sec * USECS_IN_SECOND;
10836                 if (bbr_is_ratio == 0) {
10837                         if (fval > bbr_rtt_probe_limit)
10838                                 newval = cur_rttp + (fval - bbr_rtt_probe_limit);
10839                         else
10840                                 newval = cur_rttp;
10841                 } else {
10842                         int mul;
10843
10844                         mul = fval / bbr_rtt_probe_limit;
10845                         newval = cur_rttp * mul;
10846                 }
10847                 if (cur_rttp >  bbr->r_ctl.rc_probertt_int) {
10848                         bbr->r_ctl.rc_probertt_int = cur_rttp;
10849                         reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10850                         val = 1;
10851                 } else {
10852                         /*
10853                          * No adjustments were made
10854                          * do we need to shrink it?
10855                          */
10856                         if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) {
10857                                 if (cur_rttp <= bbr_rtt_probe_limit) {
10858                                         /*
10859                                          * Things have calmed down lets
10860                                          * shrink all the way to default
10861                                          */
10862                                         bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10863                                         reset_time_small(&bbr->r_ctl.rc_rttprop,
10864                                                          (bbr_filter_len_sec * USECS_IN_SECOND));
10865                                         cur_rttp = bbr_rtt_probe_limit;
10866                                         newval = (bbr_filter_len_sec * USECS_IN_SECOND);
10867                                         val = 2;
10868                                 } else {
10869                                         /*
10870                                          * Well does some adjustment make sense?
10871                                          */
10872                                         if (cur_rttp < bbr->r_ctl.rc_probertt_int) {
10873                                                 /* We can reduce interval time some */
10874                                                 bbr->r_ctl.rc_probertt_int = cur_rttp;
10875                                                 reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10876                                                 val = 3;
10877                                         }
10878                                 }
10879                         }
10880                 }
10881                 if (val)
10882                         bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val);
10883         }
10884 }
10885
10886 static void
10887 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
10888 {
10889         /* Exit probe-rtt */
10890
10891         if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) {
10892                 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10893                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10894         }
10895         bbr_log_exit_gain(bbr, cts, 1);
10896         bbr->rc_hit_state_1 = 0;
10897         bbr->r_ctl.rc_rtt_shrinks = cts;
10898         bbr->r_ctl.last_in_probertt = cts;
10899         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0);
10900         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10901         bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp,
10902                                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
10903                                           bbr->r_ctl.rc_delivered);
10904         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10905                 uint32_t time_in;
10906
10907                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10908                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10909         }
10910         if (bbr->rc_filled_pipe) {
10911                 /* Switch to probe_bw */
10912                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
10913                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
10914                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10915                 bbr_substate_change(bbr, cts, __LINE__, 0);
10916                 bbr_log_type_statechange(bbr, cts, __LINE__);
10917         } else {
10918                 /* Back to startup */
10919                 bbr->rc_bbr_state = BBR_STATE_STARTUP;
10920                 bbr->r_ctl.rc_bbr_state_time = cts;
10921                 /*
10922                  * We don't want to give a complete free 3
10923                  * measurements until we exit, so we use
10924                  * the number of pe's we were in probe-rtt
10925                  * to add to the startup_epoch. That way
10926                  * we will still retain the old state.
10927                  */
10928                 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt);
10929                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10930                 /* Make sure to use the lower pg when shifting back in */
10931                 if (bbr->r_ctl.rc_lost &&
10932                     bbr_use_lower_gain_in_startup &&
10933                     (bbr->rc_use_google == 0))
10934                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
10935                 else
10936                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
10937                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
10938                 /* Probably not needed but set it anyway */
10939                 bbr_set_state_target(bbr, __LINE__);
10940                 bbr_log_type_statechange(bbr, cts, __LINE__);
10941                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10942                     bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0);
10943         }
10944         bbr_check_probe_rtt_limits(bbr, cts);
10945 }
10946
10947 static int32_t inline
10948 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts)
10949 {
10950         if ((bbr->rc_past_init_win == 1) &&
10951             (bbr->rc_in_persist == 0) &&
10952             (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) {
10953                 return (1);
10954         }
10955         if (bbr_can_force_probertt &&
10956             (bbr->rc_in_persist == 0) &&
10957             (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
10958             ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
10959                 return (1);
10960         }
10961         return (0);
10962 }
10963
10964 static int32_t
10965 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t  pkt_epoch)
10966 {
10967         uint64_t btlbw, gain;
10968         if (pkt_epoch == 0) {
10969                 /*
10970                  * Need to be on a pkt-epoch to continue.
10971                  */
10972                 return (0);
10973         }
10974         btlbw = bbr_get_full_bw(bbr);
10975         gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
10976                  (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
10977         if (btlbw >= gain) {
10978                 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
10979                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10980                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
10981                 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
10982         }
10983         if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)
10984                 return (1);
10985         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10986                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
10987         return(0);
10988 }
10989
10990 static int32_t inline
10991 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch)
10992 {
10993         /* Have we gained 25% in the last 3 packet based epoch's? */
10994         uint64_t btlbw, gain;
10995         int do_exit;
10996         int delta, rtt_gain;
10997
10998         if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
10999             (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11000                 /*
11001                  * This qualifies as a RTT_PROBE session since we drop the
11002                  * data outstanding to nothing and waited more than
11003                  * bbr_rtt_probe_time.
11004                  */
11005                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11006                 bbr_set_reduced_rtt(bbr, cts, __LINE__);
11007         }
11008         if (bbr_should_enter_probe_rtt(bbr, cts)) {
11009                 bbr_enter_probe_rtt(bbr, cts, __LINE__);
11010                 return (0);
11011         }
11012         if (bbr->rc_use_google)
11013                 return (bbr_google_startup(bbr, cts,  pkt_epoch));
11014
11015         if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11016             (bbr_use_lower_gain_in_startup)) {
11017                 /* Drop to a lower gain 1.5 x since we saw loss */
11018                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
11019         }
11020         if (pkt_epoch == 0) {
11021                 /*
11022                  * Need to be on a pkt-epoch to continue.
11023                  */
11024                 return (0);
11025         }
11026         if (bbr_rtt_gain_thresh) {
11027                 /*
11028                  * Do we allow a flow to stay
11029                  * in startup with no loss and no
11030                  * gain in rtt over a set threshold?
11031                  */
11032                 if (bbr->r_ctl.rc_pkt_epoch_rtt &&
11033                     bbr->r_ctl.startup_last_srtt &&
11034                     (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) {
11035                         delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt;
11036                         rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt;
11037                 } else
11038                         rtt_gain = 0;
11039                 if ((bbr->r_ctl.startup_last_srtt == 0)  ||
11040                     (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt))
11041                         /* First time or new lower value */
11042                         bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
11043
11044                 if ((bbr->r_ctl.rc_lost == 0) &&
11045                     (rtt_gain < bbr_rtt_gain_thresh)) {
11046                         /*
11047                          * No loss, and we are under
11048                          * our gain threhold for
11049                          * increasing RTT.
11050                          */
11051                         if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11052                                 bbr->r_ctl.rc_bbr_last_startup_epoch++;
11053                         bbr_log_startup_event(bbr, cts, rtt_gain,
11054                                               delta, bbr->r_ctl.startup_last_srtt, 10);
11055                         return (0);
11056                 }
11057         }
11058         if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) &&
11059             (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) &&
11060             (!IN_RECOVERY(bbr->rc_tp->t_flags))) {
11061                 /*
11062                  * We only assess if we have a new measurment when
11063                  * we have no loss and are not in recovery.
11064                  * Drag up by one our last_startup epoch so we will hold
11065                  * the number of non-gain we have already accumulated.
11066                  */
11067                 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11068                         bbr->r_ctl.rc_bbr_last_startup_epoch++;
11069                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11070                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9);
11071                 return (0);
11072         }
11073         /* Case where we reduced the lost (bad retransmit) */
11074         if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost)
11075                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11076         bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count;
11077         btlbw = bbr_get_full_bw(bbr);
11078         if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower)
11079                 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11080                          (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11081         else
11082                 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11083                          (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11084         do_exit = 0;
11085         if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw)
11086                 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
11087         if (btlbw >= gain) {
11088                 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11089                 /* Update the lost so we won't exit in next set of tests */
11090                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11091                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11092                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11093         }
11094         if ((bbr->rc_loss_exit &&
11095              (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11096              (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) &&
11097             ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) {
11098                 /*
11099                  * If we had no gain,  we had loss and that loss was above
11100                  * our threshould, the rwnd is not constrained, and we have
11101                  * had at least 3 packet epochs exit. Note that this is
11102                  * switched off by sysctl. Google does not do this by the
11103                  * way.
11104                  */
11105                 if ((ctf_flight_size(bbr->rc_tp,
11106                          (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
11107                      (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) {
11108                         do_exit = 1;
11109                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11110                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4);
11111                 } else {
11112                         /* Just record an updated loss value */
11113                         bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11114                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11115                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5);
11116                 }
11117         } else
11118                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11119         if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) ||
11120             do_exit) {
11121                 /* Return 1 to exit the startup state. */
11122                 return (1);
11123         }
11124         /* Stay in startup */
11125         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11126                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
11127         return (0);
11128 }
11129
11130 static void
11131 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses)
11132 {
11133         /*
11134          * A tick occurred in the rtt epoch do we need to do anything?
11135          */
11136 #ifdef BBR_INVARIANTS
11137         if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
11138             (bbr->rc_bbr_state != BBR_STATE_DRAIN) &&
11139             (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) &&
11140             (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
11141             (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) {
11142                 /* Debug code? */
11143                 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state);
11144         }
11145 #endif
11146         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
11147                 /* Do we exit the startup state? */
11148                 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) {
11149                         uint32_t time_in;
11150
11151                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11152                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6);
11153                         bbr->rc_filled_pipe = 1;
11154                         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11155                         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11156                                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11157                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11158                         } else
11159                                 time_in = 0;
11160                         if (bbr->rc_no_pacing)
11161                                 bbr->rc_no_pacing = 0;
11162                         bbr->r_ctl.rc_bbr_state_time = cts;
11163                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg;
11164                         bbr->rc_bbr_state = BBR_STATE_DRAIN;
11165                         bbr_set_state_target(bbr, __LINE__);
11166                         if ((bbr->rc_use_google == 0) &&
11167                             bbr_slam_cwnd_in_main_drain) {
11168                                 /* Here we don't have to worry about probe-rtt */
11169                                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
11170                                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11171                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11172                         }
11173                         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
11174                         bbr_log_type_statechange(bbr, cts, __LINE__);
11175                         if (ctf_flight_size(bbr->rc_tp,
11176                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
11177                             bbr->r_ctl.rc_target_at_state) {
11178                                 /*
11179                                  * Switch to probe_bw if we are already
11180                                  * there
11181                                  */
11182                                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11183                                 bbr_substate_change(bbr, cts, __LINE__, 0);
11184                                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11185                                 bbr_log_type_statechange(bbr, cts, __LINE__);
11186                         }
11187                 }
11188         } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) {
11189                 uint32_t inflight;
11190                 struct tcpcb *tp;
11191
11192                 tp = bbr->rc_tp;
11193                 inflight = ctf_flight_size(tp,
11194                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11195                 if (inflight >= bbr->r_ctl.rc_target_at_state) {
11196                         /* We have reached a flight of the cwnd target */
11197                         bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11198                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11199                         bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
11200                         bbr_set_state_target(bbr, __LINE__);
11201                         /*
11202                          * Rig it so we don't do anything crazy and
11203                          * start fresh with a new randomization.
11204                          */
11205                         bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
11206                         bbr->rc_bbr_substate = BBR_SUB_LEVEL6;
11207                         bbr_substate_change(bbr, cts, __LINE__, 1);
11208                 }
11209         } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) {
11210                 /* Has in-flight reached the bdp (or less)? */
11211                 uint32_t inflight;
11212                 struct tcpcb *tp;
11213
11214                 tp = bbr->rc_tp;
11215                 inflight = ctf_flight_size(tp,
11216                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11217                 if ((bbr->rc_use_google == 0) &&
11218                     bbr_slam_cwnd_in_main_drain &&
11219                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11220                         /*
11221                          * Here we don't have to worry about probe-rtt
11222                          * re-slam it, but keep it slammed down.
11223                          */
11224                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11225                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11226                 }
11227                 if (inflight <= bbr->r_ctl.rc_target_at_state) {
11228                         /* We have drained */
11229                         bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11230                         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11231                         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11232                                 uint32_t time_in;
11233
11234                                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11235                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11236                         }
11237                         if ((bbr->rc_use_google == 0) &&
11238                             bbr_slam_cwnd_in_main_drain &&
11239                             (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
11240                                 /* Restore the cwnd */
11241                                 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
11242                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11243                         }
11244                         /* Setup probe-rtt has being done now RRS-HERE */
11245                         bbr->r_ctl.rc_rtt_shrinks = cts;
11246                         bbr->r_ctl.last_in_probertt = cts;
11247                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0);
11248                         /* Randomly pick a sub-state */
11249                         bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11250                         bbr_substate_change(bbr, cts, __LINE__, 0);
11251                         bbr_log_type_statechange(bbr, cts, __LINE__);
11252                 }
11253         } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) {
11254                 uint32_t flight;
11255
11256                 flight = ctf_flight_size(bbr->rc_tp,
11257                              (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11258                 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered);
11259                 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) &&
11260                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11261                         /*
11262                          * We must keep cwnd at the desired MSS.
11263                          */
11264                         bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
11265                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11266                 } else if ((bbr_prtt_slam_cwnd) &&
11267                            (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11268                         /* Re-slam it */
11269                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11270                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11271                 }
11272                 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) {
11273                         /* Has outstanding reached our target? */
11274                         if (flight <= bbr->r_ctl.rc_target_at_state) {
11275                                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0);
11276                                 bbr->r_ctl.rc_bbr_enters_probertt = cts;
11277                                 /* If time is exactly 0, be 1usec off */
11278                                 if (bbr->r_ctl.rc_bbr_enters_probertt == 0)
11279                                         bbr->r_ctl.rc_bbr_enters_probertt = 1;
11280                                 if (bbr->rc_use_google == 0) {
11281                                         /*
11282                                          * Restore any lowering that as occurred to
11283                                          * reach here
11284                                          */
11285                                         if (bbr->r_ctl.bbr_rttprobe_gain_val)
11286                                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
11287                                         else
11288                                                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11289                                 }
11290                         }
11291                         if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) &&
11292                             (bbr->rc_use_google == 0) &&
11293                             bbr->r_ctl.bbr_rttprobe_gain_val &&
11294                             (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) ||
11295                              (flight >= bbr->r_ctl.flightsize_at_drain))) {
11296                                 /*
11297                                  * We have doddled with our current hptsi
11298                                  * gain an srtt and have still not made it
11299                                  * to target, or we have increased our flight.
11300                                  * Lets reduce the gain by xx%
11301                                  * flooring the reduce at DRAIN (based on
11302                                  * mul/div)
11303                                  */
11304                                 int red;
11305
11306                                 bbr->r_ctl.flightsize_at_drain = flight;
11307                                 bbr->r_ctl.rc_probertt_srttchktim = cts;
11308                                 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1);
11309                                 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) {
11310                                         /* Reduce our gain again */
11311                                         bbr->r_ctl.rc_bbr_hptsi_gain -= red;
11312                                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0);
11313                                 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) {
11314                                         /* one more chance before we give up */
11315                                         bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
11316                                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0);
11317                                 } else {
11318                                         /* At the very bottom */
11319                                         bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1);
11320                                 }
11321                         }
11322                 }
11323                 if (bbr->r_ctl.rc_bbr_enters_probertt &&
11324                     (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) &&
11325                     ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) {
11326                         /* Time to exit probe RTT normally */
11327                         bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts);
11328                 }
11329         } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
11330                 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11331                     (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11332                         /*
11333                          * This qualifies as a RTT_PROBE session since we
11334                          * drop the data outstanding to nothing and waited
11335                          * more than bbr_rtt_probe_time.
11336                          */
11337                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11338                         bbr_set_reduced_rtt(bbr, cts, __LINE__);
11339                 }
11340                 if (bbr_should_enter_probe_rtt(bbr, cts)) {
11341                         bbr_enter_probe_rtt(bbr, cts, __LINE__);
11342                 } else {
11343                         bbr_set_probebw_gains(bbr, cts, losses);
11344                 }
11345         }
11346 }
11347
11348 static void
11349 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses)
11350 {
11351         int32_t epoch = 0;
11352
11353         if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
11354                 bbr_set_epoch(bbr, cts, line);
11355                 /* At each epoch doe lt bw sampling */
11356                 epoch = 1;
11357         }
11358         bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses);
11359 }
11360
11361 static int
11362 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
11363     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
11364     int32_t nxt_pkt, struct timeval *tv)
11365 {
11366         int32_t thflags, retval;
11367         uint32_t cts, lcts;
11368         uint32_t tiwin;
11369         struct tcpopt to;
11370         struct tcp_bbr *bbr;
11371         struct bbr_sendmap *rsm;
11372         struct timeval ltv;
11373         int32_t did_out = 0;
11374         int32_t in_recovery;
11375         uint16_t nsegs;
11376         int32_t prev_state;
11377         uint32_t lost;
11378
11379         nsegs = max(1, m->m_pkthdr.lro_nsegs);
11380         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11381         /* add in our stats */
11382         kern_prefetch(bbr, &prev_state);
11383         prev_state = 0;
11384         thflags = th->th_flags;
11385         /*
11386          * If this is either a state-changing packet or current state isn't
11387          * established, we require a write lock on tcbinfo.  Otherwise, we
11388          * allow the tcbinfo to be in either alocked or unlocked, as the
11389          * caller may have unnecessarily acquired a write lock due to a
11390          * race.
11391          */
11392         INP_WLOCK_ASSERT(tp->t_inpcb);
11393         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
11394             __func__));
11395         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
11396             __func__));
11397
11398         tp->t_rcvtime = ticks;
11399         /*
11400          * Unscale the window into a 32-bit value. For the SYN_SENT state
11401          * the scale is zero.
11402          */
11403         tiwin = th->th_win << tp->snd_scale;
11404 #ifdef STATS
11405         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
11406 #endif
11407
11408         if (m->m_flags & M_TSTMP) {
11409                 /* Prefer the hardware timestamp if present */
11410                 struct timespec ts;
11411
11412                 mbuf_tstmp2timespec(m, &ts);
11413                 bbr->rc_tv.tv_sec = ts.tv_sec;
11414                 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11415                 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11416         } else if (m->m_flags & M_TSTMP_LRO) {
11417                 /* Next the arrival timestamp */
11418                 struct timespec ts;
11419
11420                 mbuf_tstmp2timespec(m, &ts);
11421                 bbr->rc_tv.tv_sec = ts.tv_sec;
11422                 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11423                 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11424         } else {
11425                 /*
11426                  * Ok just get the current time.
11427                  */
11428                 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv);
11429         }
11430         /*
11431          * Parse options on any incoming segment.
11432          */
11433         tcp_dooptions(&to, (u_char *)(th + 1),
11434             (th->th_off << 2) - sizeof(struct tcphdr),
11435             (thflags & TH_SYN) ? TO_SYN : 0);
11436
11437         /*
11438          * If timestamps were negotiated during SYN/ACK and a
11439          * segment without a timestamp is received, silently drop
11440          * the segment, unless it is a RST segment or missing timestamps are
11441          * tolerated.
11442          * See section 3.2 of RFC 7323.
11443          */
11444         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
11445             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
11446                 retval = 0;
11447                 m_freem(m);
11448                 goto done_with_input;
11449         }
11450         /*
11451          * If echoed timestamp is later than the current time, fall back to
11452          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
11453          * were used when this connection was established.
11454          */
11455         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
11456                 to.to_tsecr -= tp->ts_offset;
11457                 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv)))
11458                         to.to_tsecr = 0;
11459         }
11460         /*
11461          * If its the first time in we need to take care of options and
11462          * verify we can do SACK for rack!
11463          */
11464         if (bbr->r_state == 0) {
11465                 /*
11466                  * Process options only when we get SYN/ACK back. The SYN
11467                  * case for incoming connections is handled in tcp_syncache.
11468                  * According to RFC1323 the window field in a SYN (i.e., a
11469                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
11470                  * this is traditional behavior, may need to be cleaned up.
11471                  */
11472                 if (bbr->rc_inp == NULL) {
11473                         bbr->rc_inp = tp->t_inpcb;
11474                 }
11475                 /*
11476                  * We need to init rc_inp here since its not init'd when
11477                  * bbr_init is called
11478                  */
11479                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
11480                         if ((to.to_flags & TOF_SCALE) &&
11481                             (tp->t_flags & TF_REQ_SCALE)) {
11482                                 tp->t_flags |= TF_RCVD_SCALE;
11483                                 tp->snd_scale = to.to_wscale;
11484                         } else
11485                                 tp->t_flags &= ~TF_REQ_SCALE;
11486                         /*
11487                          * Initial send window.  It will be updated with the
11488                          * next incoming segment to the scaled value.
11489                          */
11490                         tp->snd_wnd = th->th_win;
11491                         if ((to.to_flags & TOF_TS) &&
11492                             (tp->t_flags & TF_REQ_TSTMP)) {
11493                                 tp->t_flags |= TF_RCVD_TSTMP;
11494                                 tp->ts_recent = to.to_tsval;
11495                                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
11496                         } else
11497                             tp->t_flags &= ~TF_REQ_TSTMP;
11498                         if (to.to_flags & TOF_MSS)
11499                                 tcp_mss(tp, to.to_mss);
11500                         if ((tp->t_flags & TF_SACK_PERMIT) &&
11501                             (to.to_flags & TOF_SACKPERM) == 0)
11502                                 tp->t_flags &= ~TF_SACK_PERMIT;
11503                         if (IS_FASTOPEN(tp->t_flags)) {
11504                                 if (to.to_flags & TOF_FASTOPEN) {
11505                                         uint16_t mss;
11506
11507                                         if (to.to_flags & TOF_MSS)
11508                                                 mss = to.to_mss;
11509                                         else
11510                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
11511                                                         mss = TCP6_MSS;
11512                                                 else
11513                                                         mss = TCP_MSS;
11514                                         tcp_fastopen_update_cache(tp, mss,
11515                                             to.to_tfo_len, to.to_tfo_cookie);
11516                                 } else
11517                                         tcp_fastopen_disable_path(tp);
11518                         }
11519                 }
11520                 /*
11521                  * At this point we are at the initial call. Here we decide
11522                  * if we are doing RACK or not. We do this by seeing if
11523                  * TF_SACK_PERMIT is set, if not rack is *not* possible and
11524                  * we switch to the default code.
11525                  */
11526                 if ((tp->t_flags & TF_SACK_PERMIT) == 0) {
11527                         /* Bail */
11528                         tcp_switch_back_to_default(tp);
11529                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
11530                             tlen, iptos);
11531                         return (1);
11532                 }
11533                 /* Set the flag */
11534                 bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
11535                 tcp_set_hpts(tp->t_inpcb);
11536                 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack);
11537         }
11538         if (thflags & TH_ACK) {
11539                 /* Track ack types */
11540                 if (to.to_flags & TOF_SACK)
11541                         BBR_STAT_INC(bbr_acks_with_sacks);
11542                 else
11543                         BBR_STAT_INC(bbr_plain_acks);
11544         }
11545         /*
11546          * This is the one exception case where we set the rack state
11547          * always. All other times (timers etc) we must have a rack-state
11548          * set (so we assure we have done the checks above for SACK).
11549          */
11550         if (thflags & TH_FIN)
11551                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
11552         if (bbr->r_state != tp->t_state)
11553                 bbr_set_state(tp, bbr, tiwin);
11554
11555         if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL)
11556                 kern_prefetch(rsm, &prev_state);
11557         prev_state = bbr->r_state;
11558         bbr->rc_ack_was_delayed = 0;
11559         lost = bbr->r_ctl.rc_lost;
11560         bbr->rc_is_pkt_epoch_now = 0;
11561         if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) {
11562                 /* Get the real time into lcts and figure the real delay */
11563                 lcts = tcp_get_usecs(&ltv);
11564                 if (TSTMP_GT(lcts, cts)) {
11565                         bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts;
11566                         bbr->rc_ack_was_delayed = 1;
11567                         if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay,
11568                                      bbr->r_ctl.highest_hdwr_delay))
11569                                 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay;
11570                 } else {
11571                         bbr->r_ctl.rc_ack_hdwr_delay = 0;
11572                         bbr->rc_ack_was_delayed = 0;
11573                 }
11574         } else {
11575                 bbr->r_ctl.rc_ack_hdwr_delay = 0;
11576                 bbr->rc_ack_was_delayed = 0;
11577         }
11578         bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m);
11579         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
11580                 retval = 0;
11581                 m_freem(m);
11582                 goto done_with_input;
11583         }
11584         /*
11585          * If a segment with the ACK-bit set arrives in the SYN-SENT state
11586          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
11587          */
11588         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
11589             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
11590                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11591                 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11592                 return (1);
11593         }
11594         in_recovery = IN_RECOVERY(tp->t_flags);
11595         if (tiwin > bbr->r_ctl.rc_high_rwnd)
11596                 bbr->r_ctl.rc_high_rwnd = tiwin;
11597 #ifdef BBR_INVARIANTS
11598         if ((tp->t_inpcb->inp_flags & INP_DROPPED) ||
11599             (tp->t_inpcb->inp_flags2 & INP_FREED)) {
11600                 panic("tp:%p bbr:%p given a dropped inp:%p",
11601                     tp, bbr, tp->t_inpcb);
11602         }
11603 #endif
11604         bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp,
11605                                             (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11606         bbr->rtt_valid = 0;
11607         if (to.to_flags & TOF_TS) {
11608                 bbr->rc_ts_valid = 1;
11609                 bbr->r_ctl.last_inbound_ts = to.to_tsval;
11610         } else {
11611                 bbr->rc_ts_valid = 0;
11612                 bbr->r_ctl.last_inbound_ts = 0;
11613         }
11614         retval = (*bbr->r_substate) (m, th, so,
11615             tp, &to, drop_hdrlen,
11616             tlen, tiwin, thflags, nxt_pkt, iptos);
11617 #ifdef BBR_INVARIANTS
11618         if ((retval == 0) &&
11619             (tp->t_inpcb == NULL)) {
11620                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
11621                     retval, tp, prev_state);
11622         }
11623 #endif
11624         if (nxt_pkt == 0)
11625                 BBR_STAT_INC(bbr_rlock_left_ret0);
11626         else
11627                 BBR_STAT_INC(bbr_rlock_left_ret1);
11628         if (retval == 0) {
11629                 /*
11630                  * If retval is 1 the tcb is unlocked and most likely the tp
11631                  * is gone.
11632                  */
11633                 INP_WLOCK_ASSERT(tp->t_inpcb);
11634                 tcp_bbr_xmit_timer_commit(bbr, tp, cts);
11635                 if (bbr->rc_is_pkt_epoch_now)
11636                         bbr_set_pktepoch(bbr, cts, __LINE__);
11637                 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost));
11638                 if (nxt_pkt == 0) {
11639                         if (bbr->r_wanted_output != 0) {
11640                                 bbr->rc_output_starts_timer = 0;
11641                                 did_out = 1;
11642                                 (void)tp->t_fb->tfb_tcp_output(tp);
11643                         } else
11644                                 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0);
11645                 }
11646                 if ((nxt_pkt == 0) &&
11647                     ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
11648                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
11649                      (tp->t_flags & TF_DELACK) ||
11650                      ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
11651                       (tp->t_state <= TCPS_CLOSING)))) {
11652                         /*
11653                          * We could not send (probably in the hpts but
11654                          * stopped the timer)?
11655                          */
11656                         if ((tp->snd_max == tp->snd_una) &&
11657                             ((tp->t_flags & TF_DELACK) == 0) &&
11658                             (bbr->rc_inp->inp_in_hpts) &&
11659                             (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11660                                 /*
11661                                  * keep alive not needed if we are hptsi
11662                                  * output yet
11663                                  */
11664                                 ;
11665                         } else {
11666                                 if (bbr->rc_inp->inp_in_hpts) {
11667                                         tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
11668                                         if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11669                                             (TSTMP_GT(lcts, bbr->rc_pacer_started))) {
11670                                                 uint32_t del;
11671
11672                                                 del = lcts - bbr->rc_pacer_started;
11673                                                 if (bbr->r_ctl.rc_last_delay_val > del) {
11674                                                         BBR_STAT_INC(bbr_force_timer_start);
11675                                                         bbr->r_ctl.rc_last_delay_val -= del;
11676                                                         bbr->rc_pacer_started = lcts;
11677                                                 } else {
11678                                                         /* We are late */
11679                                                         bbr->r_ctl.rc_last_delay_val = 0;
11680                                                         BBR_STAT_INC(bbr_force_output);
11681                                                         (void)tp->t_fb->tfb_tcp_output(tp);
11682                                                 }
11683                                         }
11684                                 }
11685                                 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val,
11686                                     0);
11687                         }
11688                 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) {
11689                         /* Do we have the correct timer running? */
11690                         bbr_timer_audit(tp, bbr, lcts, &so->so_snd);
11691                 }
11692                 /* Do we have a new state */
11693                 if (bbr->r_state != tp->t_state)
11694                         bbr_set_state(tp, bbr, tiwin);
11695 done_with_input:
11696                 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out);
11697                 if (did_out)
11698                         bbr->r_wanted_output = 0;
11699 #ifdef BBR_INVARIANTS
11700                 if (tp->t_inpcb == NULL) {
11701                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
11702                             did_out,
11703                             retval, tp, prev_state);
11704                 }
11705 #endif
11706         }
11707         return (retval);
11708 }
11709
11710 static void
11711 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
11712     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
11713 {
11714         struct timeval tv;
11715         int retval;
11716
11717         /* First lets see if we have old packets */
11718         if (tp->t_in_pkt) {
11719                 if (ctf_do_queued_segments(so, tp, 1)) {
11720                         m_freem(m);
11721                         return;
11722                 }
11723         }
11724         if (m->m_flags & M_TSTMP_LRO) {
11725                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
11726                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
11727         } else {
11728                 /* Should not be should we kassert instead? */
11729                 tcp_get_usecs(&tv);
11730         }
11731         retval = bbr_do_segment_nounlock(m, th, so, tp,
11732                                          drop_hdrlen, tlen, iptos, 0, &tv);
11733         if (retval == 0) {
11734                 INP_WUNLOCK(tp->t_inpcb);
11735         }
11736 }
11737
11738 /*
11739  * Return how much data can be sent without violating the
11740  * cwnd or rwnd.
11741  */
11742
11743 static inline uint32_t
11744 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin,
11745     uint32_t avail, int32_t sb_offset, uint32_t cts)
11746 {
11747         uint32_t len;
11748
11749         if (ctf_outstanding(tp) >= tp->snd_wnd) {
11750                 /* We never want to go over our peers rcv-window */
11751                 len = 0;
11752         } else {
11753                 uint32_t flight;
11754
11755                 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11756                 if (flight >= sendwin) {
11757                         /*
11758                          * We have in flight what we are allowed by cwnd (if
11759                          * it was rwnd blocking it would have hit above out
11760                          * >= tp->snd_wnd).
11761                          */
11762                         return (0);
11763                 }
11764                 len = sendwin - flight;
11765                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
11766                         /* We would send too much (beyond the rwnd) */
11767                         len = tp->snd_wnd - ctf_outstanding(tp);
11768                 }
11769                 if ((len + sb_offset) > avail) {
11770                         /*
11771                          * We don't have that much in the SB, how much is
11772                          * there?
11773                          */
11774                         len = avail - sb_offset;
11775                 }
11776         }
11777         return (len);
11778 }
11779
11780 static inline void
11781 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11782 {
11783 #ifdef NETFLIX_STATS
11784         KMOD_TCPSTAT_INC(tcps_sndpack_error);
11785         KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len);
11786 #endif
11787 }
11788
11789 static inline void
11790 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11791 {
11792         if (error) {
11793                 bbr_do_error_accounting(tp, bbr, rsm, len, error);
11794                 return;
11795         }
11796         if (rsm) {
11797                 if (rsm->r_flags & BBR_TLP) {
11798                         /*
11799                          * TLP should not count in retran count, but in its
11800                          * own bin
11801                          */
11802 #ifdef NETFLIX_STATS
11803                         KMOD_TCPSTAT_INC(tcps_tlpresends);
11804                         KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len);
11805 #endif
11806                 } else {
11807                         /* Retransmit */
11808                         tp->t_sndrexmitpack++;
11809                         KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
11810                         KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
11811 #ifdef STATS
11812                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
11813                             len);
11814 #endif
11815                 }
11816                 /*
11817                  * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is
11818                  * sub-state
11819                  */
11820                 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len);
11821                 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) {
11822                         /* Non probe_bw log in 1, 2, or 4. */
11823                         counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len);
11824                 } else {
11825                         /*
11826                          * Log our probe state 3, and log also 5-13 to show
11827                          * us the recovery sub-state for the send. This
11828                          * means that 3 == (5+6+7+8+9+10+11+12+13)
11829                          */
11830                         counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len);
11831                         counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len);
11832                 }
11833                 /* Place in both 16's the totals of retransmitted */
11834                 counter_u64_add(bbr_state_lost[16], len);
11835                 counter_u64_add(bbr_state_resend[16], len);
11836                 /* Place in 17's the total sent */
11837                 counter_u64_add(bbr_state_resend[17], len);
11838                 counter_u64_add(bbr_state_lost[17], len);
11839
11840         } else {
11841                 /* New sends */
11842                 KMOD_TCPSTAT_INC(tcps_sndpack);
11843                 KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
11844                 /* Place in 17's the total sent */
11845                 counter_u64_add(bbr_state_resend[17], len);
11846                 counter_u64_add(bbr_state_lost[17], len);
11847 #ifdef STATS
11848                 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
11849                     len);
11850 #endif
11851         }
11852 }
11853
11854 static void
11855 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level)
11856 {
11857         if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) {
11858                 /*
11859                  * Limit the cwnd to not be above N x the target plus whats
11860                  * is outstanding. The target is based on the current b/w
11861                  * estimate.
11862                  */
11863                 uint32_t target;
11864
11865                 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT);
11866                 target += ctf_outstanding(tp);
11867                 target *= bbr_target_cwnd_mult_limit;
11868                 if (tp->snd_cwnd > target)
11869                         tp->snd_cwnd = target;
11870                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__);
11871         }
11872 }
11873
11874 static int
11875 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg)
11876 {
11877         /*
11878          * "adv" is the amount we could increase the window, taking into
11879          * account that we are limited by TCP_MAXWIN << tp->rcv_scale.
11880          */
11881         int32_t adv;
11882         int32_t oldwin;
11883
11884         adv = recwin;
11885         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
11886                 oldwin = (tp->rcv_adv - tp->rcv_nxt);
11887                 if (adv > oldwin)
11888                         adv -= oldwin;
11889                 else {
11890                         /* We can't increase the window */
11891                         adv = 0;
11892                 }
11893         } else
11894                 oldwin = 0;
11895
11896         /*
11897          * If the new window size ends up being the same as or less
11898          * than the old size when it is scaled, then don't force
11899          * a window update.
11900          */
11901         if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
11902                 return (0);
11903
11904         if (adv >= (2 * maxseg) &&
11905             (adv >= (so->so_rcv.sb_hiwat / 4) ||
11906             recwin <= (so->so_rcv.sb_hiwat / 8) ||
11907             so->so_rcv.sb_hiwat <= 8 * maxseg)) {
11908                 return (1);
11909         }
11910         if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat)
11911                 return (1);
11912         return (0);
11913 }
11914
11915 /*
11916  * Return 0 on success and a errno on failure to send.
11917  * Note that a 0 return may not mean we sent anything
11918  * if the TCB was on the hpts. A non-zero return
11919  * does indicate the error we got from ip[6]_output.
11920  */
11921 static int
11922 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv)
11923 {
11924         struct socket *so;
11925         int32_t len;
11926         uint32_t cts;
11927         uint32_t recwin, sendwin;
11928         int32_t sb_offset;
11929         int32_t flags, abandon, error = 0;
11930         struct tcp_log_buffer *lgb = NULL;
11931         struct mbuf *m;
11932         struct mbuf *mb;
11933         uint32_t if_hw_tsomaxsegcount = 0;
11934         uint32_t if_hw_tsomaxsegsize = 0;
11935         uint32_t if_hw_tsomax = 0;
11936         struct ip *ip = NULL;
11937 #ifdef TCPDEBUG
11938         struct ipovly *ipov = NULL;
11939 #endif
11940         struct tcp_bbr *bbr;
11941         struct tcphdr *th;
11942         struct udphdr *udp = NULL;
11943         u_char opt[TCP_MAXOLEN];
11944         unsigned ipoptlen, optlen, hdrlen;
11945         unsigned ulen;
11946         uint32_t bbr_seq;
11947         uint32_t delay_calc=0;
11948         uint8_t doing_tlp = 0;
11949         uint8_t local_options;
11950 #ifdef BBR_INVARIANTS
11951         uint8_t doing_retran_from = 0;
11952         uint8_t picked_up_retran = 0;
11953 #endif
11954         uint8_t wanted_cookie = 0;
11955         uint8_t more_to_rxt=0;
11956         int32_t prefetch_so_done = 0;
11957         int32_t prefetch_rsm = 0;
11958         uint32_t what_we_can = 0;
11959         uint32_t tot_len = 0;
11960         uint32_t rtr_cnt = 0;
11961         uint32_t maxseg, pace_max_segs, p_maxseg;
11962         int32_t csum_flags = 0;
11963         int32_t hw_tls;
11964 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
11965         unsigned ipsec_optlen = 0;
11966
11967 #endif
11968         volatile int32_t sack_rxmit;
11969         struct bbr_sendmap *rsm = NULL;
11970         int32_t tso, mtu;
11971         struct tcpopt to;
11972         int32_t slot = 0;
11973         struct inpcb *inp;
11974         struct sockbuf *sb;
11975         uint32_t hpts_calling;
11976 #ifdef INET6
11977         struct ip6_hdr *ip6 = NULL;
11978         int32_t isipv6;
11979 #endif
11980         uint8_t app_limited = BBR_JR_SENT_DATA;
11981         uint8_t filled_all = 0;
11982         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11983         /* We take a cache hit here */
11984         memcpy(&bbr->rc_tv, tv, sizeof(struct timeval));
11985         cts = tcp_tv_to_usectick(&bbr->rc_tv);
11986         inp = bbr->rc_inp;
11987         so = inp->inp_socket;
11988         sb = &so->so_snd;
11989         if (sb->sb_flags & SB_TLS_IFNET)
11990                 hw_tls = 1;
11991         else
11992                 hw_tls = 0;
11993         kern_prefetch(sb, &maxseg);
11994         maxseg = tp->t_maxseg - bbr->rc_last_options;
11995         if (bbr_minseg(bbr) < maxseg) {
11996                 tcp_bbr_tso_size_check(bbr, cts);
11997         }
11998         /* Remove any flags that indicate we are pacing on the inp  */
11999         pace_max_segs = bbr->r_ctl.rc_pace_max_segs;
12000         p_maxseg = min(maxseg, pace_max_segs);
12001         INP_WLOCK_ASSERT(inp);
12002 #ifdef TCP_OFFLOAD
12003         if (tp->t_flags & TF_TOE)
12004                 return (tcp_offload_output(tp));
12005 #endif
12006
12007 #ifdef INET6
12008         if (bbr->r_state) {
12009                 /* Use the cache line loaded if possible */
12010                 isipv6 = bbr->r_is_v6;
12011         } else {
12012                 isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
12013         }
12014 #endif
12015         if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
12016             inp->inp_in_hpts) {
12017                 /*
12018                  * We are on the hpts for some timer but not hptsi output.
12019                  * Possibly remove from the hpts so we can send/recv etc.
12020                  */
12021                 if ((tp->t_flags & TF_ACKNOW) == 0) {
12022                         /*
12023                          * No immediate demand right now to send an ack, but
12024                          * the user may have read, making room for new data
12025                          * (a window update). If so we may want to cancel
12026                          * whatever timer is running (KEEP/DEL-ACK?) and
12027                          * continue to send out a window update. Or we may
12028                          * have gotten more data into the socket buffer to
12029                          * send.
12030                          */
12031                         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12032                                       (long)TCP_MAXWIN << tp->rcv_scale);
12033                         if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) &&
12034                             ((tcp_outflags[tp->t_state] & TH_RST) == 0) &&
12035                             ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <=
12036                             (tp->snd_max - tp->snd_una))) {
12037                                 /*
12038                                  * Nothing new to send and no window update
12039                                  * is needed to send. Lets just return and
12040                                  * let the timer-run off.
12041                                  */
12042                                 return (0);
12043                         }
12044                 }
12045                 tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12046                 bbr_timer_cancel(bbr, __LINE__, cts);
12047         }
12048         if (bbr->r_ctl.rc_last_delay_val) {
12049                 /* Calculate a rough delay for early escape to sending  */
12050                 if (SEQ_GT(cts, bbr->rc_pacer_started))
12051                         delay_calc = cts - bbr->rc_pacer_started;
12052                 if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12053                         delay_calc -= bbr->r_ctl.rc_last_delay_val;
12054                 else
12055                         delay_calc = 0;
12056         }
12057         /* Mark that we have called bbr_output(). */
12058         if ((bbr->r_timer_override) ||
12059             (tp->t_state < TCPS_ESTABLISHED)) {
12060                 /* Timeouts or early states are exempt */
12061                 if (inp->inp_in_hpts)
12062                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12063         } else if (inp->inp_in_hpts) {
12064                 if ((bbr->r_ctl.rc_last_delay_val) &&
12065                     (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
12066                     delay_calc) {
12067                         /*
12068                          * We were being paced for output and the delay has
12069                          * already exceeded when we were supposed to be
12070                          * called, lets go ahead and pull out of the hpts
12071                          * and call output.
12072                          */
12073                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1);
12074                         bbr->r_ctl.rc_last_delay_val = 0;
12075                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12076                 } else if (tp->t_state == TCPS_CLOSED) {
12077                         bbr->r_ctl.rc_last_delay_val = 0;
12078                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12079                 } else {
12080                         /*
12081                          * On the hpts, you shall not pass! even if ACKNOW
12082                          * is on, we will when the hpts fires, unless of
12083                          * course we are overdue.
12084                          */
12085                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1);
12086                         return (0);
12087                 }
12088         }
12089         bbr->rc_cwnd_limited = 0;
12090         if (bbr->r_ctl.rc_last_delay_val) {
12091                 /* recalculate the real delay and deal with over/under  */
12092                 if (SEQ_GT(cts, bbr->rc_pacer_started))
12093                         delay_calc = cts - bbr->rc_pacer_started;
12094                 else
12095                         delay_calc = 0;
12096                 if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12097                         /* Setup the delay which will be added in */
12098                         delay_calc -= bbr->r_ctl.rc_last_delay_val;
12099                 else {
12100                         /*
12101                          * We are early setup to adjust
12102                          * our slot time.
12103                          */
12104                         uint64_t merged_val;
12105
12106                         bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc);
12107                         bbr->r_agg_early_set = 1;
12108                         if (bbr->r_ctl.rc_hptsi_agg_delay) {
12109                                 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) {
12110                                         /* Nope our previous late cancels out the early */
12111                                         bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early;
12112                                         bbr->r_agg_early_set = 0;
12113                                         bbr->r_ctl.rc_agg_early = 0;
12114                                 } else {
12115                                         bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay;
12116                                         bbr->r_ctl.rc_hptsi_agg_delay = 0;
12117                                 }
12118                         }
12119                         merged_val = bbr->rc_pacer_started;
12120                         merged_val <<= 32;
12121                         merged_val |= bbr->r_ctl.rc_last_delay_val;
12122                         bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls,
12123                                                  bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val,
12124                                                  bbr->r_agg_early_set, 3);
12125                         bbr->r_ctl.rc_last_delay_val = 0;
12126                         BBR_STAT_INC(bbr_early);
12127                         delay_calc = 0;
12128                 }
12129         } else {
12130                 /* We were not delayed due to hptsi */
12131                 if (bbr->r_agg_early_set)
12132                         bbr->r_ctl.rc_agg_early = 0;
12133                 bbr->r_agg_early_set = 0;
12134                 delay_calc = 0;
12135         }
12136         if (delay_calc) {
12137                 /*
12138                  * We had a hptsi delay which means we are falling behind on
12139                  * sending at the expected rate. Calculate an extra amount
12140                  * of data we can send, if any, to put us back on track.
12141                  */
12142                 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay)
12143                         bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff;
12144                 else
12145                         bbr->r_ctl.rc_hptsi_agg_delay += delay_calc;
12146         }
12147         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12148         if ((tp->snd_una == tp->snd_max) &&
12149             (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
12150             (sbavail(sb))) {
12151                 /*
12152                  * Ok we have been idle with nothing outstanding
12153                  * we possibly need to start fresh with either a new
12154                  * suite of states or a fast-ramp up.
12155                  */
12156                 bbr_restart_after_idle(bbr,
12157                                        cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time));
12158         }
12159         /*
12160          * Now was there a hptsi delay where we are behind? We only count
12161          * being behind if: a) We are not in recovery. b) There was a delay.
12162          * <and> c) We had room to send something.
12163          *
12164          */
12165         hpts_calling = inp->inp_hpts_calls;
12166         inp->inp_hpts_calls = 0;
12167         if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
12168                 if (bbr_process_timers(tp, bbr, cts, hpts_calling)) {
12169                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1);
12170                         return (0);
12171                 }
12172         }
12173         bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12174         if (hpts_calling &&
12175             (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12176                 bbr->r_ctl.rc_last_delay_val = 0;
12177         }
12178         bbr->r_timer_override = 0;
12179         bbr->r_wanted_output = 0;
12180         /*
12181          * For TFO connections in SYN_RECEIVED, only allow the initial
12182          * SYN|ACK and those sent by the retransmit timer.
12183          */
12184         if (IS_FASTOPEN(tp->t_flags) &&
12185             ((tp->t_state == TCPS_SYN_RECEIVED) ||
12186              (tp->t_state == TCPS_SYN_SENT)) &&
12187             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
12188             (tp->t_rxtshift == 0)) {    /* not a retransmit */
12189                 len = 0;
12190                 goto just_return_nolock;
12191         }
12192         /*
12193          * Before sending anything check for a state update. For hpts
12194          * calling without input this is important. If its input calling
12195          * then this was already done.
12196          */
12197         if (bbr->rc_use_google == 0)
12198                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12199 again:
12200         /*
12201          * If we've recently taken a timeout, snd_max will be greater than
12202          * snd_max. BBR in general does not pay much attention to snd_nxt
12203          * for historic reasons the persist timer still uses it. This means
12204          * we have to look at it. All retransmissions that are not persits
12205          * use the rsm that needs to be sent so snd_nxt is ignored. At the
12206          * end of this routine we pull snd_nxt always up to snd_max.
12207          */
12208         doing_tlp = 0;
12209 #ifdef BBR_INVARIANTS
12210         doing_retran_from = picked_up_retran = 0;
12211 #endif
12212         error = 0;
12213         tso = 0;
12214         slot = 0;
12215         mtu = 0;
12216         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12217         sb_offset = tp->snd_max - tp->snd_una;
12218         flags = tcp_outflags[tp->t_state];
12219         sack_rxmit = 0;
12220         len = 0;
12221         rsm = NULL;
12222         if (flags & TH_RST) {
12223                 SOCKBUF_LOCK(sb);
12224                 goto send;
12225         }
12226 recheck_resend:
12227         while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
12228                 /* We need to always have one in reserve */
12229                 rsm = bbr_alloc(bbr);
12230                 if (rsm == NULL) {
12231                         error = ENOMEM;
12232                         /* Lie to get on the hpts */
12233                         tot_len = tp->t_maxseg;
12234                         if (hpts_calling)
12235                                 /* Retry in a ms */
12236                                 slot = 1001;
12237                         goto just_return_nolock;
12238                 }
12239                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
12240                 bbr->r_ctl.rc_free_cnt++;
12241                 rsm = NULL;
12242         }
12243         /* What do we send, a resend? */
12244         if (bbr->r_ctl.rc_resend == NULL) {
12245                 /* Check for rack timeout */
12246                 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
12247                 if (bbr->r_ctl.rc_resend) {
12248 #ifdef BBR_INVARIANTS
12249                         picked_up_retran = 1;
12250 #endif
12251                         bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend);
12252                 }
12253         }
12254         if (bbr->r_ctl.rc_resend) {
12255                 rsm = bbr->r_ctl.rc_resend;
12256 #ifdef BBR_INVARIANTS
12257                 doing_retran_from = 1;
12258 #endif
12259                 /* Remove any TLP flags its a RACK or T-O */
12260                 rsm->r_flags &= ~BBR_TLP;
12261                 bbr->r_ctl.rc_resend = NULL;
12262                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
12263 #ifdef BBR_INVARIANTS
12264                         panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n",
12265                             tp, bbr, rsm, rsm->r_start, tp->snd_una);
12266                         goto recheck_resend;
12267 #else
12268                         /* TSNH */
12269                         rsm = NULL;
12270                         goto recheck_resend;
12271 #endif
12272                 }
12273                 rtr_cnt++;
12274                 if (rsm->r_flags & BBR_HAS_SYN) {
12275                         /* Only retransmit a SYN by itself */
12276                         len = 0;
12277                         if ((flags & TH_SYN) == 0) {
12278                                 /* Huh something is wrong */
12279                                 rsm->r_start++;
12280                                 if (rsm->r_start == rsm->r_end) {
12281                                         /* Clean it up, somehow we missed the ack? */
12282                                         bbr_log_syn(tp, NULL);
12283                                 } else {
12284                                         /* TFO with data? */
12285                                         rsm->r_flags &= ~BBR_HAS_SYN;
12286                                         len = rsm->r_end - rsm->r_start;
12287                                 }
12288                         } else {
12289                                 /* Retransmitting SYN */
12290                                 rsm = NULL;
12291                                 SOCKBUF_LOCK(sb);
12292                                 goto send;
12293                         }
12294                 } else
12295                         len = rsm->r_end - rsm->r_start;
12296                 if ((bbr->rc_resends_use_tso == 0) &&
12297                     (len > maxseg)) {
12298                         len = maxseg;
12299                         more_to_rxt = 1;
12300                 }
12301                 sb_offset = rsm->r_start - tp->snd_una;
12302                 if (len > 0) {
12303                         sack_rxmit = 1;
12304                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
12305                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
12306                             min(len, maxseg));
12307                 } else {
12308                         /* I dont think this can happen */
12309                         rsm = NULL;
12310                         goto recheck_resend;
12311                 }
12312                 BBR_STAT_INC(bbr_resends_set);
12313         } else if (bbr->r_ctl.rc_tlp_send) {
12314                 /*
12315                  * Tail loss probe
12316                  */
12317                 doing_tlp = 1;
12318                 rsm = bbr->r_ctl.rc_tlp_send;
12319                 bbr->r_ctl.rc_tlp_send = NULL;
12320                 sack_rxmit = 1;
12321                 len = rsm->r_end - rsm->r_start;
12322                 rtr_cnt++;
12323                 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12324                         len = maxseg;
12325
12326                 if (SEQ_GT(tp->snd_una, rsm->r_start)) {
12327 #ifdef BBR_INVARIANTS
12328                         panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u",
12329                             tp, bbr, tp->snd_una, rsm, rsm->r_start);
12330 #else
12331                         /* TSNH */
12332                         rsm = NULL;
12333                         goto recheck_resend;
12334 #endif
12335                 }
12336                 sb_offset = rsm->r_start - tp->snd_una;
12337                 BBR_STAT_INC(bbr_tlp_set);
12338         }
12339         /*
12340          * Enforce a connection sendmap count limit if set
12341          * as long as we are not retransmiting.
12342          */
12343         if ((rsm == NULL) &&
12344             (V_tcp_map_entries_limit > 0) &&
12345             (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
12346                 BBR_STAT_INC(bbr_alloc_limited);
12347                 if (!bbr->alloc_limit_reported) {
12348                         bbr->alloc_limit_reported = 1;
12349                         BBR_STAT_INC(bbr_alloc_limited_conns);
12350                 }
12351                 goto just_return_nolock;
12352         }
12353 #ifdef BBR_INVARIANTS
12354         if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) {
12355                 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u",
12356                     tp, bbr, rsm, sb_offset, len);
12357         }
12358 #endif
12359         /*
12360          * Get standard flags, and add SYN or FIN if requested by 'hidden'
12361          * state flags.
12362          */
12363         if (tp->t_flags & TF_NEEDFIN && (rsm == NULL))
12364                 flags |= TH_FIN;
12365         if (tp->t_flags & TF_NEEDSYN)
12366                 flags |= TH_SYN;
12367
12368         if (rsm && (rsm->r_flags & BBR_HAS_FIN)) {
12369                 /* we are retransmitting the fin */
12370                 len--;
12371                 if (len) {
12372                         /*
12373                          * When retransmitting data do *not* include the
12374                          * FIN. This could happen from a TLP probe if we
12375                          * allowed data with a FIN.
12376                          */
12377                         flags &= ~TH_FIN;
12378                 }
12379         } else if (rsm) {
12380                 if (flags & TH_FIN)
12381                         flags &= ~TH_FIN;
12382         }
12383         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
12384                 void *end_rsm;
12385
12386                 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
12387                 if (end_rsm)
12388                         kern_prefetch(end_rsm, &prefetch_rsm);
12389                 prefetch_rsm = 1;
12390         }
12391         SOCKBUF_LOCK(sb);
12392         /*
12393          * If snd_nxt == snd_max and we have transmitted a FIN, the
12394          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
12395          * negative length.  This can also occur when TCP opens up its
12396          * congestion window while receiving additional duplicate acks after
12397          * fast-retransmit because TCP will reset snd_nxt to snd_max after
12398          * the fast-retransmit.
12399          *
12400          * In the normal retransmit-FIN-only case, however, snd_nxt will be
12401          * set to snd_una, the sb_offset will be 0, and the length may wind
12402          * up 0.
12403          *
12404          * If sack_rxmit is true we are retransmitting from the scoreboard
12405          * in which case len is already set.
12406          */
12407         if (sack_rxmit == 0) {
12408                 uint32_t avail;
12409
12410                 avail = sbavail(sb);
12411                 if (SEQ_GT(tp->snd_max, tp->snd_una))
12412                         sb_offset = tp->snd_max - tp->snd_una;
12413                 else
12414                         sb_offset = 0;
12415                 if (bbr->rc_tlp_new_data) {
12416                         /* TLP is forcing out new data */
12417                         uint32_t tlplen;
12418
12419                         doing_tlp = 1;
12420                         tlplen = maxseg;
12421
12422                         if (tlplen > (uint32_t)(avail - sb_offset)) {
12423                                 tlplen = (uint32_t)(avail - sb_offset);
12424                         }
12425                         if (tlplen > tp->snd_wnd) {
12426                                 len = tp->snd_wnd;
12427                         } else {
12428                                 len = tlplen;
12429                         }
12430                         bbr->rc_tlp_new_data = 0;
12431                 } else {
12432                         what_we_can = len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts);
12433                         if ((len < p_maxseg) &&
12434                             (bbr->rc_in_persist == 0) &&
12435                             (ctf_outstanding(tp) >= (2 * p_maxseg)) &&
12436                             ((avail - sb_offset) >= p_maxseg)) {
12437                                 /*
12438                                  * We are not completing whats in the socket
12439                                  * buffer (i.e. there is at least a segment
12440                                  * waiting to send) and we have 2 or more
12441                                  * segments outstanding. There is no sense
12442                                  * of sending a little piece. Lets defer and
12443                                  * and wait until we can send a whole
12444                                  * segment.
12445                                  */
12446                                 len = 0;
12447                         }
12448                         if (bbr->rc_in_persist) {
12449                                 /*
12450                                  * We are in persists, figure out if
12451                                  * a retransmit is available (maybe the previous
12452                                  * persists we sent) or if we have to send new
12453                                  * data.
12454                                  */
12455                                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
12456                                 if (rsm) {
12457                                         len = rsm->r_end - rsm->r_start;
12458                                         if (rsm->r_flags & BBR_HAS_FIN)
12459                                                 len--;
12460                                         if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12461                                                 len = maxseg;
12462                                         if (len > 1)
12463                                                 BBR_STAT_INC(bbr_persist_reneg);
12464                                         /*
12465                                          * XXXrrs we could force the len to
12466                                          * 1 byte here to cause the chunk to
12467                                          * split apart.. but that would then
12468                                          * mean we always retransmit it as
12469                                          * one byte even after the window
12470                                          * opens.
12471                                          */
12472                                         sack_rxmit = 1;
12473                                         sb_offset = rsm->r_start - tp->snd_una;
12474                                 } else {
12475                                         /*
12476                                          * First time through in persists or peer
12477                                          * acked our one byte. Though we do have
12478                                          * to have something in the sb.
12479                                          */
12480                                         len = 1;
12481                                         sb_offset = 0;
12482                                         if (avail == 0)
12483                                             len = 0;
12484                                 }
12485                         }
12486                 }
12487         }
12488         if (prefetch_so_done == 0) {
12489                 kern_prefetch(so, &prefetch_so_done);
12490                 prefetch_so_done = 1;
12491         }
12492         /*
12493          * Lop off SYN bit if it has already been sent.  However, if this is
12494          * SYN-SENT state and if segment contains data and if we don't know
12495          * that foreign host supports TAO, suppress sending segment.
12496          */
12497         if ((flags & TH_SYN) && (rsm == NULL) &&
12498             SEQ_GT(tp->snd_max, tp->snd_una)) {
12499                 if (tp->t_state != TCPS_SYN_RECEIVED)
12500                         flags &= ~TH_SYN;
12501                 /*
12502                  * When sending additional segments following a TFO SYN|ACK,
12503                  * do not include the SYN bit.
12504                  */
12505                 if (IS_FASTOPEN(tp->t_flags) &&
12506                     (tp->t_state == TCPS_SYN_RECEIVED))
12507                         flags &= ~TH_SYN;
12508                 sb_offset--, len++;
12509                 if (sbavail(sb) == 0)
12510                         len = 0;
12511         } else if ((flags & TH_SYN) && rsm) {
12512                 /*
12513                  * Subtract one from the len for the SYN being
12514                  * retransmitted.
12515                  */
12516                 len--;
12517         }
12518         /*
12519          * Be careful not to send data and/or FIN on SYN segments. This
12520          * measure is needed to prevent interoperability problems with not
12521          * fully conformant TCP implementations.
12522          */
12523         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
12524                 len = 0;
12525                 flags &= ~TH_FIN;
12526         }
12527         /*
12528          * On TFO sockets, ensure no data is sent in the following cases:
12529          *
12530          *  - When retransmitting SYN|ACK on a passively-created socket
12531          *  - When retransmitting SYN on an actively created socket
12532          *  - When sending a zero-length cookie (cookie request) on an
12533          *    actively created socket
12534          *  - When the socket is in the CLOSED state (RST is being sent)
12535          */
12536         if (IS_FASTOPEN(tp->t_flags) &&
12537             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
12538              ((tp->t_state == TCPS_SYN_SENT) &&
12539               (tp->t_tfo_client_cookie_len == 0)) ||
12540              (flags & TH_RST))) {
12541                 len = 0;
12542                 sack_rxmit = 0;
12543                 rsm = NULL;
12544         }
12545         /* Without fast-open there should never be data sent on a SYN */
12546         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags)))
12547                 len = 0;
12548         if (len <= 0) {
12549                 /*
12550                  * If FIN has been sent but not acked, but we haven't been
12551                  * called to retransmit, len will be < 0.  Otherwise, window
12552                  * shrank after we sent into it.  If window shrank to 0,
12553                  * cancel pending retransmit, pull snd_nxt back to (closed)
12554                  * window, and set the persist timer if it isn't already
12555                  * going.  If the window didn't close completely, just wait
12556                  * for an ACK.
12557                  *
12558                  * We also do a general check here to ensure that we will
12559                  * set the persist timer when we have data to send, but a
12560                  * 0-byte window. This makes sure the persist timer is set
12561                  * even if the packet hits one of the "goto send" lines
12562                  * below.
12563                  */
12564                 len = 0;
12565                 if ((tp->snd_wnd == 0) &&
12566                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12567                     (tp->snd_una == tp->snd_max) &&
12568                     (sb_offset < (int)sbavail(sb))) {
12569                         /*
12570                          * Not enough room in the rwnd to send
12571                          * a paced segment out.
12572                          */
12573                         bbr_enter_persist(tp, bbr, cts, __LINE__);
12574                 }
12575         } else if ((rsm == NULL) &&
12576                    (doing_tlp == 0) &&
12577                    (len < bbr->r_ctl.rc_pace_max_segs)) {
12578                 /*
12579                  * We are not sending a full segment for
12580                  * some reason. Should we not send anything (think
12581                  * sws or persists)?
12582                  */
12583                 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12584                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12585                     (len < (int)(sbavail(sb) - sb_offset))) {
12586                         /*
12587                          * Here the rwnd is less than
12588                          * the pacing size, this is not a retransmit,
12589                          * we are established and
12590                          * the send is not the last in the socket buffer
12591                          * lets not send, and possibly enter persists.
12592                          */
12593                         len = 0;
12594                         if (tp->snd_max == tp->snd_una)
12595                                 bbr_enter_persist(tp, bbr, cts, __LINE__);
12596                 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) &&
12597                            (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12598                                                  bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12599                            (len < (int)(sbavail(sb) - sb_offset)) &&
12600                            (len < bbr_minseg(bbr))) {
12601                         /*
12602                          * Here we are not retransmitting, and
12603                          * the cwnd is not so small that we could
12604                          * not send at least a min size (rxt timer
12605                          * not having gone off), We have 2 segments or
12606                          * more already in flight, its not the tail end
12607                          * of the socket buffer  and the cwnd is blocking
12608                          * us from sending out minimum pacing segment size.
12609                          * Lets not send anything.
12610                          */
12611                         bbr->rc_cwnd_limited = 1;
12612                         len = 0;
12613                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
12614                             min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12615                            (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12616                                                  bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12617                            (len < (int)(sbavail(sb) - sb_offset)) &&
12618                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
12619                         /*
12620                          * Here we have a send window but we have
12621                          * filled it up and we can't send another pacing segment.
12622                          * We also have in flight more than 2 segments
12623                          * and we are not completing the sb i.e. we allow
12624                          * the last bytes of the sb to go out even if
12625                          * its not a full pacing segment.
12626                          */
12627                         len = 0;
12628                 }
12629         }
12630         /* len will be >= 0 after this point. */
12631         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
12632         tcp_sndbuf_autoscale(tp, so, sendwin);
12633         /*
12634          *
12635          */
12636         if (bbr->rc_in_persist &&
12637             len &&
12638             (rsm == NULL) &&
12639             (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) {
12640                 /*
12641                  * We are in persist, not doing a retransmit and don't have enough space
12642                  * yet to send a full TSO. So is it at the end of the sb
12643                  * if so we need to send else nuke to 0 and don't send.
12644                  */
12645                 int sbleft;
12646                 if (sbavail(sb) > sb_offset)
12647                         sbleft = sbavail(sb) - sb_offset;
12648                 else
12649                         sbleft = 0;
12650                 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) {
12651                         /* not at end of sb lets not send */
12652                         len = 0;
12653                 }
12654         }
12655         /*
12656          * Decide if we can use TCP Segmentation Offloading (if supported by
12657          * hardware).
12658          *
12659          * TSO may only be used if we are in a pure bulk sending state.  The
12660          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
12661          * options prevent using TSO.  With TSO the TCP header is the same
12662          * (except for the sequence number) for all generated packets.  This
12663          * makes it impossible to transmit any options which vary per
12664          * generated segment or packet.
12665          *
12666          * IPv4 handling has a clear separation of ip options and ip header
12667          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen()
12668          * does the right thing below to provide length of just ip options
12669          * and thus checking for ipoptlen is enough to decide if ip options
12670          * are present.
12671          */
12672 #ifdef INET6
12673         if (isipv6)
12674                 ipoptlen = ip6_optlen(inp);
12675         else
12676 #endif
12677         if (inp->inp_options)
12678                 ipoptlen = inp->inp_options->m_len -
12679                     offsetof(struct ipoption, ipopt_list);
12680         else
12681                 ipoptlen = 0;
12682 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12683         /*
12684          * Pre-calculate here as we save another lookup into the darknesses
12685          * of IPsec that way and can actually decide if TSO is ok.
12686          */
12687 #ifdef INET6
12688         if (isipv6 && IPSEC_ENABLED(ipv6))
12689                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
12690 #ifdef INET
12691         else
12692 #endif
12693 #endif                          /* INET6 */
12694 #ifdef INET
12695         if (IPSEC_ENABLED(ipv4))
12696                 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
12697 #endif                          /* INET */
12698 #endif                          /* IPSEC */
12699 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12700         ipoptlen += ipsec_optlen;
12701 #endif
12702         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
12703             (len > maxseg) &&
12704             (tp->t_port == 0) &&
12705             ((tp->t_flags & TF_SIGNATURE) == 0) &&
12706             tp->rcv_numsacks == 0 &&
12707             ipoptlen == 0)
12708                 tso = 1;
12709
12710         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12711             (long)TCP_MAXWIN << tp->rcv_scale);
12712         /*
12713          * Sender silly window avoidance.   We transmit under the following
12714          * conditions when len is non-zero:
12715          *
12716          * - We have a full segment (or more with TSO) - This is the last
12717          * buffer in a write()/send() and we are either idle or running
12718          * NODELAY - we've timed out (e.g. persist timer) - we have more
12719          * then 1/2 the maximum send window's worth of data (receiver may be
12720          * limited the window size) - we need to retransmit
12721          */
12722         if (rsm)
12723                 goto send;
12724         if (len) {
12725                 if (sack_rxmit)
12726                         goto send;
12727                 if (len >= p_maxseg)
12728                         goto send;
12729                 /*
12730                  * NOTE! on localhost connections an 'ack' from the remote
12731                  * end may occur synchronously with the output and cause us
12732                  * to flush a buffer queued with moretocome.  XXX
12733                  *
12734                  */
12735                 if (((tp->t_flags & TF_MORETOCOME) == 0) &&     /* normal case */
12736                     ((tp->t_flags & TF_NODELAY) ||
12737                     ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) &&
12738                     (tp->t_flags & TF_NOPUSH) == 0) {
12739                         goto send;
12740                 }
12741                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
12742                         goto send;
12743                 }
12744                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
12745                         goto send;
12746                 }
12747         }
12748         /*
12749          * Sending of standalone window updates.
12750          *
12751          * Window updates are important when we close our window due to a
12752          * full socket buffer and are opening it again after the application
12753          * reads data from it.  Once the window has opened again and the
12754          * remote end starts to send again the ACK clock takes over and
12755          * provides the most current window information.
12756          *
12757          * We must avoid the silly window syndrome whereas every read from
12758          * the receive buffer, no matter how small, causes a window update
12759          * to be sent.  We also should avoid sending a flurry of window
12760          * updates when the socket buffer had queued a lot of data and the
12761          * application is doing small reads.
12762          *
12763          * Prevent a flurry of pointless window updates by only sending an
12764          * update when we can increase the advertized window by more than
12765          * 1/4th of the socket buffer capacity.  When the buffer is getting
12766          * full or is very small be more aggressive and send an update
12767          * whenever we can increase by two mss sized segments. In all other
12768          * situations the ACK's to new incoming data will carry further
12769          * window increases.
12770          *
12771          * Don't send an independent window update if a delayed ACK is
12772          * pending (it will get piggy-backed on it) or the remote side
12773          * already has done a half-close and won't send more data.  Skip
12774          * this if the connection is in T/TCP half-open state.
12775          */
12776         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
12777             !(tp->t_flags & TF_DELACK) &&
12778             !TCPS_HAVERCVDFIN(tp->t_state)) {
12779                 /* Check to see if we should do a window update */
12780                 if (bbr_window_update_needed(tp, so, recwin, maxseg))
12781                         goto send;
12782         }
12783         /*
12784          * Send if we owe the peer an ACK, RST, SYN.  ACKNOW
12785          * is also a catch-all for the retransmit timer timeout case.
12786          */
12787         if (tp->t_flags & TF_ACKNOW) {
12788                 goto send;
12789         }
12790         if (flags & TH_RST) {
12791                 /* Always send a RST if one is due */
12792                 goto send;
12793         }
12794         if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) {
12795                 goto send;
12796         }
12797         /*
12798          * If our state indicates that FIN should be sent and we have not
12799          * yet done so, then we need to send.
12800          */
12801         if (flags & TH_FIN &&
12802             ((tp->t_flags & TF_SENTFIN) == 0)) {
12803                 goto send;
12804         }
12805         /*
12806          * No reason to send a segment, just return.
12807          */
12808 just_return:
12809         SOCKBUF_UNLOCK(sb);
12810 just_return_nolock:
12811         if (tot_len)
12812                 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
12813         if (bbr->rc_no_pacing)
12814                 slot = 0;
12815         if (tot_len == 0) {
12816                 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >=
12817                     tp->snd_wnd) {
12818                         BBR_STAT_INC(bbr_rwnd_limited);
12819                         app_limited = BBR_JR_RWND_LIMITED;
12820                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12821                         if ((bbr->rc_in_persist == 0) &&
12822                             TCPS_HAVEESTABLISHED(tp->t_state) &&
12823                             (tp->snd_max == tp->snd_una) &&
12824                             sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
12825                                 /* No send window.. we must enter persist */
12826                                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
12827                         }
12828                 } else if (ctf_outstanding(tp) >= sbavail(sb)) {
12829                         BBR_STAT_INC(bbr_app_limited);
12830                         app_limited = BBR_JR_APP_LIMITED;
12831                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12832                 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12833                                                  bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) {
12834                         BBR_STAT_INC(bbr_cwnd_limited);
12835                         app_limited = BBR_JR_CWND_LIMITED;
12836                         bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12837                                                                         bbr->r_ctl.rc_lost_bytes)));
12838                         bbr->rc_cwnd_limited = 1;
12839                 } else {
12840                         BBR_STAT_INC(bbr_app_limited);
12841                         app_limited = BBR_JR_APP_LIMITED;
12842                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12843                 }
12844                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
12845                 bbr->r_agg_early_set = 0;
12846                 bbr->r_ctl.rc_agg_early = 0;
12847                 bbr->r_ctl.rc_last_delay_val = 0;
12848         } else if (bbr->rc_use_google == 0)
12849                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12850         /* Are we app limited? */
12851         if ((app_limited == BBR_JR_APP_LIMITED) ||
12852             (app_limited == BBR_JR_RWND_LIMITED)) {
12853                 /**
12854                  * We are application limited.
12855                  */
12856                 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12857                                                                        bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered);
12858         }
12859         if (tot_len == 0)
12860                 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1);
12861         /* Dont update the time if we did not send */
12862         bbr->r_ctl.rc_last_delay_val = 0;
12863         bbr->rc_output_starts_timer = 1;
12864         bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len);
12865         bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len);
12866         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12867                 /* Make sure snd_nxt is drug up */
12868                 tp->snd_nxt = tp->snd_max;
12869         }
12870         return (error);
12871
12872 send:
12873         if (doing_tlp == 0) {
12874                 /*
12875                  * Data not a TLP, and its not the rxt firing. If it is the
12876                  * rxt firing, we want to leave the tlp_in_progress flag on
12877                  * so we don't send another TLP. It has to be a rack timer
12878                  * or normal send (response to acked data) to clear the tlp
12879                  * in progress flag.
12880                  */
12881                 bbr->rc_tlp_in_progress = 0;
12882                 bbr->rc_tlp_rtx_out = 0;
12883         } else {
12884                 /*
12885                  * Its a TLP.
12886                  */
12887                 bbr->rc_tlp_in_progress = 1;
12888         }
12889         bbr_timer_cancel(bbr, __LINE__, cts);
12890         if (rsm == NULL) {
12891                 if (sbused(sb) > 0) {
12892                         /*
12893                          * This is sub-optimal. We only send a stand alone
12894                          * FIN on its own segment.
12895                          */
12896                         if (flags & TH_FIN) {
12897                                 flags &= ~TH_FIN;
12898                                 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) {
12899                                         /* Lets not send this */
12900                                         slot = 0;
12901                                         goto just_return;
12902                                 }
12903                         }
12904                 }
12905         } else {
12906                 /*
12907                  * We do *not* send a FIN on a retransmit if it has data.
12908                  * The if clause here where len > 1 should never come true.
12909                  */
12910                 if ((len > 0) &&
12911                     (((rsm->r_flags & BBR_HAS_FIN) == 0) &&
12912                     (flags & TH_FIN))) {
12913                         flags &= ~TH_FIN;
12914                         len--;
12915                 }
12916         }
12917         SOCKBUF_LOCK_ASSERT(sb);
12918         if (len > 0) {
12919                 if ((tp->snd_una == tp->snd_max) &&
12920                     (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
12921                         /*
12922                          * This qualifies as a RTT_PROBE session since we
12923                          * drop the data outstanding to nothing and waited
12924                          * more than bbr_rtt_probe_time.
12925                          */
12926                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
12927                         bbr_set_reduced_rtt(bbr, cts, __LINE__);
12928                 }
12929                 if (len >= maxseg)
12930                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
12931                 else
12932                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
12933         }
12934         /*
12935          * Before ESTABLISHED, force sending of initial options unless TCP
12936          * set not to do any options. NOTE: we assume that the IP/TCP header
12937          * plus TCP options always fit in a single mbuf, leaving room for a
12938          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
12939          * + optlen <= MCLBYTES
12940          */
12941         optlen = 0;
12942 #ifdef INET6
12943         if (isipv6)
12944                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12945         else
12946 #endif
12947                 hdrlen = sizeof(struct tcpiphdr);
12948
12949         /*
12950          * Compute options for segment. We only have to care about SYN and
12951          * established connection segments.  Options for SYN-ACK segments
12952          * are handled in TCP syncache.
12953          */
12954         to.to_flags = 0;
12955         local_options = 0;
12956         if ((tp->t_flags & TF_NOOPT) == 0) {
12957                 /* Maximum segment size. */
12958                 if (flags & TH_SYN) {
12959                         to.to_mss = tcp_mssopt(&inp->inp_inc);
12960                         if (tp->t_port)
12961                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
12962                         to.to_flags |= TOF_MSS;
12963                         /*
12964                          * On SYN or SYN|ACK transmits on TFO connections,
12965                          * only include the TFO option if it is not a
12966                          * retransmit, as the presence of the TFO option may
12967                          * have caused the original SYN or SYN|ACK to have
12968                          * been dropped by a middlebox.
12969                          */
12970                         if (IS_FASTOPEN(tp->t_flags) &&
12971                             (tp->t_rxtshift == 0)) {
12972                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
12973                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
12974                                         to.to_tfo_cookie =
12975                                             (u_int8_t *)&tp->t_tfo_cookie.server;
12976                                         to.to_flags |= TOF_FASTOPEN;
12977                                         wanted_cookie = 1;
12978                                 } else if (tp->t_state == TCPS_SYN_SENT) {
12979                                         to.to_tfo_len =
12980                                             tp->t_tfo_client_cookie_len;
12981                                         to.to_tfo_cookie =
12982                                             tp->t_tfo_cookie.client;
12983                                         to.to_flags |= TOF_FASTOPEN;
12984                                         wanted_cookie = 1;
12985                                 }
12986                         }
12987                 }
12988                 /* Window scaling. */
12989                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
12990                         to.to_wscale = tp->request_r_scale;
12991                         to.to_flags |= TOF_SCALE;
12992                 }
12993                 /* Timestamps. */
12994                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
12995                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
12996                         to.to_tsval =   tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset;
12997                         to.to_tsecr = tp->ts_recent;
12998                         to.to_flags |= TOF_TS;
12999                         local_options += TCPOLEN_TIMESTAMP + 2;
13000                 }
13001                 /* Set receive buffer autosizing timestamp. */
13002                 if (tp->rfbuf_ts == 0 &&
13003                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
13004                         tp->rfbuf_ts =  tcp_tv_to_mssectick(&bbr->rc_tv);
13005                 /* Selective ACK's. */
13006                 if (flags & TH_SYN)
13007                         to.to_flags |= TOF_SACKPERM;
13008                 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13009                     tp->rcv_numsacks > 0) {
13010                         to.to_flags |= TOF_SACK;
13011                         to.to_nsacks = tp->rcv_numsacks;
13012                         to.to_sacks = (u_char *)tp->sackblks;
13013                 }
13014 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13015                 /* TCP-MD5 (RFC2385). */
13016                 if (tp->t_flags & TF_SIGNATURE)
13017                         to.to_flags |= TOF_SIGNATURE;
13018 #endif                          /* TCP_SIGNATURE */
13019
13020                 /* Processing the options. */
13021                 hdrlen += (optlen = tcp_addoptions(&to, opt));
13022                 /*
13023                  * If we wanted a TFO option to be added, but it was unable
13024                  * to fit, ensure no data is sent.
13025                  */
13026                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
13027                     !(to.to_flags & TOF_FASTOPEN))
13028                         len = 0;
13029         }
13030         if (tp->t_port) {
13031                 if (V_tcp_udp_tunneling_port == 0) {
13032                         /* The port was removed?? */
13033                         SOCKBUF_UNLOCK(&so->so_snd);
13034                         return (EHOSTUNREACH);
13035                 }
13036                 hdrlen += sizeof(struct udphdr);
13037         }
13038 #ifdef INET6
13039         if (isipv6)
13040                 ipoptlen = ip6_optlen(tp->t_inpcb);
13041         else
13042 #endif
13043         if (tp->t_inpcb->inp_options)
13044                 ipoptlen = tp->t_inpcb->inp_options->m_len -
13045                     offsetof(struct ipoption, ipopt_list);
13046         else
13047                 ipoptlen = 0;
13048         ipoptlen = 0;
13049 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
13050         ipoptlen += ipsec_optlen;
13051 #endif
13052         if (bbr->rc_last_options != local_options) {
13053                 /*
13054                  * Cache the options length this generally does not change
13055                  * on a connection. We use this to calculate TSO.
13056                  */
13057                 bbr->rc_last_options = local_options;
13058         }
13059         maxseg = tp->t_maxseg - (ipoptlen + optlen);
13060         p_maxseg = min(maxseg, pace_max_segs);
13061         /*
13062          * Adjust data length if insertion of options will bump the packet
13063          * length beyond the t_maxseg length. Clear the FIN bit because we
13064          * cut off the tail of the segment.
13065          */
13066         if (len > maxseg) {
13067                 if (len != 0 && (flags & TH_FIN)) {
13068                         flags &= ~TH_FIN;
13069                 }
13070                 if (tso) {
13071                         uint32_t moff;
13072                         int32_t max_len;
13073
13074                         /* extract TSO information */
13075                         if_hw_tsomax = tp->t_tsomax;
13076                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
13077                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
13078                         KASSERT(ipoptlen == 0,
13079                             ("%s: TSO can't do IP options", __func__));
13080
13081                         /*
13082                          * Check if we should limit by maximum payload
13083                          * length:
13084                          */
13085                         if (if_hw_tsomax != 0) {
13086                                 /* compute maximum TSO length */
13087                                 max_len = (if_hw_tsomax - hdrlen -
13088                                     max_linkhdr);
13089                                 if (max_len <= 0) {
13090                                         len = 0;
13091                                 } else if (len > max_len) {
13092                                         len = max_len;
13093                                 }
13094                         }
13095                         /*
13096                          * Prevent the last segment from being fractional
13097                          * unless the send sockbuf can be emptied:
13098                          */
13099                         if ((sb_offset + len) < sbavail(sb)) {
13100                                 moff = len % (uint32_t)maxseg;
13101                                 if (moff != 0) {
13102                                         len -= moff;
13103                                 }
13104                         }
13105                         /*
13106                          * In case there are too many small fragments don't
13107                          * use TSO:
13108                          */
13109                         if (len <= maxseg) {
13110                                 len = maxseg;
13111                                 tso = 0;
13112                         }
13113                 } else {
13114                         /* Not doing TSO */
13115                         if (optlen + ipoptlen >= tp->t_maxseg) {
13116                                 /*
13117                                  * Since we don't have enough space to put
13118                                  * the IP header chain and the TCP header in
13119                                  * one packet as required by RFC 7112, don't
13120                                  * send it. Also ensure that at least one
13121                                  * byte of the payload can be put into the
13122                                  * TCP segment.
13123                                  */
13124                                 SOCKBUF_UNLOCK(&so->so_snd);
13125                                 error = EMSGSIZE;
13126                                 sack_rxmit = 0;
13127                                 goto out;
13128                         }
13129                         len = maxseg;
13130                 }
13131         } else {
13132                 /* Not doing TSO */
13133                 if_hw_tsomaxsegcount = 0;
13134                 tso = 0;
13135         }
13136         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
13137             ("%s: len > IP_MAXPACKET", __func__));
13138 #ifdef DIAGNOSTIC
13139 #ifdef INET6
13140         if (max_linkhdr + hdrlen > MCLBYTES)
13141 #else
13142         if (max_linkhdr + hdrlen > MHLEN)
13143 #endif
13144                 panic("tcphdr too big");
13145 #endif
13146         /*
13147          * This KASSERT is here to catch edge cases at a well defined place.
13148          * Before, those had triggered (random) panic conditions further
13149          * down.
13150          */
13151 #ifdef BBR_INVARIANTS
13152         if (sack_rxmit) {
13153                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
13154                         panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u",
13155                             rsm, tp, bbr, rsm->r_start, tp->snd_una);
13156                 }
13157         }
13158 #endif
13159         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
13160         if ((len == 0) &&
13161             (flags & TH_FIN) &&
13162             (sbused(sb))) {
13163                 /*
13164                  * We have outstanding data, don't send a fin by itself!.
13165                  */
13166                 slot = 0;
13167                 goto just_return;
13168         }
13169         /*
13170          * Grab a header mbuf, attaching a copy of data to be transmitted,
13171          * and initialize the header from the template for sends on this
13172          * connection.
13173          */
13174         if (len) {
13175                 uint32_t moff;
13176                 uint32_t orig_len;
13177
13178                 /*
13179                  * We place a limit on sending with hptsi.
13180                  */
13181                 if ((rsm == NULL) && len > pace_max_segs)
13182                         len = pace_max_segs;
13183                 if (len <= maxseg)
13184                         tso = 0;
13185 #ifdef INET6
13186                 if (MHLEN < hdrlen + max_linkhdr)
13187                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
13188                 else
13189 #endif
13190                         m = m_gethdr(M_NOWAIT, MT_DATA);
13191
13192                 if (m == NULL) {
13193                         BBR_STAT_INC(bbr_failed_mbuf_aloc);
13194                         bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13195                         SOCKBUF_UNLOCK(sb);
13196                         error = ENOBUFS;
13197                         sack_rxmit = 0;
13198                         goto out;
13199                 }
13200                 m->m_data += max_linkhdr;
13201                 m->m_len = hdrlen;
13202                 /*
13203                  * Start the m_copy functions from the closest mbuf to the
13204                  * sb_offset in the socket buffer chain.
13205                  */
13206                 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) {
13207 #ifdef BBR_INVARIANTS
13208                         if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0)))
13209                                 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u",
13210                                     tp, bbr, len, sb_offset, sbavail(sb), rsm,
13211                                     doing_retran_from,
13212                                     picked_up_retran,
13213                                     doing_tlp);
13214
13215 #endif
13216                         /*
13217                          * In this messed up situation we have two choices,
13218                          * a) pretend the send worked, and just start timers
13219                          * and what not (not good since that may lead us
13220                          * back here a lot). <or> b) Send the lowest segment
13221                          * in the map. <or> c) Drop the connection. Lets do
13222                          * <b> which if it continues to happen will lead to
13223                          * <c> via timeouts.
13224                          */
13225                         BBR_STAT_INC(bbr_offset_recovery);
13226                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
13227                         sb_offset = 0;
13228                         if (rsm == NULL) {
13229                                 sack_rxmit = 0;
13230                                 len = sbavail(sb);
13231                         } else {
13232                                 sack_rxmit = 1;
13233                                 if (rsm->r_start != tp->snd_una) {
13234                                         /*
13235                                          * Things are really messed up, <c>
13236                                          * is the only thing to do.
13237                                          */
13238                                         BBR_STAT_INC(bbr_offset_drop);
13239                                         tcp_set_inp_to_drop(inp, EFAULT);
13240                                         SOCKBUF_UNLOCK(sb);
13241                                         (void)m_free(m);
13242                                         return (0);
13243                                 }
13244                                 len = rsm->r_end - rsm->r_start;
13245                         }
13246                         if (len > sbavail(sb))
13247                                 len = sbavail(sb);
13248                         if (len > maxseg)
13249                                 len = maxseg;
13250                 }
13251                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
13252                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
13253                         m_copydata(mb, moff, (int)len,
13254                             mtod(m, caddr_t)+hdrlen);
13255                         if (rsm == NULL)
13256                                 sbsndptr_adv(sb, mb, len);
13257                         m->m_len += len;
13258                 } else {
13259                         struct sockbuf *msb;
13260
13261                         if (rsm)
13262                                 msb = NULL;
13263                         else
13264                                 msb = sb;
13265 #ifdef BBR_INVARIANTS
13266                         if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) {
13267                                 if (rsm) {
13268                                         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 ",
13269                                             tp, bbr, len, moff,
13270                                             sbavail(sb), rsm,
13271                                             tp->snd_una, rsm->r_flags, rsm->r_start,
13272                                             doing_retran_from,
13273                                             picked_up_retran,
13274                                             doing_tlp, sack_rxmit);
13275                                 } else {
13276                                         panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u",
13277                                             tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una);
13278                                 }
13279                         }
13280 #endif
13281                         orig_len = len;
13282                         m->m_next = tcp_m_copym(
13283                                 mb, moff, &len,
13284                                 if_hw_tsomaxsegcount,
13285                                 if_hw_tsomaxsegsize, msb,
13286                                 ((rsm == NULL) ? hw_tls : 0)
13287 #ifdef NETFLIX_COPY_ARGS
13288                                 , &filled_all
13289 #endif
13290                                 );
13291                         if (len <= maxseg) {
13292                                 /*
13293                                  * Must have ran out of mbufs for the copy
13294                                  * shorten it to no longer need tso. Lets
13295                                  * not put on sendalot since we are low on
13296                                  * mbufs.
13297                                  */
13298                                 tso = 0;
13299                         }
13300                         if (m->m_next == NULL) {
13301                                 SOCKBUF_UNLOCK(sb);
13302                                 (void)m_free(m);
13303                                 error = ENOBUFS;
13304                                 sack_rxmit = 0;
13305                                 goto out;
13306                         }
13307                 }
13308 #ifdef BBR_INVARIANTS
13309                 if (tso && len < maxseg) {
13310                         panic("tp:%p tso on, but len:%d < maxseg:%d",
13311                             tp, len, maxseg);
13312                 }
13313                 if (tso && if_hw_tsomaxsegcount) {
13314                         int32_t seg_cnt = 0;
13315                         struct mbuf *foo;
13316
13317                         foo = m;
13318                         while (foo) {
13319                                 seg_cnt++;
13320                                 foo = foo->m_next;
13321                         }
13322                         if (seg_cnt > if_hw_tsomaxsegcount) {
13323                                 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount);
13324                         }
13325                 }
13326 #endif
13327                 /*
13328                  * If we're sending everything we've got, set PUSH. (This
13329                  * will keep happy those implementations which only give
13330                  * data to the user when a buffer fills or a PUSH comes in.)
13331                  */
13332                 if (sb_offset + len == sbused(sb) &&
13333                     sbused(sb) &&
13334                     !(flags & TH_SYN)) {
13335                         flags |= TH_PUSH;
13336                 }
13337                 SOCKBUF_UNLOCK(sb);
13338         } else {
13339                 SOCKBUF_UNLOCK(sb);
13340                 if (tp->t_flags & TF_ACKNOW)
13341                         KMOD_TCPSTAT_INC(tcps_sndacks);
13342                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
13343                         KMOD_TCPSTAT_INC(tcps_sndctrl);
13344                 else
13345                         KMOD_TCPSTAT_INC(tcps_sndwinup);
13346
13347                 m = m_gethdr(M_NOWAIT, MT_DATA);
13348                 if (m == NULL) {
13349                         BBR_STAT_INC(bbr_failed_mbuf_aloc);
13350                         bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13351                         error = ENOBUFS;
13352                         /* Fudge the send time since we could not send */
13353                         sack_rxmit = 0;
13354                         goto out;
13355                 }
13356 #ifdef INET6
13357                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
13358                     MHLEN >= hdrlen) {
13359                         M_ALIGN(m, hdrlen);
13360                 } else
13361 #endif
13362                         m->m_data += max_linkhdr;
13363                 m->m_len = hdrlen;
13364         }
13365         SOCKBUF_UNLOCK_ASSERT(sb);
13366         m->m_pkthdr.rcvif = (struct ifnet *)0;
13367 #ifdef MAC
13368         mac_inpcb_create_mbuf(inp, m);
13369 #endif
13370 #ifdef INET6
13371         if (isipv6) {
13372                 ip6 = mtod(m, struct ip6_hdr *);
13373                 if (tp->t_port) {
13374                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
13375                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13376                         udp->uh_dport = tp->t_port;
13377                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
13378                         udp->uh_ulen = htons(ulen);
13379                         th = (struct tcphdr *)(udp + 1);
13380                 } else {
13381                         th = (struct tcphdr *)(ip6 + 1);
13382                 }
13383                 tcpip_fillheaders(inp, tp->t_port, ip6, th);
13384         } else
13385 #endif                          /* INET6 */
13386         {
13387                 ip = mtod(m, struct ip *);
13388 #ifdef TCPDEBUG
13389                 ipov = (struct ipovly *)ip;
13390 #endif
13391                 if (tp->t_port) {
13392                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
13393                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13394                         udp->uh_dport = tp->t_port;
13395                         ulen = hdrlen + len - sizeof(struct ip);
13396                         udp->uh_ulen = htons(ulen);
13397                         th = (struct tcphdr *)(udp + 1);
13398                 } else {
13399                         th = (struct tcphdr *)(ip + 1);
13400                 }
13401                 tcpip_fillheaders(inp, tp->t_port, ip, th);
13402         }
13403         /*
13404          * If we are doing retransmissions, then snd_nxt will not reflect
13405          * the first unsent octet.  For ACK only packets, we do not want the
13406          * sequence number of the retransmitted packet, we want the sequence
13407          * number of the next unsent octet.  So, if there is no data (and no
13408          * SYN or FIN), use snd_max instead of snd_nxt when filling in
13409          * ti_seq.  But if we are in persist state, snd_max might reflect
13410          * one byte beyond the right edge of the window, so use snd_nxt in
13411          * that case, since we know we aren't doing a retransmission.
13412          * (retransmit and persist are mutually exclusive...)
13413          */
13414         if (sack_rxmit == 0) {
13415                 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) {
13416                         /* New data (including new persists) */
13417                         th->th_seq = htonl(tp->snd_max);
13418                         bbr_seq = tp->snd_max;
13419                 } else if (flags & TH_SYN) {
13420                         /* Syn's always send from iss */
13421                         th->th_seq = htonl(tp->iss);
13422                         bbr_seq = tp->iss;
13423                 } else if (flags & TH_FIN) {
13424                         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) {
13425                                 /*
13426                                  * If we sent the fin already its 1 minus
13427                                  * snd_max
13428                                  */
13429                                 th->th_seq = (htonl(tp->snd_max - 1));
13430                                 bbr_seq = (tp->snd_max - 1);
13431                         } else {
13432                                 /* First time FIN use snd_max */
13433                                 th->th_seq = htonl(tp->snd_max);
13434                                 bbr_seq = tp->snd_max;
13435                         }
13436                 } else {
13437                         /*
13438                          * len == 0 and not persist we use snd_max, sending
13439                          * an ack unless we have sent the fin then its 1
13440                          * minus.
13441                          */
13442                         /*
13443                          * XXXRRS Question if we are in persists and we have
13444                          * nothing outstanding to send and we have not sent
13445                          * a FIN, we will send an ACK. In such a case it
13446                          * might be better to send (tp->snd_una - 1) which
13447                          * would force the peer to ack.
13448                          */
13449                         if (tp->t_flags & TF_SENTFIN) {
13450                                 th->th_seq = htonl(tp->snd_max - 1);
13451                                 bbr_seq = (tp->snd_max - 1);
13452                         } else {
13453                                 th->th_seq = htonl(tp->snd_max);
13454                                 bbr_seq = tp->snd_max;
13455                         }
13456                 }
13457         } else {
13458                 /* All retransmits use the rsm to guide the send */
13459                 th->th_seq = htonl(rsm->r_start);
13460                 bbr_seq = rsm->r_start;
13461         }
13462         th->th_ack = htonl(tp->rcv_nxt);
13463         if (optlen) {
13464                 bcopy(opt, th + 1, optlen);
13465                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
13466         }
13467         th->th_flags = flags;
13468         /*
13469          * Calculate receive window.  Don't shrink window, but avoid silly
13470          * window syndrome.
13471          */
13472         if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) &&
13473                                   recwin < maxseg)))
13474                 recwin = 0;
13475         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
13476             recwin < (tp->rcv_adv - tp->rcv_nxt))
13477                 recwin = (tp->rcv_adv - tp->rcv_nxt);
13478         if (recwin > TCP_MAXWIN << tp->rcv_scale)
13479                 recwin = TCP_MAXWIN << tp->rcv_scale;
13480
13481         /*
13482          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
13483          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
13484          * handled in syncache.
13485          */
13486         if (flags & TH_SYN)
13487                 th->th_win = htons((u_short)
13488                     (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
13489         else {
13490                 /* Avoid shrinking window with window scaling. */
13491                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
13492                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
13493         }
13494         /*
13495          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
13496          * window.  This may cause the remote transmitter to stall.  This
13497          * flag tells soreceive() to disable delayed acknowledgements when
13498          * draining the buffer.  This can occur if the receiver is
13499          * attempting to read more data than can be buffered prior to
13500          * transmitting on the connection.
13501          */
13502         if (th->th_win == 0) {
13503                 tp->t_sndzerowin++;
13504                 tp->t_flags |= TF_RXWIN0SENT;
13505         } else
13506                 tp->t_flags &= ~TF_RXWIN0SENT;
13507         /*
13508          * We don't support urgent data, but drag along
13509          * the pointer in case of a stack switch.
13510          */
13511         tp->snd_up = tp->snd_una;
13512
13513 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13514         if (to.to_flags & TOF_SIGNATURE) {
13515                 /*
13516                  * Calculate MD5 signature and put it into the place
13517                  * determined before. NOTE: since TCP options buffer doesn't
13518                  * point into mbuf's data, calculate offset and use it.
13519                  */
13520                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
13521                     (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
13522                         /*
13523                          * Do not send segment if the calculation of MD5
13524                          * digest has failed.
13525                          */
13526                         goto out;
13527                 }
13528         }
13529 #endif
13530
13531         /*
13532          * Put TCP length in extended header, and then checksum extended
13533          * header and data.
13534          */
13535         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
13536 #ifdef INET6
13537         if (isipv6) {
13538                 /*
13539                  * ip6_plen is not need to be filled now, and will be filled
13540                  * in ip6_output.
13541                  */
13542                 if (tp->t_port) {
13543                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
13544                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13545                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
13546                         th->th_sum = htons(0);
13547                         UDPSTAT_INC(udps_opackets);
13548                 } else {
13549                         csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
13550                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13551                         th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
13552                             optlen + len, IPPROTO_TCP, 0);
13553                 }
13554         }
13555 #endif
13556 #if defined(INET6) && defined(INET)
13557         else
13558 #endif
13559 #ifdef INET
13560         {
13561                 if (tp->t_port) {
13562                         m->m_pkthdr.csum_flags = CSUM_UDP;
13563                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13564                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
13565                             ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
13566                         th->th_sum = htons(0);
13567                         UDPSTAT_INC(udps_opackets);
13568                 } else {
13569                         csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP;
13570                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13571                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
13572                             ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
13573                             IPPROTO_TCP + len + optlen));
13574                 }
13575                 /* IP version must be set here for ipv4/ipv6 checking later */
13576                 KASSERT(ip->ip_v == IPVERSION,
13577                     ("%s: IP version incorrect: %d", __func__, ip->ip_v));
13578         }
13579 #endif
13580
13581         /*
13582          * Enable TSO and specify the size of the segments. The TCP pseudo
13583          * header checksum is always provided. XXX: Fixme: This is currently
13584          * not the case for IPv6.
13585          */
13586         if (tso) {
13587                 KASSERT(len > maxseg,
13588                     ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg));
13589                 m->m_pkthdr.csum_flags |= CSUM_TSO;
13590                 csum_flags |= CSUM_TSO;
13591                 m->m_pkthdr.tso_segsz = maxseg;
13592         }
13593         KASSERT(len + hdrlen == m_length(m, NULL),
13594             ("%s: mbuf chain different than expected: %d + %u != %u",
13595             __func__, len, hdrlen, m_length(m, NULL)));
13596
13597 #ifdef TCP_HHOOK
13598         /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */
13599         hhook_run_tcp_est_out(tp, th, &to, len, tso);
13600 #endif
13601 #ifdef TCPDEBUG
13602         /*
13603          * Trace.
13604          */
13605         if (so->so_options & SO_DEBUG) {
13606                 u_short save = 0;
13607
13608 #ifdef INET6
13609                 if (!isipv6)
13610 #endif
13611                 {
13612                         save = ipov->ih_len;
13613                         ipov->ih_len = htons(m->m_pkthdr.len    /* - hdrlen +
13614                               * (th->th_off << 2) */ );
13615                 }
13616                 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
13617 #ifdef INET6
13618                 if (!isipv6)
13619 #endif
13620                         ipov->ih_len = save;
13621         }
13622 #endif                          /* TCPDEBUG */
13623
13624         /* Log to the black box */
13625         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13626                 union tcp_log_stackspecific log;
13627
13628                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
13629                 /* Record info on type of transmission */
13630                 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay;
13631                 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3);
13632                 log.u_bbr.flex3 = maxseg;
13633                 log.u_bbr.flex4 = delay_calc;
13634                 /* Encode filled_all into the upper flex5 bit */
13635                 log.u_bbr.flex5 = bbr->rc_past_init_win;
13636                 log.u_bbr.flex5 <<= 1;
13637                 log.u_bbr.flex5 |= bbr->rc_no_pacing;
13638                 log.u_bbr.flex5 <<= 29;
13639                 if (filled_all)
13640                         log.u_bbr.flex5 |= 0x80000000;
13641                 log.u_bbr.flex5 |= tp->t_maxseg;
13642                 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs;
13643                 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr);
13644                 /* lets poke in the low and the high here for debugging */
13645                 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
13646                 if (rsm || sack_rxmit) {
13647                         if (doing_tlp)
13648                                 log.u_bbr.flex8 = 2;
13649                         else
13650                                 log.u_bbr.flex8 = 1;
13651                 } else {
13652                         log.u_bbr.flex8 = 0;
13653                 }
13654                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
13655                     len, &log, false, NULL, NULL, 0, tv);
13656         } else {
13657                 lgb = NULL;
13658         }
13659         /*
13660          * Fill in IP length and desired time to live and send to IP level.
13661          * There should be a better way to handle ttl and tos; we could keep
13662          * them in the template, but need a way to checksum without them.
13663          */
13664         /*
13665          * m->m_pkthdr.len should have been set before cksum calcuration,
13666          * because in6_cksum() need it.
13667          */
13668 #ifdef INET6
13669         if (isipv6) {
13670                 /*
13671                  * we separately set hoplimit for every segment, since the
13672                  * user might want to change the value via setsockopt. Also,
13673                  * desired default hop limit might be changed via Neighbor
13674                  * Discovery.
13675                  */
13676                 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
13677
13678                 /*
13679                  * Set the packet size here for the benefit of DTrace
13680                  * probes. ip6_output() will set it properly; it's supposed
13681                  * to include the option header lengths as well.
13682                  */
13683                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
13684
13685                 if (V_path_mtu_discovery && maxseg > V_tcp_minmss)
13686                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13687                 else
13688                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13689
13690                 if (tp->t_state == TCPS_SYN_SENT)
13691                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
13692
13693                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
13694                 /* TODO: IPv6 IP6TOS_ECT bit on */
13695                 error = ip6_output(m, inp->in6p_outputopts,
13696                     &inp->inp_route6,
13697                     ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
13698                     NULL, NULL, inp);
13699
13700                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
13701                         mtu = inp->inp_route6.ro_nh->nh_mtu;
13702         }
13703 #endif                          /* INET6 */
13704 #if defined(INET) && defined(INET6)
13705         else
13706 #endif
13707 #ifdef INET
13708         {
13709                 ip->ip_len = htons(m->m_pkthdr.len);
13710 #ifdef INET6
13711                 if (isipv6)
13712                         ip->ip_ttl = in6_selecthlim(inp, NULL);
13713 #endif                          /* INET6 */
13714                 /*
13715                  * If we do path MTU discovery, then we set DF on every
13716                  * packet. This might not be the best thing to do according
13717                  * to RFC3390 Section 2. However the tcp hostcache migitates
13718                  * the problem so it affects only the first tcp connection
13719                  * with a host.
13720                  *
13721                  * NB: Don't set DF on small MTU/MSS to have a safe
13722                  * fallback.
13723                  */
13724                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
13725                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13726                         if (tp->t_port == 0 || len < V_tcp_minmss) {
13727                                 ip->ip_off |= htons(IP_DF);
13728                         }
13729                 } else {
13730                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13731                 }
13732
13733                 if (tp->t_state == TCPS_SYN_SENT)
13734                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
13735
13736                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
13737
13738                 error = ip_output(m, inp->inp_options, &inp->inp_route,
13739                     ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
13740                     inp);
13741                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
13742                         mtu = inp->inp_route.ro_nh->nh_mtu;
13743         }
13744 #endif                          /* INET */
13745 out:
13746
13747         if (lgb) {
13748                 lgb->tlb_errno = error;
13749                 lgb = NULL;
13750         }
13751         /*
13752          * In transmit state, time the transmission and arrange for the
13753          * retransmit.  In persist state, just set snd_max.
13754          */
13755         if (error == 0) {
13756                 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
13757                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13758                     (tp->t_flags & TF_SACK_PERMIT) &&
13759                     tp->rcv_numsacks > 0)
13760                         tcp_clean_dsack_blocks(tp);
13761                 /* We sent an ack clear the bbr_segs_rcvd count */
13762                 bbr->output_error_seen = 0;
13763                 bbr->oerror_cnt = 0;
13764                 bbr->bbr_segs_rcvd = 0;
13765                 if (len == 0)
13766                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1);
13767                 /* Do accounting for new sends */
13768                 if ((len > 0) && (rsm == NULL)) {
13769                         int idx;
13770                         if (tp->snd_una == tp->snd_max) {
13771                                 /*
13772                                  * Special case to match google, when
13773                                  * nothing is in flight the delivered
13774                                  * time does get updated to the current
13775                                  * time (see tcp_rate_bsd.c).
13776                                  */
13777                                 bbr->r_ctl.rc_del_time = cts;
13778                         }
13779                         if (len >= maxseg) {
13780                                 idx = (len / maxseg) + 3;
13781                                 if (idx >= TCP_MSS_ACCT_ATIMER)
13782                                         counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1);
13783                                 else
13784                                         counter_u64_add(bbr_out_size[idx], 1);
13785                         } else {
13786                                 /* smaller than a MSS */
13787                                 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options);
13788                                 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV)
13789                                         idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1);
13790                                 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1);
13791                         }
13792                 }
13793         }
13794         abandon = 0;
13795         /*
13796          * We must do the send accounting before we log the output,
13797          * otherwise the state of the rsm could change and we account to the
13798          * wrong bucket.
13799          */
13800         if (len > 0) {
13801                 bbr_do_send_accounting(tp, bbr, rsm, len, error);
13802                 if (error == 0) {
13803                         if (tp->snd_una == tp->snd_max)
13804                                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
13805                 }
13806         }
13807         bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error,
13808             cts, mb, &abandon, rsm, 0, sb);
13809         if (abandon) {
13810                 /*
13811                  * If bbr_log_output destroys the TCB or sees a TH_RST being
13812                  * sent we should hit this condition.
13813                  */
13814                 return (0);
13815         }
13816         if (bbr->rc_in_persist == 0) {
13817                 /*
13818                  * Advance snd_nxt over sequence space of this segment.
13819                  */
13820                 if (error)
13821                         /* We don't log or do anything with errors */
13822                         goto skip_upd;
13823
13824                 if (tp->snd_una == tp->snd_max &&
13825                     (len || (flags & (TH_SYN | TH_FIN)))) {
13826                         /*
13827                          * Update the time we just added data since none was
13828                          * outstanding.
13829                          */
13830                         bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
13831                         bbr->rc_tp->t_acktime  = ticks;
13832                 }
13833                 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) {
13834                         if (flags & TH_SYN) {
13835                                 /*
13836                                  * Smack the snd_max to iss + 1
13837                                  * if its a FO we will add len below.
13838                                  */
13839                                 tp->snd_max = tp->iss + 1;
13840                         }
13841                         if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
13842                                 tp->snd_max++;
13843                                 tp->t_flags |= TF_SENTFIN;
13844                         }
13845                 }
13846                 if (sack_rxmit == 0)
13847                         tp->snd_max += len;
13848 skip_upd:
13849                 if ((error == 0) && len)
13850                         tot_len += len;
13851         } else {
13852                 /* Persists case */
13853                 int32_t xlen = len;
13854
13855                 if (error)
13856                         goto nomore;
13857
13858                 if (flags & TH_SYN)
13859                         ++xlen;
13860                 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
13861                         ++xlen;
13862                         tp->t_flags |= TF_SENTFIN;
13863                 }
13864                 if (xlen && (tp->snd_una == tp->snd_max)) {
13865                         /*
13866                          * Update the time we just added data since none was
13867                          * outstanding.
13868                          */
13869                         bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
13870                         bbr->rc_tp->t_acktime = ticks;
13871                 }
13872                 if (sack_rxmit == 0)
13873                         tp->snd_max += xlen;
13874                 tot_len += (len + optlen + ipoptlen);
13875         }
13876 nomore:
13877         if (error) {
13878                 /*
13879                  * Failures do not advance the seq counter above. For the
13880                  * case of ENOBUFS we will fall out and become ack-clocked.
13881                  * capping the cwnd at the current flight.
13882                  * Everything else will just have to retransmit with the timer
13883                  * (no pacer).
13884                  */
13885                 SOCKBUF_UNLOCK_ASSERT(sb);
13886                 BBR_STAT_INC(bbr_saw_oerr);
13887                 /* Clear all delay/early tracks */
13888                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
13889                 bbr->r_ctl.rc_agg_early = 0;
13890                 bbr->r_agg_early_set = 0;
13891                 bbr->output_error_seen = 1;
13892                 if (bbr->oerror_cnt < 0xf)
13893                         bbr->oerror_cnt++;
13894                 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) {
13895                         /* drop the session */
13896                         tcp_set_inp_to_drop(inp, ENETDOWN);
13897                 }
13898                 switch (error) {
13899                 case ENOBUFS:
13900                         /*
13901                          * Make this guy have to get ack's to send
13902                          * more but lets make sure we don't
13903                          * slam him below a T-O (1MSS).
13904                          */
13905                         if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
13906                                 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13907                                                                     bbr->r_ctl.rc_lost_bytes)) - maxseg;
13908                                 if (tp->snd_cwnd < maxseg)
13909                                         tp->snd_cwnd = maxseg;
13910                         }
13911                         slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt;
13912                         BBR_STAT_INC(bbr_saw_enobuf);
13913                         if (bbr->bbr_hdrw_pacing)
13914                                 counter_u64_add(bbr_hdwr_pacing_enobuf, 1);
13915                         else
13916                                 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1);
13917                         /*
13918                          * Here even in the enobuf's case we want to do our
13919                          * state update. The reason being we may have been
13920                          * called by the input function. If so we have had
13921                          * things change.
13922                          */
13923                         error = 0;
13924                         goto enobufs;
13925                 case EMSGSIZE:
13926                         /*
13927                          * For some reason the interface we used initially
13928                          * to send segments changed to another or lowered
13929                          * its MTU. If TSO was active we either got an
13930                          * interface without TSO capabilits or TSO was
13931                          * turned off. If we obtained mtu from ip_output()
13932                          * then update it and try again.
13933                          */
13934                         /* Turn on tracing (or try to) */
13935                         {
13936                                 int old_maxseg;
13937
13938                                 old_maxseg = tp->t_maxseg;
13939                                 BBR_STAT_INC(bbr_saw_emsgsiz);
13940                                 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts);
13941                                 if (mtu != 0)
13942                                         tcp_mss_update(tp, -1, mtu, NULL, NULL);
13943                                 if (old_maxseg <= tp->t_maxseg) {
13944                                         /* Huh it did not shrink? */
13945                                         tp->t_maxseg = old_maxseg - 40;
13946                                         bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts);
13947                                 }
13948                                 /*
13949                                  * Nuke all other things that can interfere
13950                                  * with slot
13951                                  */
13952                                 if ((tot_len + len) && (len >= tp->t_maxseg)) {
13953                                         slot = bbr_get_pacing_delay(bbr,
13954                                             bbr->r_ctl.rc_bbr_hptsi_gain,
13955                                             (tot_len + len), cts, 0);
13956                                         if (slot < bbr_error_base_paceout)
13957                                                 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
13958                                 } else
13959                                         slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
13960                                 bbr->rc_output_starts_timer = 1;
13961                                 bbr_start_hpts_timer(bbr, tp, cts, 10, slot,
13962                                     tot_len);
13963                                 return (error);
13964                         }
13965                 case EPERM:
13966                         tp->t_softerror = error;
13967                         /* Fall through */
13968                 case EHOSTDOWN:
13969                 case EHOSTUNREACH:
13970                 case ENETDOWN:
13971                 case ENETUNREACH:
13972                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
13973                                 tp->t_softerror = error;
13974                         }
13975                         /* FALLTHROUGH */
13976                 default:
13977                         slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt;
13978                         bbr->rc_output_starts_timer = 1;
13979                         bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0);
13980                         return (error);
13981                 }
13982 #ifdef STATS
13983         } else if (((tp->t_flags & TF_GPUTINPROG) == 0) &&
13984                     len &&
13985                     (rsm == NULL) &&
13986             (bbr->rc_in_persist == 0)) {
13987                 tp->gput_seq = bbr_seq;
13988                 tp->gput_ack = bbr_seq +
13989                     min(sbavail(&so->so_snd) - sb_offset, sendwin);
13990                 tp->gput_ts = cts;
13991                 tp->t_flags |= TF_GPUTINPROG;
13992 #endif
13993         }
13994         KMOD_TCPSTAT_INC(tcps_sndtotal);
13995         if ((bbr->bbr_hdw_pace_ena) &&
13996             (bbr->bbr_attempt_hdwr_pace == 0) &&
13997             (bbr->rc_past_init_win) &&
13998             (bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
13999             (get_filter_value(&bbr->r_ctl.rc_delrate)) &&
14000             (inp->inp_route.ro_nh &&
14001              inp->inp_route.ro_nh->nh_ifp)) {
14002                 /*
14003                  * We are past the initial window and
14004                  * have at least one measurement so we
14005                  * could use hardware pacing if its available.
14006                  * We have an interface and we have not attempted
14007                  * to setup hardware pacing, lets try to now.
14008                  */
14009                 uint64_t rate_wanted;
14010                 int err = 0;
14011
14012                 rate_wanted = bbr_get_hardware_rate(bbr);
14013                 bbr->bbr_attempt_hdwr_pace = 1;
14014                 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp,
14015                                                       inp->inp_route.ro_nh->nh_ifp,
14016                                                       rate_wanted,
14017                                                       (RS_PACING_GEQ|RS_PACING_SUB_OK),
14018                                                       &err, NULL);
14019                 if (bbr->r_ctl.crte) {
14020                         bbr_type_log_hdwr_pacing(bbr,
14021                                                  bbr->r_ctl.crte->ptbl->rs_ifp,
14022                                                  rate_wanted,
14023                                                  bbr->r_ctl.crte->rate,
14024                                                  __LINE__, cts, err);
14025                         BBR_STAT_INC(bbr_hdwr_rl_add_ok);
14026                         counter_u64_add(bbr_flows_nohdwr_pacing, -1);
14027                         counter_u64_add(bbr_flows_whdwr_pacing, 1);
14028                         bbr->bbr_hdrw_pacing = 1;
14029                         /* Now what is our gain status? */
14030                         if (bbr->r_ctl.crte->rate < rate_wanted) {
14031                                 /* We have a problem */
14032                                 bbr_setup_less_of_rate(bbr, cts,
14033                                                        bbr->r_ctl.crte->rate, rate_wanted);
14034                         } else {
14035                                 /* We are good */
14036                                 bbr->gain_is_limited = 0;
14037                                 bbr->skip_gain = 0;
14038                         }
14039                         tcp_bbr_tso_size_check(bbr, cts);
14040                 } else {
14041                         bbr_type_log_hdwr_pacing(bbr,
14042                                                  inp->inp_route.ro_nh->nh_ifp,
14043                                                  rate_wanted,
14044                                                  0,
14045                                                  __LINE__, cts, err);
14046                         BBR_STAT_INC(bbr_hdwr_rl_add_fail);
14047                 }
14048         }
14049         if (bbr->bbr_hdrw_pacing) {
14050                 /*
14051                  * Worry about cases where the route
14052                  * changes or something happened that we
14053                  * lost our hardware pacing possibly during
14054                  * the last ip_output call.
14055                  */
14056                 if (inp->inp_snd_tag == NULL) {
14057                         /* A change during ip output disabled hw pacing? */
14058                         bbr->bbr_hdrw_pacing = 0;
14059                 } else if ((inp->inp_route.ro_nh == NULL) ||
14060                     (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) {
14061                         /*
14062                          * We had an interface or route change,
14063                          * detach from the current hdwr pacing
14064                          * and setup to re-attempt next go
14065                          * round.
14066                          */
14067                         bbr->bbr_hdrw_pacing = 0;
14068                         bbr->bbr_attempt_hdwr_pace = 0;
14069                         tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
14070                         tcp_bbr_tso_size_check(bbr, cts);
14071                 }
14072         }
14073         /*
14074          * Data sent (as far as we can tell). If this advertises a larger
14075          * window than any other segment, then remember the size of the
14076          * advertised window. Any pending ACK has now been sent.
14077          */
14078         if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
14079                 tp->rcv_adv = tp->rcv_nxt + recwin;
14080
14081         tp->last_ack_sent = tp->rcv_nxt;
14082         if ((error == 0) &&
14083             (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) &&
14084             (doing_tlp == 0) &&
14085             (tso == 0) &&
14086             (len > 0) &&
14087             ((flags & TH_RST) == 0) &&
14088             ((flags & TH_SYN) == 0) &&
14089             (IN_RECOVERY(tp->t_flags) == 0) &&
14090             (bbr->rc_in_persist == 0) &&
14091             (tot_len < bbr->r_ctl.rc_pace_max_segs)) {
14092                 /*
14093                  * For non-tso we need to goto again until we have sent out
14094                  * enough data to match what we are hptsi out every hptsi
14095                  * interval.
14096                  */
14097                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14098                         /* Make sure snd_nxt is drug up */
14099                         tp->snd_nxt = tp->snd_max;
14100                 }
14101                 if (rsm != NULL) {
14102                         rsm = NULL;
14103                         goto skip_again;
14104                 }
14105                 rsm = NULL;
14106                 sack_rxmit = 0;
14107                 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14108                 goto again;
14109         }
14110 skip_again:
14111         if ((error == 0) && (flags & TH_FIN))
14112                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
14113         if ((error == 0) && (flags & TH_RST))
14114                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
14115         if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) {
14116                 /*
14117                  * Calculate/Re-Calculate the hptsi slot in usecs based on
14118                  * what we have sent so far
14119                  */
14120                 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
14121                 if (bbr->rc_no_pacing)
14122                         slot = 0;
14123         }
14124         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14125 enobufs:
14126         if (bbr->rc_use_google == 0)
14127                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
14128         bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
14129                                                         bbr->r_ctl.rc_lost_bytes)));
14130         bbr->rc_output_starts_timer = 1;
14131         if (bbr->bbr_use_rack_cheat &&
14132             (more_to_rxt ||
14133              ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) {
14134                 /* Rack cheats and shotguns out all rxt's 1ms apart */
14135                 if (slot > 1000)
14136                         slot = 1000;
14137         }
14138         if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) {
14139                 /*
14140                  * We don't change the tso size until some number of sends
14141                  * to give the hardware commands time to get down
14142                  * to the interface.
14143                  */
14144                 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++;
14145                 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) {
14146                         bbr->hw_pacing_set = 1;
14147                         tcp_bbr_tso_size_check(bbr, cts);
14148                 }
14149         }
14150         bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len);
14151         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14152                 /* Make sure snd_nxt is drug up */
14153                 tp->snd_nxt = tp->snd_max;
14154         }
14155         return (error);
14156
14157 }
14158
14159 /*
14160  * See bbr_output_wtime() for return values.
14161  */
14162 static int
14163 bbr_output(struct tcpcb *tp)
14164 {
14165         int32_t ret;
14166         struct timeval tv;
14167         struct tcp_bbr *bbr;
14168
14169         NET_EPOCH_ASSERT();
14170
14171         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14172         INP_WLOCK_ASSERT(tp->t_inpcb);
14173         (void)tcp_get_usecs(&tv);
14174         ret = bbr_output_wtime(tp, &tv);
14175         return (ret);
14176 }
14177
14178 static void
14179 bbr_mtu_chg(struct tcpcb *tp)
14180 {
14181         struct tcp_bbr *bbr;
14182         struct bbr_sendmap *rsm, *frsm = NULL;
14183         uint32_t maxseg;
14184
14185         /*
14186          * The MTU has changed. a) Clear the sack filter. b) Mark everything
14187          * over the current size as SACK_PASS so a retransmit will occur.
14188          */
14189
14190         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14191         maxseg = tp->t_maxseg - bbr->rc_last_options;
14192         sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
14193         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
14194                 /* Don't mess with ones acked (by sack?) */
14195                 if (rsm->r_flags & BBR_ACKED)
14196                         continue;
14197                 if ((rsm->r_end - rsm->r_start) > maxseg) {
14198                         /*
14199                          * We mark sack-passed on all the previous large
14200                          * sends we did. This will force them to retransmit.
14201                          */
14202                         rsm->r_flags |= BBR_SACK_PASSED;
14203                         if (((rsm->r_flags & BBR_MARKED_LOST) == 0) &&
14204                             bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) {
14205                                 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
14206                                 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
14207                                 rsm->r_flags |= BBR_MARKED_LOST;
14208                         }
14209                         if (frsm == NULL)
14210                                 frsm = rsm;
14211                 }
14212         }
14213         if (frsm) {
14214                 bbr->r_ctl.rc_resend = frsm;
14215         }
14216 }
14217
14218 static int
14219 bbr_pru_options(struct tcpcb *tp, int flags)
14220 {
14221         if (flags & PRUS_OOB)
14222                 return (EOPNOTSUPP);
14223         return (0);
14224 }
14225
14226 struct tcp_function_block __tcp_bbr = {
14227         .tfb_tcp_block_name = __XSTRING(STACKNAME),
14228         .tfb_tcp_output = bbr_output,
14229         .tfb_do_queued_segments = ctf_do_queued_segments,
14230         .tfb_do_segment_nounlock = bbr_do_segment_nounlock,
14231         .tfb_tcp_do_segment = bbr_do_segment,
14232         .tfb_tcp_ctloutput = bbr_ctloutput,
14233         .tfb_tcp_fb_init = bbr_init,
14234         .tfb_tcp_fb_fini = bbr_fini,
14235         .tfb_tcp_timer_stop_all = bbr_stopall,
14236         .tfb_tcp_timer_activate = bbr_timer_activate,
14237         .tfb_tcp_timer_active = bbr_timer_active,
14238         .tfb_tcp_timer_stop = bbr_timer_stop,
14239         .tfb_tcp_rexmit_tmr = bbr_remxt_tmr,
14240         .tfb_tcp_handoff_ok = bbr_handoff_ok,
14241         .tfb_tcp_mtu_chg = bbr_mtu_chg,
14242         .tfb_pru_options = bbr_pru_options,
14243 };
14244
14245 /*
14246  * bbr_ctloutput() must drop the inpcb lock before performing copyin on
14247  * socket option arguments.  When it re-acquires the lock after the copy, it
14248  * has to revalidate that the connection is still valid for the socket
14249  * option.
14250  */
14251 static int
14252 bbr_set_sockopt(struct socket *so, struct sockopt *sopt,
14253                 struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14254 {
14255         struct epoch_tracker et;
14256         int32_t error = 0, optval;
14257
14258         switch (sopt->sopt_name) {
14259         case TCP_RACK_PACE_MAX_SEG:
14260         case TCP_RACK_MIN_TO:
14261         case TCP_RACK_REORD_THRESH:
14262         case TCP_RACK_REORD_FADE:
14263         case TCP_RACK_TLP_THRESH:
14264         case TCP_RACK_PKT_DELAY:
14265         case TCP_BBR_ALGORITHM:
14266         case TCP_BBR_TSLIMITS:
14267         case TCP_BBR_IWINTSO:
14268         case TCP_BBR_RECFORCE:
14269         case TCP_BBR_STARTUP_PG:
14270         case TCP_BBR_DRAIN_PG:
14271         case TCP_BBR_RWND_IS_APP:
14272         case TCP_BBR_PROBE_RTT_INT:
14273         case TCP_BBR_PROBE_RTT_GAIN:
14274         case TCP_BBR_PROBE_RTT_LEN:
14275         case TCP_BBR_STARTUP_LOSS_EXIT:
14276         case TCP_BBR_USEDEL_RATE:
14277         case TCP_BBR_MIN_RTO:
14278         case TCP_BBR_MAX_RTO:
14279         case TCP_BBR_PACE_PER_SEC:
14280         case TCP_DELACK:
14281         case TCP_BBR_PACE_DEL_TAR:
14282         case TCP_BBR_SEND_IWND_IN_TSO:
14283         case TCP_BBR_EXTRA_STATE:
14284         case TCP_BBR_UTTER_MAX_TSO:
14285         case TCP_BBR_MIN_TOPACEOUT:
14286         case TCP_BBR_FLOOR_MIN_TSO:
14287         case TCP_BBR_TSTMP_RAISES:
14288         case TCP_BBR_POLICER_DETECT:
14289         case TCP_BBR_USE_RACK_CHEAT:
14290         case TCP_DATA_AFTER_CLOSE:
14291         case TCP_BBR_HDWR_PACE:
14292         case TCP_BBR_PACE_SEG_MAX:
14293         case TCP_BBR_PACE_SEG_MIN:
14294         case TCP_BBR_PACE_CROSS:
14295         case TCP_BBR_PACE_OH:
14296 #ifdef NETFLIX_PEAKRATE
14297         case TCP_MAXPEAKRATE:
14298 #endif
14299         case TCP_BBR_TMR_PACE_OH:
14300         case TCP_BBR_RACK_RTT_USE:
14301         case TCP_BBR_RETRAN_WTSO:
14302                 break;
14303         default:
14304                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14305                 break;
14306         }
14307         INP_WUNLOCK(inp);
14308         error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
14309         if (error)
14310                 return (error);
14311         INP_WLOCK(inp);
14312         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
14313                 INP_WUNLOCK(inp);
14314                 return (ECONNRESET);
14315         }
14316         tp = intotcpcb(inp);
14317         if (tp->t_fb != &__tcp_bbr) {
14318                 INP_WUNLOCK(inp);
14319                 return (ENOPROTOOPT);
14320         }
14321         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14322         switch (sopt->sopt_name) {
14323         case TCP_BBR_PACE_PER_SEC:
14324                 BBR_OPTS_INC(tcp_bbr_pace_per_sec);
14325                 bbr->r_ctl.bbr_hptsi_per_second = optval;
14326                 break;
14327         case TCP_BBR_PACE_DEL_TAR:
14328                 BBR_OPTS_INC(tcp_bbr_pace_del_tar);
14329                 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval;
14330                 break;
14331         case TCP_BBR_PACE_SEG_MAX:
14332                 BBR_OPTS_INC(tcp_bbr_pace_seg_max);
14333                 bbr->r_ctl.bbr_hptsi_segments_max = optval;
14334                 break;
14335         case TCP_BBR_PACE_SEG_MIN:
14336                 BBR_OPTS_INC(tcp_bbr_pace_seg_min);
14337                 bbr->r_ctl.bbr_hptsi_bytes_min = optval;
14338                 break;
14339         case TCP_BBR_PACE_CROSS:
14340                 BBR_OPTS_INC(tcp_bbr_pace_cross);
14341                 bbr->r_ctl.bbr_cross_over = optval;
14342                 break;
14343         case TCP_BBR_ALGORITHM:
14344                 BBR_OPTS_INC(tcp_bbr_algorithm);
14345                 if (optval && (bbr->rc_use_google == 0)) {
14346                         /* Turn on the google mode */
14347                         bbr_google_mode_on(bbr);
14348                         if ((optval > 3) && (optval < 500)) {
14349                                 /*
14350                                  * Must be at least greater than .3%
14351                                  * and must be less than 50.0%.
14352                                  */
14353                                 bbr->r_ctl.bbr_google_discount = optval;
14354                         }
14355                 } else if ((optval == 0) && (bbr->rc_use_google == 1)) {
14356                         /* Turn off the google mode */
14357                         bbr_google_mode_off(bbr);
14358                 }
14359                 break;
14360         case TCP_BBR_TSLIMITS:
14361                 BBR_OPTS_INC(tcp_bbr_tslimits);
14362                 if (optval == 1)
14363                         bbr->rc_use_ts_limit = 1;
14364                 else if (optval == 0)
14365                         bbr->rc_use_ts_limit = 0;
14366                 else
14367                         error = EINVAL;
14368                 break;
14369
14370         case TCP_BBR_IWINTSO:
14371                 BBR_OPTS_INC(tcp_bbr_iwintso);
14372                 if ((optval >= 0) && (optval < 128)) {
14373                         uint32_t twin;
14374
14375                         bbr->rc_init_win = optval;
14376                         twin = bbr_initial_cwnd(bbr, tp);
14377                         if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd))
14378                                 tp->snd_cwnd = twin;
14379                         else
14380                                 error = EBUSY;
14381                 } else
14382                         error = EINVAL;
14383                 break;
14384         case TCP_BBR_STARTUP_PG:
14385                 BBR_OPTS_INC(tcp_bbr_startup_pg);
14386                 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) {
14387                         bbr->r_ctl.rc_startup_pg = optval;
14388                         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
14389                                 bbr->r_ctl.rc_bbr_hptsi_gain = optval;
14390                         }
14391                 } else
14392                         error = EINVAL;
14393                 break;
14394         case TCP_BBR_DRAIN_PG:
14395                 BBR_OPTS_INC(tcp_bbr_drain_pg);
14396                 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE))
14397                         bbr->r_ctl.rc_drain_pg = optval;
14398                 else
14399                         error = EINVAL;
14400                 break;
14401         case TCP_BBR_PROBE_RTT_LEN:
14402                 BBR_OPTS_INC(tcp_bbr_probertt_len);
14403                 if (optval <= 1)
14404                         reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND));
14405                 else
14406                         error = EINVAL;
14407                 break;
14408         case TCP_BBR_PROBE_RTT_GAIN:
14409                 BBR_OPTS_INC(tcp_bbr_probertt_gain);
14410                 if (optval <= BBR_UNIT)
14411                         bbr->r_ctl.bbr_rttprobe_gain_val = optval;
14412                 else
14413                         error = EINVAL;
14414                 break;
14415         case TCP_BBR_PROBE_RTT_INT:
14416                 BBR_OPTS_INC(tcp_bbr_probe_rtt_int);
14417                 if (optval > 1000)
14418                         bbr->r_ctl.rc_probertt_int = optval;
14419                 else
14420                         error = EINVAL;
14421                 break;
14422         case TCP_BBR_MIN_TOPACEOUT:
14423                 BBR_OPTS_INC(tcp_bbr_topaceout);
14424                 if (optval == 0) {
14425                         bbr->no_pacing_until = 0;
14426                         bbr->rc_no_pacing = 0;
14427                 } else if (optval <= 0x00ff) {
14428                         bbr->no_pacing_until = optval;
14429                         if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) &&
14430                             (bbr->rc_bbr_state == BBR_STATE_STARTUP)){
14431                                 /* Turn on no pacing */
14432                                 bbr->rc_no_pacing = 1;
14433                         }
14434                 } else
14435                         error = EINVAL;
14436                 break;
14437         case TCP_BBR_STARTUP_LOSS_EXIT:
14438                 BBR_OPTS_INC(tcp_bbr_startup_loss_exit);
14439                 bbr->rc_loss_exit = optval;
14440                 break;
14441         case TCP_BBR_USEDEL_RATE:
14442                 error = EINVAL;
14443                 break;
14444         case TCP_BBR_MIN_RTO:
14445                 BBR_OPTS_INC(tcp_bbr_min_rto);
14446                 bbr->r_ctl.rc_min_rto_ms = optval;
14447                 break;
14448         case TCP_BBR_MAX_RTO:
14449                 BBR_OPTS_INC(tcp_bbr_max_rto);
14450                 bbr->rc_max_rto_sec = optval;
14451                 break;
14452         case TCP_RACK_MIN_TO:
14453                 /* Minimum time between rack t-o's in ms */
14454                 BBR_OPTS_INC(tcp_rack_min_to);
14455                 bbr->r_ctl.rc_min_to = optval;
14456                 break;
14457         case TCP_RACK_REORD_THRESH:
14458                 /* RACK reorder threshold (shift amount) */
14459                 BBR_OPTS_INC(tcp_rack_reord_thresh);
14460                 if ((optval > 0) && (optval < 31))
14461                         bbr->r_ctl.rc_reorder_shift = optval;
14462                 else
14463                         error = EINVAL;
14464                 break;
14465         case TCP_RACK_REORD_FADE:
14466                 /* Does reordering fade after ms time */
14467                 BBR_OPTS_INC(tcp_rack_reord_fade);
14468                 bbr->r_ctl.rc_reorder_fade = optval;
14469                 break;
14470         case TCP_RACK_TLP_THRESH:
14471                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
14472                 BBR_OPTS_INC(tcp_rack_tlp_thresh);
14473                 if (optval)
14474                         bbr->rc_tlp_threshold = optval;
14475                 else
14476                         error = EINVAL;
14477                 break;
14478         case TCP_BBR_USE_RACK_CHEAT:
14479                 BBR_OPTS_INC(tcp_use_rackcheat);
14480                 if (bbr->rc_use_google) {
14481                         error = EINVAL;
14482                         break;
14483                 }
14484                 BBR_OPTS_INC(tcp_rack_cheat);
14485                 if (optval)
14486                         bbr->bbr_use_rack_cheat = 1;
14487                 else
14488                         bbr->bbr_use_rack_cheat = 0;
14489                 break;
14490         case TCP_BBR_FLOOR_MIN_TSO:
14491                 BBR_OPTS_INC(tcp_utter_max_tso);
14492                 if ((optval >= 0) && (optval < 40))
14493                         bbr->r_ctl.bbr_hptsi_segments_floor = optval;
14494                 else
14495                         error = EINVAL;
14496                 break;
14497         case TCP_BBR_UTTER_MAX_TSO:
14498                 BBR_OPTS_INC(tcp_utter_max_tso);
14499                 if ((optval >= 0) && (optval < 0xffff))
14500                         bbr->r_ctl.bbr_utter_max = optval;
14501                 else
14502                         error = EINVAL;
14503                 break;
14504
14505         case TCP_BBR_EXTRA_STATE:
14506                 BBR_OPTS_INC(tcp_extra_state);
14507                 if (optval)
14508                         bbr->rc_use_idle_restart = 1;
14509                 else
14510                         bbr->rc_use_idle_restart = 0;
14511                 break;
14512         case TCP_BBR_SEND_IWND_IN_TSO:
14513                 BBR_OPTS_INC(tcp_iwnd_tso);
14514                 if (optval) {
14515                         bbr->bbr_init_win_cheat = 1;
14516                         if (bbr->rc_past_init_win == 0) {
14517                                 uint32_t cts;
14518                                 cts = tcp_get_usecs(&bbr->rc_tv);
14519                                 tcp_bbr_tso_size_check(bbr, cts);
14520                         }
14521                 } else
14522                         bbr->bbr_init_win_cheat = 0;
14523                 break;
14524         case TCP_BBR_HDWR_PACE:
14525                 BBR_OPTS_INC(tcp_hdwr_pacing);
14526                 if (optval){
14527                         bbr->bbr_hdw_pace_ena = 1;
14528                         bbr->bbr_attempt_hdwr_pace = 0;
14529                 } else {
14530                         bbr->bbr_hdw_pace_ena = 0;
14531 #ifdef RATELIMIT
14532                         if (bbr->r_ctl.crte != NULL) {
14533                                 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp);
14534                                 bbr->r_ctl.crte = NULL;
14535                         }
14536 #endif
14537                 }
14538                 break;
14539
14540         case TCP_DELACK:
14541                 BBR_OPTS_INC(tcp_delack);
14542                 if (optval < 100) {
14543                         if (optval == 0) /* off */
14544                                 tp->t_delayed_ack = 0;
14545                         else if (optval == 1) /* on which is 2 */
14546                                 tp->t_delayed_ack = 2;
14547                         else /* higher than 2 and less than 100 */
14548                                 tp->t_delayed_ack = optval;
14549                         if (tp->t_flags & TF_DELACK) {
14550                                 tp->t_flags &= ~TF_DELACK;
14551                                 tp->t_flags |= TF_ACKNOW;
14552                                 NET_EPOCH_ENTER(et);
14553                                 bbr_output(tp);
14554                                 NET_EPOCH_EXIT(et);
14555                         }
14556                 } else
14557                         error = EINVAL;
14558                 break;
14559         case TCP_RACK_PKT_DELAY:
14560                 /* RACK added ms i.e. rack-rtt + reord + N */
14561                 BBR_OPTS_INC(tcp_rack_pkt_delay);
14562                 bbr->r_ctl.rc_pkt_delay = optval;
14563                 break;
14564 #ifdef NETFLIX_PEAKRATE
14565         case TCP_MAXPEAKRATE:
14566                 BBR_OPTS_INC(tcp_maxpeak);
14567                 error = tcp_set_maxpeakrate(tp, optval);
14568                 if (!error)
14569                         tp->t_peakrate_thr = tp->t_maxpeakrate;
14570                 break;
14571 #endif
14572         case TCP_BBR_RETRAN_WTSO:
14573                 BBR_OPTS_INC(tcp_retran_wtso);
14574                 if (optval)
14575                         bbr->rc_resends_use_tso = 1;
14576                 else
14577                         bbr->rc_resends_use_tso = 0;
14578                 break;
14579         case TCP_DATA_AFTER_CLOSE:
14580                 BBR_OPTS_INC(tcp_data_ac);
14581                 if (optval)
14582                         bbr->rc_allow_data_af_clo = 1;
14583                 else
14584                         bbr->rc_allow_data_af_clo = 0;
14585                 break;
14586         case TCP_BBR_POLICER_DETECT:
14587                 BBR_OPTS_INC(tcp_policer_det);
14588                 if (bbr->rc_use_google == 0)
14589                         error = EINVAL;
14590                 else if (optval)
14591                         bbr->r_use_policer = 1;
14592                 else
14593                         bbr->r_use_policer = 0;
14594                 break;
14595
14596         case TCP_BBR_TSTMP_RAISES:
14597                 BBR_OPTS_INC(tcp_ts_raises);
14598                 if (optval)
14599                         bbr->ts_can_raise = 1;
14600                 else
14601                         bbr->ts_can_raise = 0;
14602                 break;
14603         case TCP_BBR_TMR_PACE_OH:
14604                 BBR_OPTS_INC(tcp_pacing_oh_tmr);
14605                 if (bbr->rc_use_google) {
14606                         error = EINVAL;
14607                 } else {
14608                         if (optval)
14609                                 bbr->r_ctl.rc_incr_tmrs = 1;
14610                         else
14611                                 bbr->r_ctl.rc_incr_tmrs = 0;
14612                 }
14613                 break;
14614         case TCP_BBR_PACE_OH:
14615                 BBR_OPTS_INC(tcp_pacing_oh);
14616                 if (bbr->rc_use_google) {
14617                         error = EINVAL;
14618                 } else {
14619                         if (optval > (BBR_INCL_TCP_OH|
14620                                       BBR_INCL_IP_OH|
14621                                       BBR_INCL_ENET_OH)) {
14622                                 error = EINVAL;
14623                                 break;
14624                         }
14625                         if (optval & BBR_INCL_TCP_OH)
14626                                 bbr->r_ctl.rc_inc_tcp_oh = 1;
14627                         else
14628                                 bbr->r_ctl.rc_inc_tcp_oh = 0;
14629                         if (optval & BBR_INCL_IP_OH)
14630                                 bbr->r_ctl.rc_inc_ip_oh = 1;
14631                         else
14632                                 bbr->r_ctl.rc_inc_ip_oh = 0;
14633                         if (optval & BBR_INCL_ENET_OH)
14634                                 bbr->r_ctl.rc_inc_enet_oh = 1;
14635                         else
14636                                 bbr->r_ctl.rc_inc_enet_oh = 0;
14637                 }
14638                 break;
14639         default:
14640                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14641                 break;
14642         }
14643 #ifdef NETFLIX_STATS
14644         tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
14645 #endif
14646         INP_WUNLOCK(inp);
14647         return (error);
14648 }
14649
14650 /*
14651  * return 0 on success, error-num on failure
14652  */
14653 static int
14654 bbr_get_sockopt(struct socket *so, struct sockopt *sopt,
14655     struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14656 {
14657         int32_t error, optval;
14658
14659         /*
14660          * Because all our options are either boolean or an int, we can just
14661          * pull everything into optval and then unlock and copy. If we ever
14662          * add a option that is not a int, then this will have quite an
14663          * impact to this routine.
14664          */
14665         switch (sopt->sopt_name) {
14666         case TCP_BBR_PACE_PER_SEC:
14667                 optval = bbr->r_ctl.bbr_hptsi_per_second;
14668                 break;
14669         case TCP_BBR_PACE_DEL_TAR:
14670                 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar;
14671                 break;
14672         case TCP_BBR_PACE_SEG_MAX:
14673                 optval = bbr->r_ctl.bbr_hptsi_segments_max;
14674                 break;
14675         case TCP_BBR_MIN_TOPACEOUT:
14676                 optval = bbr->no_pacing_until;
14677                 break;
14678         case TCP_BBR_PACE_SEG_MIN:
14679                 optval = bbr->r_ctl.bbr_hptsi_bytes_min;
14680                 break;
14681         case TCP_BBR_PACE_CROSS:
14682                 optval = bbr->r_ctl.bbr_cross_over;
14683                 break;
14684         case TCP_BBR_ALGORITHM:
14685                 optval = bbr->rc_use_google;
14686                 break;
14687         case TCP_BBR_TSLIMITS:
14688                 optval = bbr->rc_use_ts_limit;
14689                 break;
14690         case TCP_BBR_IWINTSO:
14691                 optval = bbr->rc_init_win;
14692                 break;
14693         case TCP_BBR_STARTUP_PG:
14694                 optval = bbr->r_ctl.rc_startup_pg;
14695                 break;
14696         case TCP_BBR_DRAIN_PG:
14697                 optval = bbr->r_ctl.rc_drain_pg;
14698                 break;
14699         case TCP_BBR_PROBE_RTT_INT:
14700                 optval = bbr->r_ctl.rc_probertt_int;
14701                 break;
14702         case TCP_BBR_PROBE_RTT_LEN:
14703                 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND);
14704                 break;
14705         case TCP_BBR_PROBE_RTT_GAIN:
14706                 optval = bbr->r_ctl.bbr_rttprobe_gain_val;
14707                 break;
14708         case TCP_BBR_STARTUP_LOSS_EXIT:
14709                 optval = bbr->rc_loss_exit;
14710                 break;
14711         case TCP_BBR_USEDEL_RATE:
14712                 error = EINVAL;
14713                 break;
14714         case TCP_BBR_MIN_RTO:
14715                 optval = bbr->r_ctl.rc_min_rto_ms;
14716                 break;
14717         case TCP_BBR_MAX_RTO:
14718                 optval = bbr->rc_max_rto_sec;
14719                 break;
14720         case TCP_RACK_PACE_MAX_SEG:
14721                 /* Max segments in a pace */
14722                 optval = bbr->r_ctl.rc_pace_max_segs;
14723                 break;
14724         case TCP_RACK_MIN_TO:
14725                 /* Minimum time between rack t-o's in ms */
14726                 optval = bbr->r_ctl.rc_min_to;
14727                 break;
14728         case TCP_RACK_REORD_THRESH:
14729                 /* RACK reorder threshold (shift amount) */
14730                 optval = bbr->r_ctl.rc_reorder_shift;
14731                 break;
14732         case TCP_RACK_REORD_FADE:
14733                 /* Does reordering fade after ms time */
14734                 optval = bbr->r_ctl.rc_reorder_fade;
14735                 break;
14736         case TCP_BBR_USE_RACK_CHEAT:
14737                 /* Do we use the rack cheat for rxt */
14738                 optval = bbr->bbr_use_rack_cheat;
14739                 break;
14740         case TCP_BBR_FLOOR_MIN_TSO:
14741                 optval = bbr->r_ctl.bbr_hptsi_segments_floor;
14742                 break;
14743         case TCP_BBR_UTTER_MAX_TSO:
14744                 optval = bbr->r_ctl.bbr_utter_max;
14745                 break;
14746         case TCP_BBR_SEND_IWND_IN_TSO:
14747                 /* Do we send TSO size segments initially */
14748                 optval = bbr->bbr_init_win_cheat;
14749                 break;
14750         case TCP_BBR_EXTRA_STATE:
14751                 optval = bbr->rc_use_idle_restart;
14752                 break;
14753         case TCP_RACK_TLP_THRESH:
14754                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
14755                 optval = bbr->rc_tlp_threshold;
14756                 break;
14757         case TCP_RACK_PKT_DELAY:
14758                 /* RACK added ms i.e. rack-rtt + reord + N */
14759                 optval = bbr->r_ctl.rc_pkt_delay;
14760                 break;
14761         case TCP_BBR_RETRAN_WTSO:
14762                 optval = bbr->rc_resends_use_tso;
14763                 break;
14764         case TCP_DATA_AFTER_CLOSE:
14765                 optval = bbr->rc_allow_data_af_clo;
14766                 break;
14767         case TCP_DELACK:
14768                 optval = tp->t_delayed_ack;
14769                 break;
14770         case TCP_BBR_HDWR_PACE:
14771                 optval = bbr->bbr_hdw_pace_ena;
14772                 break;
14773         case TCP_BBR_POLICER_DETECT:
14774                 optval = bbr->r_use_policer;
14775                 break;
14776         case TCP_BBR_TSTMP_RAISES:
14777                 optval = bbr->ts_can_raise;
14778                 break;
14779         case TCP_BBR_TMR_PACE_OH:
14780                 optval = bbr->r_ctl.rc_incr_tmrs;
14781                 break;
14782         case TCP_BBR_PACE_OH:
14783                 optval = 0;
14784                 if (bbr->r_ctl.rc_inc_tcp_oh)
14785                         optval |= BBR_INCL_TCP_OH;
14786                 if (bbr->r_ctl.rc_inc_ip_oh)
14787                         optval |= BBR_INCL_IP_OH;
14788                 if (bbr->r_ctl.rc_inc_enet_oh)
14789                         optval |= BBR_INCL_ENET_OH;
14790                 break;
14791         default:
14792                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14793                 break;
14794         }
14795         INP_WUNLOCK(inp);
14796         error = sooptcopyout(sopt, &optval, sizeof optval);
14797         return (error);
14798 }
14799
14800 /*
14801  * return 0 on success, error-num on failure
14802  */
14803 static int
14804 bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
14805 {
14806         int32_t error = EINVAL;
14807         struct tcp_bbr *bbr;
14808
14809         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14810         if (bbr == NULL) {
14811                 /* Huh? */
14812                 goto out;
14813         }
14814         if (sopt->sopt_dir == SOPT_SET) {
14815                 return (bbr_set_sockopt(so, sopt, inp, tp, bbr));
14816         } else if (sopt->sopt_dir == SOPT_GET) {
14817                 return (bbr_get_sockopt(so, sopt, inp, tp, bbr));
14818         }
14819 out:
14820         INP_WUNLOCK(inp);
14821         return (error);
14822 }
14823
14824 static const char *bbr_stack_names[] = {
14825         __XSTRING(STACKNAME),
14826 #ifdef STACKALIAS
14827         __XSTRING(STACKALIAS),
14828 #endif
14829 };
14830
14831 static bool bbr_mod_inited = false;
14832
14833 static int
14834 tcp_addbbr(module_t mod, int32_t type, void *data)
14835 {
14836         int32_t err = 0;
14837         int num_stacks;
14838
14839         switch (type) {
14840         case MOD_LOAD:
14841                 printf("Attempting to load " __XSTRING(MODNAME) "\n");
14842                 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
14843                     sizeof(struct bbr_sendmap),
14844                     NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
14845                 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
14846                     sizeof(struct tcp_bbr),
14847                     NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
14848                 sysctl_ctx_init(&bbr_sysctl_ctx);
14849                 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
14850                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
14851                     OID_AUTO,
14852 #ifdef STACKALIAS
14853                     __XSTRING(STACKALIAS),
14854 #else
14855                     __XSTRING(STACKNAME),
14856 #endif
14857                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
14858                     "");
14859                 if (bbr_sysctl_root == NULL) {
14860                         printf("Failed to add sysctl node\n");
14861                         err = EFAULT;
14862                         goto free_uma;
14863                 }
14864                 bbr_init_sysctls();
14865                 num_stacks = nitems(bbr_stack_names);
14866                 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK,
14867                     bbr_stack_names, &num_stacks);
14868                 if (err) {
14869                         printf("Failed to register %s stack name for "
14870                             "%s module\n", bbr_stack_names[num_stacks],
14871                             __XSTRING(MODNAME));
14872                         sysctl_ctx_free(&bbr_sysctl_ctx);
14873         free_uma:
14874                         uma_zdestroy(bbr_zone);
14875                         uma_zdestroy(bbr_pcb_zone);
14876                         bbr_counter_destroy();
14877                         printf("Failed to register " __XSTRING(MODNAME)
14878                             " module err:%d\n", err);
14879                         return (err);
14880                 }
14881                 tcp_lro_reg_mbufq();
14882                 bbr_mod_inited = true;
14883                 printf(__XSTRING(MODNAME) " is now available\n");
14884                 break;
14885         case MOD_QUIESCE:
14886                 err = deregister_tcp_functions(&__tcp_bbr, true, false);
14887                 break;
14888         case MOD_UNLOAD:
14889                 err = deregister_tcp_functions(&__tcp_bbr, false, true);
14890                 if (err == EBUSY)
14891                         break;
14892                 if (bbr_mod_inited) {
14893                         uma_zdestroy(bbr_zone);
14894                         uma_zdestroy(bbr_pcb_zone);
14895                         sysctl_ctx_free(&bbr_sysctl_ctx);
14896                         bbr_counter_destroy();
14897                         printf(__XSTRING(MODNAME)
14898                             " is now no longer available\n");
14899                         bbr_mod_inited = false;
14900                 }
14901                 tcp_lro_dereg_mbufq();
14902                 err = 0;
14903                 break;
14904         default:
14905                 return (EOPNOTSUPP);
14906         }
14907         return (err);
14908 }
14909
14910 static moduledata_t tcp_bbr = {
14911         .name = __XSTRING(MODNAME),
14912             .evhand = tcp_addbbr,
14913             .priv = 0
14914 };
14915
14916 MODULE_VERSION(MODNAME, 1);
14917 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
14918 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);